DiffractGPT

Powder XRD ↔ crystal structure with machine learning, in the AtomGPT spectroscopy family (cf. IRGPT, RamanGPT). Two directions plus a consistency loop:

• Inverse (XRD pattern → structure): a causal LLM is fine-tuned (LoRA) on text prompts — chemical formula + the main 2θ peaks — to generate the crystal structure. Any base LLM works; this repo ships adapters for six bases trained on an identical split.
• Forward (structure → XRD pattern): the powder pattern is simulated analytically (jarvis XRD, Cu Kα), so no learned forward model is needed.
• Consistency loop: inverse → relax with ALIGNN-FF → simulate XRD → rank candidates by cosine similarity to the input pattern.

Data and trained models are on the Hugging Face Hub:

• Dataset: knc6/diffractgpt_dft (73,374 train / 8,152 val, Alpaca format)
• Models: knc6/diffractgpt_dft_models (LoRA adapters, one subfolder per base)

Install

pip install -r requirements.txt
pip install unsloth          # optional: ~2x faster train

Data

huggingface-cli download knc6/diffractgpt_dft --repo-type dataset --local-dir data/diffractgpt_split
# or rebuild (reproduces the knc6/diffractgpt_jarvis_dft text format on a fixed split):
python scripts/build_dgpt_data.py

Records are Alpaca-style (instruction, input, id, output); input holds the formula and top-20 2θ peaks <deg>(<rel intensity>), output is a b c / angles / El x y z.

Train (any base LLM)

scripts/train_dgpt.py fine-tunes a LoRA adapter; --backend hf (plain transformers+peft) or --backend unsloth, --base = any HF causal LM.

python scripts/train_dgpt.py --base unsloth/mistral-7b     --data data/diffractgpt_split --output runs/mistral7b
python scripts/train_dgpt.py --base unsloth/Qwen3-4B-Base  --data data/diffractgpt_split --output runs/qwen3_4b
# other bases trained here: unsloth/Qwen3-8B-Base, unsloth/Meta-Llama-3.1-8B, unsloth/Ministral-3-8B-Base-2512

Recipe: LoRA r16/α16; lr 2e-4 linear; effective batch 8; 2 epochs; bf16; seed 3407.

Evaluate

scripts/eval_dgpt.py generates (greedy + 8-sample), ranks candidates by XRD cosine, and scores match rate / lattice MAE against the targets.

python scripts/eval_dgpt.py runs/mistral7b/final unsloth/mistral-7b

Indicative accuracy (1000 validation targets, pymatgen StructureMatcher)

base	greedy top-1	8-sample, selected
Qwen3-4B	61.4%	67.1%
Mistral-7B	68.6%	72.4%

Full multi-base comparison (greedy/sample, raw/refined, lattice MAE, space group, incl. the 8B-class bases) is in the DiffractGPT benchmark write-up.

Inference

python scripts/infer.py --adapter knc6/diffractgpt_dft_models --subfolder model_mistral7b \
  --formula SiS --peaks "3.7deg(1.0), 7.4deg(0.14)"

Forward (structure → XRD pattern)

python scripts/predict_xrd.py --poscar POSCAR     # prints <2theta>  <intensity>

Extending

• Other base LLMs: change --base.
• Other spectroscopies: the same pipeline builds IRGPT, RamanGPT, UVVisGPT.

Citation

DiffractGPT / AtomGPT inverse-design framework; built on JARVIS-DFT (https://doi.org/10.1038/s41524-020-00440-1).