matterviz 0.3.4 → 0.3.6

package/dist/src/lib/convex-hull/thermodynamics.ts

@@ -0,0 +1,2192 @@
+import {
+  count_atoms_in_composition,
+  extract_formula_elements,
+  sort_by_electronegativity,
+} from '$lib/composition'
+import type { ElementSymbol } from '$lib/element'
+import * as math from '$lib/math'
+import type { Point2D, Point3D } from '$lib/math'
+import {
+  barycentric_to_ternary_xyz,
+  barycentric_to_tetrahedral,
+  composition_to_barycentric_3d,
+  composition_to_barycentric_4d,
+  composition_to_barycentric_nd,
+} from './barycentric-coords'
+import type {
+  ConvexHullEntry,
+  ConvexHullFace,
+  ConvexHullTriangle,
+  PhaseData,
+  PhaseStats,
+  Plane,
+  ProcessedPhaseData,
+} from './types'
+import { get_arity, HULL_STABILITY_TOL, is_on_hull, is_unary_entry } from './helpers'
+
+// Track warned keys to avoid log spam on large datasets with repeated invalid keys
+const warned_keys = new Set<string>()
+const cross_point_2d = (origin: Point2D, point_a: Point2D, point_b: Point2D) =>
+  (point_a.x - origin.x) * (point_b.y - origin.y) -
+  (point_a.y - origin.y) * (point_b.x - origin.x)
+
+// Normalize convex hull composition keys by stripping oxidation states (e.g. "V4+" -> "V")
+// and merging amounts for keys that map to the same element. Filters non-positive amounts.
+// Only extracts FIRST valid element from each key (e.g. "Fe2O3" -> "Fe", not both Fe and O).
+export function normalize_hull_composition_keys(
+  composition: Record<string, number>,
+): Partial<Record<ElementSymbol, number>> {
+  const normalized: Partial<Record<ElementSymbol, number>> = {}
+  for (const [key, amount] of Object.entries(composition)) {
+    if (typeof amount !== `number` || !Number.isFinite(amount) || amount <= 0) continue
+    // Extract first valid element symbol from key (handles oxidation states like "V4+", "Fe2+")
+    const elem = extract_formula_elements(key, { unique: false })[0]
+    if (!elem) {
+      // Dedupe warnings to avoid log spam on large datasets
+      if (!warned_keys.has(key)) {
+        warned_keys.add(key)
+        console.warn(`Skipping unrecognized composition key: "${key}"`)
+      }
+      continue
+    }
+    normalized[elem] = (normalized[elem] ?? 0) + amount
+  }
+  return normalized
+}
+
+export function process_hull_entries(entries: PhaseData[]): ProcessedPhaseData {
+  // Normalize composition keys to strip oxidation states (e.g. "Fe3+" -> "Fe")
+  // Filter out entries whose composition normalizes to {} (all keys invalid or non-positive)
+  const normalized_entries = entries
+    .map((entry) => ({
+      ...entry,
+      composition: normalize_hull_composition_keys(entry.composition),
+    }))
+    .filter((entry) => Object.keys(entry.composition).length > 0)
+
+  // Single-pass partition instead of two filter passes
+  const stable_entries: PhaseData[] = []
+  const unstable_entries: PhaseData[] = []
+  for (const entry of normalized_entries) {
+    const stable =
+      typeof entry.is_stable === `boolean`
+        ? entry.is_stable
+        : (entry.e_above_hull ?? Infinity) <= HULL_STABILITY_TOL
+    ;(stable ? stable_entries : unstable_entries).push(entry)
+  }
+
+  // Extract unique element symbols from normalized compositions
+  const elements = Array.from(
+    new Set(normalized_entries.flatMap((entry) => Object.keys(entry.composition))),
+  ).sort() as ElementSymbol[]
+
+  const el_refs = Object.fromEntries(
+    stable_entries
+      .filter(is_unary_entry)
+      .map((entry) => [Object.keys(entry.composition)[0], entry]),
+  )
+
+  return {
+    entries: normalized_entries,
+    stable_entries,
+    unstable_entries,
+    elements,
+    el_refs,
+  }
+}
+
+// Get energy per atom with correction applied, or fallback to raw energy_per_atom/energy.
+// Note: correction is expected to be a total-entry value (eV), not per-atom.
+// This matches the Materials Project convention where corrections are applied to total energies.
+function get_energy_per_atom(entry: PhaseData): number {
+  // Use Math.max instead of || to prevent pathological negative totals from flipping sign
+  const atoms = Math.max(count_atoms_in_composition(entry.composition), 1e-12)
+  if (typeof entry.correction === `number`) {
+    const total =
+      typeof entry.energy_per_atom === `number`
+        ? entry.energy_per_atom * atoms
+        : (entry.energy ?? 0)
+    return (total + entry.correction) / atoms
+  }
+  return entry.energy_per_atom ?? (entry.energy ?? 0) / atoms
+}
+
+export function compute_e_form_per_atom(
+  entry: PhaseData,
+  el_refs: Record<string, PhaseData>,
+): number | null {
+  const atoms = count_atoms_in_composition(entry.composition)
+  if (atoms <= 0) return null
+  let ref_sum = 0
+  for (const [el, amt] of Object.entries(entry.composition)) {
+    const ref = el_refs[el]
+    if (!ref) return null
+    ref_sum += (amt / atoms) * get_energy_per_atom(ref)
+  }
+  return get_energy_per_atom(entry) - ref_sum
+}
+
+export function find_lowest_energy_unary_refs(
+  entries: PhaseData[],
+): Record<string, PhaseData> {
+  const refs: Record<string, PhaseData> = {}
+  for (const entry of entries) {
+    if (!is_unary_entry(entry)) continue
+    const el = Object.keys(entry.composition).find(
+      (key) => (entry.composition[key as ElementSymbol] ?? 0) > 0,
+    )
+    if (!el) continue
+    const current = refs[el]
+    if (!current || get_energy_per_atom(entry) < get_energy_per_atom(current)) {
+      refs[el] = entry
+    }
+  }
+  return refs
+}
+
+// Calculate energy above hull (eV/atom). Missing pure element refs default to E_form = 0.
+export function calculate_e_above_hull(
+  entry: PhaseData,
+  reference_entries: PhaseData[],
+): number
+export function calculate_e_above_hull(
+  entries: PhaseData[],
+  reference_entries: PhaseData[],
+): Record<string, number>
+export function calculate_e_above_hull(
+  input: PhaseData | PhaseData[],
+  reference_entries: PhaseData[],
+): number | Record<string, number> {
+  const is_single = !Array.isArray(input)
+  const entries_of_interest = is_single ? [input] : input
+
+  if (entries_of_interest.length === 0) return {} // Empty input → empty result (not an error)
+  if (reference_entries.length === 0) {
+    throw new Error(`Reference entries cannot be empty`)
+  }
+
+  // 1. Identify chemical system
+  const elements = Array.from(
+    new Set(reference_entries.flatMap((entry) => Object.keys(entry.composition))),
+  ).sort() as ElementSymbol[]
+
+  // 2. Validate subset
+  const element_set = new Set(elements)
+  for (const entry of entries_of_interest) {
+    for (const el of Object.keys(entry.composition)) {
+      if (!element_set.has(el as ElementSymbol)) {
+        throw new Error(
+          `Entry contains element ${el} not present in reference system: ${elements.join(`-`)}`,
+        )
+      }
+    }
+  }
+
+  // 3. Compute formation energies
+  const refs = find_lowest_energy_unary_refs(reference_entries)
+  const compute_e_form = (entry: PhaseData) =>
+    typeof entry.e_form_per_atom === `number`
+      ? entry.e_form_per_atom
+      : compute_e_form_per_atom(entry, refs)
+
+  const interest_data = entries_of_interest.map((entry) => ({
+    entry,
+    e_form: compute_e_form(entry),
+  }))
+
+  // 4. Branch by arity
+  const arity = elements.length
+  const results: Record<string, number> = {}
+
+  if (arity === 1) {
+    // Unary system
+    for (const { entry, e_form } of interest_data) {
+      const id = entry.entry_id ?? JSON.stringify(entry.composition)
+      // For unary, e_above_hull is simply e_form (since stable state is 0)
+      // Unless we have multiple polymorphs, in which case the hull is at min(e_form) which should be 0
+      // But compute_e_form_per_atom already subtracts the stable unary reference energy.
+      // So e_form IS e_above_hull for unary systems if correction logic holds.
