v0.0.1 - Initial Release

Chemeleon2 v0.0.1 - Initial Release

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

Highlights

• Three-stage training pipeline: VAE → LDM → RL
• Group Relative Policy Optimization (GRPO) for RL fine-tuning
• Custom reward system for material property optimization

Added

Core Modules

• VAE module for encoding crystal structures into latent space
• LDM module with diffusion Transformer (DiT) architecture
• RL module with Group Relative Policy Optimization (GRPO)

Features

• Custom reward system for material property optimization
• Support for multiple datasets: MP-20, Alex-MP-20, MP-120
• CrystalBatch data schema for crystal structure handling
• WandB integration for experiment tracking
• Configuration management with Hydra

Documentation

• Comprehensive documentation with Jupyter Book
• Training guide for VAE, LDM, RL, and predictor models
• Evaluation guide for sampling and metrics
• Custom reward implementation guide
• API reference documentation
• Tutorial notebook for sampling and evaluation

Benchmarks

• 10,000 generated structures from MP-20 RL model
• 10,000 generated structures from Alex-MP-20 RL model

Development

• Testing suite with pytest (baseline, unit, integration tests)
• Pre-commit hooks for code quality (ruff, pyright)
• Contributing guidelines

Links

• Documentation: https://hspark1212.github.io/chemeleon2/
• arXiv Paper: https://arxiv.org/abs/2511.07158