Source code for pde.solvers.base

"""Package that contains base classes for solvers.

Beside the abstract base class defining the interfaces, we also provide
:class:`AdaptiveSolverBase`, which contains methods for implementing adaptive solvers.

.. codeauthor:: David Zwicker <david.zwicker@ds.mpg.de>
"""

from __future__ import annotations

import logging
import warnings
from collections.abc import Callable
from inspect import isabstract
from typing import TYPE_CHECKING, Any, Literal

import numpy as np

from ..tools.math import OnlineStatistics
from ..tools.misc import classproperty
from ..tools.typing import BackendType, NumericArray, StepperHook, TField

if TYPE_CHECKING:
    from ..backends.base import BackendBase
    from ..pdes.base import PDEBase


_base_logger = logging.getLogger(__name__.rsplit(".", 1)[0])
""":class:`logging.Logger`: Base logger for solvers."""


[docs] class ConvergenceError(RuntimeError): """Indicates that an implicit step did not converge."""
FixedStepperType = Callable[[NumericArray, float, int, Any], float] AdaptiveStepperType = Callable[ [NumericArray, float, float, float, OnlineStatistics | None, Any], tuple[float, float, int], ]
[docs] class SolverBase: """Base class for PDE solvers.""" dt_default: float = 1e-3 """float: default time step used if no time step was specified""" _use_post_step_hook: bool = True """bool: flag choosing whether the post-step hook of the PDE is called""" _mpi_synchronization: bool = False """bool: Flag indicating whether MPI synchronization is required. This is never the case for serial solvers and even parallelized solvers might set this flag to False if no synchronization between nodes is required""" _subclasses: dict[str, type[SolverBase]] = {} """dict: dictionary of all inheriting classes""" _logger: logging.Logger _backend_name: BackendType | Literal["auto"] __backend_obj: BackendBase | None def __init__( self, pde: PDEBase, *, backend: BackendBase | BackendType | Literal["auto"] = "auto", ): """ Args: pde (:class:`~pde.pdes.base.PDEBase`): The partial differential equation that should be solved backend (str or :class:`~pde.backends.base.BackendBase`): Determines how the function is created. Accepted values are 'numpy` and 'numba'. Alternatively, 'auto' lets the code decide for the most optimal backend. """ self.pde = pde self.info: dict[str, Any] = {"class": self.__class__.__name__} if self.pde: self.info["pde_class"] = self.pde.__class__.__name__ if isinstance(backend, str): self._backend_name = backend self.__backend_obj = None else: # assume that `backend` is of type BackendBase self._backend_name = backend.name # type: ignore self.__backend_obj = backend def __init_subclass__(cls, **kwargs): """Initialize class-level attributes of subclasses.""" super().__init_subclass__(**kwargs) # create logger for this specific field class cls._logger = _base_logger.getChild(cls.__qualname__) # register all subclasses to reconstruct them later if not isabstract(cls): if cls.__name__ in cls._subclasses: warnings.warn(f"Redefining class {cls.__name__}", stacklevel=2) cls._subclasses[cls.__name__] = cls if hasattr(cls, "name") and cls.name: if cls.name in cls._subclasses: _base_logger.warning("Solver `%s` is already registered", cls.name) cls._subclasses[cls.name] = cls
[docs] @classmethod def from_name(cls, name: str, pde: PDEBase, **kwargs) -> SolverBase: r"""Create solver class based on its name. Solver classes are automatically registered when they inherit from :class:`SolverBase`. Note that this also requires that the respective python module containing the solver has been loaded before it is attempted to be used. Args: name (str): The name of the solver to construct pde (:class:`~pde.pdes.base.PDEBase`): The partial differential equation that should be solved \**kwargs: Additional arguments for the constructor of the solver Returns: An instance of a subclass of :class:`SolverBase` """ try: # obtain the solver class associated with `name` solver_class = cls._subclasses[name] except KeyError: # solver was not registered solvers = ( f"'{solver}'" for solver in sorted(cls._subclasses) if not solver.endswith("Solver") ) raise ValueError( f"Unknown solver method '{name}'. Registered solvers are " + ", ".join(solvers) ) from None return solver_class(pde, **kwargs)
