AI for Science: From Microscopic Structures and Dynamics to Macroscopic Functions

Modern scientific problems often involve systems with intricate structures and nonlinear dynamics interacting across multiple scales. To address these challenges, I develop AI and multiscale computational frameworks that connect microscopic structures and dynamics to macroscopic functions. In this talk, I will present two representative methods. First, I introduce Neuron-based Multifractal Analysis (NeuroMFA), a graph-theoretic approach that quantifies the multiscale properties of systems such as brain networks, materials, and large-scale AI models, and maps them to functional properties. Second, I will introduce the Multiwavelet Neural Operator (MWNO), a scalable framework for learning PDE-governed dynamics via multiresolution decomposition, enabling efficient simulation and forecasting. Together, these tools support the modeling of scientific systems in a principled, interpretable, and scalable way, with applications in neuroscience, materials science, and cyber-physical systems.
 

Xiongye Xiao, PhD candidate in Electrical and Computer Engineering at the University of Southern California, earned his BS in control science and engineering from Zhejiang University. His research focuses on neural operators, complex networks, and AI for scientific discovery. His work has been published in top-tier venues, including NeurIPS, ICLR, ICML, Science Robotics, Nature Computational Science, and Advanced Materials. He is a recipient of the USC Best Research Assistant Award, USC ECE Outstanding Research Assistant Award, and the Best Presentation Award at the Fall Fourier Talks.