voxcity 0.7.0__py3-none-any.whl → 1.0.13__py3-none-any.whl

voxcity/simulator_gpu/solar/radiation.py

@@ -0,0 +1,3019 @@
+"""Radiation solver for palm-solar.
+
+Main module that integrates all components to compute
+shortwave (solar) radiation fluxes with multi-bounce reflections
+following PALM's RTM (Radiative Transfer Model) methodology.
+
+PALM Alignment Notes:
+- Solar position: Uses PALM's calc_zenith formula exactly (solar.py)
+- SVF calculation: Uses PALM's vffrac_up formula with proper weighting (svf.py)
+- Reflection steps: Default nrefsteps=3 matches PALM
+- Extinction coefficient: Default ext_coef=0.6 matches PALM
+- Beer-Lambert law: Same exponential attenuation through canopy
+- Direction indices: IUP=0, IDOWN=1, etc. match PALM convention
+
+Key differences from PALM:
+- GPU-accelerated via Taichi (PALM uses Fortran+MPI)
+- Real-time view factor computation (PALM pre-computes sparse matrix)
+- Shortwave only (PALM includes longwave radiation)
+- Axis-aligned surfaces only (PALM supports slant surfaces)
+
+Input convention:
+- sw_direct: Direct Normal Irradiance (DNI) in W/m²
+- sw_diffuse: Diffuse Horizontal Irradiance (DHI) in W/m²
+"""
+
+import taichi as ti
+import numpy as np
+from typing import Optional, Tuple
+from dataclasses import dataclass, field
+
+from .core import (
+    Vector3, Point3, SOLAR_CONSTANT, EXT_COEF,
+    PI, TWO_PI
+)
+from .domain import Domain, Surfaces, extract_surfaces_from_domain
+from .solar import SolarPosition, calc_zenith, SolarCalculator
+from .raytracing import RayTracer, ray_point_to_point_transmissivity, ray_voxel_first_hit
+from .svf import SVFCalculator
+from .csf import CSFCalculator
+from .volumetric import VolumetricFluxCalculator
+
+
+# Direction indices (matching PALM convention from radiation_model_mod.f90)
+# PALM: iup=0, idown=1, inorth=2, isouth=3, ieast=4, iwest=5
+IUP = 0
+IDOWN = 1
+INORTH = 2
+ISOUTH = 3
+IEAST = 4
+IWEST = 5
+
+
+@dataclass
+class RadiationConfig:
+    """
+    Configuration for radiation model.
+    
+    Attributes:
+        albedo_ground: Default ground albedo
+        albedo_wall: Default wall albedo
+        albedo_roof: Default roof albedo
+        albedo_leaf: Tree/leaf albedo (PALM default: 0.15)
+        n_azimuth: Number of azimuthal divisions for SVF
+        n_elevation: Number of elevation divisions for SVF
+        ext_coef: Extinction coefficient for canopy
+        skip_svf: Skip SVF calculation (use 1.0)
+        n_reflection_steps: Number of reflection iterations (PALM default: 3)
+        surface_reflections: Enable surface-to-surface reflections
+        canopy_reflections: Enable reflection attenuation through canopy
+        volumetric_flux: Enable volumetric flux calculation
+        volumetric_n_azimuth: Number of azimuths for volumetric horizon tracing
+        min_opaque_lad: Minimum LAD considered opaque for volumetric shadows
+        canopy_radiation: Enable plant canopy radiation absorption (CSF)
+        canopy_to_canopy: Enable canopy-to-canopy scattering (not in PALM, 
+            improves accuracy for dense vegetation where leaves scatter light
+            to neighboring leaves)
+    """
+    albedo_ground: float = 0.2
+    albedo_wall: float = 0.3
+    albedo_roof: float = 0.3
+    albedo_leaf: float = 0.15  # PALM default tree albedo
+    n_azimuth: int = 80         # PALM: raytrace_discrete_azims = 80
+    n_elevation: int = 40       # PALM: raytrace_discrete_elevs = 40
+    ext_coef: float = EXT_COEF  # PALM: ext_coef = 0.6
+    skip_svf: bool = False
+    n_reflection_steps: int = 3  # PALM: nrefsteps = 3
+    surface_reflections: bool = True
+    canopy_reflections: bool = True  # Enable LAD attenuation in reflections
+    volumetric_flux: bool = False
+    volumetric_n_azimuth: int = 36
+    min_opaque_lad: float = 0.5
+    canopy_radiation: bool = True  # Enable CSF-based canopy absorption
+    canopy_to_canopy: bool = True  # Enable canopy-to-canopy scattering (improves accuracy in dense vegetation)
+    # SVF matrix caching for multi-timestep efficiency (PALM-like approach)
+    cache_svf_matrix: bool = True   # Pre-compute SVF matrix for fast reflections
+    svf_min_threshold: float = 0.01  # Minimum VF to store (sparsity threshold, 0.01 sufficient for 1% accuracy)
+    # FP16 optimization for intermediate calculations (reduces memory bandwidth, ~2x faster)
+    # fp16 range: ±65,504 with ~3 decimal digits precision - sufficient for W/m² irradiance values
+    use_fp16_intermediate: bool = True  # Use float16 for intermediate reflection buffers
+
+
+@ti.data_oriented
+class RadiationModel:
+    """
+    GPU-accelerated solar radiation transfer model.
+    
+    Computes shortwave (direct and diffuse) radiation
+    for all surface elements in the domain.
+    """
+    
+    def __init__(
+        self,
+        domain: Domain,
+        config: Optional[RadiationConfig] = None
+    ):
+        """
+        Initialize radiation model.
+        
+        Args:
+            domain: Domain object with geometry
+            config: Radiation configuration (uses defaults if None)
+        """
+        self.domain = domain
+        self.config = config or RadiationConfig()
+        
+        # Extract surfaces from domain
+        print("Extracting surfaces from domain...")
+        self.surfaces = extract_surfaces_from_domain(domain)
+        self.n_surfaces = self.surfaces.count
+        print(f"Found {self.n_surfaces} surface elements")
+        
+        # Initialize sub-components
+        self.solar_calc = SolarCalculator(
+            domain.origin_lat or 0.0, 
+            domain.origin_lon or 0.0
+        )
+        
+        self.ray_tracer = RayTracer(domain)
+        self.ray_tracer.ext_coef = self.config.ext_coef
+        
+        self.svf_calc = SVFCalculator(
+            domain,
+            self.config.n_azimuth,
+            self.config.n_elevation
+        )
+        
+        self.csf_calc = CSFCalculator(
+            domain,
+            self.config.n_azimuth,
+            self.config.n_elevation
+        )
+        
+        # Set default surface properties based on direction
+        self._set_default_properties()
+        
+        # Determine dtype for intermediate buffers (fp16 reduces memory bandwidth ~2x)
+        # fp16: range ±65,504, ~3 decimal digits - sufficient for W/m² irradiance (0-1400 range)
+        inter_dtype = ti.f16 if config.use_fp16_intermediate else ti.f32
+        
+        # Allocate arrays for multi-bounce reflections
+        # These store radiation fluxes during reflection iterations
+        self._surfins = ti.field(dtype=inter_dtype, shape=(self.n_surfaces,))  # Incoming SW per reflection step
+        self._surfinl = ti.field(dtype=inter_dtype, shape=(self.n_surfaces,))  # Incoming LW per reflection step
+        self._surfoutsl = ti.field(dtype=inter_dtype, shape=(self.n_surfaces,))  # Outgoing SW per reflection step
+        self._surfoutll = ti.field(dtype=inter_dtype, shape=(self.n_surfaces,))  # Outgoing LW per reflection step
+        
+        # Ping-pong buffers for optimized reflection iterations
+        # Using separate kernels without internal ti.sync() is ~100x faster than fused kernels
+        self._surfins_ping = ti.field(dtype=inter_dtype, shape=(self.n_surfaces,))
+        self._surfins_pong = ti.field(dtype=inter_dtype, shape=(self.n_surfaces,))
+        
+        # Total accumulated radiation - keep as f32 for precision in cumulative sums
+        self._surfinsw = ti.field(dtype=ti.f32, shape=(self.n_surfaces,))  # Total incoming SW
+        self._surfinlw = ti.field(dtype=ti.f32, shape=(self.n_surfaces,))  # Total incoming LW
+        self._surfoutsw = ti.field(dtype=ti.f32, shape=(self.n_surfaces,))  # Total outgoing SW
+        self._surfoutlw = ti.field(dtype=ti.f32, shape=(self.n_surfaces,))  # Total outgoing LW
+        
+        # Direct and diffuse components - keep as f32 for final output precision
+        self._surfinswdir = ti.field(dtype=ti.f32, shape=(self.n_surfaces,))  # Direct SW
+        self._surfinswdif = ti.field(dtype=ti.f32, shape=(self.n_surfaces,))  # Diffuse SW from sky
+        
+        # For optimized reflection computation - store weighted totals
+        self._total_reflected_flux = ti.field(dtype=ti.f32, shape=())  # Sum of (surfout * area)
+        self._total_reflecting_area = ti.field(dtype=ti.f32, shape=())  # Sum of areas
+        
+        # Surface-to-surface view factors (sparse representation)
+        # svf_matrix[i, j] = view factor from surface j to surface i
+        # For efficiency, we'll compute this on-demand during reflections
+        self._svf_matrix_computed = False
+        
+        # SVF computed flag
+        self._svf_computed = False
+        
+        # Initialize volumetric flux calculator if enabled
+        self.volumetric_calc = None
+        if self.config.volumetric_flux:
+            self.volumetric_calc = VolumetricFluxCalculator(
+                domain,
+                n_azimuth=self.config.volumetric_n_azimuth,
+                min_opaque_lad=self.config.min_opaque_lad
+            )
+        
+        # Plant canopy absorption arrays (like PALM's pcbinsw, etc.)
+        # Indexed by (i, j, k) grid coordinates
+        # Units: W/m³ (power absorbed per unit volume)
+        # Total arrays use f32 for precision in cumulative sums
+        self._pcbinsw = ti.field(dtype=ti.f32, shape=(domain.nx, domain.ny, domain.nz))     # Total absorbed SW
+        self._pcbinswdir = ti.field(dtype=ti.f32, shape=(domain.nx, domain.ny, domain.nz))  # Direct SW absorbed
+        self._pcbinswdif = ti.field(dtype=ti.f32, shape=(domain.nx, domain.ny, domain.nz))  # Diffuse SW absorbed
+        
+        # Received radiation (before absorption) in W/m²
+        self._pcinsw = ti.field(dtype=ti.f32, shape=(domain.nx, domain.ny, domain.nz))      # Total received SW
+        self._pcinswdir = ti.field(dtype=ti.f32, shape=(domain.nx, domain.ny, domain.nz))   # Direct SW received
+        self._pcinswdif = ti.field(dtype=ti.f32, shape=(domain.nx, domain.ny, domain.nz))   # Diffuse SW received
+        
+        # Canopy scattered/reflected radiation (W/m³) - intermediate uses inter_dtype
+        self._pcbinswref = ti.field(dtype=inter_dtype, shape=(domain.nx, domain.ny, domain.nz))  # Reflected SW from canopy
+        
+        # Canopy-to-canopy scattering contribution (W/m³)
+        # Stores radiation scattered from other canopy cells to this cell (per iteration)
+        self._pcbinswc2c = ti.field(dtype=inter_dtype, shape=(domain.nx, domain.ny, domain.nz))  # Canopy-to-canopy SW (temp)
+        # Cumulative total of c2c contribution (for output) - f32 for precision
+        self._pcbinswc2c_total = ti.field(dtype=ti.f32, shape=(domain.nx, domain.ny, domain.nz))  # Canopy-to-canopy SW (total)
+        
+        # Canopy-to-surface contribution per reflection step - intermediate
+        self._surfinswpc = ti.field(dtype=inter_dtype, shape=(self.n_surfaces,))  # SW from plant canopy
+        
+        # ========== SVF Matrix Caching (PALM-like approach) ==========
+        # For multi-timestep efficiency, pre-compute surface-to-surface view factors
+        # Stored in sparse COO format: (source_idx, target_idx, vf_value, transmissivity)
+        # This makes reflection iterations O(nnz) instead of O(n²)
+        self._svf_matrix_cached = False
+        self._svf_nnz = 0  # Number of non-zero entries
+        
+        # Estimate max non-zeros based on threshold:
+        # With svf_min_threshold=0.01, ~99.7% of entries are filtered out
+        # For 250k surfaces: from 585M entries down to ~1.8M entries
+        # 
+        # Memory per entry: 12 bytes (2 int32 + 2 float16)
+        # At 0.01 threshold: ~22 MB for 1.8M entries
+        # At 0.001 threshold: ~260 MB for 21M entries
+        # At 0.0001 threshold: ~1.8 GB for 148M entries
+        #
+        # Estimate based on threshold (empirically derived from 250k surface domain):
+        threshold = config.svf_min_threshold
+        if threshold >= 0.01:
+            entries_per_surface = 8  # ~0.3% of 2500 avg neighbors
+        elif threshold >= 0.001:
+            entries_per_surface = 85  # ~3.5% of avg neighbors  
+        elif threshold >= 0.0001:
+            entries_per_surface = 600  # ~25% of avg neighbors
+        else:
+            entries_per_surface = 2500  # Full density
+            
+        estimated = self.n_surfaces * entries_per_surface
+        
+        # For small domains, allow full N*N
+        if self.n_surfaces < 5000:
+            self._max_svf_entries = min(self.n_surfaces * self.n_surfaces, estimated * 2)
+        else:
+            # Add 50% buffer for safety
+            self._max_svf_entries = int(estimated * 1.5)
+            
+            # Sanity check against GPU memory (each entry = 12 bytes with f16)
+            try:
+                import subprocess
+                result = subprocess.run(
+                    ['nvidia-smi', '--query-gpu=memory.free', '--format=csv,noheader,nounits'],
+                    capture_output=True, text=True, timeout=5
+                )
+                if result.returncode == 0:
+                    free_mb = int(result.stdout.strip().split('\n')[0])
+                    # Use up to 50% of free memory for SVF matrix
+                    available_bytes = free_mb * 1024 * 1024 * 0.5
+                    # Each entry = 12 bytes (2 int32 + 2 float16)
+                    memory_based_limit = int(available_bytes / 12)
+                    
+                    if self._max_svf_entries > memory_based_limit:
+                        import warnings
+                        warnings.warn(
+                            f"SVF buffer limited to {memory_based_limit:,} entries due to GPU memory. "
+                            f"Estimated {self._max_svf_entries:,} entries needed.",
+                            RuntimeWarning
+                        )
+                        self._max_svf_entries = memory_based_limit
+            except Exception:
+                pass  # Keep estimated size
+        
+        # Pre-allocate sparse COO arrays upfront to avoid CUDA issues with dynamic allocation
+        # These are allocated during __init__ to ensure proper CUDA memory management
+        if config.cache_svf_matrix and config.surface_reflections:
+            self._svf_source = ti.field(dtype=ti.i32, shape=(self._max_svf_entries,))
+            self._svf_target = ti.field(dtype=ti.i32, shape=(self._max_svf_entries,))
+            # Use float16 to reduce memory by 50% (sufficient for VF accuracy ~0.01)
+            self._svf_vf = ti.field(dtype=ti.f16, shape=(self._max_svf_entries,))
+            self._svf_trans = ti.field(dtype=ti.f16, shape=(self._max_svf_entries,))
+            self._svf_count = ti.field(dtype=ti.i32, shape=())
+            
+            # CSR format arrays for optimized sparse matmul
+            # NOTE: CSR is currently disabled because:
+            # 1. COO with modern GPU atomics is actually faster for this workload
+            # 2. CSR building from numpy adds significant overhead (12+ seconds)
+            # 3. CSR suffers from load imbalance (row sizes vary from 33 to 4789)
+            # Keep the arrays allocated for potential future use with GPU-based sorting
+            self._svf_csr_row_ptr = None  # Disabled
+            self._svf_csr_col_idx = None  # Disabled  
+            self._svf_csr_val = None  # Disabled
+            self._svf_csr_ready = False
+        else:
+            self._svf_source = None
+            self._svf_target = None
+            self._svf_vf = None
+            self._svf_trans = None
+            self._svf_count = None
+            self._svf_csr_row_ptr = None
+            self._svf_csr_col_idx = None
+            self._svf_csr_val = None
+            self._svf_csr_ready = False
+            
+        # ========== CSF Matrix Caching (Canopy-Surface Factors) ==========
+        # For efficient canopy-surface interactions during reflections
+        # Stored in sparse COO format: (canopy_idx, surface_idx, csf_value)
+        self._csf_matrix_cached = False
+        self._csf_nnz = 0
+        
+        # Estimate max non-zeros for CSF
+        # Assume each canopy cell sees ~100 surfaces on average
+        # This is a rough heuristic
+        self._max_csf_entries = 10_000_000  # 10M entries ~ 120MB
+        
+        if config.cache_svf_matrix and config.canopy_radiation and domain.lad is not None:
+            self._csf_canopy_idx = ti.field(dtype=ti.i32, shape=(self._max_csf_entries,))
+            self._csf_surface_idx = ti.field(dtype=ti.i32, shape=(self._max_csf_entries,))
+            # Use float16 for CSF values (sufficient for ~0.01 accuracy)
+            self._csf_val = ti.field(dtype=ti.f16, shape=(self._max_csf_entries,))
+            self._csf_count = ti.field(dtype=ti.i32, shape=())
+            
+            # Lookup table for surface index from grid position and direction
+            # (nx, ny, nz, 6) -> surface_index
+            self._grid_to_surf = ti.field(dtype=ti.i32, shape=(domain.nx, domain.ny, domain.nz, 6))
+        else:
+            self._csf_canopy_idx = None
+            self._csf_surface_idx = None
+            self._csf_val = None
+            self._csf_count = None
+            self._grid_to_surf = None
+        
+        # Canopy radiation computed flag
+        self._canopy_radiation_computed = False
+    
+    def _set_default_properties(self):
+        """Set default albedo for surfaces."""
