mishapadidar/diffusion_for_fusion

Data

This repo contains the data accompanying the paper "Diffusion for Fusion: Designing Stellarators with Generative AI" by Misha Padidar, Teresa Huang, Andrew Giuliani, Marina Spivak in the AI4Science workshop at Neurips, 2025.

There are two different copies of the dataset with minor differences. It was found that approximately ~3k, of the 370k, feature vectors contained some errors in them during the canonicalization process. The original data used to train the model is contained in dofs.npy, QUASR.pkl. The corrected dataset is dofs_06082025.npy, QUASR_06082025.pkl. While we attach both datasets here for transparency, in all future work the later dataset should be used.

• dofs.npy contains the Fourier coefficients of the surfaces, i.e. the x values.
• QUASR.pkl contains features (iota, nfp, etc) corresponding to each configuration in dofs.npy, i.e. the conditions y.
• dofs_06082025.npy contains the Fourier coefficients of the surfaces, i.e. the x values. The ~3k feature vectors were corrected in this version.
• QUASR_06082025.pkl contains features (iota, nfp, etc) corresponding to each configuration in dofs_06082025.npy, i.e. the conditions y. The ~3k feature vectors were corrected in this version.