+      results[id] = e_form ?? NaN
+    }
+  } else if (arity === 2) {
+    // Binary system
+    const [_el1, el2] = elements
+    // Build hull points from references
+    const hull_input_map = new Map<number, number>() // x -> min_e_form
+
+    for (const ref of reference_entries) {
+      const e_form = compute_e_form(ref)
+      if (typeof e_form !== `number`) continue
+      const total = count_atoms_in_composition(ref.composition)
+      if (total <= 0) continue
+      const x = (ref.composition[el2] ?? 0) / total
+      const current = hull_input_map.get(x)
+      if (current === undefined || e_form < current) {
+        hull_input_map.set(x, e_form)
+      }
+    }
+    // Ensure endpoints (pure elements default to e_form = 0)
+    if (!hull_input_map.has(0)) hull_input_map.set(0, 0)
+    if (!hull_input_map.has(1)) hull_input_map.set(1, 0)
+
+    const hull_points: Point2D[] = Array.from(hull_input_map, ([x, y]) => ({ x, y }))
+    const lower_hull = compute_lower_hull_2d(hull_points)
+
+    for (const { entry, e_form } of interest_data) {
+      const id = entry.entry_id ?? JSON.stringify(entry.composition)
+      if (typeof e_form !== `number`) {
+        results[id] = NaN
+        continue
+      }
+      const total = count_atoms_in_composition(entry.composition)
+      // Guard for degenerate compositions (mirror the refs loop check)
+      if (total <= 0) {
+        results[id] = NaN
+        continue
+      }
+      const x = (entry.composition[el2] ?? 0) / total
+      const y_hull = interpolate_hull_2d(lower_hull, x)
+      results[id] = y_hull === null ? NaN : Math.max(0, e_form - y_hull)
+    }
+  } else if (arity === 3) {
+    // Ternary system
+    const ref_points: Point3D[] = []
+    for (const ref of reference_entries) {
+      const e_form = compute_e_form(ref)
+      if (typeof e_form !== `number`) continue
+      try {
+        const bary = composition_to_barycentric_3d(ref.composition, elements)
+        const point = barycentric_to_ternary_xyz(bary, e_form)
+        ref_points.push(point)
+      } catch {
+        // Ignore invalid compositions
+      }
+    }
+    // Ensure corner points (pure elements default to e_form = 0)
+    for (const el of elements) {
+      const corner = barycentric_to_ternary_xyz(
+        composition_to_barycentric_3d({ [el]: 1 }, elements),
+        0,
+      )
+      if (
+        !ref_points.some(
+          (point) =>
+            Math.hypot(point.x - corner.x, point.y - corner.y, point.z - corner.z) < 1e-9,
+        )
+      ) {
+        ref_points.push(corner)
+      }
+    }
+
+    const hull_triangles = compute_lower_hull_triangles(ref_points)
+    const hull_models = build_lower_hull_model(hull_triangles)
+
+    for (const { entry, e_form } of interest_data) {
+      const id = entry.entry_id ?? JSON.stringify(entry.composition)
+      if (typeof e_form !== `number`) {
+        results[id] = NaN
+        continue
+      }
+      try {
+        const bary = composition_to_barycentric_3d(entry.composition, elements)
+        const point = barycentric_to_ternary_xyz(bary, e_form)
+        const z_hull = e_hull_at_xy(hull_models, point.x, point.y)
+        results[id] = z_hull === null ? NaN : Math.max(0, point.z - z_hull)
+      } catch {
+        results[id] = NaN
+      }
+    }
+  } else if (arity === 4) {
+    // Quaternary system
+    const ref_points: Point4D[] = []
+    for (const ref of reference_entries) {
+      const e_form = compute_e_form(ref)
+      if (typeof e_form !== `number`) continue
+      try {
+        const bary = composition_to_barycentric_4d(ref.composition, elements)
+        const tet = barycentric_to_tetrahedral(bary)
+        ref_points.push({ ...tet, w: e_form })
+      } catch {
+        // Ignore invalid
+      }
+    }
+
+    // Ensure corner points (pure elements default to e_form = 0)
+    for (const el of elements) {
+      const tet = barycentric_to_tetrahedral(
+        composition_to_barycentric_4d({ [el]: 1 }, elements),
+      )
+      const corner = { ...tet, w: 0 }
+      const dist = (point: Point4D) =>
+        Math.hypot(point.x - corner.x, point.y - corner.y, point.z - corner.z, point.w)
+      if (!ref_points.some((point) => dist(point) < 1e-9)) ref_points.push(corner)
+    }
+
+    const hull_tetrahedra = compute_lower_hull_4d(ref_points)
+    const interest_points: Point4D[] = []
+    const interest_indices: number[] = []
+
+    interest_data.forEach(({ entry, e_form }, idx) => {
+      if (typeof e_form === `number`) {
+        try {
+          const bary = composition_to_barycentric_4d(entry.composition, elements)
+          const tet = barycentric_to_tetrahedral(bary)
+          interest_points.push({ ...tet, w: e_form })
+          interest_indices.push(idx)
+        } catch {
+          // Skip
+        }
+      }
+    })
+
+    const distances = compute_e_above_hull_4d(interest_points, hull_tetrahedra)
+
+    // Build reverse lookup for O(1) access
+    const idx_to_point_idx = new Map<number, number>()
+    interest_indices.forEach((original_idx, point_idx) => {
+      idx_to_point_idx.set(original_idx, point_idx)
+    })
+
+    // Map back
+    for (let idx = 0; idx < interest_data.length; idx++) {
+      const { entry } = interest_data[idx]
+      const id = entry.entry_id ?? JSON.stringify(entry.composition)
+      const point_idx = idx_to_point_idx.get(idx) ?? -1
+      if (point_idx !== -1) {
+        results[id] = Math.max(0, distances[point_idx])
+      } else {
+        results[id] = NaN
+      }
+    }
+  } else {
+    // Arity 5+ uses generalized N-dimensional convex hull
+    // Helper to convert entry to hull point, returns null on expected errors
+    const to_hull_point = (entry: PhaseData, e_form: number): number[] | null => {
+      try {
+        return [...composition_to_barycentric_nd(entry.composition, elements), e_form]
+      } catch (err) {
+        // Skip expected errors (missing elements), warn on unexpected
+        if (err instanceof Error && !err.message.includes(`no elements from the system`)) {
+          console.warn(`Skipping entry: ${err.message}`)
+        }
+        return null
+      }
+    }
+
+    // Build reference points
+    const ref_points: number[][] = []
+    for (const ref of reference_entries) {
+      const e_form = compute_e_form(ref)
+      if (typeof e_form !== `number`) continue
+      const point = to_hull_point(ref, e_form)
+      if (point) ref_points.push(point)
+    }
+
+    // Ensure corner points (pure elements default to e_form = 0)
+    for (let el_idx = 0; el_idx < arity; el_idx++) {
+      const corner = Array(arity + 1).fill(0)
+      corner[el_idx] = 1 // ith barycentric coord = 1
+      if (!ref_points.some((pt) => norm_nd(subtract_nd(pt, corner)) < EPS)) {
+        ref_points.push(corner)
+      }
+    }
+
+    const hull_facets = compute_lower_hull_nd(compute_quickhull_nd(ref_points))
+
+    // Warn if hull is degenerate (all points coplanar or insufficient spread)
+    if (hull_facets.length === 0 && ref_points.length >= arity + 1) {
+      console.warn(
+        `N-dimensional hull for ${arity}-element system is degenerate. ` +
+          `Falling back to tie-hyperplane at energy 0. ` +
+          `Consider using pymatgen for complex high-dimensional phase diagrams.`,
+      )
+    }
+
+    // Build query points with mapping back to original indices
+    const interest_points: number[][] = []
+    const idx_to_point_idx = new Map<number, number>()
+
+    for (let idx = 0; idx < interest_data.length; idx++) {
+      const { entry, e_form } = interest_data[idx]
+      if (typeof e_form !== `number`) continue
+      const point = to_hull_point(entry, e_form)
+      if (point) {
+        idx_to_point_idx.set(idx, interest_points.length)
+        interest_points.push(point)
+      }
+    }
+
+    // Compute hull distances (empty array if degenerate hull)
+    const distances =
+      hull_facets.length > 0
+        ? compute_e_above_hull_nd(interest_points, hull_facets, ref_points)
+        : []
+
+    // Map results back to entries
+    for (let idx = 0; idx < interest_data.length; idx++) {
+      const { entry, e_form } = interest_data[idx]
+      const id = entry.entry_id ?? JSON.stringify(entry.composition)
+      const point_idx = idx_to_point_idx.get(idx)
+
+      if (point_idx === undefined) {
+        results[id] = NaN
+      } else if (hull_facets.length === 0 && typeof e_form === `number`) {
+        // Degenerate case: hull is tie-hyperplane at energy 0
+        results[id] = Math.max(0, e_form)
+      } else {
+        results[id] = Math.max(0, distances[point_idx])
+      }
+    }
+  }
+
+  if (is_single) {
+    const id =
+      entries_of_interest[0].entry_id ?? JSON.stringify(entries_of_interest[0].composition)
+    return results[id]
+  }
+  return results
+}
+
+export function get_convex_hull_stats(
+  processed_entries: PhaseData[],
+  elements: ElementSymbol[],
+  max_arity: number = 4,
+): PhaseStats | null {
+  if (processed_entries.length === 0) return null
+  max_arity = Math.max(1, max_arity)
+
+  let unary = 0
+  let binary = 0
+  let ternary = 0
+  let quaternary = 0
+  let quinary_plus = 0
+  for (const entry of processed_entries) {
+    const arity = get_arity(entry)
+    if (arity === 1) unary++
+    else if (arity === 2) binary++
+    else if (arity === 3) ternary++
+    else if (arity === 4) quaternary++
+    else if (arity >= 5) quinary_plus++
+  }
+  // Zero out counts beyond system dimensionality for cleaner display
+  // quinary_plus is intentionally not zeroed — it's a catch-all bucket that
+  // is naturally 0 for systems with fewer than 5 elements
+  if (max_arity < 4) quaternary = 0
+  if (max_arity < 3) ternary = 0
+  if (max_arity < 2) binary = 0
+
+  const stable_count = processed_entries.filter((entry) => is_on_hull(entry)).length
+  const unstable_count = processed_entries.length - stable_count
+
+  const energies = processed_entries
+    .map(
+      (entry) => entry.e_form_per_atom ?? entry.energy_per_atom ?? get_energy_per_atom(entry),
+    )
+    .filter(Number.isFinite)
+
+  // Use reduce instead of Math.min/max(...arr) to avoid stack overflow on large datasets
+  const energy_range =
+    energies.length > 0
+      ? {
+          min: energies.reduce((min, val) => (val < min ? val : min), Infinity),
+          max: energies.reduce((max, val) => (val > max ? val : max), -Infinity),
+          avg: energies.reduce((sum, val) => sum + val, 0) / energies.length,
+        }
+      : { min: 0, max: 0, avg: 0 }
+
+  const hull_distances = processed_entries
+    .map((entry) => entry.e_above_hull)
+    .filter((val): val is number => typeof val === `number` && val >= 0)
+  const hull_distance =
+    hull_distances.length > 0
+      ? {
+          max: hull_distances.reduce((max, val) => (val > max ? val : max), -Infinity),
+          avg: hull_distances.reduce((sum, val) => sum + val, 0) / hull_distances.length,
+        }
+      : { max: 0, avg: 0 }
+
+  return {
+    total: processed_entries.length,
+    unary,
+    binary,
+    ternary,
+    quaternary,
+    quinary_plus,
+    stable: stable_count,
+    unstable: unstable_count,
+    energy_range,
+    hull_distance,
+    elements: elements.length,
+    chemical_system: sort_by_electronegativity([...elements]).join(`-`),
+    max_arity,
+  }
+}
+
+// Result type for process_hull_for_stats
+export interface HighDimHullResult {
+  stable_entries: ConvexHullEntry[]
+  unstable_entries: ConvexHullEntry[]
+  phase_stats: PhaseStats | null
+}
+
+// Convert a PhaseData entry to a ConvexHullEntry with default visual fields.
+// x/y/z default to 0 since high-dim systems aren't visually plotted.
+function to_hull_entry(entry: PhaseData): ConvexHullEntry {
+  return {
+    ...entry,
+    is_element: get_arity(entry) === 1,
+    x: 0,
+    y: 0,
+    z: 0,
+  }
+}
+
+// Process raw hull entries for high-dimensional systems (5+ elements) where the
+// ConvexHull visual component can't render. Computes formation energies, hull distances,
+// stable/unstable classification, and phase stats. Returns null on failure.
+// Optionally accepts `elements` to scope the chemical system; if omitted, elements
+// are derived from the entries' compositions.
