awesome-ai-cae

A curated list of AI-powered tools for Computer-Aided Engineering.

Simulation · Visualization · Design · Manufacturing

_{AI can theoretically handle 94% of computational tasks — but only 33% is actually used (Anthropic, 2026).
In CAE, this gap is even wider. This list curates tools that close it:
tools that agents can call (MCP/API/CLI), and AI/ML that transforms how we simulate.}

_{113 open-source tools across 16 categories. Every tool is AI-ready: programmable via Python API, CLI, or MCP.}

한국어 · 中文 · 日本語 · Deutsch · Français · Español · Português

Contents

• Core Engine Readiness — AI-readiness of foundational solvers
• MCP Servers — AI-native tool interfaces
• CFD — Computational Fluid Dynamics
• FEA — Finite Element Analysis
• SPH — Smoothed Particle Hydrodynamics
• DEM — Discrete Element Method
• Visualization & Post-processing
• CAD & Geometry
• Mesh Generation
• Differentiable Simulation
• AI/ML for Simulation
• Surrogate Models & PINNs
• Optimization
• Data Formats & I/O
• Datasets & Benchmarks
• Learning Resources
• FAQ

Core Engine Readiness

How AI-ready are the foundational CAE solvers? We assessed 17 engines on Python API maturity, headless operation, Docker support, and AI agent integration. Only 2 of 17 have AI-native MCP integration. The most-cited are MuJoCo (6250+), Gmsh (5100+), and OpenFOAM (4500+).

Engine	Domain	⭐	Python API	Headless	Docker	🤖 AI-Native	Paper
OpenFOAM	CFD		PyFoam, PythonFOAM	✅	✅	✅ Foam-Agent, MetaOpenFOAM, MCP	Weller 1998
FEniCS	FEA		✅ Native	✅	✅	—	Baratta 2023
Gmsh	Mesh	—	✅ Native	✅	✅	—	Geuzaine 2009
VTK / ParaView	Viz		✅ Native	✅	✅	✅ ParaView-MCP (LLNL)	Schroeder 2006
SU2	CFD/MDO		pySU2 (SWIG)	✅	✅	—	Economon 2016
MFEM	FEA		PyMFEM	✅	✅	—	Kolev 2024
deal.II	FEA		Limited	✅	✅	—	Arndt 2020
DualSPHysics	SPH		Inductiva API	✅	✅	—	Crespo 2015
Taichi	Diff. Sim		✅ Native	✅	✅	—	Hu 2020 (ICLR)
PyFR	CFD		✅ Native	✅	✅	—	Witherden 2014
CalculiX	FEA	—	pycalculix	✅	✅	—	Dhondt 2004
Elmer	FEA		PyElmer	✅	✅	—	Malinen 2013
OpenCASCADE	CAD		pythonOCC	✅	✅	—	—
MOOSE	FEA		Python scripting	✅	✅	—	Gaston 2009
FreeFEM	FEA		FreeFem++	✅	✅	—	Hecht 2012
SfePy	FEA		✅ Native	✅	✅	—	Cimrman 2019
MuJoCo	Diff. Sim		✅ Native	✅	✅	—	Todorov 2012

Legend: ✅ Available — 🤖 AI-Native = MCP server or LLM agent framework exists — Citations from Google Scholar

^{back to top}

MCP Servers

Tools that AI agents can call directly via Model Context Protocol.

MCP servers are the AI-native interface layer for CAE tools. They allow AI agents like Claude, ChatGPT, and Copilot to directly invoke simulation, visualization, and meshing operations without human GUI interaction. As of 2026, MCP servers exist for OpenFOAM (CFD), ParaView (visualization), and VTK-based rendering, making these the first CAE tools with true AI agent integration.

