pathsim 0.22.1__tar.gz → 0.23.0__tar.gz

{pathsim-0.22.1 → pathsim-0.23.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pathsim
-Version: 0.22.1
+Version: 0.23.0
 Summary: A differentiable block based hybrid system simulation framework.
 Author-email: Milan Rother <milan.rother@gmx.de>
 License: MIT
@@ -18,14 +18,14 @@ Requires-Dist: matplotlib>=3.1
 Requires-Dist: scipy>=1.2
 Provides-Extra: rf
 Requires-Dist: scikit-rf; extra == "rf"
+Provides-Extra: fmi
+Requires-Dist: FMPy; extra == "fmi"
 Provides-Extra: test
 Requires-Dist: pytest; extra == "test"
 Requires-Dist: pytest-cov; extra == "test"
 Requires-Dist: pytest-xdist; extra == "test"
 Requires-Dist: FMPy; extra == "test"
 Requires-Dist: scikit-rf; extra == "test"
-Provides-Extra: fmi
-Requires-Dist: FMPy; extra == "fmi"
 Dynamic: license-file
 
 <p align="center">

{pathsim-0.22.1 → pathsim-0.23.0}/pyproject.toml

@@ -19,9 +19,12 @@ requires-python = ">=3.8"
 dependencies = ["numpy>=1.15", "matplotlib>=3.1", "scipy>=1.2"]
 
 [project.optional-dependencies]
+# Both `rf` and `fmi` are deprecated as of v0.22 — the corresponding blocks
+# have moved to their own toolboxes (pathsim-rf, pathsim-fmi) and the extras
+# will be removed in v1.0.0. Install the toolboxes directly instead.
 rf = ["scikit-rf"]
-test = ["pytest", "pytest-cov", "pytest-xdist", "FMPy", "scikit-rf"]
 fmi = ["FMPy"]
+test = ["pytest", "pytest-cov", "pytest-xdist", "FMPy", "scikit-rf"]
 
 [project.urls]
 Homepage = "https://github.com/pathsim/pathsim"

{pathsim-0.22.1 → pathsim-0.23.0}/src/pathsim/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.22.1'
-__version_tuple__ = version_tuple = (0, 22, 1)
+__version__ = version = '0.23.0'
+__version_tuple__ = version_tuple = (0, 23, 0)
 
-__commit_id__ = commit_id = 'g33dfc7866'
+__commit_id__ = commit_id = 'g5fc53f553'

{pathsim-0.22.1 → pathsim-0.23.0}/src/pathsim/blocks/__init__.py

@@ -1,4 +1,7 @@
 from .differentiator import *
+from .constraint import *
+from .bvp import *
+from .dae import *
 from .integrator import *
 from .multiplier import *
 from .divider import *

{pathsim-0.22.1 → pathsim-0.23.0}/src/pathsim/blocks/_block.py

@@ -707,10 +707,34 @@ class Block:
         return 0.0 
 
 
+    def derivative(self, t):
+        """Return the time derivative of the block state at time 't'.
+
+        This is the right hand side that the block feeds into its integration
+        engine. It is used to estimate a good initial timestep before a run.
+        The default implementation covers blocks whose dynamic path is defined
+        by an internal dynamic operator ('op_dyn'); blocks with a different
+        dynamic path reimplement this method, and stateless blocks return 'None'.
+
+        Parameters
+        ----------
+        t : float
+            evaluation time
+
+        Returns
+        -------
+        dxdt : None | float | np.ndarray
+            time derivative of the block state, or 'None' if undefined
+        """
+        if self.engine is None or self.op_dyn is None:
+            return None
+        return self.op_dyn(self.engine.state, self.inputs.to_array(), t)
+
+
     def step(self, t, dt):
-        """The 'step' method is used in transient simulations and performs an action 
-        (numeric integration timestep, recording data, etc.) based on the current 
-        inputs and the current internal state. 
+        """The 'step' method is used in transient simulations and performs an action
+        (numeric integration timestep, recording data, etc.) based on the current
+        inputs and the current internal state.
 
         It performs one timestep for the internal states. For instant time blocks,
         the 'step' method does not has to be implemented specifically. 