+export function process_hull_for_stats(
+  entries: PhaseData[],
+  elements?: ElementSymbol[],
+): HighDimHullResult | null {
+  if (!entries.length) return null
+
+  const processed = process_hull_entries(entries)
+  if (!processed.entries.length) return null
+  const hull_elements = elements ?? processed.elements
+
+  // Compute formation energies
+  const el_refs = find_lowest_energy_unary_refs(processed.entries)
+  for (const entry of processed.entries) {
+    if (entry.e_form_per_atom === undefined) {
+      const e_form = compute_e_form_per_atom(entry, el_refs)
+      if (e_form !== null) entry.e_form_per_atom = e_form
+    }
+  }
+
+  // Compute hull distances. Note: entries without entry_id are keyed by
+  // JSON.stringify(composition), so polymorphs at the same composition
+  // collide — the last-processed entry's distance wins for all of them.
+  try {
+    const hull_distances = calculate_e_above_hull(processed.entries, processed.entries)
+
+    for (const entry of processed.entries) {
+      const dist = hull_distances[entry.entry_id ?? JSON.stringify(entry.composition)]
+      if (typeof dist === `number` && Number.isFinite(dist)) {
+        entry.e_above_hull = dist
+        entry.is_stable = dist < HULL_STABILITY_TOL
+      } else {
+        // Clear stale hull metadata so previous values don't persist
+        entry.e_above_hull = undefined
+        entry.is_stable = undefined
+      }
+    }
+  } catch (err) {
+    console.warn(`Failed to compute high-dim hull:`, err)
+    return null
+  }
+
+  const hull_entries = processed.entries.map(to_hull_entry)
+  return {
+    stable_entries: hull_entries.filter((entry) => is_on_hull(entry)),
+    unstable_entries: hull_entries.filter((entry) => !is_on_hull(entry)),
+    phase_stats: get_convex_hull_stats(processed.entries, hull_elements, hull_elements.length),
+  }
+}
+
+// --- 2D Convex Hull (Binary Phase Diagrams) ---
+
+export function compute_lower_hull_2d(points: Point2D[]): Point2D[] {
+  // Andrew's monotone chain for lower hull
+  // Sort by x then y
+  const sorted = [...points].sort((p1, p2) => p1.x - p2.x || p1.y - p2.y)
+  const lower: Point2D[] = []
+
+  for (const point of sorted) {
+    while (
+      lower.length >= 2 &&
+      cross_point_2d(lower[lower.length - 2], lower[lower.length - 1], point) <= 0
+    )
+      lower.pop()
+    lower.push(point)
+  }
+  return lower
+}
+
+export function interpolate_hull_2d(hull: Point2D[], x: number): number | null {
+  if (hull.length < 2) return null
+  // Handle out of bounds by clamping to endpoints
+  if (x <= hull[0].x) return hull[0].y
+  if (x >= hull[hull.length - 1].x) return hull[hull.length - 1].y
+
+  for (let idx = 0; idx < hull.length - 1; idx++) {
+    const p1 = hull[idx]
+    const p2 = hull[idx + 1]
+    if (x >= p1.x && x <= p2.x) {
+      const t = (x - p1.x) / Math.max(1e-12, p2.x - p1.x)
+      return p1.y * (1 - t) + p2.y * t
+    }
+  }
+  return null
+}
+
+// --- Convex hull geometry ---
+
+const EPS = 1e-9
+
+const subtract = (pt1: Point3D, pt2: Point3D): Point3D => ({
+  x: pt1.x - pt2.x,
+  y: pt1.y - pt2.y,
+  z: pt1.z - pt2.z,
+})
+
+const cross = (vec1: Point3D, vec2: Point3D): Point3D => ({
+  x: vec1.y * vec2.z - vec1.z * vec2.y,
+  y: vec1.z * vec2.x - vec1.x * vec2.z,
+  z: vec1.x * vec2.y - vec1.y * vec2.x,
+})
+
+const norm = (point: Point3D): number =>
+  Math.sqrt(point.x * point.x + point.y * point.y + point.z * point.z)
+
+function normalize(point: Point3D): Point3D {
+  const length = norm(point)
+  if (length < EPS) return { x: 0, y: 0, z: 0 }
+  return { x: point.x / length, y: point.y / length, z: point.z / length }
+}
+
+function compute_plane(p1: Point3D, p2: Point3D, p3: Point3D): Plane {
+  const edge_12 = subtract(p2, p1)
+  const edge_13 = subtract(p3, p1)
+  const normal = normalize(cross(edge_12, edge_13))
+  const offset = -(normal.x * p1.x + normal.y * p1.y + normal.z * p1.z)
+  return { normal, offset }
+}
+
+const point_plane_signed_distance = (plane: Plane, point: Point3D): number =>
+  plane.normal.x * point.x + plane.normal.y * point.y + plane.normal.z * point.z + plane.offset
+
+const compute_centroid = (p1: Point3D, p2: Point3D, p3: Point3D): Point3D => ({
+  x: (p1.x + p2.x + p3.x) / 3,
+  y: (p1.y + p2.y + p3.y) / 3,
+  z: (p1.z + p2.z + p3.z) / 3,
+})
+
+function distance_point_to_line(
+  line_start: Point3D,
+  line_end: Point3D,
+  point: Point3D,
+): number {
+  const line_vec = subtract(line_end, line_start)
+  const to_point = subtract(point, line_start)
+  const cross_prod = cross(line_vec, to_point)
+  const line_len = norm(line_vec)
+  if (line_len < EPS) return 0
+  return norm(cross_prod) / line_len
+}
+
+function choose_initial_tetrahedron(
+  points: Point3D[],
+): [number, number, number, number] | null {
+  if (points.length < 4) return null
+  let idx_min_x = 0
+  let idx_max_x = 0
+  for (let idx = 1; idx < points.length; idx++) {
+    if (points[idx].x < points[idx_min_x].x) idx_min_x = idx
+    if (points[idx].x > points[idx_max_x].x) idx_max_x = idx
+  }
+  if (idx_min_x === idx_max_x) return null
+  let idx_far_line = -1
+  let best_dist_line = -1
+  for (let idx = 0; idx < points.length; idx++) {
+    if (idx === idx_min_x || idx === idx_max_x) continue
+    const dist = distance_point_to_line(points[idx_min_x], points[idx_max_x], points[idx])
+    if (dist > best_dist_line) {
+      best_dist_line = dist
+      idx_far_line = idx
+    }
+  }
+  if (idx_far_line === -1 || best_dist_line < EPS) return null
+  const plane0 = compute_plane(points[idx_min_x], points[idx_max_x], points[idx_far_line])
+  let idx_far_plane = -1
+  let best_dist_plane = -1
+  for (let idx = 0; idx < points.length; idx++) {
+    if (idx === idx_min_x || idx === idx_max_x || idx === idx_far_line) continue
+    const dist = Math.abs(point_plane_signed_distance(plane0, points[idx]))
+    if (dist > best_dist_plane) {
+      best_dist_plane = dist
+      idx_far_plane = idx
+    }
+  }
+  if (idx_far_plane === -1 || best_dist_plane < EPS) return null
+  return [idx_min_x, idx_max_x, idx_far_line, idx_far_plane]
+}
+
+function make_face(
+  points: Point3D[],
+  a: number,
+  b: number,
+  c: number,
+  interior_point: Point3D,
+): ConvexHullFace {
+  let plane = compute_plane(points[a], points[b], points[c])
+  let centroid = compute_centroid(points[a], points[b], points[c])
+  const dist_interior = point_plane_signed_distance(plane, interior_point)
+  if (dist_interior > 0) {
+    plane = compute_plane(points[a], points[c], points[b])
+    centroid = compute_centroid(points[a], points[c], points[b])
+    return { vertices: [a, c, b], plane, centroid, outside_points: new Set<number>() }
+  }
+  return { vertices: [a, b, c], plane, centroid, outside_points: new Set<number>() }
+}
+
+function assign_outside_points(
+  face: ConvexHullFace,
+  points: Point3D[],
+  candidate_indices: number[],
+): void {
+  face.outside_points.clear()
+  for (const idx of candidate_indices) {
+    const distance = point_plane_signed_distance(face.plane, points[idx])
+    if (distance > EPS) face.outside_points.add(idx)
+  }
+}
+
+function collect_candidate_points(faces: ConvexHullFace[]): number[] {
+  const set = new Set<number>()
+  for (const face of faces) for (const idx of face.outside_points) set.add(idx)
+  return Array.from(set)
+}
+
+function farthest_point_for_face(
+  points: Point3D[],
+  face: ConvexHullFace,
+): { idx: number; distance: number } | null {
+  let best_idx = -1
+  let best_distance = -1
+  for (const idx of face.outside_points) {
+    const distance = point_plane_signed_distance(face.plane, points[idx])
+    if (distance > best_distance) {
+      best_distance = distance
+      best_idx = idx
+    }
+  }
+  if (best_idx === -1) return null
+  return { idx: best_idx, distance: best_distance }
+}
+
+function build_horizon(
+  faces: ConvexHullFace[],
+  visible_face_indices: Set<number>,
+): [number, number][] {
+  const edge_count = new Map<string, [number, number]>()
+  for (const face_idx of visible_face_indices) {
+    const face = faces[face_idx]
+    const [a, b, c] = face.vertices
+    const edges: [number, number][] = [
+      [a, b],
+      [b, c],
+      [c, a],
+    ]
+    for (const [u, v] of edges) {
+      const key = u < v ? `${u}|${v}` : `${v}|${u}`
+      if (!edge_count.has(key)) edge_count.set(key, [u, v])
+      else edge_count.set(key, [Number.NaN, Number.NaN])
+    }
+  }
+  const horizon: [number, number][] = []
+  for (const uv of edge_count.values()) {
