Title：Building Geometric Foundations for AI in Molecular Modeling
Date：2025/11/13 11:10
Location：R210, CSIE
Speakers：王士欣先生
Abstract:
Recent advances in AI for science, such as AlphaFold and AI-driven protein design, demonstrate that deep learning can reveal and generate complex molecular structures. However, a key challenge remains: molecules and proteins are multiscale and highly flexible. They contain repeating local patterns (like amino acids or functional groups), but the same repeating elements can assemble into thousands of different global conformations, making structure prediction and learning difficult.
In this talk, I will present my work in geometric deep learning for addressing this challenge. I will introduce a sparse and rigid graph representation that allows models to capture global 3D structure efficiently, and its extension—a hierarchical framework for proteins that connects local secondary structures to overall molecular shape, achieving both computational and memory efficiency.
Biography:
Mr. Shih-Hsin Wang focused on developing mathematically rigorous methods that reduce wasted resources and costs by minimizing errors while maximizing efficiency. His research bridges theory and application across geometric deep learning, generative models (e.g., flow matching & diffusion models), and algebraic geometry. Through AI for Science, I work to translate complex mathematical insights into practical, reliable solutions for molecular and biological applications.