@classproperty def registered_solvers(cls) -> list[str]: """list of str: the names of the registered solvers""" return sorted(cls._subclasses) @property def backend(self) -> BackendType | Literal["auto"]: """str: The name of the backend used for this solver.""" return self._backend_name @backend.setter def backend(self, value: BackendBase | BackendType) -> None: """Sets a new backend for the solver. This setter tries to ensure consistency and make sure that backends are not changed after the object has been loaded (i.e., after _backend_obj has been accessed). The method also raises a warning when the backend is changed (except if it was set to `auto` before). This allows solvers to react flexibly to changes in the backend, e.g., demanded by the PDE implementation. Args: value: The backend object or name """ # determine the name of the new backend if isinstance(value, str): new_backend = value else: # assume value is of type BackendBase new_backend = value.name # type: ignore # check whether the new name contradicts the old backend name if self._backend_name in {"auto", new_backend}: pass # nothing to do else: self._logger.warning( "Changing the backend of the solver from `%s` to `%s`", self._backend_name, new_backend, ) # check whether the new backend contradicts the old backend object if self.__backend_obj is not None and self.__backend_obj.name != new_backend: msg = "Tried changing the loaded backend of the solver from `%s` to `%s`" raise TypeError( msg, self.__backend_obj.name, new_backend, ) # set the new backend self._backend_name = new_backend if not isinstance(value, str): self.__backend_obj = value @property def _backend_obj(self) -> BackendBase: """:class:`~pde.backends.base.BackendBase`: The backend used for this solver.""" if self.__backend_obj is None: from ..backends import backends if self._backend_name == "auto": self._logger.warning( "Automatic backend has not been selected yet. Choosing `numpy` " "backend as a safe option." ) self._backend_name = "numpy" # conservative fall-back self.__backend_obj = backends[self._backend_name] return self.__backend_obj def _make_pde_rhs( self, state: TField ) -> Callable[[NumericArray, float], NumericArray]: """Return a function for evaluating the right hand side of the PDE. This function also decides which backend will be used. Args: state (:class:`~pde.fields.FieldBase`): An example for the state from which the grid and other information can be extracted. backend (str): Determines how the function is created. Accepted values are 'numpy' and 'numba'. Alternatively, 'auto' lets the code pick the optimal backend. Returns: callable: Function determining the right hand side of the PDE """ from ..backends import backends if self.backend == "auto": # try using the numba backend, if it implemented try: rhs = backends["numba"].make_pde_rhs(self.pde, state) except NotImplementedError: rhs = backends["numpy"].make_pde_rhs(self.pde, state) self.backend = "numpy" else: self.backend = "numba" return rhs # get a function evaluating the rhs of the PDE return self._backend_obj.make_pde_rhs(self.pde, state) def _make_error_synchronizer( self, backend: str = "numpy", *, operator: int | str = "MAX" ) -> Callable[[float], float]: """Return function that synchronizes errors between multiple processes. Args: backend (str): The backend to use for making the synchronizer operator (str or int): Flag determining how the value from multiple nodes is combined. Possible values include "MAX", "MIN", and "SUM". Returns: Function that can be used to synchronize errors across nodes """ from ..backends import backends # deprecated on 2025-12-07 warnings.warn( "`_make_error_synchronizer` is deprecated. Use `make_mpi_synchronizer` " "from an appropriate backend instead.", DeprecationWarning, stacklevel=2, ) return backends[backend].make_mpi_synchronizer(operator=operator) def _make_post_step_hook(self, state: TField) -> StepperHook: """Create a function that calls the post-step hook of the PDE. A no-op function is returned if :attr:`SolverBase._use_post_step_hook` is `False` or the PDE does not provide :meth:`PDEBase.make_post_step_hook`. Args: state (:class:`~pde.fields.FieldBase`): An example for the state from which the grid and other information can be extracted. Returns: callable: The function that calls the post-step hook """ post_step_hook: StepperHook | None = None if self._use_post_step_hook: if self._backend_name == "auto": msg = ( "Backend has not been chosen automatically, yet. Please specify a " "backend