pathsim-0.23.0/src/pathsim/blocks/bvp.py

@@ -0,0 +1,281 @@
+#########################################################################################
+##
+##                           BOUNDARY VALUE PROBLEM BLOCK
+##                                (pathsim/blocks/bvp.py)
+##
+#########################################################################################
+
+# IMPORTS ===============================================================================
+
+import numpy as np
+
+from scipy.integrate import solve_bvp
+
+from ._block import Block
+
+
+# BLOCKS ================================================================================
+
+class BVP1D(Block):
+    """One-dimensional two-point boundary value problem (BVP) block.
+
+    Solves a first-order system of ordinary differential equations on a spatial
+    domain :math:`[a, b]` subject to two-point boundary conditions
+
+    .. math::
+
+        \\begin{align}
+            y'(x) &= \\mathrm{fun}(x, y, p, u) \\\\
+                0 &= \\mathrm{bc}(y(a), y(b), p, u)
+        \\end{align}
+
+
+    with optional free parameters :math:`p` that are determined together with
+    the solution (e.g. eigenvalue problems). Here :math:`u` is the block input,
+    so the boundary data and right hand side can depend on external signals, and
+    the prime denotes the derivative with respect to the spatial coordinate
+    :math:`x`.
+
+    The problem is solved with :func:`scipy.integrate.solve_bvp` (a 4th-order
+    collocation method with residual based mesh refinement). At every evaluation
+    the solver is warm-started with the mesh, solution and parameters of the
+    previous solve, so slowly varying inputs are tracked cheaply. When the input
+    is unchanged since the last successful solve the re-solve is skipped
+    entirely, so repeated evaluations within a timestep (e.g. during the
+    algebraic loop) do not recompute the solution. The block output is the
+    solution sampled at the query points `x_eval` (row-major, equation by
+    equation) followed by the converged free parameters.
+
+    Note
+    ----
+    Interior / multipoint conditions are not supported by
+    :func:`scipy.integrate.solve_bvp` and are therefore not available in this
+    block. If a solve does not converge, the previous output and warm-start are
+    retained and the `success` attribute is set to `False`.
+
+    Example
+    -------
+    Solve :math:`y'' = -y` with :math:`y(0) = 0`, :math:`y(\\pi/2) = 1`, whose
+    solution is :math:`y(x) = \\sin(x)`:
+
+    .. code-block:: python
+
+        import numpy as np
+        from pathsim.blocks import BVP1D
+
+        #first-order system y0' = y1, y1' = -y0
+        def fun(x, y, p, u):
+            return np.vstack([y[1], -y[0]])
+
+        def bc(ya, yb, p, u):
+            return np.array([ya[0], yb[0] - 1.0])
+
+        bvp = BVP1D(fun, bc, n=2, domain=(0.0, np.pi/2))
+
+
+    An eigenvalue problem :math:`y'' + p^2 y = 0`, :math:`y(0) = y(1) = 0`,
+    :math:`y'(0) = 1`, with the free parameter `p` converging to :math:`\\pi`:
+
+    .. code-block:: python
+
+        import numpy as np
+        from pathsim.blocks import BVP1D
+
+        def fun(x, y, p, u):
+            return np.vstack([y[1], -p[0]**2 * y[0]])
+
+        def bc(ya, yb, p, u):
+            return np.array([ya[0], yb[0], ya[1] - 1.0])
+
+        bvp = BVP1D(fun, bc, n=2, domain=(0.0, 1.0), p0=[3.0],
+                    y0=lambda x: np.vstack([np.sin(np.pi*x), np.pi*np.cos(np.pi*x)]))
+
+
+    Parameters
+    ----------
+    fun : callable
+        right hand side of the spatial ODE system with signature
+        `fun(x, y, p, u)` where `x` is the mesh `(m,)`, `y` is `(n, m)`, `p` are
+        the free parameters and `u` is the block input, returning `(n, m)`
+    bc : callable
+        boundary condition residuals with signature `bc(ya, yb, p, u)` returning
+        an array of dimension `n + n_p`
+    n : int
+        number of first-order equations
+    domain : tuple[float, float]
+        spatial domain `(a, b)`
+    n_nodes : int
+        number of nodes in the initial mesh
+    x_eval : array[float], None
+        query points for the output, defaults to `n_nodes` points spanning the
+        domain
+    y0 : array[float], callable, None
+        initial guess for the solution, either an `(n, n_nodes)` array, a
+        callable `y0(x)` returning `(n, m)`, or `None` for an all-zero guess
+    p0 : array[float], None
+        initial guess for the free parameters, `None` if there are none
+    tol : float
+        solver tolerance passed to :func:`scipy.integrate.solve_bvp`
+
+    Attributes
+    ----------
+    success : bool
+        whether the most recent solve converged
+    x : array[float]
+        current (refined) spatial mesh
+    """
+
+    def __init__(
+        self,
+        fun=lambda x, y, p, u: np.zeros_like(y),