+    if (Number.isNaN(uv[0])) continue
+    horizon.push(uv)
+  }
+  return horizon
+}
+
+export function compute_quickhull_triangles(points: Point3D[]): ConvexHullTriangle[] {
+  if (points.length < 4) return [] // hull needs at least 4 non-coplanar points, bail if not provided
+  const initial = choose_initial_tetrahedron(points)
+  if (!initial) return []
+  const [i0, i1, i2, i3] = initial
+  const interior_point = {
+    x: (points[i0].x + points[i1].x + points[i2].x + points[i3].x) / 4,
+    y: (points[i0].y + points[i1].y + points[i2].y + points[i3].y) / 4,
+    z: (points[i0].z + points[i1].z + points[i2].z + points[i3].z) / 4,
+  }
+  const faces: ConvexHullFace[] = [
+    make_face(points, i0, i1, i2, interior_point),
+    make_face(points, i0, i2, i3, interior_point),
+    make_face(points, i0, i3, i1, interior_point),
+    make_face(points, i1, i3, i2, interior_point),
+  ]
+  const all_indices: number[] = []
+  for (let idx = 0; idx < points.length; idx++) {
+    if (idx === i0 || idx === i1 || idx === i2 || idx === i3) continue
+    all_indices.push(idx)
+  }
+  for (const face of faces) assign_outside_points(face, points, all_indices)
+  while (true) {
+    let chosen_face_idx = -1
+    let chosen_point_idx = -1
+    let max_distance = -1
+    for (let face_idx = 0; face_idx < faces.length; face_idx++) {
+      const face = faces[face_idx]
+      if (face.outside_points.size === 0) continue
+      const far = farthest_point_for_face(points, face)
+      if (far && far.distance > max_distance) {
+        max_distance = far.distance
+        chosen_face_idx = face_idx
+        chosen_point_idx = far.idx
+      }
+    }
+    if (chosen_face_idx === -1) break
+    const eye_idx = chosen_point_idx
+    const visible_face_indices = new Set<number>()
+    for (let face_idx = 0; face_idx < faces.length; face_idx++) {
+      const face = faces[face_idx]
+      const dist = point_plane_signed_distance(face.plane, points[eye_idx])
+      if (dist > EPS) visible_face_indices.add(face_idx)
+    }
+    const horizon_edges = build_horizon(faces, visible_face_indices)
+    const visible_faces = Array.from(visible_face_indices).sort((a, b) => b - a)
+    const candidate_points = collect_candidate_points(visible_faces.map((idx) => faces[idx]))
+    for (const idx of visible_faces) faces.splice(idx, 1)
+    const new_faces: ConvexHullFace[] = []
+    for (const [u, v] of horizon_edges) {
+      const new_face = make_face(points, u, v, eye_idx, interior_point)
+      new_faces.push(new_face)
+    }
+    for (const face of new_faces) face.outside_points.clear()
+    for (const idx of candidate_points) {
+      if (idx === eye_idx) continue
+      let best_face: ConvexHullFace | null = null
+      let best_distance = EPS
+      for (const face of new_faces) {
+        const dist = point_plane_signed_distance(face.plane, points[idx])
+        if (dist > best_distance) {
+          best_distance = dist
+          best_face = face
+        }
+      }
+      if (best_face) best_face.outside_points.add(idx)
+    }
+    faces.push(...new_faces)
+  }
+  return faces.map((face) => {
+    const [a, b, c] = face.vertices
+    const normal = face.plane.normal
+    const centroid = face.centroid
+    return {
+      vertices: [points[a], points[b], points[c]] as [Point3D, Point3D, Point3D],
+      normal,
+      centroid,
+    }
+  })
+}
+
+export function compute_lower_hull_triangles(points: Point3D[]): ConvexHullTriangle[] {
+  const all_faces = compute_quickhull_triangles(points)
+  return all_faces.filter((face) => face.normal.z < 0 - EPS)
+}
+
+// ---------- Lower hull face models and energy-above-hull ----------
+
+export interface HullFaceModel {
+  a: number
+  b: number
+  c: number
+  x1: number
+  y1: number
+  x2: number
+  y2: number
+  x3: number
+  y3: number
+  min_x: number
+  max_x: number
+  min_y: number
+  max_y: number
+  denom: number
+}
+
+export const build_lower_hull_model = (faces: ConvexHullTriangle[]): HullFaceModel[] =>
+  faces.map((tri) => {
+    const [p1, p2, p3] = tri.vertices
+    const plane = (() => {
+      const x1 = p1.x,
+        y1 = p1.y,
+        z1 = p1.z
+      const x2 = p2.x,
+        y2 = p2.y,
+        z2 = p2.z
+      const x3 = p3.x,
+        y3 = p3.y,
+        z3 = p3.z
+      const det = x1 * (y2 - y3) + x2 * (y3 - y1) + x3 * (y1 - y2)
+      if (Math.abs(det) < 1e-12) return { a: 0, b: 0, c: (z1 + z2 + z3) / 3 }
+      const a = (z1 * (y2 - y3) + z2 * (y3 - y1) + z3 * (y1 - y2)) / det
+      const b = (z1 * (x3 - x2) + z2 * (x1 - x3) + z3 * (x2 - x1)) / det
+      const c =
+        (z1 * (x2 * y3 - x3 * y2) + z2 * (x3 * y1 - x1 * y3) + z3 * (x1 * y2 - x2 * y1)) / det
+      return { a, b, c }
+    })()
+    const [min_x, _mx, max_x] = [p1.x, p2.x, p3.x].sort((a, b) => a - b)
+    const [min_y, _my, max_y] = [p1.y, p2.y, p3.y].sort((a, b) => a - b)
+    const { x: x1, y: y1 } = p1
+    const { x: x2, y: y2 } = p2
+    const { x: x3, y: y3 } = p3
+    const denom = (y2 - y3) * (x1 - x3) + (x3 - x2) * (y1 - y3)
+    return {
+      a: plane.a,
+      b: plane.b,
+      c: plane.c,
+      x1,
+      y1,
+      x2,
+      y2,
+      x3,
+      y3,
+      min_x,
+      max_x,
+      min_y,
+      max_y,
+      denom,
+    }
+  })
+
+function point_in_triangle_xy(model: HullFaceModel, x: number, y: number): boolean {
+  const { x1, y1, x2, y2, x3, y3, denom } = model
+  if (Math.abs(denom) < 1e-14) return false
+  const l1 = ((y2 - y3) * (x - x3) + (x3 - x2) * (y - y3)) / denom
+  const l2 = ((y3 - y1) * (x - x3) + (x1 - x3) * (y - y3)) / denom
+  const l3 = 1 - l1 - l2
+  const eps = -1e-9
+  return l1 >= eps && l2 >= eps && l3 >= eps
+}
+
+export function e_hull_at_xy(models: HullFaceModel[], x: number, y: number) {
+  let z: number | null = null
+  for (const model of models) {
+    if (
+      x < model.min_x - 1e-9 ||
+      x > model.max_x + 1e-9 ||
+      y < model.min_y - 1e-9 ||
+      y > model.max_y + 1e-9
+    )
+      continue
+    if (!point_in_triangle_xy(model, x, y)) continue
+    const z_face = model.a * x + model.b * y + model.c
+    z = z === null ? z_face : Math.min(z, z_face)
+  }
+  return z
+}
+
+export const compute_e_above_hull_for_points = (points: Point3D[], models: HullFaceModel[]) =>
+  points.map((point) => {
+    const z_hull = e_hull_at_xy(models, point.x, point.y)
+    if (z_hull === null) return 0
+    return Math.max(0, point.z - z_hull)
+  })
+
+// --- 4D Convex Hull (Quaternary Phase Diagrams) ---
+
+export interface Point4D {
+  x: number
+  y: number
+  z: number
+  w: number // formation energy dimension
+}
+
+export interface ConvexHullTetrahedron {
+  vertices: [Point4D, Point4D, Point4D, Point4D]
+  normal: Point4D // 4D normal vector
+  centroid: Point4D
+}
+
+interface ConvexHullFace4D {
+  vertices: [number, number, number, number] // indices
+  plane: Plane4D
+  centroid: Point4D
+  outside_points: Set<number>
+}
+
+interface Plane4D {
+  normal: Point4D
+  offset: number
+}
+
+const subtract_4d = (pt1: Point4D, pt2: Point4D): Point4D => ({
+  x: pt1.x - pt2.x,
+  y: pt1.y - pt2.y,
+  z: pt1.z - pt2.z,
+  w: pt1.w - pt2.w,
+})
+
+const dot_4d = (vec_a: Point4D, vec_b: Point4D): number =>
+  vec_a.x * vec_b.x + vec_a.y * vec_b.y + vec_a.z * vec_b.z + vec_a.w * vec_b.w
+
+const norm_4d = (point: Point4D): number =>
+  Math.sqrt(point.x * point.x + point.y * point.y + point.z * point.z + point.w * point.w)
+
+function normalize_4d(point: Point4D): Point4D {
+  const length = norm_4d(point)
+  if (length < EPS) return { x: 0, y: 0, z: 0, w: 0 }
+  return {
+    x: point.x / length,
+    y: point.y / length,
+    z: point.z / length,
+    w: point.w / length,
+  }
+}
+
+// Compute normal to a 3D hyperplane in 4D space defined by 4 points
+//
+// Mathematical Background:
+// A 3D hyperplane (tetrahedral facet) in 4D is defined by 4 points. The normal vector
+// must be orthogonal to all three edge vectors spanning the hyperplane. This is the
+// 4D analog of computing a cross product of two vectors in 3D.
+//
+// Approach:
+// 1. Form three edge vectors v1, v2, v3 from point p1
+// 2. Find vector n such that: n · v1 = 0, n · v2 = 0, n · v3 = 0
+// 3. This is equivalent to finding the null space of the 3×4 matrix [v1; v2; v3]
+//
+// Implementation:
+// The normal components (nx, ny, nz, nw) are computed using Laplace expansion
+// (cofactor method) along each column of the matrix. Each component is the determinant
+// of the 3×3 submatrix obtained by removing that column, with alternating signs.
