Equivariant Neural Field Networks for steady PDE surrogates on general geometries.

This repository contains the Pytorch implementation of enf2enf Geometry aware inference of steady state PDEs using Equivariant Neural Fields representations, a neural operator approach to solve steady state PDEs on general geometries. It translates in Pytorch some of the models found in the JAX implementation of the Equivariant Neural Fields architecture: https://github.com/david-knigge/enf-pde.

Architecture Overview

Our architecture leverages equivariant neural fields to learn PDE solutions across different domains. The network processes geometric features and produces accurate field predictions while maintaining equivariance properties.

Experiments

For the moment this repo implements the experiment on the Aifranns dataset shape encoding. More experiments will be added son. For the whole experiments refer to the original jax implementation.

Airfrans Dataset

For airfoil simulations, we use the AirFRANS dataset to predict flow fields around airfoils at different angles of attack and flow conditions.

To run experiments:

python airfrans_full.py

Data

Data for Airfrans dataset can be found by downloading the airfrans package.

pip install airfrans