• kimimgo/viznoir Python MCP - Cinema-quality science visualization. 22 tools for rendering, slicing, contouring, volume rendering, and animating OpenFOAM/VTK/CGNS data via VTK. Headless EGL/OSMesa.
• llnl/paraview_mcp Python MCP - Natural language control of ParaView via MCP. Multimodal LLM observes viewport for visual feedback (LLNL).
• webworn/openfoam-mcp-server C++ MCP - OpenFOAM MCP server with Socratic questioning for CFD education and expert error resolution.

^{back to top}

CFD — Computational Fluid Dynamics

Six open-source solvers for fluid flow, heat transfer, and multiphysics. OpenFOAM is the only CFD solver with 3 AI agent frameworks.

Computational fluid dynamics is the CAE domain with the most advanced AI integration. Our assessment found that OpenFOAM leads with three LLM agent frameworks (Foam-Agent at 88% success rate, MetaOpenFOAM, and an MCP server), making it the only CFD solver with full AI-native support. PyFR provides GPU-accelerated high-order methods, while preCICE enables AI-orchestrated multi-physics coupling for fluid-structure interaction.

• OpenFOAM/OpenFOAM-dev C++ - The open source CFD toolbox. Finite volume solvers for incompressible/compressible flow, multiphase, combustion, heat transfer.
• su2code/SU2 C++ Python - Multiphysics simulation and design optimization. Compressible/incompressible flow, structural analysis, adjoint-based design.
• LLNL/Nek5000 Fortran - High-order spectral element CFD solver. DNS/LES of turbulent flows. Scalable to millions of cores.
• pyvista/pyvista Python - Pythonic VTK interface for 3D plotting and mesh analysis. CFD post-processing, streamlines, glyphs.
• PyFR/PyFR Python - High-order flux reconstruction CFD on mixed unstructured grids. GPU-accelerated (CUDA/OpenCL/HIP).
• precice/precice C++ Python - Coupling library for multi-physics simulations. Fluid-structure interaction, conjugate heat transfer.

^{back to top}

FEA — Finite Element Analysis

Eleven structural, thermal, and multiphysics FEM solvers. Combined 5000+ citations. No MCP servers yet — a major opportunity.

AI for finite element analysis is rapidly evolving. Our assessment covers 11 solvers totaling over 5000 citations. FEniCS offers the most Pythonic FEA interface, making it ideal for AI agent integration. MOOSE (1290+ citations, Idaho National Lab) handles coupled multiphysics, FreeFEM (3040+ citations) provides a scripting language for PDEs, and MFEM from LLNL provides GPU-accelerated high-order elements. None yet have MCP servers, representing a major opportunity for AI-native FEA automation.

• CalculiX Fortran C - Free 3D structural FEM. Linear/nonlinear static, dynamic, thermal analysis. Abaqus INP compatible.
• dealii/dealii C++ - Adaptive finite elements. Supports hp-refinement, multigrid, and parallel distributed computing.
• ElmerCSC/elmerfem Fortran C++ - Multiphysics FEM solver. Fluid dynamics, structural mechanics, electromagnetics, heat transfer. CSC Finland.
• FEniCS/dolfinx C++ Python - Next-generation FEniCS. Automated PDE solving with high-level Python/C++ interface. Parallel, scalable.
• FreeFem/FreeFem-sources C++ - Partial differential equation solver using finite element method. High-level scripting language for 2D/3D problems (Hecht 2012, 3040+ citations).
• idaholab/moose C++ Python - Multiphysics Object-Oriented Simulation Environment. Coupled physics FEM framework (Gaston 2009, 1290+ citations).
• Kratos-Multiphysics C++ Python - Framework for multi-physics FEM. Structural, fluid, thermal, contact, FSI.
• mfem/mfem C++ - High-order finite element library. Supports GPU acceleration, AMR, and dozens of physics applications.
• nschloe/meshio Python - I/O for mesh formats. Convert between Abaqus, Gmsh, VTK, XDMF, Exodus, and 30+ formats.
• OpenSees/OpenSees C++ - Open system for earthquake engineering simulation. Structural and geotechnical response analysis. Berkeley.
• sfepy/sfepy Python - Simple Finite Elements in Python. Solve PDEs by FEM in 1D, 2D, and 3D with plain Python scripting.