+//
+// References:
+// - Barber et al. (1996) "The Quickhull Algorithm for Convex Hulls"
+// - https://en.wikipedia.org/wiki/Cross_product#Multilinear_algebra
+// - https://mathworld.wolfram.com/Nullspace.html
+function compute_plane_4d(p1: Point4D, p2: Point4D, p3: Point4D, p4: Point4D): Plane4D {
+  // Three edge vectors from p1
+  const v1 = subtract_4d(p2, p1)
+  const v2 = subtract_4d(p3, p1)
+  const v3 = subtract_4d(p4, p1)
+
+  // Build matrix [v1; v2; v3] and compute normal via cofactor expansion
+  const matrix = [
+    [v1.x, v1.y, v1.z, v1.w],
+    [v2.x, v2.y, v2.z, v2.w],
+    [v3.x, v3.y, v3.z, v3.w],
+  ]
+
+  // Helper: extract 3×3 submatrix by removing column col_skip, then compute determinant
+  const det_submatrix = (col_skip: number): number => {
+    const cols = [0, 1, 2, 3].filter((col) => col !== col_skip)
+    const submatrix: math.Matrix3x3 = [
+      [matrix[0][cols[0]], matrix[0][cols[1]], matrix[0][cols[2]]],
+      [matrix[1][cols[0]], matrix[1][cols[1]], matrix[1][cols[2]]],
+      [matrix[2][cols[0]], matrix[2][cols[1]], matrix[2][cols[2]]],
+    ]
+    return math.det_3x3(submatrix)
+  }
+
+  // Compute normal components using Laplace expansion along each column
+  // Alternating signs: +, -, +, -
+  const signs = [1, -1, 1, -1]
+  const normal_components = [0, 1, 2, 3].map(
+    (col_idx) => signs[col_idx] * det_submatrix(col_idx),
+  )
+
+  const [x, y, z, w] = normal_components
+  const normal = normalize_4d({ x, y, z, w })
+
+  // Guard against degenerate (nearly co-planar) points
+  const normal_magnitude =
+    Math.abs(normal.x) + Math.abs(normal.y) + Math.abs(normal.z) + Math.abs(normal.w)
+  if (normal_magnitude < EPS) {
+    return { normal: { x: 0, y: 0, z: 0, w: 0 }, offset: 0 }
+  }
+
+  const offset = -dot_4d(normal, p1)
+
+  return { normal, offset }
+}
+
+const point_plane_signed_distance_4d = (plane: Plane4D, point: Point4D): number =>
+  dot_4d(plane.normal, point) + plane.offset
+
+const compute_centroid_4d = (p1: Point4D, p2: Point4D, p3: Point4D, p4: Point4D): Point4D => ({
+  x: (p1.x + p2.x + p3.x + p4.x) / 4,
+  y: (p1.y + p2.y + p3.y + p4.y) / 4,
+  z: (p1.z + p2.z + p3.z + p4.z) / 4,
+  w: (p1.w + p2.w + p3.w + p4.w) / 4,
+})
+
+function distance_point_to_hyperplane_4d(
+  p1: Point4D,
+  p2: Point4D,
+  p3: Point4D,
+  point: Point4D,
+): number {
+  // Distance from point to the 2D hyperplane spanned by p1, p2, p3
+  const v1 = subtract_4d(p2, p1)
+  const v2 = subtract_4d(p3, p1)
+  const vp = subtract_4d(point, p1)
+
+  // Project vp onto the plane spanned by v1 and v2
+  // Use Gram-Schmidt to find orthogonal component
+  const v1_norm_sq = dot_4d(v1, v1)
+  const v2_norm_sq = dot_4d(v2, v2)
+  const v1_dot_v2 = dot_4d(v1, v2)
+
+  if (v1_norm_sq < EPS || v2_norm_sq < EPS) return 0
+
+  const vp_dot_v1 = dot_4d(vp, v1)
+  const vp_dot_v2 = dot_4d(vp, v2)
+
+  // Solve linear system for projection coefficients
+  const det = v1_norm_sq * v2_norm_sq - v1_dot_v2 * v1_dot_v2
+  if (Math.abs(det) < EPS) return 0
+
+  const alpha = (v2_norm_sq * vp_dot_v1 - v1_dot_v2 * vp_dot_v2) / det
+  const beta = (v1_norm_sq * vp_dot_v2 - v1_dot_v2 * vp_dot_v1) / det
+
+  // Compute projection
+  const proj_x = p1.x + alpha * v1.x + beta * v2.x
+  const proj_y = p1.y + alpha * v1.y + beta * v2.y
+  const proj_z = p1.z + alpha * v1.z + beta * v2.z
+  const proj_w = p1.w + alpha * v1.w + beta * v2.w
+
+  // Distance is the length of (point - projection)
+  const dx = point.x - proj_x
+  const dy = point.y - proj_y
+  const dz = point.z - proj_z
+  const dw = point.w - proj_w
+
+  return Math.sqrt(dx * dx + dy * dy + dz * dz + dw * dw)
+}
+
+// Distance from point to line in 4D
+function distance_point_to_line_4d(a: Point4D, b: Point4D, p: Point4D): number {
+  const ab = subtract_4d(b, a)
+  const ap = subtract_4d(p, a)
+
+  const ab_len_sq = dot_4d(ab, ab)
+  if (ab_len_sq < EPS) return norm_4d(ap)
+
+  // Project ap onto ab
+  const t = dot_4d(ap, ab) / ab_len_sq
+  const projection = {
+    x: a.x + t * ab.x,
+    y: a.y + t * ab.y,
+    z: a.z + t * ab.z,
+    w: a.w + t * ab.w,
+  }
+
+  return norm_4d(subtract_4d(p, projection))
+}
+
+// Maximum sample size for initial simplex selection in 4D hulls (avoids O(n²) for large datasets)
+const INITIAL_SIMPLEX_SAMPLE_SIZE = 100
+
+function choose_initial_4_simplex(
+  points: Point4D[],
+): [number, number, number, number, number] | null {
+  if (points.length < 5) return null
+
+  // Find two points farthest apart across all dimensions for better numerical stability
+  // Sample a small subset if dataset is large to avoid O(n²) scaling
+  const sample_size = Math.min(points.length, INITIAL_SIMPLEX_SAMPLE_SIZE)
+  const sample_indices =
+    points.length <= sample_size
+      ? points.map((_, idx) => idx)
+      : Array.from({ length: sample_size }, (_, idx) =>
+          Math.floor((idx * points.length) / sample_size),
+        )
+
+  let idx_far_a = 0
+  let idx_far_b = 0
+  let max_dist_sq = -1
+
+  for (const idx_a of sample_indices) {
+    for (const idx_b of sample_indices) {
+      if (idx_a >= idx_b) continue
+      const pa = points[idx_a]
+      const pb = points[idx_b]
+      const dist_sq =
+        (pa.x - pb.x) ** 2 + (pa.y - pb.y) ** 2 + (pa.z - pb.z) ** 2 + (pa.w - pb.w) ** 2
+      if (dist_sq > max_dist_sq) {
+        max_dist_sq = dist_sq
+        idx_far_a = idx_a
+        idx_far_b = idx_b
+      }
+    }
+  }
+  if (idx_far_a === idx_far_b || max_dist_sq < EPS) return null
+
+  // Find point farthest from line through idx_far_a and idx_far_b
+  let idx_far_line = -1
+  let best_dist_line = -1
+  for (let idx = 0; idx < points.length; idx++) {
+    if (idx === idx_far_a || idx === idx_far_b) continue
+    const dist = distance_point_to_line_4d(points[idx_far_a], points[idx_far_b], points[idx])
+    if (dist > best_dist_line) {
+      best_dist_line = dist
+      idx_far_line = idx
+    }
+  }
+  if (idx_far_line === -1 || best_dist_line < EPS) return null
+
+  // Find point farthest from 2D plane through first three points
+  let idx_far_plane = -1
+  let best_dist_plane = -1
+  for (let idx = 0; idx < points.length; idx++) {
+    if (idx === idx_far_a || idx === idx_far_b || idx === idx_far_line) continue
+    const dist = distance_point_to_hyperplane_4d(
+      points[idx_far_a],
+      points[idx_far_b],
+      points[idx_far_line],
+      points[idx],
+    )
+    if (dist > best_dist_plane) {
+      best_dist_plane = dist
+      idx_far_plane = idx
+    }
+  }
+  if (idx_far_plane === -1 || best_dist_plane < EPS) return null
+
+  // Find point farthest from 3D hyperplane through first four points
+  const plane0 = compute_plane_4d(
+    points[idx_far_a],
+    points[idx_far_b],
+    points[idx_far_line],
+    points[idx_far_plane],
+  )
+  let idx_far_hyperplane = -1
+  let best_dist_hyperplane = -1
+  for (let idx = 0; idx < points.length; idx++) {
+    if (
+      idx === idx_far_a ||
+      idx === idx_far_b ||
+      idx === idx_far_line ||
+      idx === idx_far_plane
+    )
+      continue
+    const dist = Math.abs(point_plane_signed_distance_4d(plane0, points[idx]))
+    if (dist > best_dist_hyperplane) {
+      best_dist_hyperplane = dist
+      idx_far_hyperplane = idx
+    }
+  }
+  if (idx_far_hyperplane === -1 || best_dist_hyperplane < EPS) return null
+
+  return [idx_far_a, idx_far_b, idx_far_line, idx_far_plane, idx_far_hyperplane]
+}
+
+function make_face_4d(
+  points: Point4D[],
+  a: number,
+  b: number,
+  c: number,
+  d: number,
+  interior_point: Point4D,
+): ConvexHullFace4D {
+  let plane = compute_plane_4d(points[a], points[b], points[c], points[d])
+  let centroid = compute_centroid_4d(points[a], points[b], points[c], points[d])
+
+  const dist_interior = point_plane_signed_distance_4d(plane, interior_point)
+
+  // Ensure normal points outward (away from interior)
+  if (dist_interior > 0) {
+    // Swap two vertices to flip normal
+    plane = compute_plane_4d(points[a], points[c], points[b], points[d])
+    centroid = compute_centroid_4d(points[a], points[c], points[b], points[d])
+    return { vertices: [a, c, b, d], plane, centroid, outside_points: new Set<number>() }
+  }
+
+  return { vertices: [a, b, c, d], plane, centroid, outside_points: new Set<number>() }
+}
+
+function assign_outside_points_4d(
+  face: ConvexHullFace4D,
+  points: Point4D[],
+  candidate_indices: number[],
+): void {
+  face.outside_points.clear()
+  for (const idx of candidate_indices) {
+    const distance = point_plane_signed_distance_4d(face.plane, points[idx])
+    if (distance > EPS) face.outside_points.add(idx)
+  }
+}
+
+function collect_candidate_points_4d(faces: ConvexHullFace4D[]): number[] {
+  const set = new Set<number>()
+  for (const face of faces) {
+    for (const idx of face.outside_points) set.add(idx)
+  }
+  return Array.from(set)
+}
+
+function farthest_point_for_face_4d(
+  points: Point4D[],
+  face: ConvexHullFace4D,
+): { idx: number; distance: number } | null {
+  let best_idx = -1
+  let best_distance = -1
+  for (const idx of face.outside_points) {
+    const distance = point_plane_signed_distance_4d(face.plane, points[idx])
+    if (distance > best_distance) {
+      best_distance = distance
+      best_idx = idx
+    }
+  }
+  if (best_idx === -1) return null
+  return { idx: best_idx, distance: best_distance }
+}
+
+function build_horizon_4d(
+  faces: ConvexHullFace4D[],
+  visible_face_indices: Set<number>,
+): math.Vec3[] {
+  // In 4D, horizon "ridges" are triangles (3 vertices)
+  const ridge_count = new Map<string, math.Vec3>()
