Shijie Xiao (肖师杰)

A. French Building 207 F

Georgia Tech

I am a Ph.D. student in the School of Interactive Computing at Georgia Tech. My research focuses on Structure-aware learning for prediction and decision-making in dynamical systems, combining machine learning with structured dynamics learning, probabilistic prediction and inference and sequential decision-making.

I received my M.S. in Electrical and Computer Engineering from Georgia Tech and my B.S. from the School of Artificial Intelligence and Automation at Huazhong University of Science and Technology (HUST). Before starting my Ph.D., I worked with Prof. Anqi Wu on multi-agent behavior analysis. This work became my master’s thesis, A Computational Framework for Multi-Agent Behavior Analysis: Integrating Geometric Vision Extraction and Social Attention Inference. I have also worked on diffusion models with Prof. Yincong Chen at HKUST (Guangzhou) and Prof. Wu Yuan at CUHK.

Outside research, I serve as president of the GT Chinese Soccer Club. Soccer fans are always welcome to join us.

I’m open to research collaborations and conversations. Feel free to reach out by email.

Selected Research

FAST-ML: Hybrid Physics–Machine Learning for Tropical Cyclone Intensity Forecasting

Shijie Xiao, Jonathan Lin, Thomas Ehrmann, Ali Sarhadi

arXiv:2609.25505 · submitted to Journal of Advances in Modeling Earth Systems (JAMES)

Keeps the mechanistic intensity model in the loop and learns systematic corrections from multi-year, hourly, ~0.25° ERA5 fields; extended to probabilistic ensembles for extreme-event risk. Held-out 2024 North Atlantic: RMSE 15.42 → 10.78 kt (−30.1%), 10/12 storms improved.

[Paper] [Code]

Physics-Informed Machine Learning for Sub-Seasonal Prediction of Tropical Cyclone Genesis

Aren Russ, Shijie Xiao, Thomas Ehrmann, Ali Sarhadi

Journal of Geophysical Research – Machine Learning and Computation, under review

Physics-guided genesis prediction, localization, and interpretation over the North Atlantic: spatiotemporal transformers over nine physically motivated ERA5 channels, with SHAP / Integrated Gradients attribution and PCMCI causal analysis under a leakage-free temporal protocol.

[Code]

A Computational Framework for Multi-Agent Behavior Analysis: Integrating Geometric Vision Extraction and Social Attention Inference

Shijie Xiao

M.S. Thesis, BRAINML Lab, Georgia Tech · in preparation for ICML

Geometric vision extraction with social attention inference for multi-agent dynamical systems: higher-order graph interactions, structured representations, multimodal forecasting, and joint future prediction, with a novel geodesic loss on the trajectory manifold.

[Code]