Analytic Poisson and planar Poiseuille benchmark data for vascular-flow surrogate evaluation

Abstract

Benchmark data, trained models and code for two flow problems with exact analytic solutions: a two-dimensional Poisson problem (Level I) and planar Poiseuille flow in a straight two-dimensional channel (Level II). The data compare supervised deep neural networks (DNNs), physics-informed neural networks (PINNs) and hybrid data-physics networks on field accuracy, physical consistency and computational cost. These are Levels I and II of the benchmark reported in "When Accuracy, Conservation, and Cost Disagree: A Single-Geometry Benchmark of Vascular-Flow Surrogates" (Dzonzi, Mishra and Ngoepe).LEVEL I: POISSON PROBLEMResults table (CSV) for DNN and PINN models trained with 10, 50, 100, 200 and 500 labelled points, three random seeds each. Variables: relative L2 error and mean squared error on the training domain and on an enlarged domain, and training time.Provenance record (JSON): training protocol, hardware and software versions.Source code and batch scripts that reproduce the table.LEVEL II: PLANAR POISEUILLE FLOWNine runs (DNN, PINN and hybrid; three seeds each) under the scaled training setup. Each run contains its training history (CSV), metrics and configuration (JSON) and trained model checkpoint (PyTorch .pt).Per-seed and summary tables (CSV) for the scaled and unscaled training setups.Variables: velocity RMSE and relative L2 error, pressure RMSE and relative L2 error (raw and gauge-aligned), continuity and momentum residuals, inlet and outlet flux, flux imbalance, cross-section flux variation, wall-velocity error, pressure drop and pressure-drop error, inference latency and training time.Training, Slurm sweep and aggregation scripts, with a regression test. SUPPORTING FILESShared neural-network library code (pinn_library, pipe_flow_2d).Software environment record (Python, PyTorch and pip package versions).README.md with instructions for rerunning both levels.MANIFEST.json listing the size and SHA-256 checksum of every file.USAGEUnzip the archive and run the scripts from its top-level folder; no paths need editing. Re-aggregating the Level II runs reproduces the per-seed table and every summary mean. Standard deviations may differ in the final digit between machines because floating-point sums are accumulated in a different order. Training times depend on hardware; the reported times were measured on an NVIDIA L40S GPU.EXCLUSIONSThe dataset contains no patient data. The patient-specific aortic geometry, its computational fluid dynamics reference simulation and all results derived from it (Level III of the benchmark) are excluded; that reference was used with permission and is not redistributed.