+
+  for (const face_idx of visible_face_indices) {
+    const face = faces[face_idx]
+    const [a, b, c, d] = face.vertices
+
+    // Each tetrahedron face has 4 triangular ridges
+    const ridges: math.Vec3[] = [
+      [a, b, c],
+      [a, b, d],
+      [a, c, d],
+      [b, c, d],
+    ]
+
+    for (const ridge of ridges) {
+      const sorted = ridge.slice().sort((x, y) => x - y)
+      const key = sorted.join(`|`)
+
+      if (!ridge_count.has(key)) {
+        ridge_count.set(key, ridge)
+      } else {
+        // Mark as seen twice (internal ridge)
+        ridge_count.set(key, [Number.NaN, Number.NaN, Number.NaN])
+      }
+    }
+  }
+
+  const horizon: math.Vec3[] = []
+  for (const ridge of ridge_count.values()) {
+    if (!Number.isNaN(ridge[0])) {
+      horizon.push(ridge)
+    }
+  }
+
+  return horizon
+}
+
+export function compute_quickhull_4d(points: Point4D[]): ConvexHullTetrahedron[] {
+  if (points.length < 5) return [] // Need at least 5 non-coplanar points for 4D hull
+
+  const initial = choose_initial_4_simplex(points)
+  if (!initial) return []
+
+  const [i0, i1, i2, i3, i4] = initial
+
+  // Interior point for orientation
+  const interior_point = {
+    x: (points[i0].x + points[i1].x + points[i2].x + points[i3].x + points[i4].x) / 5,
+    y: (points[i0].y + points[i1].y + points[i2].y + points[i3].y + points[i4].y) / 5,
+    z: (points[i0].z + points[i1].z + points[i2].z + points[i3].z + points[i4].z) / 5,
+    w: (points[i0].w + points[i1].w + points[i2].w + points[i3].w + points[i4].w) / 5,
+  }
+
+  // Initial 4-simplex has 5 tetrahedral faces
+  const faces: ConvexHullFace4D[] = [
+    make_face_4d(points, i0, i1, i2, i3, interior_point),
+    make_face_4d(points, i0, i1, i2, i4, interior_point),
+    make_face_4d(points, i0, i1, i3, i4, interior_point),
+    make_face_4d(points, i0, i2, i3, i4, interior_point),
+    make_face_4d(points, i1, i2, i3, i4, interior_point),
+  ]
+
+  const all_indices: number[] = []
+  for (let idx = 0; idx < points.length; idx++) {
+    if (idx === i0 || idx === i1 || idx === i2 || idx === i3 || idx === i4) continue
+    all_indices.push(idx)
+  }
+
+  for (const face of faces) {
+    assign_outside_points_4d(face, points, all_indices)
+  }
+
+  // Main Quick Hull iteration
+  while (true) {
+    // Step 1: Find face with farthest outside point (the "eye" point)
+    let chosen_face_idx = -1
+    let chosen_point_idx = -1
+    let max_distance = -1
+
+    for (let face_idx = 0; face_idx < faces.length; face_idx++) {
+      const face = faces[face_idx]
+      if (face.outside_points.size === 0) continue
+
+      const far = farthest_point_for_face_4d(points, face)
+      if (far && far.distance > max_distance) {
+        max_distance = far.distance
+        chosen_face_idx = face_idx
+        chosen_point_idx = far.idx
+      }
+    }
+
+    if (chosen_face_idx === -1) break // All points processed
+
+    const eye_idx = chosen_point_idx
+
+    // Step 2: Find all faces visible from the eye point
+    const visible_face_indices = new Set<number>()
+    for (let face_idx = 0; face_idx < faces.length; face_idx++) {
+      const face = faces[face_idx]
+      const dist = point_plane_signed_distance_4d(face.plane, points[eye_idx])
+      if (dist > EPS) visible_face_indices.add(face_idx)
+    }
+
+    // Step 3: Build horizon ridges (boundary between visible and non-visible faces)
+    const horizon_ridges = build_horizon_4d(faces, visible_face_indices)
+    const visible_faces = Array.from(visible_face_indices).sort((a, b) => b - a)
+    const candidate_points = collect_candidate_points_4d(
+      visible_faces.map((idx) => faces[idx]),
+    )
+
+    // Step 4: Remove visible faces (they'll be replaced by new ones through eye point)
+    for (const idx of visible_faces) {
+      faces.splice(idx, 1)
+    }
+
+    // Step 5: Create new faces connecting horizon ridges to eye point
+    const new_faces: ConvexHullFace4D[] = []
+    for (const [u, v, w] of horizon_ridges) {
+      const new_face = make_face_4d(points, u, v, w, eye_idx, interior_point)
+      new_faces.push(new_face)
+    }
+
+    // Step 6: Reassign outside points from removed faces to new faces
+    for (const face of new_faces) face.outside_points.clear()
+
+    for (const idx of candidate_points) {
+      if (idx === eye_idx) continue
+
+      let best_face: ConvexHullFace4D | null = null
+      let best_distance = EPS
+
+      for (const face of new_faces) {
+        const dist = point_plane_signed_distance_4d(face.plane, points[idx])
+        if (dist > best_distance) {
+          best_distance = dist
+          best_face = face
+        }
+      }
+
+      if (best_face) best_face.outside_points.add(idx)
+    }
+
+    faces.push(...new_faces)
+  }
+
+  return faces.map((face) => {
+    const [a, b, c, d] = face.vertices
+    return {
+      vertices: [points[a], points[b], points[c], points[d]] as [
+        Point4D,
+        Point4D,
+        Point4D,
+        Point4D,
+      ],
+      normal: face.plane.normal,
+      centroid: face.centroid,
+    }
+  })
+}
+
+export function compute_lower_hull_4d(points: Point4D[]): ConvexHullTetrahedron[] {
+  const all_faces = compute_quickhull_4d(points)
+  // Filter for "lower" faces: those with normal pointing down in w direction
+  return all_faces.filter((face) => face.normal.w < 0 - EPS)
+}
+
+// Check if 3D point (x,y,z) is inside 3D tetrahedron using barycentric coordinates
+function point_in_tetrahedron_3d(
+  p0: Point3D,
+  p1: Point3D,
+  p2: Point3D,
+  p3: Point3D,
+  point: Point3D,
+): { inside: boolean; bary: [number, number, number, number] } {
+  // Solve for barycentric coordinates: point = l0*p0 + l1*p1 + l2*p2 + l3*p3
+  // with l0 + l1 + l2 + l3 = 1
+  // Build the linear system
+  const matrix: math.Matrix4x4 = [
+    [p0.x, p1.x, p2.x, p3.x],
+    [p0.y, p1.y, p2.y, p3.y],
+    [p0.z, p1.z, p2.z, p3.z],
+    [1, 1, 1, 1],
+  ]
+  const rhs = [point.x, point.y, point.z, 1]
+
+  // Solve using Cramer's rule with 4x4 determinants
+  const det_main = math.det_4x4(matrix)
+  if (Math.abs(det_main) < EPS) {
+    return { inside: false, bary: [0, 0, 0, 0] }
+  }
+
+  // Compute barycentric coordinates using Cramer's rule
+  const bary: [number, number, number, number] = [0, 0, 0, 0]
+  for (let idx = 0; idx < 4; idx++) {
+    const m_i = matrix.map((row) => [...row]) as math.Matrix4x4
+    for (let row = 0; row < 4; row++) {
+      m_i[row][idx] = rhs[row]
+    }
+    bary[idx] = math.det_4x4(m_i) / det_main
+  }
+
+  // Check if inside: all barycentric coords must be >= 0 and sum to 1
+  const eps_bary = -1e-9
+  const inside =
+    bary.every((coord) => coord >= eps_bary) &&
+    Math.abs(bary.reduce((sum, coord) => sum + coord, 0) - 1) < 1e-6
+
+  return { inside, bary }
+}
+
+// Precomputed bounding box for fast 3D containment checks
+// Speed boost: 6 cheap comparisons (bounding box) vs expensive matrix solve (barycentric)
+// helps filter out most tetrahedra for containment checks before doing the slow calculation
+interface TetrahedronModel {
+  vertices: [Point4D, Point4D, Point4D, Point4D]
+  vertices_3d: [Point3D, Point3D, Point3D, Point3D]
+  min_x: number
+  max_x: number
+  min_y: number
+  max_y: number
+  min_z: number
+  max_z: number
+}
+
+const build_tetrahedron_models = (
+  hull_tetrahedra: ConvexHullTetrahedron[],
+): TetrahedronModel[] =>
+  hull_tetrahedra.map((tet) => {
+    const [p0, p1, p2, p3] = tet.vertices
+    const vertices_3d: [Point3D, Point3D, Point3D, Point3D] = [
+      { x: p0.x, y: p0.y, z: p0.z },
+      { x: p1.x, y: p1.y, z: p1.z },
+      { x: p2.x, y: p2.y, z: p2.z },
+      { x: p3.x, y: p3.y, z: p3.z },
+    ]
+    const xs = [p0.x, p1.x, p2.x, p3.x]
+    const ys = [p0.y, p1.y, p2.y, p3.y]
+    const zs = [p0.z, p1.z, p2.z, p3.z]
+    return {
+      vertices: tet.vertices,
+      vertices_3d,
+      min_x: Math.min(...xs),
+      max_x: Math.max(...xs),
+      min_y: Math.min(...ys),
+      max_y: Math.max(...ys),
+      min_z: Math.min(...zs),
+      max_z: Math.max(...zs),
+    }
+  })
+
+// Compute distance from point to lower hull in 4D
+export const compute_e_above_hull_4d = (
+  points: Point4D[],
+  hull_tetrahedra: ConvexHullTetrahedron[],
+) => {
+  // Precompute bounding boxes for fast prefiltering
+  const models = build_tetrahedron_models(hull_tetrahedra)
+
+  return points.map(({ x, y, z, w }) => {
+    let hull_w: number | null = null
+
+    for (const model of models) {
+      // Fast bounding box prefilter
+      if (
+        x < model.min_x - EPS ||
+        x > model.max_x + EPS ||
+        y < model.min_y - EPS ||
+        y > model.max_y + EPS ||
+        z < model.min_z - EPS ||
+        z > model.max_z + EPS
+      )
+        continue
+
+      // Check if point's (x,y,z) is inside the 3D projection of the tetrahedron
+      const { inside, bary } = point_in_tetrahedron_3d(
+        model.vertices_3d[0],
+        model.vertices_3d[1],
+        model.vertices_3d[2],
+        model.vertices_3d[3],
+        { x, y, z },
+      )
+
+      if (inside) {
+        // Compute w on the hull at this (x,y,z) using barycentric interpolation
+        const [p0, p1, p2, p3] = model.vertices
+        const w_on_hull = bary[0] * p0.w + bary[1] * p1.w + bary[2] * p2.w + bary[3] * p3.w
+        hull_w = hull_w === null ? w_on_hull : Math.min(hull_w, w_on_hull)
+      }
+    }
+
+    // If no tetrahedron contains this point's spatial projection, it's outside the valid
+    // composition domain. Return NaN to indicate invalid input.
