Chemeleon2

A reinforcement learning framework in latent diffusion models for crystal structure generation using group relative policy optimization (GRPO).

Overview¶

Chemeleon2 implements a three-stage pipeline for crystal structure generation:

1. VAE Module: Encodes crystal structures into latent space representations

2. LDM Module: Samples crystal structures in latent space using diffusion Transformer

3. RL Module: Fine-tunes the LDM with reinforcement learning

Key Features¶

• Variational Autoencoder for crystal structure encoding

• Latent Diffusion Model with DiT-based architecture

• GRPO-based Reinforcement Learning for reward optimization

• Modular reward system with built-in and custom components

• Comprehensive evaluation metrics (uniqueness, novelty, stability, diversity)

Quick Links¶

• Installation - Get started with Chemeleon2

• Training Guide - Train VAE, LDM, and RL models

• Architecture - Understand the system design

• API Reference - Detailed API documentation

Citation¶

If you use Chemeleon2 in your research, please cite:

@article{Park2025chemeleon2,
  title={Guiding Generative Models to Uncover Diverse and Novel Crystals via Reinforcement Learning},
  author={Hyunsoo Park and Aron Walsh},
  year={2025},
  url={https://arxiv.org/abs/2511.07158}
}

License¶

Chemeleon2 is licensed under the MIT License. See LICENSE for more information.