The full chain from discretisation theory to production solvers: FEM, FVM and spectral methods, industrial solvers such as OpenFOAM and PETSc, differentiable physics programming, and recent numerical-methods research.
Updated Nov 3, 2026 43 resources
Introduction
Why this collection
Every upper-layer method in physics AI — PINNs, neural operators, surrogates — ultimately has to deal with numerical methods. Some treat it as a black box to be replaced, some embed it in the training loop, and others displace it with a differentiable version. Without discretisation literacy you cannot judge whether an AI result is trustworthy.
Industrial deployment also cannot avoid production solvers. OpenFOAM, PETSc and SU2 carry real engineering cases, and their parallel strategies, memory models and convergence criteria directly decide whether your AI pipeline can run at all.
This collection links three layers:
Discretisation and classical numerics — textbooks and tools for FEM, FVM and spectral methods, i.e. how an equation becomes a matrix;
Production-grade solvers — open-source solvers and their ecosystems, deployable as-is;
Differentiable physics and programming — frameworks that put solvers on the gradient path, currently the hottest direction.
Suggested order: discretisation → production solvers → differentiable physics. Skipping the first layer and jumping straight to the third raises debugging cost by an order of magnitude.
Engineering Practice & Industrial Workflows 1 resources
General CFD / Industrial Simulation / Data Generation
A review of the industrial multiphysics pipeline around the three OpenFOAM distribution lines: conjugate heat transfer coupling, what version 14 and v2606 actually changed, the real state of GPU offloading, and where electromagnetic capability actually sits in the mainline code.
NewGPL-3.0 (all three distribution lines publish under GPL v3)OpenFOAMFinite VolumeIndustrial Simulation
Proposes a subcell-refined entropy-residual-driven limiter for high-order DG methods on Legendre-Gauss-Lobatto nodes, enabling robust simulation of under-resolved nonlinear hyperbolic systems while preserving accuracy via nearest-neighbor pairwise dissipation.
See LICENSE fileDiscontinuous Galerkinentropy residualhigh-order methods
Proposes a ROM-based preconditioner to accelerate topology optimization, reducing computational cost of high-fidelity finite element analysis in structural conceptual design.
See LICENSE fileTopology OptimizationReduced Order ModelingPreconditioner
This paper presents an unconditionally stable second-order numerical method for fluid-poroelastic structure interaction, where free fluid governed by incompressible Stokes equations couples with poroelastic medium via the Biot system, reformulated using Robin-type interface conditions.
See LICENSE filefluid-structure interactionporoelasticityRobin boundary conditions
Introduces backward logarithmic orthogonal functions constructed via terminal-endpoint logarithmic mapping of generalized Laguerre polynomials, designed for backward spectral approximation of problems with weak endpoint singularities.
See LICENSE filespectral methodsorthogonal functionsapproximation theory
Open-source FEM-based multiphysics framework; describes variational problems in UFL and auto-generates efficient C++ discretization code - a standard academic tool for solids and coupled physics, also a high-fidelity reference for surrogates.
NewLGPL-3.0FEMFinite ElementMultiphysics
💬Editor note: The FEniCS / FFCx finite element platform, programmed through variational forms — ideal for rapid validation during research prototyping.
Systematic path from classical PDE numerics to data-driven physics AI - FEM, FVM, surrogates, PINN and their evolution.
CC BY 4.0RoadmapPDENumerical Method
💬Editor note: The overall roadmap: it threads the evolution from finite element and finite volume through surrogates to PINNs — the best entry point for the whole stack.
Build a 2D linear-elastic FEM solver from NumPy/SciPy - no FEniCS/Abaqus - to deeply understand every matrix derivation and assembly step.
GPL-3.0FEMFinite ElementPython
💬Editor note: Derives and assembles a 2D elastic FEM matrix from scratch in NumPy, with no framework at all — read this to genuinely understand discretisation.
Python framework that builds spectral-method solvers symbolically (Burns et al. 2020), widely used in fluid dynamics, plasma, star and planet research.
NewGPL-3.0Spectral MethodSymbolicPython
💬Editor note: The Dedalus spectral framework pairs symbolic computation with MPI parallelism — a practical reference for the spectral-methods direction.
FEniCS/dolfinx releases a patch version that fixes a Mathjax rendering issue in the built documentation. This release is functionally identical to v0.11.0, so users do not need to upgrade.