+    if (hull_w === null) return NaN
+    return w - hull_w
+  })
+}
+
+// --- N-Dimensional Convex Hull (for 5+ element systems) ---
+
+// N-dimensional vector operations with dimension validation
+const subtract_nd = (vec_a: number[], vec_b: number[]): number[] => {
+  if (vec_a.length !== vec_b.length) {
+    throw new Error(`Vector dimension mismatch: ${vec_a.length} vs ${vec_b.length}`)
+  }
+  return vec_a.map((val, idx) => val - vec_b[idx])
+}
+
+const dot_nd = (vec_a: number[], vec_b: number[]): number => {
+  if (vec_a.length !== vec_b.length) {
+    throw new Error(`Vector dimension mismatch: ${vec_a.length} vs ${vec_b.length}`)
+  }
+  return vec_a.reduce((sum, val, idx) => sum + val * vec_b[idx], 0)
+}
+
+const norm_nd = (vec: number[]): number => Math.sqrt(dot_nd(vec, vec))
+
+const normalize_nd = (vec: number[]): number[] => {
+  const len = norm_nd(vec)
+  if (len < EPS) return vec.map(() => 0)
+  return vec.map((val) => val / len)
+}
+
+// Hyperplane in N dimensions: normal · x + offset = 0
+interface HyperplaneND {
+  normal: number[]
+  offset: number
+}
+
+// Simplex facet of the convex hull (N vertices in N-dimensional space)
+export interface SimplexFaceND {
+  vertex_indices: number[]
+  plane: HyperplaneND
+  centroid: number[]
+  outside_points: Set<number>
+}
+
+// Compute normal to hyperplane through N points in N-dimensional space
+// Uses null space computation via cofactor expansion
+function compute_hyperplane_nd(points: number[][]): HyperplaneND {
+  const n = points.length
+  if (n < 2) return { normal: [], offset: 0 }
+
+  // Build (N-1) edge vectors from points[0]
+  const edges = points.slice(1).map((pt) => subtract_nd(pt, points[0]))
+
+  // Compute normal via cofactors of the (N-1)×N edge matrix
+  // Each component is ±det of (N-1)×(N-1) submatrix with that column removed
+  const dim = points[0].length
+  const normal_components: number[] = []
+
+  for (let col = 0; col < dim; col++) {
+    // Build (N-1)×(N-1) submatrix by removing column col
+    const submatrix = edges.map((row) => [...row.slice(0, col), ...row.slice(col + 1)])
+    const sign = col % 2 === 0 ? 1 : -1
+    normal_components.push(sign * math.det_nxn(submatrix))
+  }
+
+  const normal = normalize_nd(normal_components)
+
+  // Degenerate case: co-hyperplanar points produce zero normal
+  if (norm_nd(normal) < EPS) return { normal, offset: 0 }
+
+  const offset = -dot_nd(normal, points[0])
+  return { normal, offset }
+}
+
+const point_hyperplane_signed_distance_nd = (plane: HyperplaneND, point: number[]): number =>
+  dot_nd(plane.normal, point) + plane.offset
+
+const compute_centroid_nd = (points: number[][]): number[] => {
+  if (points.length === 0) return []
+  const dim = points[0].length
+  return Array.from(
+    { length: dim },
+    (_, idx) => points.reduce((sum, pt) => sum + pt[idx], 0) / points.length,
+  )
+}
+
+// Find N+1 points that span N dimensions (initial simplex for quickhull)
+function choose_initial_simplex_nd(points: number[][]): number[] | null {
+  const n = points[0]?.length
+  if (!n || points.length < n + 1) return null
+
+  const chosen: number[] = []
+
+  // Greedily pick points that maximize distance from current affine hull
+  // Start with two points that are farthest apart
+  let [best_i, best_j, best_dist] = [0, 1, -1]
+  const sample_size = Math.min(points.length, 100)
+  const sample_indices =
+    points.length <= sample_size
+      ? points.map((_, idx) => idx)
+      : Array.from({ length: sample_size }, (_, idx) =>
+          Math.floor((idx * points.length) / sample_size),
+        )
+
+  for (const idx_a of sample_indices) {
+    for (const idx_b of sample_indices) {
+      if (idx_a >= idx_b) continue
+      const dist = norm_nd(subtract_nd(points[idx_a], points[idx_b]))
+      if (dist > best_dist) {
+        best_dist = dist
+        best_i = idx_a
+        best_j = idx_b
+      }
+    }
+  }
+
+  if (best_dist < EPS) return null
+  chosen.push(best_i, best_j)
+  const chosen_set = new Set(chosen)
+
+  // Add remaining points to span higher dimensions
+  while (chosen.length < n + 1) {
+    let [best_idx, best_distance] = [-1, -1]
+    // Hoist chosen_points computation outside inner loop for O(n) instead of O(n²)
+    const chosen_points = chosen.map((idx_c) => points[idx_c])
+
+    for (let idx = 0; idx < points.length; idx++) {
+      if (chosen_set.has(idx)) continue
+
+      // Compute distance from point to affine hull of chosen points
+      const dist = distance_to_affine_hull_nd(points[idx], chosen_points)
+
+      if (dist > best_distance) {
+        best_distance = dist
+        best_idx = idx
+      }
+    }
+
+    if (best_idx === -1 || best_distance < EPS) return null
+    chosen.push(best_idx)
+    chosen_set.add(best_idx)
+  }
+
+  return chosen
+}
+
+// Distance from point to affine hull spanned by given points
+function distance_to_affine_hull_nd(point: number[], hull_points: number[][]): number {
+  if (hull_points.length === 1) {
+    return norm_nd(subtract_nd(point, hull_points[0]))
+  }
+
+  if (hull_points.length === 2) {
+    // Distance to line
+    const [pt_a, pt_b] = hull_points
+    const ab = subtract_nd(pt_b, pt_a)
+    const ap = subtract_nd(point, pt_a)
+    const ab_len_sq = dot_nd(ab, ab)
+    if (ab_len_sq < EPS) return norm_nd(ap)
+    const t = dot_nd(ap, ab) / ab_len_sq
+    const proj = pt_a.map((val, idx) => val + t * ab[idx])
+    return norm_nd(subtract_nd(point, proj))
+  }
+
+  // For 3+ points, use orthogonal projection
+  // Build edge vectors from hull_points[0]
+  const origin = hull_points[0]
+  const edges = hull_points.slice(1).map((pt) => subtract_nd(pt, origin))
+  const vp = subtract_nd(point, origin)
+
+  // Solve least squares using Gram matrix G[i][j] = dot(edge_i, edge_j)
+  const gram = edges.map((edge_i) => edges.map((edge_j) => dot_nd(edge_i, edge_j)))
+  const rhs = edges.map((edge) => dot_nd(edge, vp))
+
+  // Solve Gram * coeffs = rhs using simple Gaussian elimination
+  const coeffs = solve_linear_system(gram, rhs)
+  if (!coeffs) {
+    // Fallback: Gram-Schmidt when Gram matrix is singular (linearly dependent edges)
+    // Build orthogonal basis and accumulate projection in single pass
+    const ortho_basis: number[][] = []
+    let projection = vp.map(() => 0)
+
+    for (const edge of edges) {
+      // Orthogonalize edge against existing basis
+      let ortho = [...edge]
+      for (const basis of ortho_basis) {
+        const norm_sq = dot_nd(basis, basis)
+        if (norm_sq > EPS) {
+          const coeff = dot_nd(ortho, basis) / norm_sq
+          ortho = ortho.map((val, idx) => val - coeff * basis[idx])
+        }
+      }
+
+      // Add to basis and update projection if linearly independent
+      const ortho_norm_sq = dot_nd(ortho, ortho)
+      if (ortho_norm_sq > EPS) {
+        ortho_basis.push(ortho)
+        const proj_coeff = dot_nd(vp, ortho) / ortho_norm_sq
+        projection = projection.map((val, idx) => val + proj_coeff * ortho[idx])
+      }
+    }
+
+    return norm_nd(subtract_nd(vp, projection))
+  }
+
+  // Compute projection: origin + sum(coeffs[i] * edges[i])
+  const proj = origin.map(
+    (val, dim) => val + coeffs.reduce((sum, coeff, idx) => sum + coeff * edges[idx][dim], 0),
+  )
+  return norm_nd(subtract_nd(point, proj))
+}
+
+// Solve linear system Ax = b using Gaussian elimination with partial pivoting
+function solve_linear_system(matrix_a: number[][], vec_b: number[]): number[] | null {
+  const n = matrix_a.length
+  if (n === 0) return []
+  if (vec_b.length !== n) return null // Dimension mismatch
+
+  // Augmented matrix
+  const aug = matrix_a.map((row, idx) => [...row, vec_b[idx]])
+
+  for (let col = 0; col < n; col++) {
+    // Find pivot
+    let max_row = col
+    for (let row = col + 1; row < n; row++) {
+      if (Math.abs(aug[row][col]) > Math.abs(aug[max_row][col])) {
+        max_row = row
+      }
+    }
+
+    if (Math.abs(aug[max_row][col]) < EPS) return null // Singular
+
+    // Swap rows if needed
+    if (max_row !== col) {
+      const temp = aug[col]
+      aug[col] = aug[max_row]
+      aug[max_row] = temp
+    }
+
+    // Eliminate
+    for (let row = col + 1; row < n; row++) {
+      const factor = aug[row][col] / aug[col][col]
+      for (let elim_col = col; elim_col <= n; elim_col++) {
+        aug[row][elim_col] -= factor * aug[col][elim_col]
+      }
+    }
+  }
+
+  // Back substitution
+  const result = Array(n).fill(0)
+  for (let row = n - 1; row >= 0; row--) {
+    let sum = aug[row][n]
+    for (let col = row + 1; col < n; col++) {
+      sum -= aug[row][col] * result[col]
+    }
+    result[row] = sum / aug[row][row]
+  }
+
+  return result
+}
+
+// Create a simplex face with correct normal orientation (outward from interior)
+function make_face_nd(
+  points: number[][],
+  vertex_indices: number[],
+  interior_point: number[],
+): SimplexFaceND {
+  const face_points = vertex_indices.map((idx) => points[idx])
+  const plane = compute_hyperplane_nd(face_points)
+  const centroid = compute_centroid_nd(face_points)
+
+  // Flip normal if it points toward interior instead of away
+  const dist_interior = point_hyperplane_signed_distance_nd(plane, interior_point)
+  if (dist_interior > 0) {
+    plane.normal = plane.normal.map((val) => -val)
+    plane.offset = -plane.offset
+  }
+
+  return { vertex_indices, plane, centroid, outside_points: new Set() }
+}
+
+function assign_outside_points_nd(