explicitly." ) raise ValueError(msg) try: # try to get hook function and initial data from PDE instance post_step_hook, self._post_step_data_init = ( self.pde.make_post_step_hook(state, backend=self._backend_name) ) self._logger.info("Created post-step hook from PDE") except NotImplementedError: pass # no hook function defined on the PDE if post_step_hook is None: # hook function is not necessary or was not supplied -> provide no-op def post_step_hook( state_data: NumericArray, t: float, post_step_data: NumericArray ): """Default hook function does nothing.""" self._post_step_data_init = None self._logger.debug("No post-step hook defined") else: # ensure that the initial values is a mutable array self._post_step_data_init = np.array(self._post_step_data_init, copy=True) return post_step_hook # type: ignore def _make_single_step_fixed_dt( self, state: TField, dt: float ) -> Callable[[NumericArray, float], None]: """Return a function doing a single step with a fixed time step. Args: state (:class:`~pde.fields.base.FieldBase`): An example for the state from which the grid and other information can be extracted dt (float): Time step of the explicit stepping. """ msg = "Fixed stepper has not been defined" raise NotImplementedError(msg) def _make_fixed_stepper(self, state: TField, dt: float) -> FixedStepperType: """Return a stepper function using an explicit scheme with fixed time steps. Args: state (:class:`~pde.fields.base.FieldBase`): An example for the state from which the grid and other information can be extracted dt (float): Time step of the explicit stepping. """ single_step = self._make_single_step_fixed_dt(state, dt) post_step_hook = self._make_post_step_hook(state) def fixed_stepper( state_data: NumericArray, t_start: float, steps: int, post_step_data ) -> float: """Perform `steps` steps with fixed time steps.""" for i in range(steps): # calculate the right hand side t = t_start + i * dt single_step(state_data, t) post_step_hook(state_data, t, post_step_data) return t + dt return fixed_stepper
[docs] def make_stepper( self, state: TField, dt: float | None = None ) -> Callable[[TField, float, float], float]: """Return a stepper function using an explicit scheme. Args: state (:class:`~pde.fields.base.FieldBase`): An example for the state from which the grid and other information can be extracted dt (float): Time step used (Uses :attr:`SolverBase.dt_default` if `None`) Returns: Function that can be called to advance the `state` from time `t_start` to time `t_end`. The function call signature is `(state: numpy.ndarray, t_start: float, t_end: float)` """ # support `None` as a default value, so the controller can signal that # the solver should use a default time step if dt is None: dt = self.dt_default self._logger.warning( "Explicit stepper with a fixed time step did not receive any " "initial value for `dt`. Using dt=%g, but specifying a value or " "enabling adaptive stepping is advisable.", dt, ) dt_float = float(dt) # explicit casting to help type checking # create stepper with fixed steps fixed_stepper: FixedStepperType = self._backend_obj.make_inner_stepper( solver=self, stepper_style="fixed", state=state, dt=dt_float ) self.info["dt"] = dt_float self.info["dt_adaptive"] = False self.info["steps"] = 0 self.info["post_step_data"] = self._post_step_data_init self.info["stochastic"] = getattr(self.pde, "is_sde", False) # We don't access self.pde directly since we might want to reuse the solver # infrastructure for more general cases where a PDE is not defined. def wrapped_stepper(state: TField, t_start: float, t_end: float) -> float: """Advance `state` from `t_start` to `t_end` using fixed steps.""" # retrieve last post_step_data and continue with this post_step_data = self.info["post_step_data"] # calculate number of steps that lead to an end time closest to t_end steps = max(1, round((t_end - t_start) / dt_float)) # call the stepper with fixed time steps t_last = fixed_stepper(state.data, t_start, steps, post_step_data) # keep some stats and data self.info["steps"] += steps return t_last return wrapped_stepper
