Reinforcement learning with formation energy feedback for material diffusion models.

Area of Science:

• Materials Science
• Artificial Intelligence
• Computational Chemistry

Background:

• Generative models, particularly diffusion models, are increasingly used for efficient material discovery.
• Incorporating physical constraints and symmetries into diffusion models improves crystal generation quality.
• Challenges remain in accurately capturing stable crystal structures due to data limitations and complexity.

Purpose of the Study:

• To propose a novel fine-tuning framework, Reinforcement Learning Fine-tuning with Energy Feedback (RLFEF), to enhance the stability of generated crystal materials.
• To improve the success rate of generative models in producing stable crystal structures.

Main Methods:

• Formulated the material diffusion process as a Markov Decision Process, using formation energy as rewards.
• Proved the equivalence between optimizing expected return in reinforcement learning and applying policy gradient updates to diffusion models.
• Demonstrated that the fine-tuned model respects the inherent symmetries of crystal materials.

Main Results:

• The RLFEF framework achieved state-of-the-art performance across multiple tasks.
• Demonstrated superior results in property optimization, ab initio generation, crystal structure prediction, and general material generation.
• Showcased improved stability and accuracy in generated crystal material structures.

Conclusions:

• The RLFEF framework offers a significant advancement in generative material science.
• This approach effectively addresses the limitations of existing models in generating stable crystal materials.
• The method holds promise for accelerating the discovery of novel functional materials.