AI4S Seminar by Wei Huang "Diffusion Models for Scientific Data"

Wei Huang of RIKEN AIP presents a research talk on how diffusion models generate scientific data, from molecular structures and biological measurements to protein and DNA sequences. The talk examines how these models exploit structure across both continuous and discrete state spaces. Two recent works anchor the session. For continuous data, Score-induced Latent Diffusion (SiLD) reveals a "collapse-and-refine" mechanism under the manifold hypothesis: score learning first locates the low-dimensional data manifold, then refines the distribution within it. Huang covers the theoretical motivation and its application to molecular generation. For discrete data, DPRM is a plug-in token-ordering module for masked diffusion models drawing on the Doob h-transform, which combines model confidence with estimates of future task reward to adapt generation order without changing the underlying model. The speaker plans to introduce the basic ideas of continuous and discrete diffusion before presenting the two works, so attendees from a range of backgrounds should be able to follow. The seminar runs online.