[docs] class AdaptiveSolverBase(SolverBase): """Base class for adaptive time steppers.""" dt_min: float = 1e-10 """float: minimal time step that the adaptive solver will use""" dt_max: float = 1e10 """float: maximal time step that the adaptive solver will use""" def __init__( self, pde: PDEBase, *, backend: BackendType | Literal["auto"] = "auto", adaptive: bool = False, tolerance: float = 1e-4, ): """ Args: pde (:class:`~pde.pdes.base.PDEBase`): The partial differential equation that should be solved backend (str): Determines how the function is created. Accepted values are 'numpy` and 'numba'. Alternatively, 'auto' lets the code decide for the most optimal backend. adaptive (bool): When enabled, the time step is adjusted during the simulation using the error tolerance set with `tolerance`. tolerance (float): The error tolerance used in adaptive time stepping. This is used in adaptive time stepping to choose a time step which is small enough so the truncation error of a single step is below `tolerance`. """ super().__init__(pde, backend=backend) self.adaptive = adaptive self.tolerance = tolerance def _make_single_step_variable_dt( self, state: TField ) -> Callable[[NumericArray, float, float], NumericArray]: """Return a function doing a single step with a variable time step. Args: state (:class:`~pde.fields.base.FieldBase`): An example for the state from which the grid and other information can be extracted Returns: Function that can be called to advance the `state` from time `t_start` to time `t_end`. The function call signature is `(state: numpy.ndarray, t_start: float, t_end: float)` """ if self.pde.is_sde: msg = "Deterministic stepper does not support stochastic equations" raise RuntimeError(msg) rhs_pde = self._make_pde_rhs(state) def single_step(state_data: NumericArray, t: float, dt: float) -> NumericArray: """Basic implementation of Euler scheme.""" return state_data + dt * rhs_pde(state_data, t) return self._backend_obj.compile_function(single_step) def _make_single_step_error_estimate( self, state: TField ) -> Callable[[NumericArray, float, float], tuple[NumericArray, float]]: """Make a stepper that also estimates the error. Args: state (:class:`~pde.fields.base.FieldBase`): An example for the state from which the grid and other information can be extracted """ if getattr(self.pde, "is_sde", False): msg = "Cannot use adaptive stepper with stochastic equation" raise RuntimeError(msg) single_step = self._make_single_step_variable_dt(state) def single_step_error_estimate( state_data: NumericArray, t: float, dt: float ) -> tuple[NumericArray, float]: """Basic stepper to estimate error.""" # single step with dt k1 = single_step(state_data, t, dt) # double step with half the time step k2a = single_step(state_data, t, 0.5 * dt) k2 = single_step(k2a, t + 0.5 * dt, 0.5 * dt) # calculate maximal error error = np.abs(k1 - k2).max() return k2, error return single_step_error_estimate def _make_adaptive_stepper( self, state: TField, *, adjust_dt: Callable[[float, float], float] | None = None ) -> AdaptiveStepperType: """Make an adaptive Euler stepper. Args: state (:class:`~pde.fields.base.FieldBase`): An example for the state from which the grid and other information can be extracted adjust_dt (callable or None): A function that is used to adjust the time step. The function takes the current time step and a relative error and returns an adjusted time step Returns: Function that can be called to advance the `state` from time `t_start` to time `t_end`. The function call signature is `(state: numpy.ndarray, t_start: float, t_end: float)` """ # obtain functions determining how the PDE is evolved single_step_error = self._make_single_step_error_estimate(state) post_step_hook = self._make_post_step_hook(state) sync_errors = self._make_error_synchronizer() # obtain auxiliary functions if adjust_dt is None: adjust_dt = _make_dt_adjuster(self.dt_min, self.dt_max) tolerance = self.tolerance dt_min = self.dt_min def adaptive_stepper( state_data: NumericArray, t_start: float, t_end: float, dt_init: float, dt_stats: OnlineStatistics | None = None, post_step_data=None, ) -> tuple[float, float, int]: """Adaptive stepper that advances the state in time.""" dt_opt = dt_init t = t_start steps = 0 while True: # use a smaller (but not too small) time step if close to t_end dt_step = max(min(dt_opt, t_end - t), dt_min) # try two different step sizes to estimate errors new_state, error = single_step_error(state_data, t, dt_step) error_rel = error / tolerance # normalize error to given tolerance # synchronize the error between all processes (necessary for MPI) error_rel = sync_errors(error_rel) # do the step if the error is sufficiently small if error_rel <= 1: steps += 1 t += dt_step state_data[...] = new_state post_step_hook(state_data, t, post_step_data) if dt_stats is not None: dt_stats.add(dt_step) if t < t_end: # adjust the time step and continue (happens in every MPI process) dt_opt = adjust_dt(dt_step, error_rel) else: break # return to the controller return t, dt_opt, steps self._logger.info("Initialized adaptive stepper") return adaptive_stepper
