Part I: CFD Hydrogen Dispersion Scenarios for Source Localization Benchmarking: 180 GASFLOW-MPI Simulations in a Confined Parking Facility

Description This dataset contains 180 computational fluid dynamics (CFD) simulation scenarios for hydrogen leak dispersion in a confined underground parking facility, generated using the validated GASFLOW-MPI solver. It serves as the ground-truth experimental database for evaluating graph-structured source localization methods, as described in the accompanying ECML-PKDD 2026 paper "Understanding Structure Transfer in Inverse Problems: When Does Pre-Learned Structure Help Source Localization?" Facility and Physics The simulated facility is a 50 m × 30 m × 3 m underground parking garage with two longitudinal and two transverse traffic lanes, structural columns on a 6 m × 6 m grid, and ceiling-mounted jet fans. Hydrogen dispersion is governed by the three-dimensional time-dependent compressible Navier–Stokes equations with k-ε turbulence modelling, buoyancy via the Boussinesq approximation, and species transport for hydrogen mass fraction. A medium-resolution orthogonal structured mesh of 288,000 cells (200 × 120 × 12) was selected based on a mesh independence study, yielding less than 3.6% deviation from the fine-mesh reference while reducing wall-clock time by approximately 9×. Scenario Matrix The 180 scenarios are a full factorial combination of: 12 leak positions (4 corner, 4 wall-adjacent, 4 mid-row; all at z = 0.5 m fuel-tank height) 5 leak rates: 1, 30, 50, 100, 150 g/s 3 ventilation rates: ACH = 3, 6, 10 h⁻¹ Each simulation begins with a 10-minute quiescent ventilation pre-run at ACH = 3, after which the leak is activated and the simulation continues for 60 seconds. Concentration fields are saved at 1-second intervals, yielding 60 time-step snapshots per scenario. Data Format Spatial fields are provided at both the native CFD resolution (288K cells) and spatially downsampled to a 50 × 30 × 10 grid (factor-of-4 reduction in x and y via bilinear interpolation) suitable for direct use as localizer input. Sensor readings at arbitrary positions can be extracted by trilinear interpolation of the 3D concentration field. Intended Use The dataset is designed to benchmark structure learning and source localization methods for gas leak detection systems. The scenario matrix supports in-distribution evaluation (random 80/10/10 splits stratified over leak rate and ventilation rate), out-of-distribution generalization studies (train on ACH ∈ {6, 10}, test on ACH = 3), and data-scaling experiments (subsets of m ∈ {10, 25, 50, 100, 150, 180} scenarios).