GNU Lesser General Public License v3.0FEniCSdolfinxFinite Element
LeVeque team finite-volume package - Godunov-type schemes for hyperbolic conservation laws, standard tool for shocks, tsunami, reservoir, and seismic waves.
FEM platform where you write variational forms in near-math UFL syntax and discretization/solving is automated - the standard learning route with its official tutorial book.
LeafNet algorithm reformulates hyperbolic conservation laws as a coupled system with smooth solutions and approximates smooth functions via physics-informed neural networks, enabling accurate computation of multidimensional hyperbolic PDEs without diffusion.
See LICENSE filehyperbolic equationsneural networkphysics-informed learning
This paper presents a fixed-feature spectral extreme learning machine solver for partial differential equations, using Chebyshev–Gauss–Lobatto differentiation matrices and mapped multi-patch formulations for curved domains.
See LICENSE filePartial Differential EquationsSpectral MethodsExtreme Learning Machine
Open-source CFD solver as a physics-AI data backbone: principles, case structure, getting started, alternatives, cost accounting and failure boundaries.
NewGPL v3CFDOpen SourceSolver
💬Editor note: One of the de facto industrial-grade open CFD solvers, and the workhorse for industrial cases and engineering workflows.
High-performance scientific library from Argonne for large linear/nonlinear systems and parallel infrastructure - the underlying solver engine of top-tier CFD and multiphysics simulation, BSD-2-Clause friendly for commercial use.
NewBSD-2-ClauseSolverLinear AlgebraSparse Matrix
💬Editor note: PETSc provides parallel linear algebra and solver suites — an unavoidable low-level library for large sparse problems.
Python 3D visualization and mesh-processing framework wrapping VTK; a concise API for field data, mesh rendering and animation. The de-facto standard for simulation post-processing, ML dataset construction and QA, used as the output layer by FEniCSx / PyMAPDL.
Open-source CFD and discrete-adjoint solver with built-in aerodynamic shape optimization; PDE-constrained gradients at a cost independent of design-variable count, a high-fidelity reference for surrogates.
NewGNU Lesser General Public License v2.1CFDAdjointShape Optimization
💬Editor note: SU2 ships adjoint methods and is common in aerodynamic optimisation, complementing the surrogate-model direction.
Robot Kinematics and Dynamics Library / JAX Autodiff Control Framework
A pure-Python, JAX-based robot kinematics and dynamics library from Stanford ASL that delivers Pinocchio-class speed with autodifferentiation and MJX-compatible parallelism for high-performance IK and inverse-dynamics control.
A free online textbook from TU Munich covering physics + deep learning: differentiable physics, PINNs, neural operators and turbulence modeling, with runnable code.
Humanoid Whole-Body Control / Differentiable World Model Training Framework
An open-source humanoid whole-body control framework from UC Berkeley that amortizes learning across all motions with one differentiable world model, surpassing a 128-GPU, 7-day baseline in 3 hours on a single GPU and winning the CoRL 2026 Best Paper Award.
Warp-Geo provides a GPU-accelerated differentiable framework for reconstructing signed distance fields and evaluating surface normals from complex 3D point clouds, enabling gradient-based shape optimization.
See LICENSE fileDifferentiable GeometrySigned Distance FieldShape Optimization
Fully auto-differentiable CFD solver built on JAX for CPU/GPU/TPU, focused on end-to-end differentiable fluid simulation that couples seamlessly with neural networks.
NewGNU General Public License v3.0JAXDifferentiable CFDDifferentiable Physics
An open-source physics engine maintained by Google DeepMind, pairing generalised coordinates with soft convex contact dynamics for robotics, biomechanics and reinforcement learning, with MJCF, MJX, MuJoCo Warp and the 4.0 release.
NewApache-2.0 for source code; CC BY 4.0 for documentation, images and videos in the doc directoryPhysics EngineContact DynamicsRobot Simulation
NVIDIA's open-source Python framework that JIT-compiles plain Python functions into CPU/CUDA kernels with automatic differentiable adjoints, letting gradients flow through entire physics simulations - the compute substrate under Isaac Lab and Newton.
💬Editor note: NVIDIA Warp provides a GPU-native differentiable simulation kernel and is the mainstream choice for GPU-accelerated differentiable physics today.
Microsoft Research's open-source PDE benchmark suite with differentiable solvers and spatiotemporal dynamics data for evaluating long-horizon rollout stability.
