NeuralOperator is an open-source PyTorch library for learning neural operators: models that map between function spaces and can solve families of partial differential equations at once. Developed by the team behind Fourier Neural Operators, it has about 4,000 stars on GitHub and is under active development.
Traditional solvers compute one solution of a PDE at a time; neural operators learn the solution operator itself, so a trained model can evaluate new parameter values, forcings or geometries almost instantly and at resolutions different from training. The library implements the key architectures in this line of research, including the Fourier Neural Operator (FNO) and tensorized variants, along with the surrounding plumbing: PDE datasets, physics-informed losses, training loops and evaluation utilities. It builds on the companion TensorLy-Torch ecosystem for factorized layers that keep these large models tractable. Documentation, tutorials and examples cover benchmark problems from fluid dynamics to materials modeling.
NeuralOperator is completely free and open source under the MIT license. There is no commercial edition or hosted service; you run the library on your own compute.
NeuralOperator is for scientific machine learning researchers and engineers working on PDE-governed systems such as fluids, weather and materials, who want production-quality implementations of neural operator methods.
The library is distributed on PyPI as the neuralop package and is designed to slot into ordinary PyTorch workflows: models are standard torch modules, training is handled by a configurable Trainer class with checkpointing, logging and distributed data-parallel support, and losses can mix data terms with PDE residuals for physics-informed training. It also ships utilities for positional embeddings, spectral convolutions and normalizers, so teams can assemble custom operator architectures rather than only using the reference models. Development is led by researchers at Caltech’s Anandkumar group together with NVIDIA and other collaborators, and the codebase is tested against a suite of benchmark problems with published reference scores.