+        self._set_defaults_kernel(
+            self.surfaces.direction,
+            self.surfaces.albedo,
+            self.n_surfaces,
+            self.config.albedo_ground,
+            self.config.albedo_wall,
+            self.config.albedo_roof
+        )
+    
+    @ti.kernel
+    def _set_defaults_kernel(
+        self,
+        direction: ti.template(),
+        albedo: ti.template(),
+        n_surf: ti.i32,
+        alb_ground: ti.f32,
+        alb_wall: ti.f32,
+        alb_roof: ti.f32
+    ):
+        for i in range(n_surf):
+            d = direction[i]
+            if d == 0:  # Up (roof or ground)
+                albedo[i] = alb_roof
+            elif d == 1:  # Down (typically building underside)
+                albedo[i] = alb_ground
+            else:  # Walls
+                albedo[i] = alb_wall
+    
+    def compute_svf(self):
+        """
+        Compute Sky View Factors for all surfaces.
+        
+        This is computationally expensive, call once per domain setup.
+        """
+        if self.config.skip_svf:
+            # Set SVF to 1.0 for all surfaces
+            self._set_svf_one()
+            self._svf_computed = True
+            return
+        
+        print("Computing per-surface Sky View Factors (for diffuse sky radiation)...")
+        
+        if self.domain.lad is not None:
+            self.svf_calc.compute_svf_with_canopy(
+                self.surfaces.center,  # Use world coordinates, not grid indices
+                self.surfaces.direction,
+                self.domain.is_solid,
+                self.domain.lad,
+                self.n_surfaces,
+                self.config.ext_coef,
+                self.surfaces.svf,
+                self.surfaces.svf_urban
+            )
+        else:
+            self.svf_calc.compute_svf(
+                self.surfaces.center,  # Use world coordinates, not grid indices
+                self.surfaces.direction,
+                self.domain.is_solid,
+                self.n_surfaces,
+                self.surfaces.svf
+            )
+            # Copy to svf_urban
+            self._copy_svf()
+        
+        self._svf_computed = True
+        print("Per-surface SVF complete.")
+        
+        # Pre-compute SVF matrix for efficient multi-timestep reflections
+        if self.config.cache_svf_matrix and self.config.surface_reflections:
+            self.compute_svf_matrix()
+            
+            # Also compute CSF matrix if canopy is present
+            if self.config.canopy_radiation and self.domain.lad is not None:
+                self.compute_csf_matrix()
+        
+            # Pre-compute CSF sky for efficient multi-timestep canopy absorption
+            # CSF sky is geometry-dependent only and doesn't change with sun position
+            # Only needed when reflections are enabled (for canopy-surface interactions)
+            if self.config.canopy_radiation and self.domain.lad is not None:
+                print("Pre-computing CSF sky (geometry-dependent, computed once)...")
+                self.csf_calc.compute_csf_sky_cached(
+                    self.domain.is_solid,
+                    self.domain.lad,
+                    self.config.n_azimuth,
+                    self.config.n_elevation
+                )
+                print("CSF sky computation complete.")
+    
+    @ti.kernel
+    def _set_svf_one(self):
+        for i in range(self.n_surfaces):
+            self.surfaces.svf[i] = 1.0
+            self.surfaces.svf_urban[i] = 1.0
+    
+    @ti.kernel
+    def _copy_svf(self):
+        for i in range(self.n_surfaces):
+            self.surfaces.svf_urban[i] = self.surfaces.svf[i]
+    
+    def update_solar_position(self, day_of_year: int, second_of_day: float):
+        """
+        Update solar position.
+        
+        Args:
+            day_of_year: Day number (1-365)
+            second_of_day: Seconds since midnight UTC
+        """
+        self.solar_calc.update(day_of_year, second_of_day)
+    
+    def compute_shortwave_radiation(
+        self,
+        sw_direct: float,
+        sw_diffuse: float
+    ):
+        """
+        Compute shortwave (solar) radiation for all surfaces.
+        
+        Args:
+            sw_direct: Direct normal irradiance (W/m²)
+            sw_diffuse: Diffuse horizontal irradiance (W/m²)
+        """
+        if not self._svf_computed:
+            print("Warning: SVF not computed, computing now...")
+            self.compute_svf()
+        
+        # Get sun direction
+        sun_dir = self.solar_calc.sun_direction[None]
+        cos_zenith = self.solar_calc.cos_zenith[None]
+        
+        if cos_zenith > 0:
+            # Compute direct shadows
+            if self.domain.lad is not None:
+                self.ray_tracer.compute_direct_with_canopy(
+                    self.surfaces.center,  # Use world coordinates, not grid indices
+                    self.surfaces.direction,
+                    sun_dir,
+                    self.domain.is_solid,
+                    self.domain.lad,
+                    self.n_surfaces,
+                    self.surfaces.shadow_factor,
+                    self.surfaces.canopy_transmissivity
+                )
+            else:
+                self.ray_tracer.compute_direct_shadows(
+                    self.surfaces.center,  # Use world coordinates, not grid indices
+                    self.surfaces.direction,
+                    sun_dir,
+                    self.domain.is_solid,
+                    self.n_surfaces,
+                    self.surfaces.shadow_factor
+                )
+                # Set canopy transmissivity to 1 (no canopy)
+                self._set_canopy_trans_one()
+        else:
+            # Night time - no direct radiation
+            self._clear_direct_radiation()
+        
+        # Compute radiation fluxes with unified reflection loop
+        # This now includes all paths: Surface↔Surface, Surface↔Canopy, Canopy↔Canopy
+        self._compute_sw_fluxes(
+            sw_direct,
+            sw_diffuse,
+            cos_zenith
+        )
+        
+        # Compute plant canopy radiation absorption from direct/diffuse (CSF)
+        # Note: If canopy_reflections is enabled, initial absorption is computed inside 
+        # _compute_sw_fluxes via _compute_canopy_radiation_initial, so we skip here.
+        # We only call _compute_canopy_radiation if NOT using unified reflection loop.
+        if self.config.canopy_radiation and self.domain.lad is not None:
+            if self.config.canopy_reflections:
+                # Unified reflection loop already computed initial absorption
+                # Just compute received radiation and mark as computed
+                grid_volume = self.domain.dx * self.domain.dy * self.domain.dz
+                self._compute_received_radiation(sw_direct, sw_diffuse, cos_zenith, grid_volume)
+                self._canopy_radiation_computed = True
+            else:
+                # Legacy path: compute canopy radiation the old way
+                self._compute_canopy_radiation(
+                    sw_direct,
+                    sw_diffuse,
+                    sun_dir,
+                    cos_zenith
+                )
+        
+        # Compute volumetric fluxes if enabled
+        if self.volumetric_calc is not None:
+            sun_dir_tuple = (float(sun_dir[0]), float(sun_dir[1]), float(sun_dir[2]))
+            self.volumetric_calc.compute_swflux_vol(
+                sw_direct,
+                sw_diffuse,
+                cos_zenith,
+                sun_dir_tuple,
+                self.domain.lad  # Pass LAD for attenuation
+            )
+    
+    @ti.kernel
+    def _set_canopy_trans_one(self):
+        for i in range(self.n_surfaces):
+            self.surfaces.canopy_transmissivity[i] = 1.0
+    
+    @ti.kernel
+    def _clear_direct_radiation(self):
+        for i in range(self.n_surfaces):
+            self.surfaces.shadow_factor[i] = 1.0
+            self.surfaces.canopy_transmissivity[i] = 0.0
+            self.surfaces.sw_in_direct[i] = 0.0
+    
+    def _compute_sw_fluxes(
+        self,
+        sw_direct: float,
+        sw_diffuse: float,
+        cos_zenith: float
+    ):
+        """
+        Compute shortwave fluxes for all surfaces with multi-bounce reflections.
+        
+        Following PALM's RTM methodology (radiation_model_mod.f90 lines ~9300-9500):
+        
+        1. Initial pass: direct + diffuse from sky
+           - PALM: surfinswdir = rad_sw_in_dir * surf_costh * dsitrans * sun_direct_factor
+           - PALM: surfinswdif = rad_sw_in_diff * skyvft
+           
+        2. Reflection loop (DO refstep = 1, nrefsteps):
+           - PALM: surfoutsl = albedo_surf * surfins
+           - PALM: surfins(isurf) += svf(1,isvf) * svf(2,isvf) * surfoutsl(isurfsrc)
+           - PALM: pcbinsw += csf * surfoutsl(isurfsrc) * asrc * grid_volume_inverse
+           
+        3. Accumulate totals:
+           - PALM: surfinsw = surfinsw + surfins
+           - PALM: surfoutsw = surfoutsw + surfoutsl
+        """
+        # Initialize all flux arrays
+        self._init_flux_arrays()
+        self._init_canopy_arrays()
+        
+        # Compute initial (first pass) radiation: direct + diffuse from sky
+        self._compute_initial_sw_pass(sw_direct, sw_diffuse, cos_zenith)
+        
+        # Compute initial canopy absorption from direct/diffuse BEFORE reflection loop
+        # This allows canopy scattering to participate in reflections
+        if self.domain.lad is not None and self.config.canopy_radiation:
+            sun_dir = self.solar_calc.sun_direction[None]
+            self._compute_canopy_radiation_initial(
+                sw_direct,
+                sw_diffuse,
+                sun_dir,
+                cos_zenith
+            )
+            # Set up initial scattered radiation for the reflection loop
+            if self.config.canopy_reflections:
+                self._update_canopy_scattered_radiation(
+                    self.domain.lad,
+                    self.config.albedo_leaf
+                )
+        
+        # Multi-bounce reflections (PALM: DO refstep = 1, nrefsteps)
+        # Extended to include canopy scattering within the loop for full path coverage:
+        # - Surface → Surface
+        # - Surface → Canopy (absorption)
+        # - Canopy → Surface (scattering)
+        # - Canopy → Canopy (scattering)
+        # This captures all multi-bounce paths like: Canopy→Surface→Canopy, Surface→Canopy→Surface, etc.
+        if self.config.surface_reflections and self.config.n_reflection_steps > 0:
+            # Check if we have canopy - use optimized fast path if not
+            has_canopy = self.domain.lad is not None and self.config.canopy_radiation
+            
+            # Determine which optimized path to use
+            canopy_ready = has_canopy and self._csf_matrix_cached and self._svf_matrix_cached
+            no_canopy_ready = not has_canopy and self._svf_matrix_cached
+            
+            if no_canopy_ready:
+                # === OPTIMIZED FAST PATH (no canopy) ===
+                # Reduces kernel launches by fusing operations
+                self._run_optimized_reflection_loop(self.config.n_reflection_steps)
+            elif canopy_ready:
+                # === OPTIMIZED FAST PATH (with canopy) ===
+                # Uses sparse matrices for both SVF and CSF
+                self._run_optimized_reflection_loop_with_canopy(self.config.n_reflection_steps)
+            else:
+                # === STANDARD PATH (fallback) ===
+                for refstep in range(self.config.n_reflection_steps):
+                    # PALM: surfoutsl = albedo_surf * surfins
+                    # PALM: surfoutsw = surfoutsw + surfoutsl
+                    # Fused: compute outgoing AND accumulate in one kernel
+                    self._compute_outgoing_and_accumulate()
+                    
+                    # PALM: pcbinsw += csf * surfoutsl(isurfsrc) * asrc * grid_volume_inverse
+                    # Surface → Canopy: absorption from surface reflections
+                    if has_canopy:
+                        self._accumulate_canopy_absorption_from_reflections(
+                            self.domain.lad,
+                            self.domain.is_solid,
+                            self.config.ext_coef,
+                            self._pcbinsw
+                        )
+                    
+                    # PALM: surfins(isurf) += svf(1,isvf) * svf(2,isvf) * surfoutsl(isurfsrc)
+                    # Surface → Surface reflections
+                    self._compute_surface_reflections()
+                    
+                    # === Canopy scattering within reflection loop ===
+                    if self.domain.lad is not None and self.config.canopy_reflections:
+                        self._update_canopy_scattered_radiation(
+                            self.domain.lad,
+                            self.config.albedo_leaf
+                        )
+                        self._compute_canopy_to_surface_scattering(
+                            self.domain.lad,
+                            self.domain.is_solid,
+                            self.config.ext_coef
+                        )
+                        if self.config.canopy_to_canopy:
+                            self._compute_canopy_to_canopy_scattering(
+                                self.domain.lad,
+                                self.domain.is_solid,
+                                self.config.albedo_leaf,
+                                self.config.ext_coef
+                            )
+                            self._accumulate_canopy_to_canopy(
+                                self.domain.lad,
+                                self.config.albedo_leaf
+                            )
+                    
+                    # PALM: surfinsw = surfinsw + surfins
+                    self._accumulate_incoming()
+        else:
+            # No reflections - just compute single-bounce outgoing
+            self._compute_surface_outgoing_no_exchange()
+        
+        # Note: Canopy scattering is applied AFTER canopy radiation is computed
+        # (see compute_shortwave_radiation which calls _apply_canopy_scattering after CSF)
+        
+        # Copy final results to surface arrays
+        self._copy_final_fluxes()
+    
+    @ti.kernel
+    def _init_flux_arrays(self):
+        """Initialize all flux arrays to zero."""
