Wenzhuo Xu

Wenzhuo Xu

Ph.D. candidate at Carnegie Mellon University. Research Scientist Intern at Adobe Research.

Wenzhuo Xu

I build agentic and vision–language systems that reason about physical processes — how they behave, and whether a generated video or the equations printed in a document actually respect them. I am a Ph.D. candidate in Mechanical Engineering at Carnegie Mellon University (expected May 2027) and currently a Research Scientist Intern at Adobe Research. My background is in learned surrogates and neural operators for large-scale fluid simulation, which is where the physics grounding comes from: I spent four years building models that had to agree with a solver, and I now build systems that have to agree with the world.

The thread through both halves is the same — a model’s claim about a physical system should be checkable, and the machinery that checks it should be something a person can read.

Adobe Research — Research Scientist Intern, with Tong Sun and Jiuxiang Gu

Carnegie Mellon University — advised by Christopher McComb and Noelia Grande Gutiérrez, Department of Mechanical Engineering, CMU

Labs — Design Research Collective · BioSiMMlab

Email / CV / Google Scholar / GitHub / LinkedIn

Projects

Adobe Research · 2026 · report in preparation Agentic physical reasoning for video evaluation Video generators are scored with one opaque number. This turns the score into an auditable verdict you can trace back to a measurement. Adobe Research · 2026 · ongoing Executable physics from static documents The physics printed in a textbook is inert. This reads the equations and the prose around them and compiles a simulation you can actually run. Three renders of the same diseased coronary artery tree coloured by time-averaged wall shear stress — the low-resolution input almost uniformly dark, the ML prediction and the high-resolution CFD reference nearly identical. CMU · BioSiMMlab · 2025 Super-resolution for cardiovascular blood flow A coarse blood-flow simulation gets wall shear stress wrong — the marker that flags a dangerous plaque. A learned correction recovers it in 20 minutes instead of 12 hours. 35×faster than high-fidelity CFD0.92–0.94wall shear stress correlation vs. the reference solver Velocity-magnitude renders of a mixing elbow: coarse input, model prediction and high-resolution reference, nearly indistinguishable. CMU · with Eaton Research Labs and ERDC · 2023 – 2026 Super-resolution for engineering duct design High-fidelity CFD answers the next morning, so engineers stop exploring and start confirming. This gets the same answer inside the design loop. 5.5×faster than the high-fidelity simulationDeployedin Eaton's jet engine duct design workflow A transitional boundary layer predicted by the model, in yellow and purple, above its near-zero error field. CMU · Manufacturing Futures Institute · 2024 Adaptive local domain decomposition One network trained on a whole flow ends up mediocre everywhere in it. Cutting the domain into patches and routing each to a specialist model fixes that. 0.66 → 0.80+R² after 20 rollout steps, undecomposed vs. decomposed3physics regimes routed to their own sub-model A cylinder wake velocity field on a coarse triangular mesh, in rainbow colours, with the mesh edges visible. CMU · Design Research Collective · 2022 – 2025 TEECNet and MegaFlow2D Instead of learning the expensive simulation's answer, learn how wrong the cheap one is — then subtract. 30train/test resolution pairs, one architecture3.2×faster than the neural-operator baseline on 12 cores

Recent publications

All publications and talks →

Education

  • Ph.D., Mechanical Engineering — Carnegie Mellon University, expected May 2027. Advised by Christopher McComb and Noelia Grande Gutiérrez.
  • M.S., Mechanical Engineering — Carnegie Mellon University, August 2024. Earned en route to the Ph.D.
  • B.Eng., Mechanical Engineering & B.A., German — Shanghai Jiao Tong University, June 2022. Dual-degree program.