汕头大学“桑浦论坛”数学学科分论坛学术报告（蔡志强博士）

报告题目：Weak Generative Sampler for Stationary Distribution

报 告 人：蔡志强 博士（香港大学）

邀 请 人：陈晓鹏 副教授

报告时间：2026年11月6日 14：00

报告地点：东海岸校区D209

报告摘要：Computing stationary distributions of stochastic dynamical systems is challenging in the presence of metastability, high dimensionality, and nonlinear interactions. In this talk, we present the Weak Generative Sampler (WGS), a framework that learns a transport map to generate independent samples from an approximation of the stationary distribution, without requiring training data from trajectory simulations. The method combines the weak formulation of the stationary Fokker–Planck equation with generative modeling and randomized test functions. By transferring differential operators to the test functions, WGS avoids density derivatives and Jacobian determinants in the training loss, while adaptive test functions help explore the relevant regions of state space. We further extend WGS to McKean–Vlasov equations through implicit and Picard iterations, with numerical experiments demonstrating its ability to handle high-dimensional mean-field interactions and recover multiple stationary states, including an unstable branch.

报告人简介：蔡志强，香港大学数学系博士后。2025年于香港城市大学获博士学位。研究方向聚焦机器学习与科学计算（生成模型与高维偏微分方程求解）。