Zeyu Xia

I build models that learn physical dynamics, generate scientific data, and solve inverse problems. My work emphasizes computational efficiency and evaluation beyond training loss.

• Efficient generative modeling: Cosmo3DFlow combines flow matching and 3D wavelet representations for up to 46× faster sampling than diffusion models on 128³ cosmological simulations (KDD 2026). ScatterPrism studies conditional generation and multi-metric evaluation for particle physics.
• Agentic benchmarking for scientific applications: In DeepVELA (Genesis Mission), I collaborate with MLCommons to explore agentic benchmarks, evaluating time, energy, space, cost, and precision. I design agent harnesses with structured outputs and verifiable results to support trustworthy evaluation.
• Exploratory world models: I have prototyped and trained JEPA models for spatiotemporal scientific data, investigating self-supervised representations for potential data assimilation and sensor placement applications.
• GPU-accelerated simulation and data pipelines: PyTorchFire enables differentiable wildfire simulation and gradient-based calibration. FireDataForge integrates 11 geospatial data sources for reproducible scientific workflows.
• Computer vision: DT-SegNet combines YOLOv5 object detection with SegFormer semantic segmentation for microscopy.

Tools: Python, C/C++, PyTorch, JAX, TensorFlow, Scikit-learn, SLURM, Docker, Git, and Linux. View research experience and skills.