A Sober Look at LLMs for Material Discovery

Official experiment repo for the "A Sober Look at LLMs for Material Discovery" paper (ICML 2024).

Tip

If you just want to use the method as a library, check out the sister repo: https://github.com/wiseodd/lapeft-bayesopt.

Setup

Important

Note that the ordering is important.

1. Install PyTorch (with CUDA): https://pytorch.org/get-started/locally/
2. Install Huggingface libraries and others: pip install transformers datasets peft tqdm
3. Install laplace-torch: pip install laplace-torch

Fixed-Feature Experiments

Cache molecules in $\mathcal{D}_{\mathrm{cand}}$ (see full parameters in the Python file):

python cache_features.py --feature_type {FEATURE_TYPE} --problem {PROBLEM} --prompt_type {PROMPT_TYPE}

Then, do BO:

python run_fixed_features.py --feature_type {FEATURE_TYPE} --method {METHOD} --randseed {RANDSEED} --problem {PROBLEM}

Similarly for the multiobjective experiments (cache_features_multiobj.py and run_multiobj.py).

Finetuning Experiments

Simply run the following.

python run_finetuning.py --foundation_model {FOUNDATION_MODEL} --randseed {RANDSEED} --problem {PROBLEM}

See the Python file for the full arguments.

BO-LIFT In Context Learning Baseline

The script is in baselines/run_bolift.py. It has similar options as the fixed-feature script.

Citation

@inproceedings{kristiadi2024sober,
  title={A Sober Look at {LLMs} for Material Discovery: {A}re They Actually Good for {B}ayesian Optimization Over Molecules?},
  author={Kristiadi, Agustinus and Strieth-Kalthoff, Felix and Skreta, Marta and Poupart, Pascal and Aspuru-Guzik, Al\'{a}n and Pleiss, Geoff},
  booktitle={ICML},
  year={2024}
}