+        # Use ti.cast for fp16 fields to avoid precision loss warnings
+        zero_f16 = ti.cast(0.0, ti.f16)
+        for i in range(self.n_surfaces):
+            self._surfins[i] = zero_f16
+            self._surfinl[i] = zero_f16
+            self._surfoutsl[i] = zero_f16
+            self._surfoutll[i] = zero_f16
+            self._surfinsw[i] = 0.0
+            self._surfinlw[i] = 0.0
+            self._surfoutsw[i] = 0.0
+            self._surfoutlw[i] = 0.0
+            self._surfinswdir[i] = 0.0
+            self._surfinswdif[i] = 0.0
+            self._surfinswpc[i] = zero_f16  # From plant canopy
+    
+    @ti.kernel
+    def _init_canopy_arrays(self):
+        """Initialize canopy radiation arrays to zero."""
+        zero_f16 = ti.cast(0.0, ti.f16)
+        for i, j, k in ti.ndrange(self.domain.nx, self.domain.ny, self.domain.nz):
+            self._pcbinswref[i, j, k] = zero_f16
+            self._pcbinswc2c[i, j, k] = zero_f16
+            self._pcbinswc2c_total[i, j, k] = 0.0
+    
+    @ti.kernel
+    def _compute_initial_sw_pass(
+        self,
+        sw_direct: ti.f32,
+        sw_diffuse: ti.f32,
+        cos_zenith: ti.f32
+    ):
+        """
+        Compute initial radiation pass: direct solar + diffuse sky radiation.
+        
+        This is the first pass before any surface reflections.
+        """
+        # Minimum stable cosine of zenith angle (PALM default)
+        min_stable_coszen = 0.0262
+        
+        for i in range(self.n_surfaces):
+            direction = self.surfaces.direction[i]
+            svf = self.surfaces.svf[i]
+            shadow = self.surfaces.shadow_factor[i]
+            canopy_trans = self.surfaces.canopy_transmissivity[i]
+            
+            # Get surface normal
+            normal = Vector3(0.0, 0.0, 0.0)
+            if direction == 0:  # Up
+                normal = Vector3(0.0, 0.0, 1.0)
+            elif direction == 1:  # Down
+                normal = Vector3(0.0, 0.0, -1.0)
+            elif direction == 2:  # North
+                normal = Vector3(0.0, 1.0, 0.0)
+            elif direction == 3:  # South
+                normal = Vector3(0.0, -1.0, 0.0)
+            elif direction == 4:  # East
+                normal = Vector3(1.0, 0.0, 0.0)
+            elif direction == 5:  # West
+                normal = Vector3(-1.0, 0.0, 0.0)
+            
+            # Sun direction
+            sun_dir = self.solar_calc.sun_direction[None]
+            
+            # Cosine of incidence angle (angle between sun and surface normal)
+            cos_incidence = (sun_dir[0] * normal[0] + 
+                            sun_dir[1] * normal[1] + 
+                            sun_dir[2] * normal[2])
+            cos_incidence = ti.max(0.0, cos_incidence)
+            
+            # Direct radiation
+            # PALM formula: surfinswdir = rad_sw_in_dir * surf_costh * dsitrans * sun_direct_factor
+            # where sun_direct_factor = 1 / max(min_stable_coszen, cos_zenith)
+            # 
+            # PALM's rad_sw_in_dir is Direct Horizontal Irradiance, so it multiplies by
+            # sun_direct_factor to convert to DNI, then by surf_costh for surface projection.
+            # 
+            # palm_solar assumes sw_direct input is already DNI (Direct Normal Irradiance),
+            # so we only need: sw_direct * cos_incidence * canopy_trans
+            # This is equivalent to PALM when: sw_direct = rad_sw_in_dir * sun_direct_factor
+            sw_in_dir = 0.0
+            if cos_zenith > min_stable_coszen and shadow < 0.5:
+                sw_in_dir = sw_direct * cos_incidence * canopy_trans
+            
+            # Diffuse radiation from sky (weighted by sky view factor)
+            # The input sw_diffuse is diffuse horizontal irradiance (DHI)
+            # 
+            # For vertical surfaces, the SVF already incorporates the geometric
+            # factor that vertical walls can only see half the sky hemisphere.
+            # The SVF calculation uses view factor fractions weighted by cos(angle)
+            # and normalizes against the theoretical maximum for that surface type.
+            # 
+            # Therefore: sw_diffuse * svf gives the correct diffuse irradiance
+            # for all surface orientations.
+            
+            # Diffuse radiation
+            # PALM formula: surfinswdif = rad_sw_in_diff * skyvft
+            # where skyvft is the transmissivity-weighted sky view factor
+            # palm_solar's svf is equivalent to PALM's skyvft (computed in svf.py)
+            sw_in_dif = 0.0
+            if direction == 0:  # Upward facing - full hemisphere (PALM: iup)
+                sw_in_dif = sw_diffuse * svf
+            elif direction == 1:  # Downward facing - cannot see sky (PALM: idown)
+                # Downward surfaces face away from sky, receive no sky diffuse
+                sw_in_dif = 0.0
+            else:  # Vertical walls (PALM: inorth, isouth, ieast, iwest)
+                sw_in_dif = sw_diffuse * svf
+            
+            # Store initial pass results
+            self._surfinswdir[i] = sw_in_dir
+            self._surfinswdif[i] = sw_in_dif
+            
+            # Initial incoming for reflection loop (cast to fp16)
+            self._surfins[i] = ti.cast(sw_in_dir + sw_in_dif, ti.f16)
+            
+            # Accumulate to totals
+            self._surfinsw[i] = self._surfins[i]
+    
+    @ti.kernel
+    def _compute_surface_outgoing(self):
+        """
+        Compute outgoing (reflected) radiation from each surface.
+        
+        PALM formula: surfoutsl = albedo * surfins
+        """
+        for i in range(self.n_surfaces):
+            albedo = self.surfaces.albedo[i]
+            self._surfoutsl[i] = ti.cast(albedo * self._surfins[i], ti.f16)
+    
+    @ti.kernel
+    def _compute_outgoing_and_accumulate(self):
+        """
+        Fused kernel: Compute outgoing AND accumulate to totals in one pass.
+        
+        Combines _compute_surface_outgoing + _accumulate_outgoing.
+        Reduces kernel launch overhead by 50% for this operation.
+        """
+        for i in range(self.n_surfaces):
+            albedo = self.surfaces.albedo[i]
+            outgoing = albedo * self._surfins[i]
+            self._surfoutsl[i] = ti.cast(outgoing, ti.f16)
+            self._surfoutsw[i] += outgoing
+    
+    @ti.kernel
+    def _compute_total_reflected(self):
+        """
+        Compute total reflected flux weighted by area (parallel reduction).
+        
+        This is O(n) and fully parallelized by Taichi.
+        """
+        self._total_reflected_flux[None] = 0.0
+        self._total_reflecting_area[None] = 0.0
+        
+        for i in range(self.n_surfaces):
+            area_i = self.surfaces.area[i]
+            flux_i = self._surfoutsl[i] * area_i
+            ti.atomic_add(self._total_reflected_flux[None], flux_i)
+            ti.atomic_add(self._total_reflecting_area[None], area_i)
+    
+    @ti.kernel
+    def _distribute_reflected_radiation(self):
+        """
+        Distribute reflected radiation based on distance-weighted view factors.
+        
+        PALM uses pre-computed sparse SVF matrix:
+            surfins(isurf) += svf(1,isvf) * svf(2,isvf) * surfoutsl(isurfsrc)
+        where svf(1,isvf) is the geometric view factor and svf(2,isvf) is transmissivity.
+        
+        palm_solar computes view factors dynamically for efficiency on GPU:
+        - Distance between surfaces (inverse square law)
+        - Orientation (dot product of normals with connection vector)
+        - Visibility (simplified - assume visible if facing each other)
+        
+        This gives equivalent physics with different numerical implementation.
+        The O(n^2) pairwise computation is fully parallelized on GPU.
+        """
+        PI = 3.14159265359
+        
+        for i in range(self.n_surfaces):
+            # Receiving surface properties
+            pos_i = self.surfaces.center[i]
+            normal_i = self.surfaces.normal[i]
+            urban_vf_i = 1.0 - self.surfaces.svf[i]
+            
+            # Skip if surface sees only sky
+            if urban_vf_i < 0.01:
+                self._surfins[i] = ti.cast(0.0, ti.f16)
+                continue
+            
+            # Accumulate contributions from all emitting surfaces
+            total_incoming = 0.0
+            
+            for j in range(self.n_surfaces):
+                if i == j:
+                    continue
+                
+                outgoing_j = self._surfoutsl[j]
+                if outgoing_j < 0.01:
+                    continue
+                
+                # Emitting surface properties
+                pos_j = self.surfaces.center[j]
+                normal_j = self.surfaces.normal[j]
+                area_j = self.surfaces.area[j]
+                
+                # Vector from j to i
+                dx = pos_i[0] - pos_j[0]
+                dy = pos_i[1] - pos_j[1]
+                dz = pos_i[2] - pos_j[2]
+                dist_sq = dx*dx + dy*dy + dz*dz
+                
+                if dist_sq < 0.1:
+                    continue
+                
+                dist = ti.sqrt(dist_sq)
+                
+                # Unit vector from j to i
+                dir_x = dx / dist
+                dir_y = dy / dist
+                dir_z = dz / dist
+                
+                # Cosine of angle at emitting surface (must be positive - facing towards i)
+                cos_emit = normal_j[0]*dir_x + normal_j[1]*dir_y + normal_j[2]*dir_z
+                
+                # Cosine of angle at receiving surface (must be positive - facing towards j)
+                cos_recv = -(normal_i[0]*dir_x + normal_i[1]*dir_y + normal_i[2]*dir_z)
+                
+                # Both surfaces must face each other
+                if cos_emit > 0.0 and cos_recv > 0.0:
+                    # Radiative view factor formula:
+                    # F_ij = (cos_emit * cos_recv * area_j) / (pi * dist^2)
+                    # Incoming irradiance from j = outgoing_j * F_ij
+                    view_factor = (cos_emit * cos_recv * area_j) / (PI * dist_sq)
+                    
+                    # Clamp to reasonable maximum
+                    view_factor = ti.min(view_factor, 1.0)
+                    
+                    total_incoming += outgoing_j * view_factor
+            
+            # Scale by urban view factor (what fraction of hemisphere sees urban surfaces)
+            self._surfins[i] = ti.cast(total_incoming * urban_vf_i, ti.f16)
+    
+    @ti.kernel
+    def _distribute_reflected_radiation_with_canopy(
+        self,
+        lad: ti.template(),
+        is_solid: ti.template(),
+        ext_coef: ti.f32,
+        albedo_leaf: ti.f32
+    ):
+        """
+        Distribute reflected radiation with LAD transmissivity.
+        
+        For each surface pair, traces ray through canopy applying Beer-Lambert
+        attenuation. Radiation absorbed by canopy is partially scattered
+        back to surfaces (based on leaf albedo).
+        
+        Args:
+            lad: 3D field of Leaf Area Density
+            is_solid: 3D field of solid cells
+            ext_coef: Extinction coefficient for canopy
+            albedo_leaf: Leaf/tree albedo for scattering
+        """
+        PI = 3.14159265359
+        nx = self.domain.nx
+        ny = self.domain.ny
+        nz = self.domain.nz
+        dx = self.domain.dx
+        dy = self.domain.dy
+        dz = self.domain.dz
+        
+        for i in range(self.n_surfaces):
+            # Receiving surface properties
+            pos_i = self.surfaces.center[i]
+            normal_i = self.surfaces.normal[i]
+            urban_vf_i = 1.0 - self.surfaces.svf[i]
+            
+            # Skip if surface sees only sky
+            if urban_vf_i < 0.01:
+                self._surfins[i] = 0.0
+                continue
+            
+            # Accumulate contributions from all emitting surfaces
+            total_incoming = 0.0
+            
+            for j in range(self.n_surfaces):
+                if i == j:
+                    continue
+                
+                outgoing_j = self._surfoutsl[j]
+                if outgoing_j < 0.01:
+                    continue
+                
+                # Emitting surface properties
+                pos_j = self.surfaces.center[j]
+                normal_j = self.surfaces.normal[j]
+                area_j = self.surfaces.area[j]
+                
+                # Vector from j to i
+                diff_x = pos_i[0] - pos_j[0]
+                diff_y = pos_i[1] - pos_j[1]
+                diff_z = pos_i[2] - pos_j[2]
+                dist_sq = diff_x*diff_x + diff_y*diff_y + diff_z*diff_z
+                
+                if dist_sq < 0.1:
+                    continue
+                
+                dist = ti.sqrt(dist_sq)
+                
+                # Unit vector from j to i
+                dir_x = diff_x / dist
+                dir_y = diff_y / dist
+                dir_z = diff_z / dist
+                
+                # Cosine of angle at emitting surface
+                cos_emit = normal_j[0]*dir_x + normal_j[1]*dir_y + normal_j[2]*dir_z
+                
+                # Cosine of angle at receiving surface
+                cos_recv = -(normal_i[0]*dir_x + normal_i[1]*dir_y + normal_i[2]*dir_z)
+                
+                # Both surfaces must face each other
+                if cos_emit > 0.0 and cos_recv > 0.0:
+                    # Compute transmissivity through canopy
+                    trans, blocked = ray_point_to_point_transmissivity(
+                        pos_j, pos_i,
+                        lad, is_solid,
+                        nx, ny, nz,
+                        dx, dy, dz,
+                        ext_coef
+                    )
+                    
+                    # Skip if blocked by solid
+                    if blocked == 1:
+                        continue
+                    
+                    # View factor
+                    view_factor = (cos_emit * cos_recv * area_j) / (PI * dist_sq)
+                    view_factor = ti.min(view_factor, 1.0)
+                    
+                    # Apply canopy transmissivity
+                    total_incoming += outgoing_j * view_factor * trans
+            
+            # Scale by urban view factor
+            self._surfins[i] = total_incoming * urban_vf_i
+    
+    def _compute_surface_reflections(self):
+        """
+        Compute radiation exchange between surfaces for one reflection step.
+        
+        If canopy_reflections is enabled and LAD exists, uses ray tracing
+        through vegetation with Beer-Lambert attenuation. Otherwise uses
+        simplified geometry-based distribution.
+        
+        Uses two-pass algorithm:
+        1. Parallel reduction to compute total reflected flux
+        2. Parallel distribution based on view factors (with optional LAD)
+        
+        Note: For optimized multi-step reflection, use _compute_surface_reflections_optimized()
+        which uses separate kernels with ping-pong buffers for ~100x GPU speedup.
+        """
+        # Pass 1: Compute total reflected flux (parallel reduction)
+        self._compute_total_reflected()
+        
+        # Pass 2: Distribute to each surface
+        if self.config.canopy_reflections and self.domain.lad is not None:
+            # Use canopy-aware version with LAD transmissivity
+            if self.config.cache_svf_matrix and self._svf_matrix_cached:
+                # Use cached SVF matrix with optimized separate kernels
+                self._distribute_reflected_cached_single_step()
+            else:
+                # Compute dynamically (O(n²))
+                self._distribute_reflected_radiation_with_canopy(
+                    self.domain.lad,
+                    self.domain.is_solid,
+                    self.config.ext_coef,
+                    self.config.albedo_leaf
+                )
+        else:
+            # Use simple version without canopy
+            if self.config.cache_svf_matrix and self._svf_matrix_cached:
+                self._distribute_reflected_cached_single_step()
+            else:
+                self._distribute_reflected_radiation()
+    
+    def _distribute_reflected_cached_single_step(self):
+        """
+        Single-step reflection distribution using optimized separate kernels.