+        bc=lambda ya, yb, p, u: ya,
+        n=1,
+        domain=(0.0, 1.0),
+        n_nodes=11,
+        x_eval=None,
+        y0=None,
+        p0=None,
+        tol=1e-6
+        ):
+
+        super().__init__()
+
+        #spatial ode and boundary condition residuals
+        self.fun = fun
+        self.bc = bc
+
+        #problem dimensions
+        self.n = int(n)
+        a, b = domain
+        self.domain = (float(a), float(b))
+
+        #free parameters
+        self._has_p = p0 is not None
+        self.p0 = None if p0 is None else np.atleast_1d(p0).astype(float)
+
+        #query points for the output
+        self.x_eval = (np.linspace(a, b, n_nodes) if x_eval is None
+                       else np.asarray(x_eval, dtype=float))
+
+        #solver tolerance
+        self.tol = tol
+
+        #initial mesh and solution guess (used as warm-start)
+        self._x0 = np.linspace(a, b, n_nodes)
+        if y0 is None:
+            self._y0 = np.zeros((self.n, n_nodes))
+        elif callable(y0):
+            self._y0 = np.atleast_2d(y0(self._x0)).astype(float)
+        else:
+            self._y0 = np.broadcast_to(
+                np.atleast_2d(y0), (self.n, n_nodes)
+                ).astype(float).copy()
+
+        #warm-start state
+        self._x = self._x0.copy()
+        self._y = self._y0.copy()
+        self._p = None if self.p0 is None else self.p0.copy()
+
+        #solve status and sampled solution of the most recent solve
+        self.success = False
+        self._y_eval = np.zeros((self.n, self.x_eval.size))
+
+        #input of the most recent successful solve, used to skip redundant
+        #re-solves when the boundary data has not changed between evaluations
+        self._u_last = None
+
+        #pre-size the output register
+        n_out = self.n * self.x_eval.size + (self._p.size if self._has_p else 0)
+        self.outputs.update_from_array(np.zeros(n_out))
+
+
+    def __len__(self):
+        #boundary data flows in through the input, so the block has a passthrough
+        return 1 if self._active else 0
+
+
+    def reset(self):
+        """Reset inputs, outputs and the warm-start mesh, solution and
+        parameters to their initial values.
+        """
+        super().reset()
+        self._x = self._x0.copy()
+        self._y = self._y0.copy()
+        self._p = None if self.p0 is None else self.p0.copy()
+        self.success = False
+        self._u_last = None
+
+
+    def solution(self):
+        """Return the most recent solution sampled at the query points.
+
+        Returns
+        -------
+        y : array[array[float]]
+            solution of shape `(n, len(x_eval))`
+        """
+        return self._y_eval.copy()
+
+
+    def parameters(self):
+        """Return the most recent converged free parameters.
+
+        Returns
+        -------
+        p : array[float], None
+            converged free parameters, `None` if there are none
+        """
+        return None if self._p is None else self._p.copy()
+
+    @property
+    def x(self):
+        return self._x
+
+
+    def update(self, t):
+        """Solve the boundary value problem for the current input and expose the
+        sampled solution and converged parameters at the output.
+
+        Parameters
+        ----------
+        t : float
+            evaluation time
+        """
+
+        #current block input
+        u = self.inputs.to_array()
+
+        #skip the solve if the boundary data is unchanged since the last
+        #successful solve; the warm-start and output are still valid
+        if self._u_last is not None and np.array_equal(u, self._u_last):
+            return
+
+        #scipy callbacks, with or without free parameters
+        if self._has_p:
+            _fun = lambda x, y, p: self.fun(x, y, p, u)
+            _bc = lambda ya, yb, p: self.bc(ya, yb, p, u)
+            sol = solve_bvp(_fun, _bc, self._x, self._y, p=self._p, tol=self.tol)
+        else:
+            _fun = lambda x, y: self.fun(x, y, None, u)
+            _bc = lambda ya, yb: self.bc(ya, yb, None, u)
+            sol = solve_bvp(_fun, _bc, self._x, self._y, tol=self.tol)
+
+        self.success = bool(sol.success)
+
+        #keep the previous output and warm-start if the solve failed
+        if not self.success:
+            return
+
+        #warm-start the next solve with the refined mesh and solution, and
+        #remember the input so an unchanged re-evaluation can be skipped
+        self._x, self._y = sol.x, sol.y
+        if self._has_p:
+            self._p = sol.p
+        self._u_last = u.copy()
+
+        #sample the solution at the query points and assemble the output
+        self._y_eval = sol.sol(self.x_eval)[:self.n]
+        y_flat = self._y_eval.ravel()
+        if self._has_p:
+            y_flat = np.concatenate([y_flat, self._p])
+        self.outputs.update_from_array(y_flat)

