CrystalTextLLM
Fine-tuned language model proposes 3D crystal CIF files where sampling retries when a CIF cannot be parsed, and validity, M3GNet and DFT stability are measured afterwards.
The loop
- Proposes
A fine-tuned LLaMA-2 model generates crystallographic information files (space group, lattice lengths/angles, fractional coordinates).
- Judged by
Pymatgen structural validators, M3GNet interatomic potential relaxations, and VASP Density Functional Theory calculations measure validity and thermodynamic stability afterwards.
- Steers
Sampling retries when a CIF cannot be parsed, and conditional infill prompts steer generation.
Headline result
Nearly 90% of sampled structures pass geometric and charge-neutrality checks, and a high fraction relax to metastable or stable states in DFT.
What this does not prove
Text-tokenized coordinates have finite numerical precision prior to force-field or DFT structural relaxation.
How it generalizes the harness
Crystallographic geometry checkers and quantum-mechanical DFT relaxations measure validity and stability of the generated crystals afterwards.
Classification
| What it proposes | Crystal structure |
|---|---|
| Generative model | Fine-tuned LLaMA-2 70B on CIF crystal strings |
| Algorithmic authority | Pymatgen validator, M3GNet potential, and VASP DFT solver |
| Feedback | Score |
| What gets updated | Population of candidates |
| When it runs | At inference time |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Code released |
| Lineage | Original system |