neuraloperator/neuraloperator

What it solves

NeuralOperator is a PyTorch library designed to learn mappings between function spaces, rather than just mapping between fixed-size vectors. This allows the model to be resolution invariant, meaning a model trained on one resolution of data can be applied to data of any other resolution.

How it works

The library provides implementations of neural operator architectures, most notably the Fourier Neural Operator (FNO) and the Tucker Tensor FNO (TFNO). TFNO uses Tucker factorization of weights to significantly reduce the number of parameters while maintaining efficiency in the forward pass by contracting inputs directly with the decomposition factors.

Who it’s for

Researchers and developers working with scientific machine learning, specifically those needing to learn operators that are independent of the data resolution.

Highlights

• Official implementation of Fourier Neural Operators (FNO).
• Resolution invariant learning across function spaces.
• Tucker Tensor FNO (TFNO) for parameter-efficient model architectures.
• Part of the PyTorch Ecosystem.