Haozhe Zhang

Modeling Engineer at Tesla · Physics-based ML · LLMs

Palo Alto, CA

USA

I trained in theory and computation during my PhD at the University of Virginia, working on closed-form models of multi-physics fields.

At Tesla I model battery cells and other engineering systems, and apply reinforcement learning to automate and scale them.

My current interest is AI that reasons with real physics and engineering, from LLMs and agents to AI for hardware.

Open to conversations on physics-based ML and AI for engineering.

news

May 11, 2026	Posted BenchCAD on arXiv: arXiv:2605.10865.
Mar 24, 2026	Accepted CADLoop, a CAD data-curation method, to the CVPR 2026 NeXD Workshop.
Aug 15, 2025	Published a buckling-mechanics model for lattice superstructures in Mechanics of Materials (Vol. 202, art. 105253).

featured

arXiv

BenchCAD: A Comprehensive, Industry-Standard Benchmark for Programmatic CAD

Haozhe Zhang†, Kaichen Liu, Miaomiao Chen, Lei Li, Shaojie Yang, Cheng Peng, and Hanjie Chen†

arXiv preprint arXiv:2605.10865 2026

† Corresponding authors

★ Reported in Anthropic system cards and OpenAI’s GPT-5.6 launch

Abs URL Project

We present BenchCAD, a comprehensive industry-standard benchmark for evaluating programmatic CAD code generation. The benchmark includes over 17,000 execution-verified CadQuery programs spanning 106 industrial part families (e.g., gears, springs, drills) and assesses models across visual question answering, code analysis, image-to-code conversion, and code editing. Our results show that current models often recover coarse outer geometry but fail to produce faithful parametric CAD programs, with common failure modes including missing structural details and oversimplified operations; generalization to unseen part families remains limited even after fine-tuning.