Problem Statement
RL-CFD studies require robust tooling for experiment control, monitoring, and reproducibility.
PROJECT DETAIL · 2024
TorchRL/Stable-Baselines3 + PyFR Coupled Control Framework
Research software framework for parallelized RL-CFD experiments with reproducible training, rollout, and evaluation workflows.
- Current Focus
- DRL
- CFD
- HPC
- Research Tools
Why It Matters
A reusable framework reduces setup overhead and improves research velocity for active flow control studies.
Methods / Numerical + ML Setup
- Implemented coupled RL-CFD execution loops with configurable experiment parameters.
- Added workflow patterns for scaling training jobs on cluster resources.
Key Results
- Enabled faster iteration on control policy experiments.
- Improved reproducibility of training/evaluation runs.
Toolchain
- TorchRL
- Stable-Baselines3
- PyFR
- Python
- HPC job orchestration
Challenges and Lessons
- Synchronizing solver timesteps with RL rollout cadence.
- Managing long-running training stability across distributed jobs.
Future Work
- Add richer diagnostics and standardized benchmark scripts.
- Integrate stronger experiment tracking for comparative studies.