Fine-Tuned Language Models Generate Stable Inorganic Materials as Text

This repository contains the code for the paper Fine-Tuned Language Models Generate Stable Inorganic Materials as Text by Nate Gruver, Anuroop Sriram, Andrea Madotto, Andrew Gordon Wilson, C. Lawrence Zitnick, and Zachary Ward Ulissi (ICLR 2024).

We show that finetuned LLMs can be used to generate stable materials using string encodings. These finetuned LLMs can match or exceed the performance of a domain specific diffusion model (CDVAE). LLMs can also be used to mutate existing materials or to sample crystal structures conditioned on text descriptions.

⚠️ Our method resembles but is not the same as CrystaLLM (https://arxiv.org/abs/2307.04340), which explores training language models from scratch on CIF-formatted crystals. You can find the code associated with that project at the following link: https://github.com/lantunes/CrystaLLM. ⚠️

🛠 Installation

Run the following command to install all dependencies.

source install.sh

After installation, activate the environment with

conda activate crystal-llm

If you prefer not using conda, you can also install the dependencies listed in install.sh manually.

🚀 Training and Sampling Models

Run training with

python llama_finetune.py --run-name 7b-test-run --model 7b

and sample from a PEFT model with

python llama_sample.py --model_name 7b --model_path=exp/7b-test-run/checkpoint-500 --out_path=llm_samples.csv

E above hull evaluation

To construct the convex hull, you must download structure entries from the following link: https://figshare.com/articles/dataset/Matbench_Discovery_v1_0_0/22715158

Then energy values can be computed with

python e_above_hull.py --structures_fn [CSV FILE WITH STRUCTURES] --entries_fn [PATH TO STRUCTURE ENTRIES .json.gz FILE]

License

The majority of crystall-llm is licensed under CC-BY-NC, however portions of the project are available under separate license terms: https://github.com/materialsproject/pymatgen is licensed under the MIT license, https://github.com/huggingface/transformers is licensed under Apache 2.0, and https://gitlab.com/ase/ase/-/ is licensed under GNU Lesser General License

Citation

Please cite our work as:

@inproceedings{gruver2023llmtime,
    title={Fine-Tuned Language Models Generate Stable Inorganic Materials as Text},
    author={Nate Gruver, Anuroop Sriram, Andrea Madotto, Andrew Gordon Wilson, C. Lawrence Zitnick, and Zachary Ward Ulissi},
    booktitle={International Conference on Learning Representations 2024},
    year={2024}
}