Reference and tools / Build your own

Frameworks

Where to start in code: open neural frameworks, plus classical surrogate and UQ toolkits.

• PhysicsNeMo (NVIDIA): A PyTorch framework for training and deploying physics-ML models on simulation data, from data loading to multi-GPU training. Formerly NVIDIA Modulus.
• NeuralOperator (Caltech / NVIDIA (PyTorch Ecosystem)): The reference library for neural operators: models that map between functions and work across discretisations.
• Neural-Solver-Library (THUML, Tsinghua University): A library of neural PDE solvers from the authors of Transolver, with unified benchmarks for comparing architectures.
• DeepXDE (Lu Lu et al.): Physics-informed neural networks and operator learning, with PyTorch, JAX and TensorFlow backends.
• PyTorch Geometric (PyG team): General-purpose graph neural network building blocks, often used to assemble mesh-based surrogates.
• SMT (Surrogate Modeling Toolbox): Classical surrogates in Python: kriging, RBF, polynomial models, gradient-enhanced and multi-fidelity kriging, sampling plans.
• DAKOTA (Sandia National Laboratories): Optimisation, uncertainty quantification and surrogate-based design studies driving any simulation code.
• UQPy (Johns Hopkins University): Uncertainty quantification with surrogates: sampling, polynomial chaos, Gaussian processes, sensitivity.

Related

• Datasets: Public simulation datasets to benchmark on before you have your own.
• Get started: From open frameworks and datasets to a surrogate trained on your own simulations.
• Model families: Neural fields, U-Nets, graph networks and transformers, and what each is good for.