+  face: SimplexFaceND,
+  points: number[][],
+  candidate_indices: number[],
+): void {
+  face.outside_points.clear()
+  for (const idx of candidate_indices) {
+    const dist = point_hyperplane_signed_distance_nd(face.plane, points[idx])
+    if (dist > EPS) face.outside_points.add(idx)
+  }
+}
+
+function farthest_outside_point_nd(
+  points: number[][],
+  face: SimplexFaceND,
+): { idx: number; distance: number } | null {
+  let best: { idx: number; distance: number } | null = null
+  for (const idx of face.outside_points) {
+    const distance = point_hyperplane_signed_distance_nd(face.plane, points[idx])
+    if (!best || distance > best.distance) best = { idx, distance }
+  }
+  return best
+}
+
+// Build horizon ridges (boundary between visible and non-visible faces)
+// In N dimensions, ridges are (N-2)-simplices with (N-1) vertices
+function build_horizon_nd(faces: SimplexFaceND[], visible_indices: Set<number>): number[][] {
+  const ridge_count = new Map<string, number[]>()
+
+  for (const face_idx of visible_indices) {
+    const face = faces[face_idx]
+    const verts = face.vertex_indices
+    const n = verts.length
+
+    // Each face has n ridges, each ridge omits one vertex
+    for (let skip = 0; skip < n; skip++) {
+      const ridge = verts.filter((_, idx) => idx !== skip)
+      const sorted = [...ridge].sort((a, b) => a - b)
+      const key = sorted.join(`|`)
+
+      if (!ridge_count.has(key)) {
+        ridge_count.set(key, ridge)
+      } else {
+        // Mark as internal (seen twice)
+        ridge_count.set(key, [])
+      }
+    }
+  }
+
+  // Return only boundary ridges (seen exactly once, not marked empty)
+  return [...ridge_count.values()].filter((ridge) => ridge.length > 0)
+}
+
+// N-dimensional quickhull algorithm
+export function compute_quickhull_nd(points: number[][]): SimplexFaceND[] {
+  if (points.length === 0) return []
+  const n = points[0].length
+  if (points.length < n + 1) return []
+
+  // Find initial n-simplex
+  const initial = choose_initial_simplex_nd(points)
+  if (!initial) return []
+
+  // Interior point for normal orientation
+  const interior = compute_centroid_nd(initial.map((idx) => points[idx]))
+
+  // Create initial n+1 facets (each omits one vertex from the simplex)
+  const faces: SimplexFaceND[] = []
+  for (let skip = 0; skip <= n; skip++) {
+    const verts = initial.filter((_, idx) => idx !== skip)
+    faces.push(make_face_nd(points, verts, interior))
+  }
+
+  // Assign outside points to initial faces
+  const remaining = points.map((_, idx) => idx).filter((idx) => !initial.includes(idx))
+  for (const face of faces) {
+    assign_outside_points_nd(face, points, remaining)
+  }
+
+  // Main quickhull loop
+  while (true) {
+    // Find face with farthest outside point
+    let [chosen_face_idx, chosen_point_idx, max_distance] = [-1, -1, -1]
+
+    for (let face_idx = 0; face_idx < faces.length; face_idx++) {
+      const face = faces[face_idx]
+      if (face.outside_points.size === 0) continue
+
+      const far = farthest_outside_point_nd(points, face)
+      if (far && far.distance > max_distance) {
+        max_distance = far.distance
+        chosen_face_idx = face_idx
+        chosen_point_idx = far.idx
+      }
+    }
+
+    if (chosen_face_idx === -1) break // All points processed
+
+    const eye_idx = chosen_point_idx
+
+    // Find all faces visible from eye point
+    const visible_indices = new Set<number>()
+    for (let face_idx = 0; face_idx < faces.length; face_idx++) {
+      const dist = point_hyperplane_signed_distance_nd(faces[face_idx].plane, points[eye_idx])
+      if (dist > EPS) visible_indices.add(face_idx)
+    }
+
+    // Build horizon ridges
+    const horizon = build_horizon_nd(faces, visible_indices)
+
+    // Collect candidate points from visible faces
+    const candidates = new Set<number>()
+    for (const face_idx of visible_indices) {
+      for (const pt_idx of faces[face_idx].outside_points) {
+        candidates.add(pt_idx)
+      }
+    }
+
+    // Remove visible faces (in reverse order to maintain indices)
+    const sorted_visible = Array.from(visible_indices).sort((a, b) => b - a)
+    for (const idx of sorted_visible) {
+      faces.splice(idx, 1)
+    }
+
+    // Create new faces from horizon ridges to eye point
+    const new_faces: SimplexFaceND[] = []
+    for (const ridge of horizon) {
+      const new_verts = [...ridge, eye_idx]
+      new_faces.push(make_face_nd(points, new_verts, interior))
+    }
+
+    // Reassign candidate points to new faces
+    for (const face of new_faces) face.outside_points.clear()
+
+    for (const pt_idx of candidates) {
+      if (pt_idx === eye_idx) continue
+
+      let best_face: SimplexFaceND | null = null
+      let best_dist = EPS
+
+      for (const face of new_faces) {
+        const dist = point_hyperplane_signed_distance_nd(face.plane, points[pt_idx])
+        if (dist > best_dist) {
+          best_dist = dist
+          best_face = face
+        }
+      }
+
+      if (best_face) best_face.outside_points.add(pt_idx)
+    }
+
+    faces.push(...new_faces)
+  }
+
+  return faces
+}
+
+// Filter for lower hull facets (normal pointing "down" in energy dimension)
+export function compute_lower_hull_nd(faces: SimplexFaceND[]): SimplexFaceND[] {
+  // Last dimension is energy; negative normal means "downward"
+  return faces.filter((face) => (face.plane.normal.at(-1) ?? 0) < -EPS)
+}
+
+// Precomputed model for fast containment checks
+interface SimplexModelND {
+  vertices: number[][]
+  vertices_spatial: number[][] // Without energy dimension
+  bbox_min: number[]
+  bbox_max: number[]
+}
+
+function build_simplex_models_nd(
+  faces: SimplexFaceND[],
+  points: number[][],
+): SimplexModelND[] {
+  return faces.map((face) => {
+    const vertices = face.vertex_indices.map((idx) => points[idx])
+    const n = vertices[0].length
+    // Spatial coords are all except last (energy)
+    const vertices_spatial = vertices.map((pt) => pt.slice(0, n - 1))
+
+    // Compute bounding box in spatial dimensions
+    const spatial_dim = n - 1
+    const bbox_min = Array.from({ length: spatial_dim }, (_, idx) =>
+      Math.min(...vertices_spatial.map((pt) => pt[idx])),
+    )
+    const bbox_max = Array.from({ length: spatial_dim }, (_, idx) =>
+      Math.max(...vertices_spatial.map((pt) => pt[idx])),
+    )
+
+    return { vertices, vertices_spatial, bbox_min, bbox_max }
+  })
+}
+
+// Check if point is inside simplex and return barycentric coordinates
+// Uses linear system solution: point = sum(bary[i] * vertex[i]) with sum(bary) = 1
+function point_in_simplex_nd(point: number[], simplex_vertices: number[][]): number[] | null {
+  const n = simplex_vertices.length // Number of vertices = spatial_dim + 1
+  if (n === 0) return null
+
+  const dim = point.length
+  if (dim !== n - 1) return null // Spatial dim should be one less than vertex count
+
+  // Build linear system: [v1-v0, v2-v0, ..., vn-v0] * [b1, b2, ..., bn] = point - v0
+  // Then b0 = 1 - sum(b1..bn)
+  const v0 = simplex_vertices[0]
+  const edges = simplex_vertices.slice(1).map((vert) => subtract_nd(vert, v0))
+  const rhs = subtract_nd(point, v0)
+
+  // For square system (dim == n-1), solve directly
+  // Build matrix where each column is an edge vector
+  const matrix: number[][] = []
+  for (let row = 0; row < dim; row++) {
+    matrix.push(edges.map((edge) => edge[row]))
+  }
+
+  const coeffs = solve_linear_system(matrix, rhs)
+  if (!coeffs) return null
+
+  // Compute b0 = 1 - sum(coeffs), ensuring sum(bary) = 1 by construction
+  const sum_coeffs = coeffs.reduce((sum, val) => sum + val, 0)
+  const bary = [1 - sum_coeffs, ...coeffs]
+
+  // Point is inside simplex if all barycentric coords are non-negative
+  return bary.every((val) => val >= -EPS) ? bary : null
+}
+
+// Compute energy above hull for N-dimensional points
+export function compute_e_above_hull_nd(
+  query_points: number[][],
+  hull_facets: SimplexFaceND[],
+  all_points: number[][],
+): number[] {
+  const models = build_simplex_models_nd(hull_facets, all_points)
+
+  return query_points.map((query) => {
+    const n = query.length
+    const spatial = query.slice(0, n - 1) // All but last coord
+    const energy = query[n - 1]
+
+    let hull_energy: number | null = null
+
+    for (const model of models) {
+      // Fast bounding box rejection
+      let outside_bbox = false
+      for (let idx = 0; idx < spatial.length; idx++) {
+        if (
+          spatial[idx] < model.bbox_min[idx] - EPS ||
+          spatial[idx] > model.bbox_max[idx] + EPS
+        ) {
+          outside_bbox = true
+          break
+        }
+      }
+      if (outside_bbox) continue
+
+      // Check if spatial coords are inside simplex projection
+      const bary = point_in_simplex_nd(spatial, model.vertices_spatial)
+      if (!bary) continue
+
+      // Interpolate energy using barycentric coords
+      const e_hull = bary.reduce(
+        (sum, coeff, idx) => sum + coeff * model.vertices[idx][n - 1],
+        0,
+      )
+      hull_energy = hull_energy === null ? e_hull : Math.min(hull_energy, e_hull)
+    }
+
+    // If no facet contains this point's spatial projection, it's outside the valid
+    // composition domain. Return NaN to indicate invalid input rather than 0 (which
+    // would falsely imply the point is stable/on-hull).
+    if (hull_energy === null) return NaN
+    return energy - hull_energy
+  })
+}