Seminars

Home Centers & Programs AI and Natural Sciences Seminars

TITLE: AI-Driven Atomistic Discovery: From ML Potentials to Diffusion-Based Structure Generation

ABSTRACT: This talk will present a unified generative-AI framework for atomistic scientific discovery that combines machine-learning-accelerated molecular dynamics (MLMD) with diffusion-based generative models. I will begin with a case study on proton transfer in nanoporous TiO2 films, where deep potential molecular dynamics (DeepMD) enables large-scale simulations that reveal how confinement reorganizes hydrogen-bond networks and reshapes proton-transfer mechanisms. Building on these mechanistic trajectories, I will show how diffusion models -- adapted from image generation to atomic structures -- serve as a general inverse engine in two settings. First, I will demonstrate spectra-conditioned diffusion models for spectroscopy interpretation, which reconstruct physically consistent atomic and molecular structures from noisy experimental signals while naturally quantifying uncertainty through sampled structural ensembles. Second, diffusion provides robust denoising of highly perturbed simulation snapshots and enables joint denoising-classification into established crystal prototypes or phases, supporting automated tracking of structural evolution and phase transitions. Together, MLMD and diffusion models form a complementary, physics-grounded pipeline for accelerating simulation, characterization, and inverse design in materials and molecular systems.