Integrated Sciences student at Claremont McKenna College. I build GPU physics simulations, and the checks that make their results trustworthy.
Summer 2026: NSF SCIPE REU scholar (Chishiki AI scholarship), GeoElements Lab, UT Austin, with simulation work on TACC's Vista (NVIDIA GH200) and Lonestar6 (NVIDIA A100) systems. PI: Dr. Krishna Kumar.
Water coloured by speed around a 1,100 kg Toyota Yaris hull, run g64_m1100 from the 17-run sweep.
Can a specific car cross a specific flooded road? I compared three answers of increasing cost: a depth rule of thumb, the published Australian Rainfall and Runoff (AR&R) vehicle hazard criterion, and a coupled material point method (MPM) simulation of water and a rigid vehicle hull on GPUs.
- 17 simulation runs with a complete provenance record. Every run records its code commit, solver version, mesh hash and grid settings: 10 of 10 fields present on all 17 runs, zero unresolved commits, zero mesh mismatches.
- Caught a rule being applied halfway. The depth x velocity product on its own is only part of the published AR&R rule, and an earlier version of this project's own code used it that way. Applying the full two-part rule for the car's class moved 23 of 70 flood scenarios to NO-FORD, and none the other way.
- 3D scene reconstruction. Trained a 1,147,694-Gaussian splat of a real drainage crossing with gsplat (30,000 iterations, PSNR 22.74).
- Open results. An interactive demo, published datasets with full data cards, and automated checks that run in GitHub Actions.
Code · Live demo · Findings · Scenario data · Load-surface data
I also contributed watertight-mesh particle seeding and content-based PLY loading to a fork of the lab's Warp-based MPM engine: jcerrell-IS/mpm-engine.
Python · NumPy · matplotlib · NVIDIA Warp (warpmpm) · gsplat · Slurm on TACC · Linux · Git and GitHub Actions · Gradio · Hugging Face Hub · Weights & Biases
LinkedIn · Hugging Face · jcerrell29@students.claremontmckenna.edu