+        
+        Uses separate kernels instead of fused kernel with internal ti.sync()
+        for ~100x GPU speedup. Each kernel is fully parallel without barriers.
+        """
+        n = self.n_surfaces
+        svf_nnz = self._svf_nnz
+        
+        # Reset incoming buffer
+        self._reset_buffer(self._surfins, n)
+        
+        # Sparse matmul for reflection distribution
+        self._sparse_matmul_step(self._surfoutsl, self._surfins, svf_nnz)
+        
+        # Scale by urban view factor
+        self._scale_by_urban_vf(self._surfins, n)
+    
+    def compute_svf_matrix(self):
+        """
+        Pre-compute surface-to-surface view factor matrix (PALM-like approach).
+        
+        This is expensive O(n²) but only needs to be done once for fixed geometry.
+        Subsequent reflection iterations become O(nnz) instead of O(n²).
+        
+        Call this before running multi-timestep simulations for efficiency.
+        """
+        if self._svf_matrix_cached:
+            print("SVF matrix already cached, skipping recomputation.")
+            return
+        
+        # Check if arrays were allocated in __init__
+        if self._svf_source is None:
+            print("Warning: SVF caching not enabled in config, skipping matrix computation.")
+            return
+        
+        print(f"Pre-computing SVF matrix for {self.n_surfaces} surfaces...")
+        print("  This is O(n²) but only runs once for fixed geometry.")
+        
+        # Compute the matrix
+        if self.domain.lad is not None and self.config.canopy_reflections:
+            self._compute_svf_matrix_with_canopy(
+                self.domain.lad,
+                self.domain.is_solid,
+                self.config.ext_coef,
+                self.config.svf_min_threshold
+            )
+        else:
+            self._compute_svf_matrix_simple(self.config.svf_min_threshold)
+        
+        # Clamp nnz to max entries to avoid out-of-bounds reads
+        computed_nnz = int(self._svf_count[None])
+        if computed_nnz > self._max_svf_entries:
+            truncated_pct = (computed_nnz - self._max_svf_entries) / computed_nnz * 100
+            print(f"Warning: SVF matrix truncated! Computed {computed_nnz:,} entries but buffer size is {self._max_svf_entries:,}.")
+            print(f"  {truncated_pct:.1f}% of surface-to-surface view factors are being discarded.")
+            print(f"  This may affect reflection accuracy. To fix: clear_all_caches() before creating this model,")
+            print(f"  or increase svf_min_threshold in RadiationConfig to reduce entries.")
+            self._svf_nnz = self._max_svf_entries
+        else:
+            self._svf_nnz = computed_nnz
+            
+        self._svf_matrix_cached = True
+        
+        sparsity = self._svf_nnz / (self.n_surfaces * self.n_surfaces) * 100
+        memory_mb = self._svf_nnz * 12 / 1e6  # 12 bytes per entry (2 int32 + 2 float16)
+        print(f"  SVF matrix computed: {self._svf_nnz:,} non-zero entries ({memory_mb:.1f} MB)")
+        print(f"  Sparsity: {sparsity:.2f}% of full matrix")
+        print(f"  Speedup factor: ~{self.n_surfaces * self.n_surfaces / max(1, self._svf_nnz):.1f}x per reflection step")
+        
+        # Build CSR format for optimized sparse matmul
+        if self._svf_csr_row_ptr is not None:
+            self._build_csr_format()
+    
+    def _build_csr_format(self):
+        """
+        Convert COO format to CSR format for optimized sparse matmul.
+        
+        CSR (Compressed Sparse Row by target) provides:
+        1. O(n_surfaces) parallelism with one thread per row
+        2. Local accumulation instead of atomic operations
+        3. Better cache locality for reading source values
+        
+        Reduces sparse matmul time by ~2-5x on GPU.
+        """
+        import numpy as np
+        
+        print("  Building CSR format for optimized sparse matmul...")
+        
+        # Copy COO data to numpy for sorting
+        coo_target = self._svf_target.to_numpy()[:self._svf_nnz]
+        coo_source = self._svf_source.to_numpy()[:self._svf_nnz]
+        coo_vf = self._svf_vf.to_numpy()[:self._svf_nnz]
+        coo_trans = self._svf_trans.to_numpy()[:self._svf_nnz]
+        
+        # Pre-multiply vf * trans
+        coo_val = coo_vf * coo_trans
+        
+        # Sort by target (row) for CSR format
+        sort_idx = np.argsort(coo_target)
+        sorted_target = coo_target[sort_idx]
+        sorted_source = coo_source[sort_idx]
+        sorted_val = coo_val[sort_idx]
+        
+        # Build row pointers
+        row_ptr = np.zeros(self.n_surfaces + 1, dtype=np.int32)
+        for t in sorted_target:
+            row_ptr[t + 1] += 1
+        row_ptr = np.cumsum(row_ptr)
+        
+        # Pad arrays to match Taichi field shape (required for from_numpy)
+        # The fields are allocated to _max_svf_entries, but we only use _svf_nnz
+        padded_col_idx = np.zeros(self._max_svf_entries, dtype=np.int32)
+        padded_val = np.zeros(self._max_svf_entries, dtype=np.float32)
+        padded_col_idx[:self._svf_nnz] = sorted_source
+        padded_val[:self._svf_nnz] = sorted_val
+        
+        # Copy to Taichi fields
+        self._svf_csr_row_ptr.from_numpy(row_ptr)
+        self._svf_csr_col_idx.from_numpy(padded_col_idx)
+        self._svf_csr_val.from_numpy(padded_val)
+        
+        self._svf_csr_ready = True
+        print(f"  CSR format ready: {self.n_surfaces} rows, {self._svf_nnz:,} entries")
+    
+    def compute_csf_matrix(self):
+        """
+        Pre-compute Canopy-Surface Factor matrix.
+        
+        Stores geometric factors for Surface <-> Canopy interactions.
+        This allows O(nnz) computation for canopy absorption and scattering
+        instead of O(N_cells * N_surfaces).
+        """
+        if self._csf_matrix_cached:
+            return
+            
+        if self._csf_canopy_idx is None:
+            return
+            
+        print(f"Pre-computing CSF matrix for canopy-surface interactions...")
+        
+        self._compute_csf_matrix_kernel(
+            self.domain.lad,
+            self.domain.is_solid,
+            self.config.ext_coef,
+            1e-7 # Threshold
+        )
+        
+        computed_nnz = int(self._csf_count[None])
+        if computed_nnz > self._max_csf_entries:
+            print(f"Warning: CSF matrix truncated! {computed_nnz} > {self._max_csf_entries}")
+            self._csf_nnz = self._max_csf_entries
+        else:
+            self._csf_nnz = computed_nnz
+            
+        self._csf_matrix_cached = True
+        print(f"  CSF matrix computed: {self._csf_nnz:,} entries")
+
+    @ti.kernel
+    def _compute_csf_matrix_kernel(
+        self,
+        lad: ti.template(),
+        is_solid: ti.template(),
+        ext_coef: ti.f32,
+        min_threshold: ti.f32
+    ):
+        """
+        Compute CSF matrix entries.
+        
+        Stores (canopy_idx, surface_idx, base_factor) where:
+        base_factor = (cell_area * cos_surf * trans) / (4 * PI * dist_sq)
+        
+        This factor is used for:
+        1. Surface->Canopy: absorbed = outgoing * area_surf * base_factor * abs_frac / grid_vol
+        2. Canopy->Surface: incoming = scattered_power * base_factor / area_surf
+        """
+        PI = 3.14159265359
+        nx = self.domain.nx
+        ny = self.domain.ny
+        nz = self.domain.nz
+        dx = self.domain.dx
+        dy = self.domain.dy
+        dz = self.domain.dz
+        
+        # Reset counter
+        self._csf_count[None] = 0
+        
+        # Iterate over all canopy cells
+        for ci, cj, ck in ti.ndrange(nx, ny, nz):
+            cell_lad = lad[ci, cj, ck]
+            if cell_lad <= 0.0:
+                continue
+                
+            # Cell center
+            pos_cell = Vector3(
+                (ci + 0.5) * dx,
+                (cj + 0.5) * dy,
+                (ck + 0.5) * dz
+            )
+            
+            # Linear index for canopy cell
+            canopy_idx = ci * (ny * nz) + cj * nz + ck
+            
+            # Iterate over all surfaces
+            for surf_i in range(self.n_surfaces):
+                pos_surf = self.surfaces.center[surf_i]
+                normal_surf = self.surfaces.normal[surf_i]
+                
+                # Vector from surface to cell
+                diff = pos_cell - pos_surf
+                dist_sq = diff[0]*diff[0] + diff[1]*diff[1] + diff[2]*diff[2]
+                
+                if dist_sq < 0.1:
+                    continue
+                    
+                dist = ti.sqrt(dist_sq)
+                
+                # Direction from surface to cell
+                dir_x = diff[0] / dist
+                dir_y = diff[1] / dist
+                dir_z = diff[2] / dist
+                
+                # Check if surface faces the cell
+                cos_emit = (normal_surf[0]*dir_x + normal_surf[1]*dir_y + normal_surf[2]*dir_z)
+                
+                if cos_emit > 0.0:
+                    # Geometric factor (solid angle / 4pi * cos)
+                    # cell_solid_angle = (dx * dy) / dist_sq  (approx)
+                    # factor = (cell_solid_angle * cos_emit) / (4 * PI)
+                    # factor = (dx * dy * cos_emit) / (4 * PI * dist_sq)
+                    
+                    base_factor = (dx * dy * cos_emit) / (4.0 * PI * dist_sq)
+                    
+                    if base_factor > min_threshold:
+                        # Check transmissivity
+                        trans, blocked = ray_point_to_point_transmissivity(
+                            pos_surf, pos_cell,
+                            lad, is_solid,
+                            nx, ny, nz,
+                            dx, dy, dz,
+                            ext_coef
+                        )
+                        
+                        if blocked == 0 and trans > 0.01:
+                            final_factor = base_factor * trans
+                            
+                            if final_factor > min_threshold:
+                                idx = ti.atomic_add(self._csf_count[None], 1)
+                                if idx < self._max_csf_entries:
+                                    self._csf_canopy_idx[idx] = canopy_idx
+                                    self._csf_surface_idx[idx] = surf_i
+                                    self._csf_val[idx] = final_factor
+
+    @ti.kernel
+    def _compute_svf_matrix_simple(self, min_threshold: ti.f32):
+        """
+        Compute SVF matrix without canopy (simple geometry).
+        
+        Stores entries where view factor > min_threshold in sparse COO format.
+        """
+        PI = 3.14159265359
+        
+        # Reset counter
+        self._svf_count[None] = 0
+        
+        # Compute all pairwise view factors
+        for i in range(self.n_surfaces):
+            pos_i = self.surfaces.center[i]
+            normal_i = self.surfaces.normal[i]
+            
+            for j in range(self.n_surfaces):
+                if i == j:
+                    continue
+                
+                pos_j = self.surfaces.center[j]
+                normal_j = self.surfaces.normal[j]
+                area_j = self.surfaces.area[j]
+                
+                # Vector from j to i
+                dx = pos_i[0] - pos_j[0]
+                dy = pos_i[1] - pos_j[1]
+                dz = pos_i[2] - pos_j[2]
+                dist_sq = dx*dx + dy*dy + dz*dz
+                
+                if dist_sq < 0.1:
+                    continue
+                
+                dist = ti.sqrt(dist_sq)
+                dir_x = dx / dist
+                dir_y = dy / dist
+                dir_z = dz / dist
+                
+                # Cosines
+                cos_emit = normal_j[0]*dir_x + normal_j[1]*dir_y + normal_j[2]*dir_z
+                cos_recv = -(normal_i[0]*dir_x + normal_i[1]*dir_y + normal_i[2]*dir_z)
+                
+                if cos_emit > 0.0 and cos_recv > 0.0:
+                    view_factor = (cos_emit * cos_recv * area_j) / (PI * dist_sq)
+                    view_factor = ti.min(view_factor, 1.0)
+                    
+                    if view_factor > min_threshold:
+                        # Atomic increment to get unique index
+                        idx = ti.atomic_add(self._svf_count[None], 1)
+                        if idx < self._max_svf_entries:
+                            self._svf_source[idx] = j
+                            self._svf_target[idx] = i
+                            self._svf_vf[idx] = view_factor
+                            self._svf_trans[idx] = 1.0  # No canopy
+    
+    @ti.kernel
+    def _compute_svf_matrix_with_canopy(
+        self,
+        lad: ti.template(),
+        is_solid: ti.template(),
+        ext_coef: ti.f32,
+        min_threshold: ti.f32
+    ):
+        """
+        Compute SVF matrix with canopy transmissivity.
+        
+        For each surface pair, computes view factor AND transmissivity
+        through intervening vegetation using Beer-Lambert law.
+        
+        Stores entries where (vf * trans) > min_threshold.
+        """
+        PI = 3.14159265359
+        nx = self.domain.nx
+        ny = self.domain.ny
+        nz = self.domain.nz
+        dx = self.domain.dx
+        dy = self.domain.dy
+        dz = self.domain.dz
+        
+        # Reset counter
+        self._svf_count[None] = 0
+        
+        for i in range(self.n_surfaces):
+            pos_i = self.surfaces.center[i]
+            normal_i = self.surfaces.normal[i]
+            
+            for j in range(self.n_surfaces):
+                if i == j:
+                    continue
+                
+                pos_j = self.surfaces.center[j]
+                normal_j = self.surfaces.normal[j]
+                area_j = self.surfaces.area[j]
+                
+                # Vector from j to i
+                diff_x = pos_i[0] - pos_j[0]
+                diff_y = pos_i[1] - pos_j[1]
+                diff_z = pos_i[2] - pos_j[2]
+                dist_sq = diff_x*diff_x + diff_y*diff_y + diff_z*diff_z
+                
+                if dist_sq < 0.1:
+                    continue
+                
+                dist = ti.sqrt(dist_sq)
+                dir_x = diff_x / dist
+                dir_y = diff_y / dist
+                dir_z = diff_z / dist
+                
+                cos_emit = normal_j[0]*dir_x + normal_j[1]*dir_y + normal_j[2]*dir_z
+                cos_recv = -(normal_i[0]*dir_x + normal_i[1]*dir_y + normal_i[2]*dir_z)
+                
+                if cos_emit > 0.0 and cos_recv > 0.0:
+                    # Compute transmissivity through canopy
+                    trans, blocked = ray_point_to_point_transmissivity(
+                        pos_j, pos_i,
+                        lad, is_solid,
+                        nx, ny, nz,
+                        dx, dy, dz,
+                        ext_coef
+                    )
+                    
+                    if blocked == 0 and trans > 0.001:
+                        view_factor = (cos_emit * cos_recv * area_j) / (PI * dist_sq)
+                        view_factor = ti.min(view_factor, 1.0)
+                        
+                        effective_vf = view_factor * trans
+                        
+                        if effective_vf > min_threshold:
+                            idx = ti.atomic_add(self._svf_count[None], 1)
+                            if idx < self._max_svf_entries:
+                                self._svf_source[idx] = j
+                                self._svf_target[idx] = i
+                                self._svf_vf[idx] = view_factor
+                                self._svf_trans[idx] = trans
+    
+    # ========== Optimized Reflection Kernels (Separate for GPU efficiency) ==========
+    # Using separate kernels without internal ti.sync() is ~100x faster than fused kernels
+    # with ti.sync() inside. Taichi handles synchronization between kernel calls efficiently.
+    
+    @ti.kernel
+    def _reset_buffer(self, buf: ti.template(), n: ti.i32):
+        """Reset a buffer to zero - vectorized for GPU efficiency."""
