GLASS: Global Latent Aggregation with Slot-based Set Decoding for Scalable All-Atom Crystal Generation
Hendrik Kraß, Seyed Mohamad Moosavi, Mathias Niepert
GLASS encodes a crystal into one permutation-invariant latent vector and decodes it with learned slots trained by Hungarian matching. A flow in the latent space generates new crystals.
Install uv, then:
git clone https://github.com/henk789/glass
cd glass
uv syncTraining and sampling run on Linux, macOS, and Windows (NVIDIA GPU, Apple GPU, or CPU).
Pretrained checkpoints are on Hugging Face.
uv run hf download henk789/glass mp20/seed0/flow_50k.pt --local-dir checkpoints
uv run python -m glass.sample --config configs/mp20.yaml --checkpoint checkpoints/mp20/seed0/flow_50k.pt --out runs/mp20/samplesuv run hf download henk789/glass qmof150/seed0/flow_1m.pt --local-dir checkpoints
uv run python -m glass.sample --config configs/qmof150.yaml --checkpoint checkpoints/qmof150/seed0/flow_1m.pt --out runs/qmof150/samplesEach run writes 10,000 structures to samples.pt and one CIF per valid structure.
uv run python -m glass.data configs/mp20.yaml
uv run python -m glass.train_autoencoder --config configs/mp20.yaml --out runs/mp20/ae
uv run python -m glass.train_flow --config configs/mp20.yaml --autoencoder runs/mp20/ae/model.pt --out runs/mp20/flowThe first command downloads and prepares the dataset. Replace mp20 with qmof150
for QMOF150. Flow checkpoints are saved every 10k steps in runs/mp20/flow/checkpoints/.
Training time on one H100: about 20 minutes for the MP20 autoencoder and 3.5 hours for the QMOF150 autoencoder; about 50 minutes for either flow (1M steps). The 50k-step MP20 checkpoint used for the MP20 benchmarks is reached after about 3 minutes.
Evaluation requires Linux and a CUDA GPU. Install the benchmark tools once:
uv run python -m eval.setup genbench
uv run python -m eval.setup crystalite
uv run python -m eval.setup mofchecker
uv run python -m eval.setup mofid
uv run python -m eval.setup esencrystalite needs the CUDA toolkit, mofid a C++ compiler, and esen access to
the gated OMAT24 models.
Evaluate the pretrained checkpoints:
scripts/evaluate_mp20.sh checkpoints/mp20/seed0/flow_50k.pt runs/mp20/benchmark
scripts/evaluate_qmof150.sh checkpoints/qmof150/seed0/flow_1m.pt runs/qmof150/benchmark- MP20: LeMat-GenBench on 2,500 samples, raw and NequIP-relaxed; Crystalite metrics on 10,000 samples.
- QMOF150: MOFChecker validity, MOFid novelty and uniqueness, and similarity to training structures on raw samples; MOFChecker validity after relaxation.
Validity and yields are computed over all requested samples, counting failed structures. Training-structure similarity is reported among valid samples. Crystalite metrics use Crystalite's own evaluation pools.
Autoencoder reconstruction error:
uv run hf download henk789/glass mp20/seed0/autoencoder.pt --local-dir checkpoints
uv run python -m eval.reconstruction --config configs/mp20.yaml --autoencoder checkpoints/mp20/seed0/autoencoder.pt --out runs/mp20/reconstructionuv run python -m toy.fixed_grid train --n 24 --seed 0 --method naive
uv run python -m toy.fixed_grid sample --run runs/toy/n24_seed0_naiveMethods are naive, eqot, and decoder. scripts/toy.sh runs the full study.
MIT