Pretrained diffusion checkpoints for GSS and RSS material datasets
Abstract
Description This record provides the pretrained CSP diffusion model used by GSS (Generative Structure Search), a framework that combines diffusion generative models with energy-based guidance from machine-learning force fields to search for stable crystal structures. The paper's main experiments target periodic crystal structure prediction; an aspirin case study demonstrates the same idea on a non-periodic molecular system. CSP diffusion model Architecture: GemNet-T denoiser (derived from MatterGen), trained in CSP mode — denoising atomic positions and lattice parameters while keeping atomic types fixed. Training data: the alex-mp-20 dataset (Materials Project + Alexandria entries with ≤ 20 atoms per unit cell). Framework: PyTorch 2.6 + PyTorch Lightning, 1000-step diffusion schedule. Aspirin diffusion model Architecture: EquiformerV2 score model (derived from EquiformerV2). Training data: the MD17 aspirin trajectory data. Framework: PyTorch 2.6 + PyTorch Lightning, 1000-step VE SDE, denoising score matching (DSM). Reference structures for the 12 benchmark systems Reference sets of stable structures for the twelve chemical systems used to evaluate coverage and efficiency in the paper: Al, C, Si, Ti (elemental); Al2O3, SiO2, TiO2, BN (binary); CaTiO3, LiCoO2, AlPN (ternary); LaCuSO (quaternary). Each set was generated by random structure search: AIRSS proposals at several stoichiometric factors, relaxation with MatterSim-v1.0.0-5M (FIRE, f_max ≤ 0.01 eV/Å, ≤ 1000 steps), removal of unconverged and high-energy structures (E_hull > 1.0 eV/atom), and deduplication with pymatgen's StructureMatcher. In total 4,919 structures in CIF format, organised as / _ .cif, with an index.csv listing the number of RSS proposals that relaxed to each structure. See the Methods section of the paper and the README inside the archive. Contents File Description last.ckpt Trained model weights (~460 MB) config.yaml Hydra training configuration required to reload the CSP model aspirin_dsm.ckpt Aspirin EquiformerV2 weights (~250 MB) gss_reference_structures.tar.gz Reference structures for the 12 benchmark systems (CIF, ~1.8 MB)