^{back to top}

SPH — Smoothed Particle Hydrodynamics

Meshless particle methods for free-surface flows, fluid-structure interaction, and wave dynamics.

SPH is a meshless Lagrangian method particularly suited for AI-driven simulation of free-surface flows, wave impacts, and fluid-structure interaction. DualSPHysics (746 citations) is the most widely used open-source SPH solver with GPU acceleration via CUDA, while PySPH offers a pure Python framework for rapid prototyping of particle-based methods.

• DualSPHysics/DualSPHysics C++ CUDA - GPU-accelerated SPH solver. Free-surface flows, wave generation, fluid-structure interaction, floating bodies.
• SPlisHSPlasH/SPlisHSPlasH C++ - Physically-based SPH fluid simulation. DFSPH, IISPH, PBF pressure solvers. Viscosity, surface tension.
• pypr/pysph Python Cython - SPH framework in Python. Compressible/incompressible flows, solid mechanics, coupled problems.

^{back to top}

DEM — Discrete Element Method

Particle-based simulation of granular materials, powders, and coupled particle-fluid systems.

• CFDEMproject/LIGGGHTS-PUBLIC C++ - Industry-standard open-source DEM for granular materials. LAMMPS-based with heat transfer and CFD coupling (CFDEM).
• SudoDEM/SudoDEM C++ Python - DEM for non-spherical particles. Polyhedra, super-ellipsoids, and cylinders for realistic granular simulations.

^{back to top}

Visualization & Post-processing

Rendering, plotting, and interactive exploration of simulation results.

AI-driven scientific visualization enables agents to render simulation results, generate animations, and create publication-quality figures without GUI interaction. ParaView-MCP from LLNL and viznoir provide MCP-based visualization, allowing AI agents to control rendering pipelines via natural language. VTK and PyVista form the foundation, with Python APIs that can be scripted for automated post-processing of CFD, FEA, and SPH results.

• kimimgo/viznoir Python MCP - Cinema-quality science visualization MCP server. 22 tools, EGL/OSMesa headless, cinematic lighting, physics animations.
• Kitware/VTK C++ Python - The Visualization Toolkit. 3D computer graphics, image processing, scientific visualization. Industry standard.
• pyvista/pyvista Python - Pythonic VTK. Streamlined 3D plotting, mesh analysis, and interactive visualization.
• Kitware/ParaView C++ Python - Multi-platform data analysis and visualization. VTK-based GUI + Python scripting + client-server architecture.
• napari/napari Python - Fast n-dimensional image viewer. Plugin ecosystem for biomedical and scientific imaging.
• marcomusy/vedo Python - Scientific analysis and visualization of 3D objects and point clouds. VTK-based with simple API.
• plotly/plotly.py Python - Interactive, publication-quality graphs. 3D scatter, surface, mesh, volume. Web-based rendering.

^{back to top}

CAD & Geometry

Parametric modeling, geometry processing, and CAD data exchange tools.

AI for CAD enables programmatic 3D modeling, geometry manipulation, and design automation. CadQuery provides a Python-first parametric CAD API built on OpenCASCADE, making it the most accessible tool for AI agents to generate STEP/STL models from code. FreeCAD offers a full parametric modeler with Python scripting, while trimesh handles mesh-level geometry operations for AI-driven design optimization.

• CadQuery/cadquery Python - Parametric 3D CAD scripting. Build models with Python, export STEP/STL/IGES. OpenCASCADE kernel.
• FreeCAD/FreeCAD C++ Python - Open-source parametric 3D CAD modeler. Part design, FEM workbench, BIM, path (CAM).
• OpenCASCADE/OCCT C++ - Open CASCADE Technology. Kernel for 3D surface and solid modeling, CAD data exchange (STEP/IGES).
• mikedh/trimesh Python - Load and manipulate triangular meshes. Boolean operations, ray tracing, convex hulls, format conversion.
• pygmsh/pygmsh Python - Python interface for Gmsh. Scripted geometry + mesh generation with parametric control.