+        for i in range(n):
+            buf[i] = 0.0
+    
+    @ti.kernel
+    def _compute_outgoing_step(self, surfins: ti.template(), surfout: ti.template(), n: ti.i32):
+        """Compute outgoing radiation: surfout = albedo * surfins."""
+        for i in range(n):
+            surfout[i] = self.surfaces.albedo[i] * surfins[i]
+    
+    @ti.kernel
+    def _sparse_matmul_step(
+        self,
+        surfout: ti.template(),
+        surfins_next: ti.template(),
+        svf_nnz: ti.i32
+    ):
+        """
+        Sparse matrix-vector multiply for reflection distribution.
+        
+        PALM equivalent: surfins(isurf) += svf(1,isvf) * svf(2,isvf) * surfoutsl(isurfsrc)
+        
+        Optimized for GPU with:
+        - Coalesced memory access through sparse COO format
+        - Minimal thread divergence with early threshold check
+        - Efficient atomic operations for parallel accumulation
+        """
+        for idx in range(svf_nnz):
+            source = self._svf_source[idx]
+            target = self._svf_target[idx]
+            vf = self._svf_vf[idx]
+            trans = self._svf_trans[idx]
+            
+            # Pre-multiply vf * trans to reduce FLOPs
+            vf_trans = vf * trans
+            outgoing = surfout[source]
+            
+            # Use threshold to skip negligible contributions
+            if outgoing * vf_trans > 0.001:
+                ti.atomic_add(surfins_next[target], outgoing * vf_trans)
+    
+    @ti.kernel
+    def _sparse_matmul_step_batched(
+        self,
+        surfout: ti.template(),
+        surfins_next: ti.template(),
+        svf_nnz: ti.i32,
+        batch_size: ti.i32
+    ):
+        """
+        Batched sparse matrix-vector multiply for improved GPU utilization.
+        
+        Processes multiple sparse entries per thread to improve memory locality
+        and reduce atomic operation contention.
+        """
+        n_batches = (svf_nnz + batch_size - 1) // batch_size
+        
+        for batch_idx in range(n_batches):
+            start_idx = batch_idx * batch_size
+            end_idx = ti.min(start_idx + batch_size, svf_nnz)
+            
+            # Local accumulator to reduce atomics
+            for idx in range(start_idx, end_idx):
+                source = self._svf_source[idx]
+                target = self._svf_target[idx]
+                vf_trans = self._svf_vf[idx] * self._svf_trans[idx]
+                outgoing = surfout[source]
+                
+                if outgoing * vf_trans > 0.001:
+                    ti.atomic_add(surfins_next[target], outgoing * vf_trans)
+    
+    @ti.kernel
+    def _scale_by_urban_vf(self, surfins: ti.template(), n: ti.i32):
+        """Scale incoming by urban view factor (1 - SVF)."""
+        for i in range(n):
+            urban_vf_i = 1.0 - self.surfaces.svf[i]
+            if urban_vf_i < 0.01:
+                surfins[i] = 0.0
+            else:
+                surfins[i] *= urban_vf_i
+    
+    @ti.kernel
+    def _sparse_matmul_csr(
+        self,
+        surfout: ti.template(),
+        surfins_next: ti.template(),
+        n_surfaces: ti.i32
+    ):
+        """
+        CSR-format sparse matrix-vector multiply for reflection distribution.
+        
+        MASSIVELY optimized for GPU:
+        - One thread per target surface (row) = perfect parallelism
+        - No atomic operations needed (each row is processed by one thread)
+        - Local accumulation in registers before final write
+        - Better cache locality from contiguous column access
+        
+        Reduces atomic operations from O(nnz) = 45M to O(n) = 42K
+        Expected speedup: 2-5x over COO format with atomics.
+        """
+        for row in range(n_surfaces):
+            row_start = self._svf_csr_row_ptr[row]
+            row_end = self._svf_csr_row_ptr[row + 1]
+            
+            # Local accumulator - no atomics needed!
+            local_sum = 0.0
+            
+            for idx in range(row_start, row_end):
+                source = self._svf_csr_col_idx[idx]
+                vf_trans = self._svf_csr_val[idx]  # Pre-multiplied vf * trans
+                outgoing = surfout[source]
+                
+                # Threshold check for negligible contributions
+                if outgoing * vf_trans > 0.001:
+                    local_sum += outgoing * vf_trans
+            
+            # Single write per row (no atomic needed since one thread per row)
+            surfins_next[row] = local_sum
+    
+    @ti.kernel
+    def _accumulate_reflected(self, surfins: ti.template(), n: ti.i32):
+        """Accumulate incoming reflected radiation to totals."""
+        for i in range(n):
+            self._surfins[i] += surfins[i]
+    
+    def _distribute_reflected_cached_optimized(self, svf_nnz: int, use_ping: bool):
+        """
+        Optimized reflection distribution using separate kernels.
+        
+        This is ~100x faster than the fused kernel approach because:
+        1. No ti.sync() inside kernels (Taichi handles sync between kernel calls)
+        2. Each kernel is fully parallel without internal barriers
+        3. GPU can overlap kernel launches with computation
+        
+        Args:
+            svf_nnz: Number of non-zero entries in SVF matrix
+            use_ping: If True, read from ping buffer and write to pong buffer
+        """
+        n = self.n_surfaces
+        
+        if use_ping:
+            # Read from ping, write to pong
+            self._compute_outgoing_step(self._surfins_ping, self._surfoutsl, n)
+            self._reset_buffer(self._surfins_pong, n)
+            self._sparse_matmul_step(self._surfoutsl, self._surfins_pong, svf_nnz)
+            self._scale_by_urban_vf(self._surfins_pong, n)
+            self._accumulate_reflected(self._surfins_pong, n)
+        else:
+            # Read from pong, write to ping
+            self._compute_outgoing_step(self._surfins_pong, self._surfoutsl, n)
+            self._reset_buffer(self._surfins_ping, n)
+            self._sparse_matmul_step(self._surfoutsl, self._surfins_ping, svf_nnz)
+            self._scale_by_urban_vf(self._surfins_ping, n)
+            self._accumulate_reflected(self._surfins_ping, n)
+        
+        # Accumulate outgoing to totals
+        self._accumulate_outgoing()
+    
+    @ti.kernel
+    def _init_ping_buffer(self, n: ti.i32):
+        """Initialize ping buffer with initial SW radiation for reflection loop."""
+        for i in range(n):
+            self._surfins_ping[i] = self._surfinswdir[i] + self._surfinswdif[i]
+            self._surfins[i] = 0.0  # Reset accumulated reflected
+    
+    def _run_optimized_reflection_loop(self, n_steps: int):
+        """
+        Optimized reflection loop using fused kernels and ping-pong buffers.
+        
+        This is the GPU-optimized fast path for surface-only reflections
+        (no canopy). Achieves ~10-20x speedup over CPU by:
+        1. Fusing outgoing + accumulate into single kernel
+        2. Using separate kernels (no internal ti.sync())
+        3. Ping-pong buffers for efficient memory access
+        4. Minimizing kernel launch count
+        5. Using ultra-fused kernel for small reflection counts
+        
+        Only 4 kernel calls per reflection step vs 5+ in standard path.
+        
+        Args:
+            n_steps: Number of reflection iterations
+        """
+        n = self.n_surfaces
+        svf_nnz = self._svf_nnz
+        
+        # Use ultra-optimized path for typical reflection counts (1-5 steps)
+        # This reduces kernel launch overhead significantly
+        if n_steps <= 5:
+            self._run_reflection_loop_ultra_fused(n_steps, n, svf_nnz)
+            return
+        
+        for step in range(n_steps):
+            use_ping = (step % 2 == 0)
+            
+            if use_ping:
+                # Step 1: Compute outgoing from current incoming (fused with accumulate)
+                self._compute_outgoing_fused(self._surfins, n)
+                
+                # Step 2: Reset next buffer
+                self._reset_buffer(self._surfins_pong, n)
+                
+                # Step 3: Sparse matmul for reflection distribution
+                self._sparse_matmul_step(self._surfoutsl, self._surfins_pong, svf_nnz)
+                
+                # Step 4: Scale by urban VF and accumulate incoming (fused)
+                self._scale_and_accumulate_incoming(self._surfins_pong, n)
+                
+                # Copy to _surfins for next iteration
+                self._copy_buffer(self._surfins_pong, self._surfins, n)
+            else:
+                self._compute_outgoing_fused(self._surfins, n)
+                self._reset_buffer(self._surfins_ping, n)
+                self._sparse_matmul_step(self._surfoutsl, self._surfins_ping, svf_nnz)
+                self._scale_and_accumulate_incoming(self._surfins_ping, n)
+                self._copy_buffer(self._surfins_ping, self._surfins, n)
+    
+    def _run_reflection_loop_ultra_fused(self, n_steps: int, n: int, svf_nnz: int):
+        """
+        Ultra-optimized reflection loop with minimal kernel launches.
+        
+        For typical 3-step reflections, this reduces from 15+ kernel calls
+        to just 3-4 by doing more work per kernel.
+        
+        Uses CSR format if available for ~2-5x faster sparse matmul.
+        """
+        # Choose sparse matmul method: CSR (fast, no atomics) or COO (fallback)
+        use_csr = self._svf_csr_ready
+        
+        def do_sparse_matmul(surfout, surfins_next):
+            if use_csr:
+                # CSR format: one thread per row, no atomics
+                self._sparse_matmul_csr(surfout, surfins_next, n)
+            else:
+                # COO format fallback with atomics
+                self._reset_buffer(surfins_next, n)
+                self._sparse_matmul_step(surfout, surfins_next, svf_nnz)
+        
+        # Step 1: First reflection iteration with combined operations
+        # Compute outgoing and prepare for sparse matmul
+        self._compute_outgoing_fused(self._surfins, n)
+        do_sparse_matmul(self._surfoutsl, self._surfins_pong)
+        self._scale_and_accumulate_incoming(self._surfins_pong, n)
+        
+        if n_steps >= 2:
+            # Step 2: Second reflection - use pong as input
+            self._compute_outgoing_fused_from_buffer(self._surfins_pong, n)
+            do_sparse_matmul(self._surfoutsl, self._surfins_ping)
+            self._scale_and_accumulate_incoming(self._surfins_ping, n)
+        
+        if n_steps >= 3:
+            # Step 3: Third reflection - use ping as input
+            self._compute_outgoing_fused_from_buffer(self._surfins_ping, n)
+            do_sparse_matmul(self._surfoutsl, self._surfins_pong)
+            self._scale_and_accumulate_incoming(self._surfins_pong, n)
+        
+        # Handle remaining steps if any
+        for step in range(3, n_steps):
+            use_ping = (step % 2 == 1)
+            src = self._surfins_ping if use_ping else self._surfins_pong
+            dst = self._surfins_pong if use_ping else self._surfins_ping
+            
+            self._compute_outgoing_fused_from_buffer(src, n)
+            do_sparse_matmul(self._surfoutsl, dst)
+            self._scale_and_accumulate_incoming(dst, n)
+    
+    @ti.kernel
+    def _compute_outgoing_fused_from_buffer(self, surfins: ti.template(), n: ti.i32):
+        """Compute outgoing from a specific buffer and accumulate."""
+        for i in range(n):
+            outgoing = self.surfaces.albedo[i] * surfins[i]
+            self._surfoutsl[i] = outgoing
+            self._surfoutsw[i] += outgoing
+    
+    def _run_optimized_reflection_loop_with_canopy(self, n_steps: int):
+        """
+        Optimized reflection loop with canopy interactions.
+        
+        Uses sparse matrices for both Surface-Surface (SVF) and Surface-Canopy (CSF)
+        interactions, achieving O(nnz) complexity instead of O(N_surf * N_cell).
+        """
+        n = self.n_surfaces
+        svf_nnz = self._svf_nnz
+        csf_nnz = self._csf_nnz
+        
+        # Initialize ping buffer with initial radiation
+        self._init_ping_buffer(n)
+        
+        for step in range(n_steps):
+            use_ping = (step % 2 == 0)
+            
+            # Input buffer for this step (contains incoming radiation)
+            src_buf = self._surfins_ping if use_ping else self._surfins_pong
+            # Output buffer for next step (will accumulate reflected radiation)
+            dst_buf = self._surfins_pong if use_ping else self._surfins_ping
+            
+            # 1. Compute outgoing from surfaces (and accumulate to total outgoing)
+            self._compute_outgoing_fused(src_buf, n)
+            
+            # 2. Surface -> Canopy Absorption (using CSF matrix)
+            self._csf_absorb_step(self._surfoutsl, csf_nnz)
+            
+            # 3. Update Canopy Scattering (based on absorbed)
+            self._update_canopy_scattered_optimized(self.domain.lad, self.config.albedo_leaf)
+            
+            # 4. Reset destination buffer and do sparse matmul
+            if self._svf_csr_ready:
+                # CSR format: one thread per row, no atomics, includes reset
+                self._sparse_matmul_csr(self._surfoutsl, dst_buf, n)
+            else:
+                # COO format fallback
+                self._reset_buffer(dst_buf, n)
+                self._sparse_matmul_step(self._surfoutsl, dst_buf, svf_nnz)
+            
+            # 6. Canopy -> Surface Scattering (CSF matrix transposed)
+            self._csf_scatter_step(dst_buf, csf_nnz)
+            
+            # 7. Scale by urban VF and accumulate to total incoming
+            self._scale_and_accumulate_incoming(dst_buf, n)
+            
+            # Copy to _surfins for consistency (though not strictly needed for loop)
+            self._copy_buffer(dst_buf, self._surfins, n)
+
+    @ti.kernel
+    def _csf_absorb_step(self, surfout: ti.template(), csf_nnz: ti.i32):
+        """
+        Surface -> Canopy absorption using sparse CSF matrix.
+        """
+        grid_vol_inv = 1.0 / (self.domain.dx * self.domain.dy * self.domain.dz)
+        ext_coef = self.config.ext_coef
+        
+        for idx in range(csf_nnz):
+            canopy_idx = self._csf_canopy_idx[idx]
+            surf_idx = self._csf_surface_idx[idx]
+            base_factor = self._csf_val[idx]
+            
+            outgoing = surfout[surf_idx]
+            if outgoing > 0.01:
+                # Reconstruct 3D indices from linear index
+                # canopy_idx = i * (ny * nz) + j * nz + k
+                tmp = canopy_idx
+                k = tmp % self.domain.nz
+                tmp //= self.domain.nz
+                j = tmp % self.domain.ny
+                i = tmp // self.domain.ny
+                
+                lad_val = self.domain.lad[i, j, k]
+                area_surf = self.surfaces.area[surf_idx]
+                
+                # Absorption fraction (approximate path = dz)
+                abs_frac = 1.0 - ti.exp(-ext_coef * lad_val * self.domain.dz)
+                
+                # absorbed = outgoing * area_surf * base_factor * abs_frac / grid_vol
+                absorbed = outgoing * area_surf * base_factor * abs_frac * grid_vol_inv
+                
+                ti.atomic_add(self._pcbinsw[i, j, k], absorbed)
+
+    @ti.kernel
+    def _csf_scatter_step(self, surfins_next: ti.template(), csf_nnz: ti.i32):
+        """
+        Canopy -> Surface scattering using sparse CSF matrix.
