ADR2D-Hybrid: Meshless Numerical-Machine Learning for Two-Dimensional Reactive Contaminant Transport

ADR2D-Hybrid is a reproducible scientific-machine-learning release for two-dimensional reactive contaminant transport. A meshless generalized finite difference solver provides a moderate-resolution trajectory, and a pointwise neural network predicts its defect relative to an accepted refined numerical reference. The selected hybrid checkpoint was evaluated once on 18 held-out controlled-pulse scenarios and reduced median peak-normalized RMSE to 0.2576 of the numerical baseline while preserving explicit boundary and non-negativity constraints. The release includes protocols, numerical corpora, model checkpoints, inference and validation interfaces, figures, and a transparent irregular-geometry audit whose perforated-domain reference was rejected rather than used to overstate transfer.