^{back to top}

Mesh Generation

Structured, unstructured, and AI-driven mesh generation for simulation preprocessing.

AI-driven mesh generation is transforming simulation preprocessing. Gmsh (5100+ citations) remains the industry-standard mesh generator with a native Python API, while MeshAnything (ICLR 2025) and MeshXL (NeurIPS 2024) represent a new generation of transformer-based mesh generators that learn to produce artist-quality meshes from any 3D input. CGAL provides the computational geometry backbone for custom meshing algorithms.

• gmsh C++ Python - Full-featured 3D finite element mesh generator. CAD engine, structured/unstructured meshing, built-in post-processing.
• CGAL/cgal C++ - Computational Geometry Algorithms Library. Mesh generation, triangulation, Boolean operations, convex hulls.
• NETGEN/NETGEN C++ Python - Automatic 3D tetrahedral mesh generator. CAD (OCC) integration, mesh optimization, parallel meshing.
• buaacyw/MeshAnything Python - Artist-quality mesh generation with autoregressive transformers. Any 3D input to mesh (ICLR 2025 spotlight).
• OpenMeshLab/MeshXL Python - Foundation model for 3D mesh generation. Pre-trained on Objaverse, text-to-mesh capable (NeurIPS 2024).
• PyMesh/PyMesh Python C++ - Geometry processing library. Boolean, convex hull, remeshing, self-intersection repair.

^{back to top}

Differentiable Simulation

Eight GPU-native differentiable frameworks. MuJoCo leads with 6250+ citations and 12.4k GitHub stars.

Differentiable simulation frameworks enable gradient-based optimization through physical simulations, bridging the gap between machine learning and physics. MuJoCo (6250+ citations, DeepMind) is the most widely used physics engine for robotics and control. JAX-based tools (XLB, Brax, jax-md, JAXFLUIDS) provide end-to-end differentiability on GPU/TPU, while NVIDIA Warp and Taichi (DiffTaichi, ICLR 2020) offer CUDA-native alternatives.

• Autodesk/XLB Python JAX - Differentiable Lattice Boltzmann for physics-ML. Scales to billions of cells on multi-GPU (CPC 2024).
• google/brax Python JAX - Massively parallel rigidbody physics on accelerator hardware. Millions of steps/second on TPU (NeurIPS 2021).
• jax-md/jax-md Python JAX - Differentiable, hardware-accelerated molecular dynamics. Runs on CPU/GPU/TPU via XLA.
• gbionics/jaxsim Python JAX - Differentiable multibody dynamics engine. Hardware-accelerated robot learning and control via JAX.
• google-deepmind/mujoco C++ Python - Multi-joint dynamics with contact. General-purpose physics engine for robotics, biomechanics, and control (Todorov 2012, 6250+ citations).
• NVIDIA/warp Python CUDA - Differentiable simulation and spatial computing. Reverse-mode AD, PyTorch/JAX interop.
• taichi-dev/taichi Python CUDA - Productive GPU programming with automatic differentiation. DiffTaichi for differentiable physics (ICLR 2020).
• tumaer/JAXFLUIDS Python JAX - Fully-differentiable CFD solver for 3D compressible single-phase and two-phase flows.

^{back to top}

AI/ML for Simulation

Thirteen tools — from neural operators to LLM agents. GraphCast (Nature 2023, 1160+ citations) and PhysicsNeMo lead the field.

AI and machine learning for simulation is the fastest-growing area in computational engineering, with 13 tools spanning neural operators, LLM agents, and foundation models. GraphCast (DeepMind, Nature 2023) produces 10-day weather forecasts in under a minute. Neural operators like FNO learn to solve entire families of PDEs in milliseconds. LLM agent frameworks like Foam-Agent (88% success rate) and MetaOpenFOAM automate end-to-end CFD workflows from natural language. NVIDIA PhysicsNeMo provides the most comprehensive physics-ML framework.

