腾讯会议ID：531-460-998

会议密码：0609

主持人：王磊 研究员

联系人：傅琦 (fuqi@iphy.ac.cn)

Abstract:

Diffusion models, the engine behind modern image and video generations, share their mathematical backbone with stochastic quantization, the 1980s formulation of Euclidean field theory by Parisi and Wu. This correspondence turns a generative AI technique into a physically motivated sampler for lattice simulations, where critical slowing down and topological freezing limit conventional Monte Carlo. I will first introduce the physics underlying diffusion models, then our recent work showing that physics-conditioned and gauge-equivariant diffusion models generalize across couplings and lattice volumes in scalar and gauge theories, and an optimal stochastic quantization perspective on designing denoising samplers without data.

Brief CV of Dr. Lingxiao Wang:

Lingxiao Wang is now the Deputy Director of AI as Science team at RIKEN, and an Assistant Professor at the University of Tokyo. He received his physics Ph.D. from Tsinghua University in 2020. From 2020 to 2023, he worked as a postdoctoral researcher at the Frankfurt Institute for Advanced Studies (FIAS), and concurrently served as a research assistant at the University of Frankfurt. In 2024, he joined the RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences (iTHEMS) and has led a working group on deep learning. His main research areas focus on machine learning and physics, including non-perturbative approches for simulating quantum field theories, especially for QCD physics. He is also involved in AI for Science from a multidisciplinary perspective.