Data and code for: Sample-efficient and stable reinforcement learning for active flow control via short-horizon solver gradients

Abstract

Trained policies, training histories and evaluation outputs behind every table and figure of the article Sample-efficient and stable reinforcement learning for active flow control via short-horizon solver gradients (J. Kang, I. Kim, S. Lee; submitted to Computer Physics Communications), together with a snapshot of the code (also at https://github.com/rider37/diffrl-flow-control, tag v1.0-cpc-submission). The study compares a differentiable-simulation policy gradient (short-horizon actor-critic, SHAC) with proximal policy optimization (PPO) in a differentiable incompressible Navier-Stokes solver written in JAX, on cylinder drag reduction (Re = 100-500) and Rayleigh-Benard convection control (Ra = 10^4-10^6), with five training seeds per configuration and evaluation on held-out initial conditions. SHAC is also compared with full-episode backpropagation through time and gradient-informed PPO (GI-PPO) in the same solver. Files cpc_data_v1.0.tar.gz: data (2.6 GB unpacked); extract in the root of the code repository. It contains runs/cylinder/ and runs/rayleigh_benard/ (history.json and final checkpoint of every run) and results/ (held-out evaluations, statistics, figure data, the numerical checks quoted in the article and the RTX 3090 training times of Figs. 5–6). diffrl-flow-control-v1.0.0.zip: code snapshot, identical to the GitHub release (MIT licence). MANIFEST.tsv: every file of the data archive with its size and SHA-256 checksum. README.md: layout of the data and naming of the run directories. The README of the code repository maps each table and figure to the script that produces it and to its input files. Data: CC BY 4.0. Code: MIT.