• csml-rpi/Foam-Agent Python API - AI agent for automated CFD workflows. LLM-driven OpenFOAM simulation setup and execution.
• deepmodeling/deepmd-kit Python C++ - Deep learning for molecular dynamics. Neural network potentials for large-scale atomistic simulations.
• dynamicslab/pykoopman Python - Data-driven Koopman operator approximation. Dynamical system analysis and prediction from time series.
• dynamicslab/pysindy Python - Sparse Identification of Nonlinear Dynamics. Data-driven discovery of governing equations from measurements.
• google-deepmind/graphcast Python - Graph neural network for medium-range weather forecasting. Ten-day forecasts in under a minute (Nature 2023, 1160+ citations).
• google/jax-cfd Python - JAX-based CFD. Differentiable Navier-Stokes solvers. GPU-accelerated, auto-differentiable.
• Koopman-Laboratory/KoopmanLab Python - Koopman Neural Operator for mesh-free nonlinear PDE solving. Multi-scale decomposition.
• lululxvi/deepxde Python - Deep learning library for PDEs. PINNs, DeepONet. Backends: TensorFlow, PyTorch, JAX, PaddlePaddle.
• microsoft/ClimaX Python - Foundation model for weather and climate. Pre-trained on CMIP6, fine-tunable for downstream tasks.
• NeuralOperator/neuraloperator Python - Neural operators in PyTorch. FNO, SFNO, UNO for learning PDE solution operators.
• NVIDIA/physicsnemo Python CUDA - Physics-ML framework (formerly Modulus). PINNs, neural operators, GNNs, diffusion models. Apache 2.0.
• Terry-cyx/MetaOpenFOAM Python API - LLM-based multi-agent framework for CFD. Automated simulation pipeline from natural language.
• tum-pbs/PhiFlow Python - Differentiable PDE simulations. Fluid dynamics with TF/PyTorch/JAX. ML-physics hybrid workflows.

^{back to top}

Surrogate Models & PINNs

Physics-informed neural networks and data-driven reduced-order models for fast PDE solving.

Physics-informed neural networks (PINNs) embed physical laws directly into neural network training, enabling PDE solving without simulation data. DeepXDE is the most widely used PINN library with multi-backend support (TensorFlow, PyTorch, JAX), while pinns-torch provides a production-ready PyTorch implementation. For reduced-order modeling, PyDMD implements Dynamic Mode Decomposition for data-driven surrogate models that can replace expensive CFD/FEA simulations with real-time predictions.

• lululxvi/deepxde Python - Physics-informed neural networks for PDEs. Multi-backend (TF, PyTorch, JAX). Inverse problems, fractional PDEs.
• mathLab/PINA Python - Physics-Informed Neural networks for Advanced modeling. PyTorch Lightning-based with multi-device training.
• mathLab/PyDMD Python - Dynamic Mode Decomposition. Data-driven reduced-order modeling for fluid dynamics and beyond.
• NeuroDiffGym/neurodiffeq Python - Neural network solver for ODEs and PDEs. Flexible architecture with native boundary condition handling.
• NVIDIA/physicsnemo-sym Python - Symbolic AI for physics. Physics-informed neural networks with symbolic equation definition.
• rezaakb/pinns-torch Python PyTorch - Production-ready PINNs in PyTorch. Multi-physics support, inverse problems, uncertainty quantification.
• sciann/sciann Python - Neural networks for scientific computing. Keras-based PINNs with custom loss and constraints.
• thuml/Neural-Solver-Library Python - Library for advanced neural PDE solvers. Benchmarking Transolver, FNO, and variants on diverse PDE families (ICML 2024, Tsinghua).

