Autonomous ML research on CFD surrogates, powered by Claude Code agents coordinated through GitHub PRs.
We're training a neural network surrogate for computational fluid dynamics (CFD) on the TandemFoilSet dataset. The task: given tandem-airfoil geometry and flow conditions, predict the full velocity (Ux, Uy) and pressure (p) field at every mesh node. Traditional CFD solvers are accurate but slow — a learned surrogate trades a small accuracy loss for orders-of-magnitude speedup. The key metric is surface MAE (especially pressure on the airfoil surface), since that's what engineers use for design decisions.
The model is a Transolver with physics-aware attention over irregular meshes.
An advisor agent (no GPU) creates hypothesis PRs with detailed instructions and assigns them to student agents (GPU nodes). Students implement, run experiments, and report results on the PR. The advisor reviews: merge winners, iterate on promising ideas, close dead ends. Coordination uses GitHub labels (<advisor-name>, student:<name>, status:wip, status:review). W&B tracks metrics.
graph TD
subgraph K8s["Kubernetes Cluster"]
A["Advisor Pod<br/>(Claude Code, no GPU)<br/>Creates hypothesis PRs<br/>Reviews results, merges/closes"]
subgraph Students["Student Deployments (one per GPU node)"]
S1["frieren<br/>8x GPU"]
S2["fern<br/>8x GPU"]
S3["tanjiro<br/>8x GPU"]
S4["..."]
end
A -->|"GitHub PRs<br/>(draft → review → merge/close)"| Students
end
K8s --> GH["GitHub<br/>PRs = hypotheses<br/>Labels = routing"]
K8s --> WB["Weights & Biases<br/>Metrics, runs, groups"]
graph TD
A["Advisor creates draft PR"] -->|"student:name + status:wip"| B["Student picks up PR"]
B --> C["Implements hypothesis, runs experiments"]
C -->|"status:review"| D["Advisor reviews"]
D -->|Merge| E["Improvement lands on advisor branch"]
D -->|Request changes| F["status:wip — student iterates"]
D -->|Close| G["Dead end, branch deleted"]
F --> B
senpai/
├── train.py # Training script + Transolver model (students modify this)
├── program.md # Research context, metrics, constraints
├── data/ # Data preparation and benchmark splits
│ ├── prepare.py # Dataset loading and collation
│ ├── prepare_multi.py # Extended preprocessing (24-dim x, foil-2 features)
│ ├── utils.py # Visualization utilities
│ ├── split.py # One-time split manifest generator
│ ├── split_manifest.json # Committed train/val indices
│ └── split_stats.json # Committed normalization stats
├── instructions/ # Role-specific Claude Code instructions
│ ├── CLAUDE-ADVISOR.md # Advisor workflow
│ ├── CLAUDE-STUDENT.md # Student workflow
│ ├── prompt-advisor.md # Advisor prompt template
│ └── prompt-student.md # Student prompt template
├── k8s/ # Kubernetes deployment
│ ├── launch.py # Deploy advisor + student pods
│ ├── advisor-deployment.yaml # Advisor pod spec (CPU only)
│ ├── student-deployment.yaml # Student pod spec (8x GPU)
│ ├── entrypoint-advisor.sh # Advisor startup script
│ └── entrypoint-student.sh # Student startup script
└── .claude/skills/ # Claude Code skills
├── wandb-primary/ # W&B + Weave queries
└── list-experiments/ # Experiment history (advisor only)
# Train locally
python train.py --agent <name> --wandb_name "<name>/<description>"
# Debug (3 epochs, tiny subset)
python train.py --debug
# Deploy to k8s
python k8s/launch.py --tag <research-tag> --n_students 4 --advisorTandemFoilSet: Datasets for Flow Field Prediction of Tandem-Airfoil Through the Reuse of Single Airfoils is distributed by CC-BY-4.0.
@inproceedings{
lim2026tandemfoilset,
title={{TandemFoilSet}: Datasets for Flow Field Prediction of Tandem-Airfoil Through the Reuse of Single Airfoils},
author={Wei Xian Lim and Loh Sher En Jessica and Zenong Li and Thant Zin Oo and Wai Lee Chan and Adams Wai-Kin Kong},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=4Z0P4Nbosn}
}