Tutorial 4. Generative AI for Physics: From Models to Materials

ANCILLARYEVENT · MAR-4T · ID: MAR-4T

Generative AI is transforming how physics problems are approached by learning probability distributions over data rather than relying solely on deterministic equations. This tutorial introduces the key foundations of this paradigm and examines core model architectures such as variational autoencoders (VAEs), generative adversarial networks (GANs), denoising diffusion probabilistic models (DDPMs) and other diffusion models, and flow-based methods. We will discuss how these models encode physical priors, enable property-guided generation, and use techniques like classifier-free guidance and targeted conditioning to enforce domain-specific constraints. Crystal structure prediction and quantum device optimization will be used as illustrative applications to show how generative AI is beginning to accelerate discovery in physical sciences.

Topics Covered: 

• Foundations of generative modeling in physics: forward vs. inverse design
• Core architectures: VAEs, GANs, diffusion models, and flow matching
• Latent space geometry and conditional generation techniques
• Training tricks and constraint enforcement, and materials gen-AI models like CDVAE, DiffCSP, MatterGen 
• Selected applications: crystal structure generation and quantum device design