^{back to top}

Optimization

Six optimization tools — from Bayesian (BoTorch) to topology (dl4to) to multidisciplinary (OpenMDAO, NASA).

AI-powered optimization combines gradient-based methods with neural networks for engineering design. BoTorch (Meta, NeurIPS 2020) provides Bayesian optimization in PyTorch for sample-efficient design. OpenMDAO (NASA) is the standard for multidisciplinary design optimization, while dl4to applies deep learning to 3D topology optimization. Multi-objective optimization via pymoo supports NSGA-II/III for Pareto-optimal engineering designs.

• meta-pytorch/botorch Python PyTorch - Bayesian optimization in PyTorch. Sequential decision making, multi-objective optimization, batch acquisition (NeurIPS 2020, Meta).
• OpenMDAO/OpenMDAO Python - Multidisciplinary design optimization. NASA-developed. Gradient-based + surrogate-assisted optimization.
• airbus/pymoo Python - Multi-objective optimization. NSGA-II/III, reference directions, constraint handling, parallelization.
• dl4to/dl4to Python PyTorch - Deep learning for 3D topology optimization. Autograd + adjoint method for efficient neural optimization.
• topology-opt/topy Python - Topology optimization with Python. Minimum compliance, heat conduction, mechanism design.
• M2DOLab/OpenAeroStruct Python - Aerostructural optimization. VLM aerodynamics + beam FEM structures + ply-level composites.

^{back to top}

Data Formats & I/O

Libraries for reading, writing, and converting simulation data across mesh and field formats.

• nschloe/meshio Python - I/O for mesh formats. Abaqus, CGNS, Gmsh, VTK, XDMF, Exodus, and 30+ more.
• h5py/h5py Python - Pythonic interface to HDF5. Read/write large numerical datasets efficiently.
• Unidata/netcdf4-python Python - Python/NumPy interface to NetCDF. Climate, ocean, atmospheric simulation data.
• CGNS/CGNS C Fortran - CFD General Notation System. Standard for CFD data storage and exchange. HDF5-based.
• pyvista/pyvista Python - Read/write VTK formats (VTI, VTP, VTU, VTS, VTR), STL, OBJ, PLY, glTF, and more.

^{back to top}

Datasets & Benchmarks

Standardized datasets and benchmarks for training and evaluating scientific ML models.

Benchmarks are critical for advancing AI for science. The Well (NeurIPS 2024) provides the largest collection of diverse physics simulations with fifteen-plus PDE systems, while PINNacle (NeurIPS 2024) offers a comprehensive PINN benchmark with 20 PDE problems. PDEBench standardizes scientific ML evaluation, and DrivAerNet provides 4000+ automotive CFD cases for data-driven aerodynamic design.

• divelab/AIRS Python - AI for science benchmarks. Molecular, protein, climate, physics datasets.
• DrivAerNet Python - Large-scale automotive CFD dataset. 4000+ car designs with drag coefficients and surface fields.
• i207M/PINNacle Python - Comprehensive PINN benchmark with 20 PDE problems across difficulty levels (NeurIPS 2024).
• NASA TMR - Turbulence Modeling Resource. Validation cases for CFD turbulence models with experimental data.
• pdebench/PDEBench Python - Benchmarks for scientific ML. Standardized PDE datasets with baseline models.
• PolymathicAI/the_well Python - Large-scale collection of diverse physics simulations for ML. Fifteen-plus PDE systems (NeurIPS 2024).

^{back to top}

Learning Resources

Tutorials, courses, and templates for learning computational engineering and scientific computing.

Learning resources for AI-powered CAE span from classical CFD fundamentals (12 Steps to Navier-Stokes) to cutting-edge surveys of machine learning for simulation. Physics-Based Deep Learning from TUM provides the most comprehensive collection of ML-physics papers with code, while Awesome-AI4CFD covers the latest data-driven surrogates, PINNs, and ML-assisted numerical methods for computational fluid dynamics.