pathsim-0.23.0/src/pathsim/blocks/constraint.py

@@ -0,0 +1,165 @@
+#########################################################################################
+##
+##                            ALGEBRAIC CONSTRAINT BLOCK
+##                              (pathsim/blocks/constraint.py)
+##
+#########################################################################################
+
+# IMPORTS ===============================================================================
+
+import numpy as np
+
+from ._block import Block
+
+from ..optim.anderson import NewtonAnderson, solve_root
+
+
+# BLOCKS ================================================================================
+
+class AlgebraicConstraint(Block):
+    """Solve a nonlinear algebraic constraint for its internal unknown.
+
+    At every evaluation the block solves the square nonlinear system
+
+    .. math::
+
+        \\mathrm{func}(x, u) = 0
+
+    for the internal unknown :math:`x` given the current block input :math:`u`,
+    and exposes the converged :math:`x` at its output:
+
+    .. math::
+
+        y = x \\quad\\text{such that}\\quad \\mathrm{func}(x, u) = 0
+
+
+    The constraint is resolved with an internal Anderson-accelerated damped
+    Newton iteration (:class:`.NewtonAnderson`) that is warm-started with the
+    solution of the previous evaluation, so typically only a couple of
+    iterations are required.
+    If no analytical Jacobian :math:`\\partial \\mathrm{func} / \\partial x` is
+    supplied it is approximated by central finite differences.
+
+    This is the building block for steady-state operating points, implicit
+    constitutive laws, chemical / phase equilibria and quasi-steady-state
+    approximations, where an output is defined implicitly rather than
+    explicitly. The input :math:`u` is wired dynamically through the block
+    ports, so no input dimension has to be declared in advance.
+
+    Note
+    ----
+    This block is purely algebraic. Its constraint solve is part of the
+    algebraic component of the global system DAE and is therefore evaluated
+    multiple times per timestep, each time `Simulation._update(t)` is called.
+    The residual `func` must be purely algebraic and must not introduce states,
+    delay or other dynamic behaviour.
+
+    Example
+    -------
+    Compute the square root of the input as the positive root of
+    :math:`x^2 - u = 0`:
+
+    .. code-block:: python
+
+        import numpy as np
+        from pathsim.blocks import AlgebraicConstraint
+
+        #residual of the constraint x**2 - u = 0
+        def func(x, u):
+            return x**2 - u
+
+        ac = AlgebraicConstraint(func, x0=1.0)
+
+
+    A vector valued example, the equilibrium of a simple reversible reaction
+    :math:`A \\rightleftharpoons B` with conservation :math:`x_A + x_B = u`:
+
+    .. code-block:: python
+
+        import numpy as np
+        from pathsim.blocks import AlgebraicConstraint
+
+        k_f, k_r = 2.0, 1.0
+
+        def func(x, u):
+            return np.array([
+                k_f*x[0] - k_r*x[1],   #reaction equilibrium
+                x[0] + x[1] - u[0]     #mass conservation
+                ])
+
+        ac = AlgebraicConstraint(func, x0=[0.5, 0.5])
+
+
+    Parameters
+    ----------
+    func : callable
+        residual function of the constraint with signature `func(x, u)` that
+        returns an array of the same dimension as `x`, where `x` is the
+        internal unknown and `u` is the block input array
+    x0 : float, array[float]
+        initial value / initial guess for the internal unknown `x`
+    jac : callable, None
+        optional analytical jacobian of `func` with respect to `x` with
+        signature `jac(x, u)`, central finite differences are used if `None`
+
+    Attributes
+    ----------
+    opt : NewtonAnderson
+        internal Newton-Anderson optimizer for the algebraic constraint
+    """
+
+    def __init__(self, func=lambda x, u: x, x0=0.0, jac=None):
+        super().__init__()
+
+        #some checks to ensure that function works correctly
+        if not callable(func):
+            raise ValueError(f"'{func}' is not callable")
+
+        #residual function and optional jacobian of the constraint
+        self.func = func
+        self.jac = jac
+
+        #initial guess and warm-start for the internal unknown
+        self.x0 = np.atleast_1d(x0).astype(float)
+        self._x = self.x0.copy()
+
+        #internal optimizer for the constraint (consistent with implicit solvers)
+        self.opt = NewtonAnderson()
+
+        #pre-size the output register to the unknown dimension
+        self.outputs.update_from_array(self._x)
+
+
+    def __len__(self):
+        return 1 if self._active else 0
+
+
+    def reset(self):
+        """Reset inputs, outputs and the warm-start of the internal unknown."""
+        super().reset()
+        self._x = self.x0.copy()
+        self.outputs.update_from_array(self._x)
+
+
+    def update(self, t):
+        """Solve the algebraic constraint for the current input and set the
+        output to the converged internal unknown.
+
+        Parameters
+        ----------
+        t : float
+            evaluation time
+        """
+
+        #current block input
+        u = self.inputs.to_array()
+
+        #residual and optional jacobian as functions of x at fixed u
+        _func = lambda x: self.func(x, u)
+        _jac = None if self.jac is None else (lambda x: self.jac(x, u))
+
+        #solve the constraint, warm-started with the previous solution
+        self._x, _, _ = solve_root(self.opt, _func, self._x, _jac)
+
+        #expose the converged unknown
+        self.outputs.update_from_array(self._x)