CrystaLLM-π

CrystaLLM-π generates crystal structures as CIF files, with optional targets such as bandgap, density, photovoltaic efficiency or XRD peaks. You can use a released model or fine-tune one on your own data.

This repo includes the package, tutorials, documentation and a containerised API, with pretrained models available on Hugging Face. The API also powers the CrystaLLM-π web application. Most models fit on a 16 GB GPU for training and need 1–2 GB for generation. Generation also runs on CPU.

Key Features

• Structure generation: generate CIFs from scratch or from a formula, with optional Z and space-group inputs.
• Property targets: guide generation with one or more numerical properties.
• Fine-tuning: train Prefix or Residual models on your own data. Residual models also support missing property values.
• Evaluation: check validity, uniqueness, novelty and stability.
• Released models: download pretrained models from Hugging Face and run them yourself.

Where to start

If you want to	Page	Notebook
Install the package	Installation
Generate with a released model	Quickstart	T2
Understand the conditional families	Models
Finetune on your own property	Train, generate & evaluate	T1
Recover a structure from XRD peaks	Quickstart
Run it as an HTTP service	API service	T3
Screen for a target property	Train, generate & evaluate	T4
Look up a function	API reference

Reproducing the paper

The notebooks and configs for "Discovery and recovery of crystalline materials with property-conditioned transformers" are in the repository made to reproduce our paper: CrystaLLM-pi-paper.

Use this repo for up to date training, generation, and access to models from our other CrystaLLM-π projects as they come out!

In the reproduce paper code, the Residual model goes buy Slider, and the Prefix model goes by PKV. This is legacy naming and has been updated across this codebase.