CrystaLLM
Autoregressive crystal language model with validity checks on generated CIFs, optional Monte Carlo tree search guided by a learned formation-energy network, and DFT on finished structures.
The loop
- Proposes
A domain-trained transformer generates full CIF crystal descriptions conditioned on composition or space group.
- Judged by
Validity checks on generated CIFs; the optional Monte Carlo tree search is scored by a learned formation-energy network (ALIGNN); DFT is run on finished structures.
- Steers
Optional Monte Carlo tree search guided by a learned formation-energy network steers generation.
Headline result
Rediscovered unseen experimental crystal structures and identified novel stable inorganic phases validated by ab initio DFT.
What this does not prove
Struggles on very large unit cells (>60 atoms) where token context length and space-group Wyckoff constraints become tight.
How it generalizes the harness
Validity checks on generated CIFs, a learned formation-energy network guiding optional tree search, and DFT on finished structures.
Classification
| What it proposes | Crystal structure |
|---|---|
| Generative model | Autoregressive CIF transformer (CrystaLLM) |
| Algorithmic authority | CIF validity checks and a learned formation-energy network (ALIGNN); DFT on finished structures |
| Feedback | Verifier errors |
| What gets updated | Population of candidates and search tree |
| When it runs | At inference time |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Code released |
| Lineage | Original system |