[docs] def make_stepper( self, state: TField, dt: float | None = None ) -> Callable[[TField, float, float], float]: """Return a stepper function using an explicit scheme. Args: state (:class:`~pde.fields.base.FieldBase`): An example for the state from which the grid and other information can be extracted dt (float): Time step used (Uses :attr:`SolverBase.dt_default` if `None`). This sets the initial time step for adaptive solvers. Returns: Function that can be called to advance the `state` from time `t_start` to time `t_end`. The function call signature is `(state: numpy.ndarray, t_start: float, t_end: float)` """ if not self.adaptive: # create stepper with fixed steps return super().make_stepper(state, dt) if getattr(self.pde, "is_sde", False): # adaptive steppers cannot deal with stochastic PDEs msg = "Cannot use adaptive stepper with stochastic equation" raise RuntimeError(msg) # Support `None` as a default value, so the controller can signal that # the solver should use a default time step. if dt is None: dt_float = self.dt_default else: dt_float = float(dt) # explicit casting to help type checking # create stepper with fixed steps adaptive_stepper: AdaptiveStepperType = self._backend_obj.make_inner_stepper( solver=self, stepper_style="adaptive", state=state, dt=dt_float ) self.info["dt"] = dt_float self.info["dt_adaptive"] = True self.info["steps"] = 0 self.info["stochastic"] = getattr(self.pde, "is_sde", False) self.info["post_step_data"] = self._post_step_data_init def wrapped_stepper(state: TField, t_start: float, t_end: float) -> float: """Advance `state` from `t_start` to `t_end` using adaptive steps.""" nonlocal dt_float # `dt_float` stores value for the next call # retrieve last post_step_data and continue with this post_step_data = self.info["post_step_data"] # call the adaptive stepper t_last, dt_float, steps = adaptive_stepper( state.data, t_start, t_end, dt_float, self.info["dt_statistics"], post_step_data, ) # keep some stats and data self.info["steps"] += steps return t_last return wrapped_stepper
def _make_dt_adjuster(dt_min: float, dt_max: float) -> Callable[[float, float], float]: """Return a function that can be used to adjust time steps. The returned function adjust_dt(dt, error_rel) adjusts the current time step `dt` based on the normalized error estimate `error_rel` with the goal of keeping `error_rel` close to 1. Behavior: - If the error is very small the time step is increased (up to a factor 4). - If the error is NaN the time step is reduced strongly. - Otherwise the time step is scaled according to error_rel**-0.2 with a conservative lower bound for the scaling factor. - The adjusted time step is clamped to the interval [dt_min, dt_max]. - If the adjusted time step falls below dt_min a RuntimeError is raised. Args: dt_min (float): Minimal allowed time step. dt_max (float): Maximal allowed time step. Returns: Callable[[float, float], float]: Function that takes (dt, error_rel) and returns the adjusted dt. """ dt_min_nan_err = f"Encountered NaN even though dt < {dt_min}" dt_min_err = f"Time step below {dt_min}" def adjust_dt(dt: float, error_rel: float) -> float: """Helper function that adjust the time step. The goal is to keep the relative error `error_rel` close to 1. Args: dt (float): Current time step error_rel (float): Current (normalized) error estimate Returns: float: Time step of the next iteration """ # adjust the time step if error_rel < 0.00057665: # error was very small => maximal increase in dt # The constant on the right hand side of the comparison is chosen to # agree with the equation for adjusting dt below dt *= 4.0 elif np.isnan(error_rel): # state contained NaN => decrease time step strongly dt *= 0.25 else: # otherwise, adjust time step according to error dt *= max(0.9 * error_rel**-0.2, 0.1) # limit time step to permissible bracket if dt > dt_max: dt = dt_max elif dt < dt_min: if np.isnan(error_rel): raise RuntimeError(dt_min_nan_err) raise RuntimeError(dt_min_err) return dt return adjust_dt