• Barba-group/CFDPython Python - Classic "12 Steps to Navier-Stokes" tutorial. Learn CFD fundamentals with Python step by step.
• ikespand/awesome-machine-learning-fluid-mechanics - Curated list of ML applications in fluid mechanics. Papers, code, and tutorials.
• jxx123/simglucose Python - Type 1 diabetes simulator. Example of AI-in-the-loop biomedical simulation.
• kks32/phd-thesis-template LaTeX - Clean PhD thesis template. Widely used in computational mechanics community.
• maziarraissi/PINNs Python - The foundational PINN reference implementation. Data-driven PDE solutions and discovery (JCP 2019).
• thunil/Physics-Based-Deep-Learning - Comprehensive collection of physics-based deep learning resources. Papers, code links, and tutorials from TUM.
• WillDreamer/Awesome-AI4CFD - Survey of ML for CFD covering data-driven surrogates, PINNs, and ML-assisted numerical solvers.

^{back to top}

FAQ

What is AI-CAE?
AI-CAE (Artificial Intelligence for Computer-Aided Engineering) is the application of AI and machine learning to engineering simulation, design, and manufacturing. It includes AI agents that automate CFD and FEA workflows, physics-informed neural networks (PINNs) that solve PDEs without mesh generation, neural operators like Fourier Neural Operator (FNO) that learn to predict simulation results in milliseconds, and differentiable simulation frameworks that enable gradient-based optimization through physics.

What are MCP servers for engineering?
Model Context Protocol (MCP) servers allow AI agents like Claude, ChatGPT, and GitHub Copilot to directly control engineering software. As of 2026, MCP servers exist for OpenFOAM (CFD simulation), ParaView (scientific visualization via LLNL's ParaView-MCP), and VTK (rendering via viznoir). These enable AI agents to set up simulations, run solvers, and render results without human GUI interaction.

What is the best open-source AI tool for CFD?
OpenFOAM is the most AI-ready CFD solver, with three LLM agent frameworks: Foam-Agent (88% success rate, MCP-native), MetaOpenFOAM (multi-agent from natural language), and an OpenFOAM MCP server. For differentiable CFD, JAX-CFD provides auto-differentiable Navier-Stokes solvers, and JAXFLUIDS offers fully-differentiable 3D compressible flow simulation. For ML-accelerated CFD, NVIDIA PhysicsNeMo includes neural operators and PINNs.

What is the best PINN library?
DeepXDE is the most widely used physics-informed neural network library, supporting TensorFlow, PyTorch, JAX, and PaddlePaddle backends. It implements PINNs, DeepONet, and handles inverse problems and fractional PDEs. For PyTorch-specific workflows, pinns-torch provides a production-ready implementation, while PINA offers PyTorch Lightning integration with multi-device training. PINNacle (NeurIPS 2024) provides a benchmark with 20 PDE problems for evaluating PINN libraries.

How do AI agents run simulations?
AI agents use MCP (Model Context Protocol) or Python APIs to control simulation software. For example, an agent can use Foam-Agent to set up an OpenFOAM case from a natural language description, execute the solver, then use viznoir or ParaView-MCP to visualize results. The Core Engine Readiness table in this list assesses which of 13 foundational solvers support this workflow. Currently, only OpenFOAM and ParaView have full AI-native integration.

What is the difference between this list and awesome-scientific-computing?
awesome-scientific-computing lists general numerical computing tools. awesome-ai-cae focuses specifically on the intersection of AI and engineering simulation, evaluating every tool for AI-readiness — whether an AI agent can call it via MCP, API, or CLI. It includes a Core Engine Readiness assessment, AI-readiness tiers (AI-Native, API-Ready, ML-Powered), and covers emerging areas like LLM agents for CFD and differentiable simulation.

^{back to top}

Contributing

Contributions welcome! Please read the contributing guidelines first.

To the extent possible under law, kimimgo has waived all copyright and related or neighboring rights to this work.