+        """
+        grid_vol = self.domain.dx * self.domain.dy * self.domain.dz
+        
+        for idx in range(csf_nnz):
+            canopy_idx = self._csf_canopy_idx[idx]
+            surf_idx = self._csf_surface_idx[idx]
+            base_factor = self._csf_val[idx]
+            
+            # Reconstruct indices
+            tmp = canopy_idx
+            k = tmp % self.domain.nz
+            tmp //= self.domain.nz
+            j = tmp % self.domain.ny
+            i = tmp // self.domain.ny
+            
+            # Scattered power from this cell (W/m^3 * m^3 = W)
+            scattered_flux_vol = self._pcbinswref[i, j, k]
+            if scattered_flux_vol > 0.001:
+                scattered_power = scattered_flux_vol * grid_vol
+                
+                area_surf = self.surfaces.area[surf_idx]
+                
+                # incoming = scattered_power * base_factor / area_surf
+                contribution = scattered_power * base_factor / area_surf
+                
+                ti.atomic_add(surfins_next[surf_idx], contribution)
+                ti.atomic_add(self._surfinswpc[surf_idx], contribution)
+
+    @ti.kernel
+    def _update_canopy_scattered_optimized(self, lad: ti.template(), albedo_leaf: ti.f32):
+        for i, j, k in ti.ndrange(self.domain.nx, self.domain.ny, self.domain.nz):
+            if lad[i, j, k] > 0.0:
+                self._pcbinswref[i, j, k] = self._pcbinsw[i, j, k] * albedo_leaf
+
+    @ti.kernel
+    def _compute_outgoing_fused(self, surfins: ti.template(), n: ti.i32):
+        """Fused: compute outgoing AND accumulate to totals."""
+        for i in range(n):
+            outgoing = self.surfaces.albedo[i] * surfins[i]
+            self._surfoutsl[i] = outgoing
+            self._surfoutsw[i] += outgoing
+    
+    @ti.kernel
+    def _scale_and_accumulate_incoming(self, surfins: ti.template(), n: ti.i32):
+        """Fused: scale by urban VF AND accumulate to incoming totals."""
+        for i in range(n):
+            urban_vf = 1.0 - self.surfaces.svf[i]
+            if urban_vf < 0.01:
+                surfins[i] = 0.0
+            else:
+                surfins[i] *= urban_vf
+                self._surfinsw[i] += surfins[i]
+    
+    @ti.kernel
+    def _copy_buffer(self, src: ti.template(), dst: ti.template(), n: ti.i32):
+        """Copy buffer contents."""
+        for i in range(n):
+            dst[i] = src[i]
+    
+    @ti.kernel
+    def _distribute_reflected_cached(self, svf_nnz: ti.i32):
+        """
+        Distribute reflected radiation using cached SVF matrix (PALM-like).
+        
+        NOTE: This is the legacy single-kernel version kept for compatibility.
+        The optimized version using separate kernels (_distribute_reflected_cached_optimized)
+        is ~100x faster and should be preferred.
+        
+        This is O(nnz) instead of O(n²), providing major speedup for
+        multi-timestep simulations with fixed geometry.
+        
+        PALM equivalent: surfins(isurf) += svf(1,isvf) * svf(2,isvf) * surfoutsl(isurfsrc)
+        """
+        # Reset incoming first
+        for i in range(self.n_surfaces):
+            self._surfins[i] = 0.0
+        
+        # Sync point - ensure reset completes before sparse matmul
+        ti.sync()
+        
+        # Apply sparse matrix-vector multiply
+        for idx in range(svf_nnz):
+            source = self._svf_source[idx]
+            target = self._svf_target[idx]
+            vf = self._svf_vf[idx]
+            trans = self._svf_trans[idx]
+            
+            outgoing = self._surfoutsl[source]
+            # Use same threshold as non-cached version (0.01)
+            if outgoing > 0.01:
+                contribution = outgoing * vf * trans
+                ti.atomic_add(self._surfins[target], contribution)
+        
+        # Sync before scaling
+        ti.sync()
+        
+        # Scale by urban view factor
+        for i in range(self.n_surfaces):
+            urban_vf_i = 1.0 - self.surfaces.svf[i]
+            # Skip surfaces that only see sky (match non-cached behavior)
+            if urban_vf_i < 0.01:
+                self._surfins[i] = 0.0
+            else:
+                self._surfins[i] *= urban_vf_i
+    
+    def invalidate_svf_cache(self):
+        """
+        Invalidate the cached SVF matrix.
+        
+        Call this if geometry (buildings, terrain, vegetation) changes.
+        The matrix will be recomputed on the next compute_svf() call.
+        """
+        self._svf_matrix_cached = False
+        self._svf_nnz = 0
+        print("SVF matrix cache invalidated. Will recompute on next compute_svf() call.")
+    
+    @property
+    def svf_matrix_cached(self) -> bool:
+        """Check if SVF matrix is currently cached."""
+        return self._svf_matrix_cached
+    
+    @property 
+    def svf_matrix_entries(self) -> int:
+        """Get number of non-zero entries in cached SVF matrix."""
+        return self._svf_nnz
+    
+    @ti.kernel
+    def _accumulate_outgoing(self):
+        """Accumulate outgoing radiation to totals."""
+        for i in range(self.n_surfaces):
+            self._surfoutsw[i] += self._surfoutsl[i]
+    
+    @ti.kernel
+    def _accumulate_incoming(self):
+        """Accumulate incoming reflected radiation to totals."""
+        for i in range(self.n_surfaces):
+            self._surfinsw[i] += self._surfins[i]
+
+    @ti.kernel
+    def _accumulate_canopy_absorption_from_reflections(
+        self,
+        lad: ti.template(),
+        is_solid: ti.template(),
+        ext_coef: ti.f32,
+        pcbinsw: ti.template()
+    ):
+        """
+        Accumulate canopy absorption from surface reflections (PALM's reflection loop).
+        
+        PALM formula: pcbinsw(ipcgb) += csf * surfoutsl(isurfsrc) * asrc * grid_volume_inverse
+        
+        For each canopy cell, computes absorption from all reflecting surfaces.
+        The CSF from surface to canopy includes:
+        - View factor (solid angle from surface to canopy cell)
+        - Transmissivity through intervening canopy (Beer-Lambert)
+        - Absorption fraction in target canopy cell
+        
+        Args:
+            lad: 3D LAD field
+            is_solid: 3D solid field
+            ext_coef: Extinction coefficient
+            pcbinsw: Output array for accumulated absorption (W/m³)
+        """
+        PI = 3.14159265359
+        nx = self.domain.nx
+        ny = self.domain.ny
+        nz = self.domain.nz
+        dx = self.domain.dx
+        dy = self.domain.dy
+        dz = self.domain.dz
+        grid_volume = dx * dy * dz
+        grid_volume_inverse = 1.0 / grid_volume
+        
+        # For each canopy cell, accumulate absorption from all reflecting surfaces
+        for ci, cj, ck in ti.ndrange(nx, ny, nz):
+            cell_lad = lad[ci, cj, ck]
+            if cell_lad <= 0.0:
+                continue
+            
+            # Cell center position
+            pos_cell = Vector3(
+                (ci + 0.5) * dx,
+                (cj + 0.5) * dy,
+                (ck + 0.5) * dz
+            )
+            
+            total_absorbed = 0.0
+            
+            # Loop over all surfaces with outgoing radiation
+            for surf_i in range(self.n_surfaces):
+                outgoing = self._surfoutsl[surf_i]
+                if outgoing < 0.01:
+                    continue
+                
+                # Surface properties
+                pos_surf = self.surfaces.center[surf_i]
+                normal_surf = self.surfaces.normal[surf_i]
+                area_surf = self.surfaces.area[surf_i]
+                
+                # Vector from surface to canopy cell
+                diff = pos_cell - pos_surf
+                dist_sq = diff[0]*diff[0] + diff[1]*diff[1] + diff[2]*diff[2]
+                
+                if dist_sq < 0.01:
+                    continue
+                
+                dist = ti.sqrt(dist_sq)
+                
+                # Direction from surface to canopy
+                dir_to_cell_x = diff[0] / dist
+                dir_to_cell_y = diff[1] / dist
+                dir_to_cell_z = diff[2] / dist
+                
+                # Check if surface faces the canopy cell
+                cos_emit = (normal_surf[0]*dir_to_cell_x + 
+                           normal_surf[1]*dir_to_cell_y + 
+                           normal_surf[2]*dir_to_cell_z)
+                
+                if cos_emit > 0.0:
+                    # Compute transmissivity from surface to canopy cell (through intervening canopy)
+                    trans, blocked = ray_point_to_point_transmissivity(
+                        pos_surf, pos_cell,
+                        lad, is_solid,
+                        nx, ny, nz,
+                        dx, dy, dz,
+                        ext_coef
+                    )
+                    
+                    if blocked == 0:
+                        # Approximate path length through target cell
+                        path_in_cell = dz  # Simplified; could be more accurate
+                        
+                        # Absorption fraction in this cell
+                        abs_frac = 1.0 - ti.exp(-ext_coef * cell_lad * path_in_cell)
+                        
+                        # View factor from surface to canopy cell (simplified)
+                        # CSF = view_factor * transmissivity * absorption_fraction
+                        cell_solid_angle = (dx * dy) / (4.0 * PI * dist_sq)
+                        csf_factor = cell_solid_angle * cos_emit * trans * abs_frac
+                        csf_factor = ti.min(csf_factor, 1.0)  # Clamp
+                        
+                        # PALM formula: pcbinsw += csf * surfoutsl * asrc * grid_volume_inverse
+                        absorbed = csf_factor * outgoing * area_surf * grid_volume_inverse
+                        total_absorbed += absorbed
+            
+            ti.atomic_add(pcbinsw[ci, cj, ck], total_absorbed)
+
+    @ti.kernel
+    def _update_canopy_scattered_radiation(self, lad: ti.template(), albedo_leaf: ti.f32):
+        """
+        Update scattered radiation field based on current absorbed radiation.
+        
+        pcbinswref = albedo_leaf * pcbinsw (fraction that gets scattered)
+        
+        This is called at each reflection step to update what's available for scattering.
+        """
+        for i, j, k in ti.ndrange(self.domain.nx, self.domain.ny, self.domain.nz):
+            if lad[i, j, k] > 0.0:
+                absorbed = self._pcbinsw[i, j, k]
+                self._pcbinswref[i, j, k] = albedo_leaf * absorbed
+
+    @ti.kernel
+    def _compute_canopy_to_surface_scattering(
+        self,
+        lad: ti.template(),
+        is_solid: ti.template(),
+        ext_coef: ti.f32
+    ):
+        """
+        Compute radiation scattered from canopy cells toward surfaces.
+        
+        For each canopy cell with scattered radiation (pcbinswref), distribute
+        to surfaces based on solid angle and transmissivity.
+        
+        This is called within the reflection loop to capture Canopy→Surface paths.
+        The contribution is added to _surfins so it participates in subsequent reflections.
+        """
+        PI = 3.14159265359
+        nx = self.domain.nx
+        ny = self.domain.ny
+        nz = self.domain.nz
+        dx = self.domain.dx
+        dy = self.domain.dy
+        dz = self.domain.dz
+        grid_volume = dx * dy * dz
+        
+        # For each canopy cell with scattered radiation
+        for ci, cj, ck in ti.ndrange(nx, ny, nz):
+            cell_lad = lad[ci, cj, ck]
+            if cell_lad <= 0.0:
+                continue
+            
+            scattered_power = self._pcbinswref[ci, cj, ck] * grid_volume
+            if scattered_power <= 0.01:
+                continue
+            
+            # Cell center position
+            pos_cell = Vector3(
+                (ci + 0.5) * dx,
+                (cj + 0.5) * dy,
+                (ck + 0.5) * dz
+            )
+            
+            # Distribute to all surfaces
+            for surf_i in range(self.n_surfaces):
+                pos_surf = self.surfaces.center[surf_i]
+                normal_surf = self.surfaces.normal[surf_i]
+                
+                # Vector from cell to surface
+                diff = pos_surf - pos_cell
+                dist_sq = diff[0]*diff[0] + diff[1]*diff[1] + diff[2]*diff[2]
+                
+                if dist_sq < 0.01:
+                    continue
+                
+                dist = ti.sqrt(dist_sq)
+                
+                # Direction from cell to surface
+                dir_x = diff[0] / dist
+                dir_y = diff[1] / dist
+                dir_z = diff[2] / dist
+                
+                # Check if surface faces the cell
+                cos_recv = (normal_surf[0]*dir_x + normal_surf[1]*dir_y + normal_surf[2]*dir_z)
+                
+                if cos_recv > 0.0:
+                    # Solid angle factor
+                    cell_cross = dx * dy
+                    solid_angle_factor = cell_cross / (4.0 * PI * dist_sq)
+                    solid_angle_factor = ti.min(solid_angle_factor, 0.25)
+                    
+                    # Transmissivity through intervening canopy
+                    trans, blocked = ray_point_to_point_transmissivity(
+                        pos_cell, pos_surf,
+                        lad, is_solid,
+                        nx, ny, nz,
+                        dx, dy, dz,
+                        ext_coef
+                    )
+                    
+                    if blocked == 0:
+                        contribution = scattered_power * solid_angle_factor * cos_recv * trans
+                        contribution /= self.surfaces.area[surf_i]
+                        # Add to surfins so it participates in next reflection step
+                        ti.atomic_add(self._surfins[surf_i], contribution)
+                        # Also track separately for output
+                        ti.atomic_add(self._surfinswpc[surf_i], contribution)
+
+    @ti.kernel
+    def _compute_surface_outgoing_no_exchange(self):
+        """Compute outgoing radiation without inter-surface exchange."""
+        for i in range(self.n_surfaces):
+            albedo = self.surfaces.albedo[i]
+            self._surfoutsl[i] = albedo * self._surfins[i]
+            self._surfoutsw[i] = self._surfoutsl[i]
+    
+    @ti.kernel
+    def _copy_final_fluxes(self):
+        """Copy final computed fluxes to surface arrays."""
+        for i in range(self.n_surfaces):
+            self.surfaces.sw_in_direct[i] = self._surfinswdir[i]
+            self.surfaces.sw_in_diffuse[i] = self._surfinswdif[i] + (
+                self._surfinsw[i] - self._surfinswdir[i] - self._surfinswdif[i]
+            ) + self._surfinswpc[i]  # Include reflected and canopy scattered as diffuse
+            self.surfaces.sw_out[i] = self._surfoutsw[i]
+    
+    @ti.kernel
+    def _compute_canopy_scattering(
+        self,
+        lad: ti.template(),
+        is_solid: ti.template(),
+        albedo_leaf: ti.f32,
+        ext_coef: ti.f32
+    ):
+        """
+        Compute radiation scattered from canopy cells toward surfaces.
+        
+        For each canopy cell that absorbs radiation, a fraction (albedo_leaf)
+        is scattered isotropically. This scattered radiation contributes to
+        nearby surfaces based on solid angle and distance.
