cv

Graduate Research Assistant (co-first author)

Climate Risk & Extreme Dynamics Lab, Georgia Tech

Physics-Guided Tropical Cyclone Genesis Prediction and Localization (TCG) — co-first author, under review at JGR: Machine Learning and Computation.

• Spatiotemporal learning: developed a TimeSformer–U-Net framework combining factorized spatial–temporal attention with regional localization, using 14-day sequences of nine atmospheric and oceanic predictors.
• Temporal validation & model diagnostics: chronological training/calibration/validation/test partitions; achieved F1 = 0.722 versus 0.402 for Random Forest under a 48-hour precursor-matching criterion; used SHAP, Integrated Gradients, and climate-regime stratification to investigate predictive drivers and failure conditions.

Graduate Research Assistant, advised by Prof. Ali Sarhadi

Climate Risk & Extreme Dynamics Lab, Georgia Tech

Physics-based Generative AI for Probabilistic Compound Flood Forecasting — first author, in preparation for Nature Machine Intelligence.

• Sequence-to-field operator learning: developed a Transformer–FNO emulator mapping low-dimensional hurricane trajectories to 1024×1024 flood-depth fields of coupled surge–rainfall–riverine flooding.
• Simulation-based learning: trained on ~13,000 synthetic hurricane scenarios with coupled hydrodynamic simulations; compared 10 operator-learning and generative configurations (FNO, U-Net, diffusion, GAN, VAE), achieving 0.0771 m domain-wide RMSE on held-out flood fields.
• Uncertainty propagation & hazard forecasting: converted ECMWF ensemble trajectories into flood-depth distributions, inundation probabilities, and threshold-exceedance maps in seconds per member versus hours for physical simulations, with retrospective 1–7-day lead-time evaluation and SHAP-based storm-forcing sensitivity analysis.

Graduate Research Assistant, advised by Prof. Ali Sarhadi

Climate Risk & Extreme Dynamics Lab, Georgia Tech (sponsored by Sandia National Laboratories)

FAST-ML: Hybrid Physics–Machine Learning for Tropical Cyclone Intensity Forecasting — first author, submitted to JAMES (arXiv:2609.25505).

• Differentiable physics / UDE modeling: embedded dual-stream CNN parameterizations within the FAST thermodynamic ODE, combining physics-guided pretraining with end-to-end differentiation through the numerical solver.
• Large-scale meteorological data engineering: harmonized TB–PB-scale ERA5 reanalysis and GEFS/ECMWF forecast archives into storm-centered, hourly, 0.25° inputs aligned with IBTrACS observations.
• Forecast accuracy & extreme events: reduced mean per-storm RMSE by 30.1% on 12 held-out 2024 North Atlantic storms; reduced 60-hour ensemble CRPS by ~31% and rapid-intensification false-alarm ratio by 45%.
• AI benchmarking & generalization: storm-level advantages over Google DeepMind's FNV3 (Beryl ensemble-mean peak-intensity error 26 kt vs 49 kt, ERA5-driven); evaluated GEFS-driven forecasting, zero-shot Eastern Pacific transfer, ablations, and gradient-based physical attribution.
• Efficient ensemble computation: 1,000-member GEFS-driven ensembles; a 200-simulation benchmark completed in under three minutes on one 32-core/A100 node, with neural inference adding ~2% to per-member runtime.

Graduate Researcher (Master's Thesis), advised by Prof. Anqi Wu

BRAINML Lab, Georgia Tech

Structured Multi-Agent Dynamics and Probabilistic Trajectory Forecasting.

• Structured sequence modeling: developed a spatiotemporal graph model with separate intra-agent and inter-agent attention, representing body geometry and social interactions to predict joint motion.
• Probabilistic forecasting: implemented autoregressive Gaussian trajectory prediction with geometry-aware distance regularization and activity-dependent loss weighting, evaluated through multi-step rollouts.
• Comparative evaluation: matched comparisons against SRNN, EqMotion, PCHGCN, and ViTE, analyzing how keypoint representations affect displacement error and joint trajectory consistency.
• Ongoing method development: geometry-constrained latent dynamics and learned skeletal decoders; master's thesis completed in 2025, with ongoing extension targeting ICML 2027.

Research Assistant, advised by Prof. Yincong Chen

The Hong Kong University of Science and Technology (Guangzhou)

Diffusion-Based Visual Generation with Noisy Supervision.

• Investigated diffusion-based methods for 2D image segmentation under noisy supervision, using noisy-labeled and unlabeled images to improve learning from imperfect annotations.
• Developed data preprocessing, model training, and evaluation pipelines to study segmentation robustness under label noise and limited supervision.

Research Assistant, advised by Prof. Wu Yuan

The Chinese University of Hong Kong

Diffusion-Based Reconstruction of Noisy OCT Images.

• Developed a DDPM-based denoising approach for multiplicative speckle noise, combining noise modeling with learned image reconstruction.
• Built comparative evaluation pipelines against DnCNN, U-Net, BM3D, and TCFL-OCT using PSNR, SSIM, and GCMSE to assess reconstruction fidelity.

Undergraduate Research Assistant, advised by Prof. Xiang Bai

Huazhong University of Science and Technology

Musical Notation Detection and Recognition.

• Built and annotated a musical-notation dataset; implemented and evaluated DBNet, DBNet++, and PANet within MMOCR for detection and extraction of notation elements.

Undergraduate Research Assistant, advised by Prof. Yang Xiao

Huazhong University of Science and Technology

eSports Gaze Estimation — CNNs, Temporal Modeling, Few-Shot Learning.

• Developed a video-based gaze estimation system combining CNN features and temporal alignment.
• Proposed few-shot user adaptation to reduce calibration requirements and support real-time inference.

Doctor of Philosophy

Georgia Institute of Technology

Computer Science, School of Interactive Computing

• Advisors: Josiah Hester and Ali Sarhadi
• Expected Aug 2030

Master of Science

Georgia Institute of Technology

Electrical and Computer Engineering

• GPA: 4.0/4.0
• Thesis: A Computational Framework for Multi-Agent Behavior Analysis: Integrating Geometric Vision Extraction and Social Attention Inference
• Thesis Advisor: Anqi Wu

Bachelor of Science

Huazhong University of Science and Technology

Artificial Intelligence, School of Artificial Intelligence and Automation

• GPA: 3.95/4.0
• Thesis: Generative Diffusion Models for Segmentation with Noisy Labels
• Thesis Advisor: Yincong Chen

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

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

Shijie Xiao, Jonathan Lin, Thomas Ehrmann, and Ali Sarhadi.

Physics-Guided Machine Learning for Tropical Cyclone Genesis Prediction, Localization, and Interpretation in North Atlantic

Under review at Journal of Geophysical Research: Machine Learning and Computation

Aren Russ†, Shijie Xiao†, Thomas Ehrmann, and Ali Sarhadi. † These authors contributed equally.

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

M.S. Thesis, Georgia Institute of Technology

Shijie Xiao. Advisor: Anqi Wu.

Multi-Object Tracking, CVPR 2023 BDD100K Challenge

Detection-and-tracking pipelines for autonomous driving scenarios.

• Combined YOLO/DINO detectors with ByteTrack and BoT-SORT, addressing temporal association, occlusion, and identity continuity in driving videos.

EEG Sequence Classification, NC State Winter School

Motor-imagery classification on BCI Competition datasets.

• Developed EALNet, combining EEGNet feature extraction, LSTM temporal modeling, and Transformer attention for motor-imagery classification.