MITPDEBenchmarkDifferentiable Solver
💬Editor note: PDEArena benchmarks differentiable solvers, letting you compare accuracy and speed across JAX-CFD, PhiFlow and similar frameworks.
Google Research's differentiable incompressible-flow solver in JAX; the whole time-stepping loop is differentiable and JIT-compilable on GPU/TPU. A PNAS 2021 learned-interpolation result reached ~80x/40x speedups at equal accuracy, but the project is now unmaintained - treat it as a research prototype and reference.
NewApache-2.0JAXDifferentiable PhysicsCFD
💬Editor note: JAX-CFD is the reference fully differentiable CFD framework, turning aerodynamic design into an end-to-end optimisation problem.
Julia SciML ecosystem library unifying PINN/DeepONet/NeuralOperator solvers with seamless DifferentialEquations.jl interoperability and GPU/autodiff out of the box.
DeepXDE v1.15.0 released with L-LAAF activation function support for JAX backend and Apple MPS GPU acceleration for PyTorch backend, improving deep learning-based PDE solving performance.
A Python-embedded parallel programming language for high-performance numerical computing, with LLVM JIT compilation of per-element kernels, sparse data structures and automatic differentiation, widely used for differentiable physics prototyping.
Google's open-source NumPy-style numerical computing library built on XLA with autodiff and functional transformations (grad/jit/vmap/pmap) - the substrate behind differentiable solvers and neural operators.
Rigid-Body Dynamics Algorithm Library / Robot Kinematics and Dynamics Engine
A rigid-body dynamics library from Inria and LAAS-CNRS implementing Featherstone's algorithms with analytical derivatives, BSD-2-Clause licensed and used as the dynamics backbone of Crocoddyl, TSID, HPP, Aligator, Jiminy and many robotics frameworks.
NewBSD 2-Clause LicenseRigid Body DynamicsKinematicsRobot Control
A beginner-friendly PINN library with thorough Chinese docs and TensorFlow/PyTorch/JAX/PaddlePaddle backends - a common starting point for scientific machine learning and PDE solving.
NewGNU Lesser General Public License v2.1PINNBeginnerChinese Docs
A unified framework combining physics-informed neural networks with finite element methods via functional Gauss-Newton problems and Petrov-Galerkin discretization.
See LICENSE filePINNfinite element methodPDE solver
This paper rigorously establishes inf-sup stability of the Scott-Vogelius finite element on three-dimensional Freudenthal meshes for polynomial degree k ≥ 4, completing the theoretical framework for this classical divergence-free mixed element used in Stokes flow computations.
See LICENSE filefinite element methodsStokes probleminf-sup stability
💬Editor note: An inf-sup analysis of Scott–Vogelius elements on Freudenthal grids, giving the theoretical stability boundary for spectral element methods.
The paper presents a parallel monolithic finite element strategy for simulating two- and three-dimensional vortex-induced vibrations of a rigid, massless cylinder mounted on translational springs.
See LICENSE filevortex-induced vibrationscylinder flowparallel solver
💬Editor note: A parallel solver for zero-mass-ratio vortex-induced vibration, involving strong coupling and partitioning strategies — high practical reference value.
A physics-conditioned neural operator targets atrioventricular valve mechanics across varying pressure and tissue properties, aiming to accelerate repeated valve assessments compared with finite element analysis.
See LICENSE filePhysics-Conditioned Neural OperatorCardiac Valve MechanicsFinite Element Analysis
BAT-NO is a structural-mechanics surrogate for autoregressively predicting transient vehicle-component responses and crashworthiness under varying boundary conditions.
See LICENSE filevehicle crashworthinessneural operatorfinite element surrogate
DeepFEAv2 applies deep learning surrogate models to transient finite element analysis on unstructured meshes, enabling joint prediction of node-based and element-based outputs to reduce computational cost in structural mechanics.
See LICENSE fileFinite Element AnalysisDeep LearningTransient Simulation
The paper presents a label-free PINN for the reliability of spatially variable slopes, combining finite-element virtual-work losses with incremental strength reduction for non-associated Mohr-Coulomb plasticity.
See LICENSE fileslope reliabilityphysics-informed neural networkneural operator
This paper develops a mesh-free Kolmogorov--Arnold physics-informed neural network framework for fourth-order elliptic boundary value problems, with applications to thin-plate deflection governed by the biharmonic equation.
See LICENSE filebiharmonic equationthin-plate deflectionmesh-free numerical method