Model Card for CrystaLLM-cXRD_alex-mp-20

Model Details

Model Description

CrystaLLM-cXRD_alex-mp-20 is a conditional generative model for crystal structure generation from powder X-ray Diffraction (XRD) data. It is a GPT-2 decoder-only model that conditions the generation of Crystallographic Information Files (CIFs) on a continuous diffraction profile rather than a peak list, using a Perceiver Resampler prefix (PrefixXRD) that compresses the pattern into learned key-value tokens injected at every transformer layer.

The model consumes the whole pattern:

1. Scattering vector (Q) on a fixed 1000-point grid, arange(0, 10, 0.01) Å⁻¹
2. Intensity (I), max-normalised to $[0, 1]$

This removes the peak-picking step from the workflow: a raw diffractometer scan is converted to the (1000, 2) [Q, I] profile automatically at generation time. Because Q is wavelength-independent, patterns from any anode can be used as long as the wavelength is supplied.

Trained on Alex_Mp_20, this is the broad-distribution cXRD model. For experimentally determined inorganic nanomaterials, its CHILI-100K finetune c-bone/CrystaLLM-cXRD_chili100k is the stronger choice.

• Developed by: Bone et al. (University College London)
• Model type: Autoregressive Transformer with Perceiver prefix conditioning (~36.5M parameters)
• Language(s): CIF (Crystallographic Information File) syntax
• License: MIT
• Finetuned from model: c-bone/CrystaLLM-pi_base, the LeMat-Bulk pretrained base

Model Sources

• Repository: GitHub: CrystaLLM-pi (generation), GitHub: CrystaLLM-cXRD (training and benchmarks)
• Paper: To add
• Dataset: HuggingFace: c-bone/cxrd_alex-mp-20

Uses

Direct Use

Structure generation from powder XRD across a broad inorganic distribution. Pass a raw two-column scan and a target composition; the model proposes candidate structures consistent with the diffraction signature. Also the recommended starting point for finetuning a cXRD model onto a new material family.

Out-of-Scope Use

• Disordered Systems: No native handling of partial occupancies or significant disorder.
• Experimental nanomaterials: Alex_Mp_20 is a computed, fully ordered distribution. Use the CHILI-100K finetune for experimental patterns.
• Unconditional generation: The model requires a diffraction profile at every forward pass and has no missing-conditioning fallback.
• Non-CuKα data without a declared wavelength: CuKα1 is assumed when none is given, which silently misplaces every peak in Q.

Bias, Risks, and Limitations

• Computed-data bias: Training patterns are simulated from relaxed DFT structures, so peak shapes are cleaner and more ordered than laboratory scans.
• Preprocessing sensitivity: Recovery depends on the background subtraction and wavelength used to build the profile. Inspect the conversion (--save_plot) before trusting a result.
• Measured range: Regions of the Q grid outside the measured scan are zero-filled; short scans give the model less to work with.
• Context window: 1024 tokens (plus 8 prefix tokens), so very large unit cells will not fit.

Getting started

For generation from a scan, use T4_XRD_continuous.ipynb in CrystaLLM-pi. The CHILI-100K model is finetuned from this checkpoint; see B2_chili100k_benchmark.ipynb.

Citation

Papers to cite:

@misc{bone2025discoveryrecoverycrystallinematerials,
      title={Discovery and recovery of crystalline materials with property-conditioned transformers},
      author={Cyprien Bone and Matthew Walker and Bradley A. A. Martin and Kuangdai Leng and Luis M. Antunes and Ricardo Grau-Crespo and Amil Aligayev and Javier Dominguez and Keith T. Butler},
      year={2025},
      eprint={2511.21299},
      archivePrefix={arXiv},
      primaryClass={cond-mat.mtrl-sci},
      url={https://arxiv.org/abs/2511.21299},
}

and

To add