DL-AMR: Heteroscedastic Residual Learning for Dynamic Adaptive Mesh Refinement in Bluff-Body Wake Flows

Abstract

GitHub-Zenodo software record (repository snapshot) accompanying the paper of the same name (Journal of Computational Physics, 2026). Contents: source code for the heteroscedastic U-Net (training, evaluation, TorchScript export), the OpenFOAM custom solver amrPimpleFoam with the in-plane refinement library hexRef4, simulation case templates for all variants of the paper (including the score-ablation, wrapper and residual-estimator cases), training configurations, the post-processing pipeline and figure scripts, and documentation. Three binary artifacts are deposited alongside the repository snapshot: pretrained_models.tar.gz (the TorchScript model and weights used in the paper, with normalisation statistics and the vorticity wrapper of the score ablation), reference_data_minimal.tar.gz (the held-out test split and per-sample predictions that reproduce the uncertainty-validation figures, Figs. 9, 10 and E.1) and reference_data_fields.tar.gz (cached time-averaged fields, probe series and tables that reproduce Figs. 4, 5, 8, D.1 and F.1). Figures 3, 6 and 7 and the tables require the raw OpenFOAM case results, which the case templates regenerate and which are available from the corresponding author on request. This repository uses a multi-license structure: Python code is MIT-licensed, OpenFOAM-derived solver code is GPLv3-licensed, and model/data/documentation artifacts are CC-BY 4.0. See LICENSE for component-specific terms.