+        
+        Following PALM's methodology:
+        - pcrad = absorbed_radiation * albedo_leaf (what gets scattered)
+        - Distribution based on CSF-like view factors to surfaces
+        
+        Args:
+            lad: 3D field of Leaf Area Density
+            is_solid: 3D field of solid cells
+            albedo_leaf: Leaf albedo (fraction scattered vs absorbed)
+            ext_coef: Extinction coefficient
+        """
+        PI = 3.14159265359
+        nx = self.domain.nx
+        ny = self.domain.ny
+        nz = self.domain.nz
+        dx = self.domain.dx
+        dy = self.domain.dy
+        dz = self.domain.dz
+        grid_volume = dx * dy * dz
+        
+        # First: compute scattered radiation for each canopy cell
+        for i, j, k in ti.ndrange(nx, ny, nz):
+            cell_lad = lad[i, j, k]
+            if cell_lad > 0.0:
+                absorbed = self._pcbinsw[i, j, k] * grid_volume
+                self._pcbinswref[i, j, k] = albedo_leaf * absorbed / grid_volume
+        
+        # Second: for each canopy cell, contribute to nearby surfaces
+        # Iterate over canopy cells (outer) and surfaces (inner)
+        for ci, cj, ck in ti.ndrange(nx, ny, nz):
+            cell_lad = lad[ci, cj, ck]
+            if cell_lad <= 0.0:
+                continue
+            
+            scattered_power = self._pcbinswref[ci, cj, ck] * grid_volume
+            if scattered_power <= 0.0:
+                continue
+            
+            # Cell center position
+            pos_cell = Vector3(
+                (ci + 0.5) * dx,
+                (cj + 0.5) * dy,
+                (ck + 0.5) * dz
+            )
+            
+            # Distribute to all surfaces
+            for surf_i in range(self.n_surfaces):
+                pos_surf = self.surfaces.center[surf_i]
+                normal_surf = self.surfaces.normal[surf_i]
+                
+                # Vector from cell to surface
+                diff = pos_surf - pos_cell
+                dist_sq = diff[0]*diff[0] + diff[1]*diff[1] + diff[2]*diff[2]
+                
+                if dist_sq < 0.01:
+                    continue
+                
+                dist = ti.sqrt(dist_sq)
+                
+                # Direction from surface to cell (for checking if surface "sees" the cell)
+                dir_to_cell_x = -diff[0] / dist
+                dir_to_cell_y = -diff[1] / dist
+                dir_to_cell_z = -diff[2] / dist
+                
+                # Check if surface faces the cell (cell is in hemisphere surface faces)
+                cos_recv = (normal_surf[0]*dir_to_cell_x + 
+                           normal_surf[1]*dir_to_cell_y + 
+                           normal_surf[2]*dir_to_cell_z)
+                
+                if cos_recv > 0.0:
+                    # Solid angle factor
+                    cell_cross = dx * dy
+                    solid_angle_factor = cell_cross / (4.0 * PI * dist_sq)
+                    solid_angle_factor = ti.min(solid_angle_factor, 0.25)
+                    
+                    # Transmissivity through intervening canopy
+                    trans, blocked = ray_point_to_point_transmissivity(
+                        pos_cell, pos_surf,
+                        lad, is_solid,
+                        nx, ny, nz,
+                        dx, dy, dz,
+                        ext_coef
+                    )
+                    
+                    if blocked == 0:
+                        contribution = scattered_power * solid_angle_factor * cos_recv * trans
+                        contribution /= self.surfaces.area[surf_i]
+                        ti.atomic_add(self._surfinswpc[surf_i], contribution)
+    
+    @ti.kernel
+    def _compute_canopy_to_canopy_scattering(
+        self,
+        lad: ti.template(),
+        is_solid: ti.template(),
+        albedo_leaf: ti.f32,
+        ext_coef: ti.f32
+    ):
+        """
+        Compute radiation scattered from one canopy cell to another.
+        
+        For each canopy cell that scatters radiation (pcbinswref), distribute
+        that scattered radiation to neighboring canopy cells based on:
+        - Solid angle (distance-based)
+        - Transmissivity through intervening canopy
+        - Absorption fraction in target cell
+        
+        This implements canopy-to-canopy scattering which PALM does not 
+        explicitly model but is important for dense vegetation canopies.
+        
+        The formula follows the same CSF methodology:
+            pcbinswc2c[target] += scattered[source] × view_factor × trans × abs_frac
+        
+        Args:
+            lad: 3D field of Leaf Area Density
+            is_solid: 3D field of solid cells
+            albedo_leaf: Leaf albedo (fraction scattered vs absorbed)
+            ext_coef: Extinction coefficient
+        """
+        PI = 3.14159265359
+        nx = self.domain.nx
+        ny = self.domain.ny
+        nz = self.domain.nz
+        dx = self.domain.dx
+        dy = self.domain.dy
+        dz = self.domain.dz
+        grid_volume = dx * dy * dz
+        
+        # For each source canopy cell with scattered radiation
+        for si, sj, sk in ti.ndrange(nx, ny, nz):
+            source_lad = lad[si, sj, sk]
+            if source_lad <= 0.0:
+                continue
+            
+            scattered_power = self._pcbinswref[si, sj, sk] * grid_volume
+            if scattered_power <= 0.01:
+                continue
+            
+            # Source cell center position
+            pos_source = Vector3(
+                (si + 0.5) * dx,
+                (sj + 0.5) * dy,
+                (sk + 0.5) * dz
+            )
+            
+            # Distribute to nearby canopy cells (limit search radius for efficiency)
+            # Use a search radius based on typical canopy interaction distance
+            search_radius_cells = 5  # Cells in each direction
+            
+            i_min = ti.max(0, si - search_radius_cells)
+            i_max = ti.min(nx, si + search_radius_cells + 1)
+            j_min = ti.max(0, sj - search_radius_cells)
+            j_max = ti.min(ny, sj + search_radius_cells + 1)
+            k_min = ti.max(0, sk - search_radius_cells)
+            k_max = ti.min(nz, sk + search_radius_cells + 1)
+            
+            for ti_idx in range(i_min, i_max):
+                for tj_idx in range(j_min, j_max):
+                    for tk_idx in range(k_min, k_max):
+                        # Skip self
+                        if ti_idx == si and tj_idx == sj and tk_idx == sk:
+                            continue
+                        
+                        target_lad = lad[ti_idx, tj_idx, tk_idx]
+                        if target_lad <= 0.0:
+                            continue
+                        
+                        # Target cell center position
+                        pos_target = Vector3(
+                            (ti_idx + 0.5) * dx,
+                            (tj_idx + 0.5) * dy,
+                            (tk_idx + 0.5) * dz
+                        )
+                        
+                        # Distance between cells
+                        diff = pos_target - pos_source
+                        dist_sq = diff[0]*diff[0] + diff[1]*diff[1] + diff[2]*diff[2]
+                        
+                        if dist_sq < 0.01:
+                            continue
+                        
+                        dist = ti.sqrt(dist_sq)
+                        
+                        # Solid angle factor (cell cross-section / distance²)
+                        # Using cell cross-sectional area perpendicular to ray
+                        cell_cross = dx * dy  # Simplified; could use projection
+                        solid_angle_factor = cell_cross / (4.0 * PI * dist_sq)
+                        solid_angle_factor = ti.min(solid_angle_factor, 0.25)
+                        
+                        # Compute transmissivity from source to target through intervening canopy
+                        trans, blocked = ray_point_to_point_transmissivity(
+                            pos_source, pos_target,
+                            lad, is_solid,
+                            nx, ny, nz,
+                            dx, dy, dz,
+                            ext_coef
+                        )
+                        
+                        if blocked == 0 and trans > 0.01:
+                            # Absorption fraction in target cell
+                            path_in_cell = dz  # Simplified path length
+                            abs_frac = 1.0 - ti.exp(-ext_coef * target_lad * path_in_cell)
+                            
+                            # Contribution to target cell (W/m³)
+                            contribution = scattered_power * solid_angle_factor * trans * abs_frac / grid_volume
+                            ti.atomic_add(self._pcbinswc2c[ti_idx, tj_idx, tk_idx], contribution)
+    
+    @ti.kernel
+    def _accumulate_canopy_to_canopy(self, lad: ti.template(), albedo_leaf: ti.f32):
+        """
+        Accumulate canopy-to-canopy contribution to total canopy absorption
+        and prepare for next iteration.
+        
+        The c2c contribution is added to pcbinsw, and a fraction (albedo_leaf)
+        is added to pcbinswref for the next scattering iteration.
+        """
+        for i, j, k in ti.ndrange(self.domain.nx, self.domain.ny, self.domain.nz):
+            if lad[i, j, k] > 0.0:
+                c2c = self._pcbinswc2c[i, j, k]
+                if c2c > 0.0:
+                    # Add to total absorbed
+                    self._pcbinsw[i, j, k] += c2c
+                    # Add to cumulative c2c total (for output)
+                    self._pcbinswc2c_total[i, j, k] += c2c
+                    # Fraction that gets scattered again
+                    self._pcbinswref[i, j, k] += albedo_leaf * c2c
+                    # Reset for next iteration
+                    self._pcbinswc2c[i, j, k] = 0.0
+    
+    def _apply_canopy_scattering(self):
+        """
+        Apply initial canopy scattering from direct/diffuse absorption.
+        
+        Note: Most canopy scattering is now handled inside the reflection loop
+        via _compute_canopy_to_surface_scattering and _compute_canopy_to_canopy_scattering.
+        This function handles the initial scattering that feeds into the loop.
+        """
+        if self.domain.lad is None:
+            return
+        
+        # Update scattered radiation based on current absorbed (from direct + diffuse)
+        self._update_canopy_scattered_radiation(
+            self.domain.lad,
+            self.config.albedo_leaf
+        )
+    
+    @ti.kernel
+    def _add_canopy_to_diffuse(self):
+        """Add canopy-scattered radiation to surface diffuse component."""
+        for i in range(self.n_surfaces):
+            self.surfaces.sw_in_diffuse[i] += self._surfinswpc[i]
+    
+    def compute_radiation(
+        self,
+        day_of_year: int,
+        second_of_day: float,
+        sw_direct: float,
+        sw_diffuse: float
+    ):
+        """
+        Compute shortwave radiation components.
+        
+        Args:
+            day_of_year: Day number (1-365)
+            second_of_day: Seconds since midnight UTC
+            sw_direct: Direct normal irradiance (W/m²)
+            sw_diffuse: Diffuse horizontal irradiance (W/m²)
+        """
+        self.update_solar_position(day_of_year, second_of_day)
+        self.compute_shortwave_radiation(sw_direct, sw_diffuse)
+    
+    def get_surface_fluxes(self) -> dict:
+        """
+        Get radiation fluxes as numpy arrays.
+        
+        Returns:
+            Dictionary with flux arrays including:
+            - position: Grid indices (i, j, k)
+            - direction: Surface direction index
+            - area: Surface area (m²)
+            - svf: Sky view factor
+            - shadow_factor: Shadow factor (0=sunlit, 1=shaded)
+            - sw_in_direct: Direct SW radiation (W/m²)
+            - sw_in_diffuse: Diffuse SW radiation including reflections (W/m²)
+            - sw_out: Outgoing (reflected) SW radiation (W/m²)
+            - sw_in_total: Total incoming SW (W/m²)
+            - sw_net: Net absorbed SW (W/m²)
+        """
+        sw_in_direct = self.surfaces.sw_in_direct.to_numpy()[:self.n_surfaces]
+        sw_in_diffuse = self.surfaces.sw_in_diffuse.to_numpy()[:self.n_surfaces]
+        sw_out = self.surfaces.sw_out.to_numpy()[:self.n_surfaces]
+        sw_in_total = sw_in_direct + sw_in_diffuse
+        sw_net = sw_in_total - sw_out
+        
+        return {
+            'position': self.surfaces.position.to_numpy()[:self.n_surfaces],
+            'direction': self.surfaces.direction.to_numpy()[:self.n_surfaces],
+            'area': self.surfaces.area.to_numpy()[:self.n_surfaces],
+            'svf': self.surfaces.svf.to_numpy()[:self.n_surfaces],
+            'shadow_factor': self.surfaces.shadow_factor.to_numpy()[:self.n_surfaces],
+            'sw_in_direct': sw_in_direct,
+            'sw_in_diffuse': sw_in_diffuse,
+            'sw_out': sw_out,
+            'sw_in_total': sw_in_total,
+            'sw_net': sw_net,
+        }
+    
+    def get_total_absorbed_sw(self) -> float:
+        """Get total absorbed shortwave radiation (W)."""
+        fluxes = self.get_surface_fluxes()
+        return float((fluxes['sw_net'] * fluxes['area']).sum())
+    
+    def get_domain_shadow_map(self) -> np.ndarray:
+        """
+        Get 2D shadow map at ground/terrain level.
+        
+        Only includes ground surfaces (k=0), not building rooftops.
+        Building footprints are marked with NaN.
+        
+        Returns:
+            2D array of shadow factors (0=shadowed, 1=sunlit, NaN=building)
+        """
+        shadow_map = np.full((self.domain.nx, self.domain.ny), np.nan)
+        fluxes = self.get_surface_fluxes()
+        
+        # Find upward-facing surfaces at ground level only (k=0)
+        for i in range(self.n_surfaces):
+            if fluxes['direction'][i] == 0:  # Upward
+                pos = fluxes['position'][i]
+                ix = int(pos[0])  # Position is already grid indices
+                iy = int(pos[1])
+                iz = int(pos[2])
+                if 0 <= ix < self.domain.nx and 0 <= iy < self.domain.ny:
+                    # Only include ground-level surfaces (terrain at k=0)
+                    if iz == 0:
+                        # Invert: shadow_factor=0 means sunlit, =1 means shaded
+                        # For display we want 1=sunlit, 0=shaded
+                        shadow_map[ix, iy] = 1.0 - fluxes['shadow_factor'][i]
+        
+        return shadow_map
+    
+    def get_irradiance_map(self) -> np.ndarray:
+        """
+        Get 2D map of total incoming shortwave irradiance at ground level.
+        
+        Returns:
+            2D array of irradiance values (W/m²), NaN for building footprints
+        """
+        irradiance_map = np.full((self.domain.nx, self.domain.ny), np.nan)
+        fluxes = self.get_surface_fluxes()
+        
+        # Find upward-facing surfaces at ground level only (k=0)
+        for i in range(self.n_surfaces):
+            if fluxes['direction'][i] == 0:  # Upward
+                pos = fluxes['position'][i]
+                ix = int(pos[0])
+                iy = int(pos[1])
+                iz = int(pos[2])
+                if 0 <= ix < self.domain.nx and 0 <= iy < self.domain.ny:
+                    if iz == 0:
+                        # Total incoming = direct + diffuse
+                        sw_in = fluxes['sw_in_direct'][i] + fluxes['sw_in_diffuse'][i]
+                        irradiance_map[ix, iy] = sw_in
+        
+        return irradiance_map
+    
+    def get_net_sw_radiation_map(self) -> np.ndarray:
+        """
+        Get 2D map of net shortwave radiation at ground level.
+        
+        Returns:
+            2D array of net SW radiation values (W/m²), NaN for building footprints
+        """
+        net_map = np.full((self.domain.nx, self.domain.ny), np.nan)
+        fluxes = self.get_surface_fluxes()
+        
+        for i in range(self.n_surfaces):
+            if fluxes['direction'][i] == 0:  # Upward
+                pos = fluxes['position'][i]
+                ix = int(pos[0])
+                iy = int(pos[1])
+                iz = int(pos[2])
+                if 0 <= ix < self.domain.nx and 0 <= iy < self.domain.ny:
+                    if iz == 0:
+                        sw_in = fluxes['sw_in_direct'][i] + fluxes['sw_in_diffuse'][i]
+                        sw_out = fluxes['sw_out'][i]
+                        net_map[ix, iy] = sw_in - sw_out
+        
+        return net_map
+    
+    # =========================================================================
+    # Volumetric flux methods
+    # =========================================================================
+    
+    def compute_volumetric_svf(self):
+        """
+        Compute volumetric sky view factors.
+        
+        This must be called before computing volumetric SW fluxes.
+        Only needed once unless domain geometry changes.
+        
+        Raises:
+            RuntimeError: If volumetric_flux is not enabled in config
+        """
+        if self.volumetric_calc is None:
+            raise RuntimeError(
+                "Volumetric flux not enabled. Set volumetric_flux=True in RadiationConfig."
+            )
+        self.volumetric_calc.compute_skyvf_vol()
+    
+    def get_volumetric_skyvf(self) -> np.ndarray:
+        """
+        Get volumetric sky view factor as 3D numpy array.
