DFTPilot

A multi-fidelity bandgap predictor powered by LLM.

Installation

It is recommended to first create and activate a virtual environment.

$ conda create -n dftpilot python=3.11
$ conda activate dftpilot

Then install the package with all dependencies:

$ pip install .[all]

Or install the package in editable mode with development dependencies:

$ pip install -e .[dev]

Data Preparation

Data preparation involves creating a JSONL file and a RAG index file from VASP calculation folder. You can use the provided script to process your raw data.

Prepare JSONL file

$ !python3 /data/parse_vasp_json.py /path/to/train_raw -o path/to/processed_data/train.jsonl

Build RAG index file using SOAP fingerprints and FAISS

$ python /data/rag_index_build.py path/to/processed_data/train.jsonl path/to/processed_data/rag.index

Training

$ python /scripts/train.py \
    --jsonl /path/to/processed_data/train.jsonl \
    --index_path /path/to/processed_data/rag.index \
    --epochs 501 \
    --lr 1e-3 \
    --batch_size 4 \
    --device cuda \
    --patience 100 \
    --val_frac 0.1 \
    --k_neighbors 8 \
    --ckpt_best_path /path/to/checkpoint/bg_llm_best.pt \
    --ckpt_last_path /path/to/checkpoint/bg_llm_last.pt \
    --log_json_path /path/to/log/bg_llm_log.json \
    --seed 42

Prediction

$ python /scripts/batch_predict.py \
  --input /home2/llmhackathon25/data/train_final_processed/test.jsonl \
  --output /home2/llmhackathon25/data/test_final/test_with_predictions_n256.jsonl \
  --checkpoint /home2/llmhackathon25/chkpt_test/bg_llm_best_final_n128.pt \
  --rag-index /home2/llmhackathon25/data/train_final_processed/rag.index \
  --device cuda \
  --k 8

App:

$ cd app/
$ uvicorn app:app --reload --port 8000

Then open your browser and go to http://127.0.0.1:8000

Now you can upload a POSCAR file and select the fidelity level to get the bandgap prediction.

Contributors:

• Chiku Parida (cparida.ai@gmail.com)
• Savyasanchi Aggarwal (savya10@gmail.com)
• Diptendu Roy (diptendu.roy.1996@gmail.com)