API

NeuralPDE's own public API — the solver algorithms (NNODE, NNDAE, BNNODE, PINOODE, NNSDE, SDEPINN, DeepGalerkin), the PhysicsInformedNN discretization, the training strategies and the adaptive loss functions — is documented in the Manual: ODE solvers, DAE solvers, PINN discretizations, Bayesian PINNs, physics-informed neural operators and neural adapters.

This page documents the reexported surface instead: names that using NeuralPDE brings into scope but that another package owns and documents.

Reexported symbolic front end

Writing down a PDESystem is the normal documented entry point to NeuralPDE, so using NeuralPDE also brings in the symbolic front end needed to express one. These names are owned and documented by ModelingToolkit, ModelingToolkitBase, Symbolics:

• Declaring symbols: @parameters, @variables, @named, @register_symbolic
• Systems: PDESystem, mtkcompile, @mtkcompile, unknowns
• Operators: Differential, Integral
• The ModelingToolkit module itself

Anything else from ModelingToolkit, ModelingToolkitBase, or Symbolics must be imported from that package directly — for example System, equations, parameters and observed are system-manipulation API that NeuralPDE does not use in its own documented workflow, so they are deliberately not reexported. Interval domains and their endpoints still come from DomainSets (using DomainSets: Interval, infimum, supremum), which NeuralPDE does not reexport.

Reexported SciML common interface

using NeuralPDE also brings in the parts of the SciML common interface needed to build a problem, solve it and inspect the result, so they do not have to be imported separately. These names are owned and documented by SciMLBase:

• Problems: ODEProblem, DAEProblem, SDEProblem, NoiseProblem, OptimizationProblem
• Functions: ODEFunction, ODEInputFunction, OptimizationFunction
• Solutions: ODESolution, PDETimeSeriesSolution
• Solving and discretizing: solve, init, remake, discretize, symbolic_discretize
• Return status: ReturnCode
• The SciMLBase module itself

Anything else from SciMLBase must be imported from SciMLBase directly. In particular the ensemble interface, the callback types and the integrator interface are not reexported: NeuralPDE's solvers train a network rather than step an integrator, so those names are not part of its documented use.