+        
+        Returns:
+            3D array of shape (nx, ny, nz) with SVF values [0, 1]
+            
+        Raises:
+            RuntimeError: If volumetric_flux is not enabled
+        """
+        if self.volumetric_calc is None:
+            raise RuntimeError(
+                "Volumetric flux not enabled. Set volumetric_flux=True in RadiationConfig."
+            )
+        return self.volumetric_calc.get_skyvf_vol()
+    
+    def get_volumetric_swflux(self) -> np.ndarray:
+        """
+        Get volumetric shortwave flux as 3D numpy array.
+        
+        This is the omnidirectional SW flux at each grid cell,
+        representing average irradiance onto an imaginary sphere (W/m²).
+        
+        Returns:
+            3D array of shape (nx, ny, nz) with SW flux values
+            
+        Raises:
+            RuntimeError: If volumetric_flux is not enabled
+        """
+        if self.volumetric_calc is None:
+            raise RuntimeError(
+                "Volumetric flux not enabled. Set volumetric_flux=True in RadiationConfig."
+            )
+        return self.volumetric_calc.get_swflux_vol()
+    
+    def get_volumetric_shadow_top(self) -> np.ndarray:
+        """
+        Get shadow top level as 2D numpy array.
+        
+        Shadow top is the highest grid level that is in shadow
+        for the current solar position.
+        
+        Returns:
+            2D array of shape (nx, ny) with vertical indices
+            
+        Raises:
+            RuntimeError: If volumetric_flux is not enabled
+        """
+        if self.volumetric_calc is None:
+            raise RuntimeError(
+                "Volumetric flux not enabled. Set volumetric_flux=True in RadiationConfig."
+            )
+        return self.volumetric_calc.get_shadow_top()
+    
+    def get_volumetric_shadow_mask(self) -> np.ndarray:
+        """
+        Get 3D shadow mask.
+        
+        Returns:
+            3D boolean array where True indicates shadowed cells
+            
+        Raises:
+            RuntimeError: If volumetric_flux is not enabled
+        """
+        if self.volumetric_calc is None:
+            raise RuntimeError(
+                "Volumetric flux not enabled. Set volumetric_flux=True in RadiationConfig."
+            )
+        return self.volumetric_calc.get_shadow_mask_3d()
+    
+    def get_volumetric_swflux_slice(
+        self,
+        level: Optional[int] = None,
+        axis: Optional[str] = None,
+        index: Optional[int] = None
+    ) -> np.ndarray:
+        """
+        Get a 2D slice of volumetric SW flux.
+        
+        Args:
+            level: Vertical level for horizontal slice (k index)
+            axis: 'x' or 'y' for vertical slices
+            index: Index along the axis for vertical slices
+        
+        Returns:
+            2D array of SW flux values (W/m²)
+            
+        Raises:
+            RuntimeError: If volumetric_flux is not enabled
+            ValueError: If invalid slice parameters
+        """
+        if self.volumetric_calc is None:
+            raise RuntimeError(
+                "Volumetric flux not enabled. Set volumetric_flux=True in RadiationConfig."
+            )
+        
+        if level is not None:
+            return self.volumetric_calc.get_horizontal_slice(level, 'swflux')
+        elif axis is not None and index is not None:
+            return self.volumetric_calc.get_vertical_slice(axis, index, 'swflux')
+        else:
+            raise ValueError("Specify either 'level' or both 'axis' and 'index'")
+
+    # =========================================================================
+    # Plant Canopy Radiation Methods
+    # =========================================================================
+    
+    def _compute_canopy_radiation(
+        self,
+        sw_direct: float,
+        sw_diffuse: float,
+        sun_dir: Vector3,
+        cos_zenith: float
+    ):
+        """
+        Compute radiation absorption in plant canopy using CSF.
+        
+        This implements PALM's plant canopy radiation balance:
+        - pcbinswdir: Direct SW absorbed per canopy cell (W/m³)
+        - pcbinswdif: Diffuse SW absorbed per canopy cell (W/m³)
+        - pcbinsw: Total SW absorbed (W/m³) - includes direct, diffuse, AND reflected
+        
+        Note: pcbinsw may already contain reflection-step contributions from
+        _compute_sw_fluxes. This method ADDS direct and diffuse to pcbinsw
+        rather than resetting it.
+        
+        Args:
+            sw_direct: Direct normal irradiance (W/m²)
+            sw_diffuse: Diffuse horizontal irradiance (W/m²)
+            sun_dir: Sun direction vector
+            cos_zenith: Cosine of solar zenith angle
+        """
+        # Reset CSF calculator (not the _pcbinsw - it may have reflection contributions)
+        self.csf_calc.reset_csf()
+        
+        # Reset only direct and diffuse arrays, NOT total pcbinsw
+        self._reset_canopy_dir_dif_arrays()
+        
+        grid_volume = self.domain.dx * self.domain.dy * self.domain.dz
+        
+        # Compute direct SW absorption if sun is up
+        # Use PALM's method: box_absorb + dsitransc + per-box absorption
+        if cos_zenith > 0.0262:  # min_stable_coszen
+            self.csf_calc.compute_canopy_absorption_direct_palm(
+                sun_dir,
+                self.domain.is_solid,
+                self.domain.lad,
+                sw_direct
+            )
+            
+            # CSF now contains absorption in W/m³, copy to pcbinswdir
+            self._copy_csf_to_pcbinswdir_direct()
+        
+        # Compute diffuse SW absorption using PALM's method
+        # Traces rays from each canopy cell to sky hemisphere (not from surfaces)
+        self.csf_calc.compute_canopy_absorption_diffuse_palm(
+            self.domain.is_solid,
+            self.domain.lad,
+            sw_diffuse,
+            self._pcbinswdif,
+            self.config.n_azimuth,
+            self.config.n_elevation
+        )
+        
+        # Total absorbed = direct + diffuse + reflected (already in pcbinsw from reflections)
+        self._sum_canopy_absorption()
+        
+        # Compute received radiation (before absorption)
+        self._compute_received_radiation(sw_direct, sw_diffuse, cos_zenith, grid_volume)
+        
+        self._canopy_radiation_computed = True
+    
+    def _compute_canopy_radiation_initial(
+        self,
+        sw_direct: float,
+        sw_diffuse: float,
+        sun_dir,  # Vector3
+        cos_zenith: float
+    ):
+        """
+        Compute initial canopy radiation absorption (direct + diffuse) BEFORE reflection loop.
+        
+        This is called before the surface reflection loop so that canopy-scattered
+        radiation can participate in reflections (enabling paths like C→S→C, S→C→S).
+        
+        After this, _update_canopy_scattered_radiation should be called to prepare
+        pcbinswref for the reflection loop.
+        
+        Args:
+            sw_direct: Direct normal irradiance (W/m²)
+            sw_diffuse: Diffuse horizontal irradiance (W/m²)
+            sun_dir: Sun direction vector
+            cos_zenith: Cosine of solar zenith angle
+        """
+        # Reset CSF calculator and all canopy arrays
+        self.csf_calc.reset_csf()
+        self._reset_canopy_arrays()
+        
+        # Compute direct SW absorption if sun is up
+        if cos_zenith > 0.0262:  # min_stable_coszen
+            self.csf_calc.compute_canopy_absorption_direct_palm(
+                sun_dir,
+                self.domain.is_solid,
+                self.domain.lad,
+                sw_direct
+            )
+            # CSF now contains absorption in W/m³, copy to pcbinswdir
+            self._copy_csf_to_pcbinswdir_direct()
+        
+        # Compute diffuse SW absorption using PALM's method
+        self.csf_calc.compute_canopy_absorption_diffuse_palm(
+            self.domain.is_solid,
+            self.domain.lad,
+            sw_diffuse,
+            self._pcbinswdif,
+            self.config.n_azimuth,
+            self.config.n_elevation
+        )
+        
+        # Total initial absorbed = direct + diffuse
+        self._sum_canopy_absorption_initial()
+    
+    @ti.kernel
+    def _sum_canopy_absorption_initial(self):
+        """Sum direct + diffuse for initial canopy absorption (resets pcbinsw first)."""
+        for i, j, k in ti.ndrange(self.domain.nx, self.domain.ny, self.domain.nz):
+            self._pcbinsw[i, j, k] = self._pcbinswdir[i, j, k] + self._pcbinswdif[i, j, k]
+    
+    @ti.kernel
+    def _reset_canopy_dir_dif_arrays(self):
+        """Reset direct and diffuse canopy absorption arrays to zero (not total)."""
+        for i, j, k in ti.ndrange(self.domain.nx, self.domain.ny, self.domain.nz):
+            self._pcbinswdir[i, j, k] = 0.0
+            self._pcbinswdif[i, j, k] = 0.0
+    
+    @ti.kernel
+    def _reset_canopy_arrays(self):
+        """Reset all canopy absorption arrays to zero."""
+        for i, j, k in ti.ndrange(self.domain.nx, self.domain.ny, self.domain.nz):
+            self._pcbinsw[i, j, k] = 0.0
+            self._pcbinsw[i, j, k] = 0.0
+            self._pcbinswdir[i, j, k] = 0.0
+            self._pcbinswdif[i, j, k] = 0.0
+            self._pcinsw[i, j, k] = 0.0
+            self._pcinswdir[i, j, k] = 0.0
+            self._pcinswdif[i, j, k] = 0.0
+    
+    @ti.kernel
+    def _copy_csf_to_pcbinswdir(self, grid_volume: ti.f32):
+        """Copy CSF field to direct absorption array, converting to W/m³."""
+        for i, j, k in ti.ndrange(self.domain.nx, self.domain.ny, self.domain.nz):
+            # CSF contains absorbed power in W, convert to W/m³
+            self._pcbinswdir[i, j, k] = self.csf_calc.csf[i, j, k] / grid_volume
+    
+    @ti.kernel
+    def _copy_csf_to_pcbinswdir_direct(self):
+        """Copy CSF field to direct absorption array (CSF already in W/m³ from PALM method)."""
+        for i, j, k in ti.ndrange(self.domain.nx, self.domain.ny, self.domain.nz):
+            self._pcbinswdir[i, j, k] = self.csf_calc.csf[i, j, k]
+    
+    @ti.kernel
+    def _copy_csf_to_pcbinswdif(self, grid_volume: ti.f32):
+        """Copy CSF field to diffuse absorption array, converting to W/m³."""
+        for i, j, k in ti.ndrange(self.domain.nx, self.domain.ny, self.domain.nz):
+            self._pcbinswdif[i, j, k] = self.csf_calc.csf[i, j, k] / grid_volume
+    
+    @ti.kernel
+    def _sum_canopy_absorption(self):
+        """Add direct and diffuse to total canopy absorption.
+        
+        Note: pcbinsw may already contain reflection contributions from
+        _accumulate_canopy_absorption_from_reflections. This method ADDS
+        direct + diffuse rather than replacing.
+        """
+        for i, j, k in ti.ndrange(self.domain.nx, self.domain.ny, self.domain.nz):
+            # Add direct + diffuse to existing value (which may have reflection contributions)
+            self._pcbinsw[i, j, k] += self._pcbinswdir[i, j, k] + self._pcbinswdif[i, j, k]
+    
+    @ti.kernel
+    def _compute_received_radiation(
+        self,
+        sw_direct: ti.f32,
+        sw_diffuse: ti.f32,
+        cos_zenith: ti.f32,
+        grid_volume: ti.f32
+    ):
+        """
+        Compute received (incident) radiation at each canopy cell.
+        
+        This is the radiation before absorption, useful for photosynthesis models.
+        """
+        for i, j, k in ti.ndrange(self.domain.nx, self.domain.ny, self.domain.nz):
+            lad = self.domain.lad[i, j, k]
+            if lad > 0.0:
+                # Leaf area in this cell
+                leaf_area = lad * grid_volume
+                
+                # Received = absorbed / absorption_fraction
+                # For small absorption: received ≈ absorbed / (ext_coef * LAD * path_length)
+                # Simplified: use absorbed * 2 as rough estimate (50% absorption typical)
+                if self._pcbinswdir[i, j, k] > 0:
+                    self._pcinswdir[i, j, k] = self._pcbinswdir[i, j, k] * grid_volume / leaf_area
+                if self._pcbinswdif[i, j, k] > 0:
+                    self._pcinswdif[i, j, k] = self._pcbinswdif[i, j, k] * grid_volume / leaf_area
+                
+                self._pcinsw[i, j, k] = self._pcinswdir[i, j, k] + self._pcinswdif[i, j, k]
+    
+    def get_canopy_absorbed_sw(self) -> np.ndarray:
+        """
+        Get total absorbed SW radiation in plant canopy.
+        
+        Returns:
+            3D array of absorbed SW radiation (W/m³)
+        """
+        return self._pcbinsw.to_numpy()
+    
+    def get_canopy_absorbed_sw_direct(self) -> np.ndarray:
+        """
+        Get direct SW radiation absorbed in plant canopy.
+        
+        Returns:
+            3D array of absorbed direct SW radiation (W/m³)
+        """
+        return self._pcbinswdir.to_numpy()
+    
+    def get_canopy_absorbed_sw_diffuse(self) -> np.ndarray:
+        """
+        Get diffuse SW radiation absorbed in plant canopy.
+        
+        Returns:
+            3D array of absorbed diffuse SW radiation (W/m³)
+        """
+        return self._pcbinswdif.to_numpy()
+    
+    def get_canopy_received_sw(self) -> np.ndarray:
+        """
+        Get total received SW radiation at plant canopy (before absorption).
+        
+        Returns:
+            3D array of received SW radiation (W/m²)
+        """
+        return self._pcinsw.to_numpy()
+    
+    def get_total_canopy_absorption(self) -> float:
+        """
+        Get total SW radiation absorbed by all plant canopy (W).
+        
+        
+        Returns:
+            Total absorbed power in Watts
+        """
+        grid_volume = self.domain.dx * self.domain.dy * self.domain.dz
+        return float(self._pcbinsw.to_numpy().sum() * grid_volume)
+    
+    def get_canopy_absorption_profile(self) -> np.ndarray:
+        """
+        Get vertical profile of canopy-averaged SW absorption.
+        
+        Returns:
+            1D array of mean absorbed SW per level (W/m³)
+        """
+        pcbinsw = self._pcbinsw.to_numpy()
+        # Mean over horizontal dimensions, excluding zero cells
+        profile = np.zeros(self.domain.nz)
+        for k in range(self.domain.nz):
+            layer = pcbinsw[:, :, k]
+            nonzero = layer[layer > 0]
+            if len(nonzero) > 0:
+                profile[k] = nonzero.mean()
+        return profile
+    
+    def get_canopy_scattered_sw(self) -> np.ndarray:
+        """
+        Get scattered (reflected) SW radiation from canopy cells.
+        
+        Returns:
+            3D array of scattered SW flux (W/m³) indexed by (i, j, k)
+        """
+        return self._pcbinswref.to_numpy()
+    
+    def get_canopy_to_canopy_sw(self) -> np.ndarray:
+        """
+        Get SW radiation received from other canopy cells (canopy-to-canopy scattering).
+        
+        This is radiation scattered by one canopy cell and absorbed by another.
+        Only non-zero if canopy_to_canopy=True in config.
+        
+        Returns:
+            3D array of canopy-to-canopy SW flux (W/m³) indexed by (i, j, k)
+        """
+        return self._pcbinswc2c_total.to_numpy()
+    
+    def get_surface_sw_from_canopy(self) -> np.ndarray:
+        """
+        Get SW radiation received by surfaces from canopy scattering.
+        
+        Returns:
+            1D array of SW from canopy (W/m²) per surface element
+        """
+        return self._surfinswpc.to_numpy()