Top

Data Driven Mesh Adaptation Using an Explainable Res2Net SE Model

Authors

Vedat Akça
Faculty of Computer and Informatics Engineering, Istanbul Technical University, Istanbul, 34469, Turkey
Galip Kahraman
Faculty of Computer and Informatics Engineering, Istanbul Technical University, Istanbul, 34469, Turkey
Faruk İskender
Faculty of Computer and Informatics Engineering, Istanbul Technical University, Istanbul, 34469, Turkey

Files

Abstract

Adaptive mesh selection requires a local error indicator, a fixed resource rule and a separate assessment of the reconstructed field. This study evaluates a Res2Net-SE predictor with sensitivity-based diagnostics for one-round center insertion in generated scalar fields. Each candidate cell supplies a 7-by-7 patch with seven channels: coarse field values, two directional gradients, a Laplacian, two coordinates and mesh spacing. Hierarchical residual groups and squeeze-and-excitation gates form a local error estimate; a prediction-consistency term regularizes the fitted mapping. One hundred fields are divided by whole field into 60 training, 15 validation, 15 held-out and 10 transfer cases. The first three groups use 8, 10 or 12 cells per direction, whereas the transfer group uses 14 cells and narrower local structures. The top quarter of candidate cells receives reference-evaluated center samples, followed by Delaunay triangulation and piecewise-linear reconstruction. On held-out fields, the full model has local mean absolute error of 0.00503 and 70.96% overlap with the evaluated-error top- quarter ranking. Its adjusted mean relative reconstruction error is 4.08%, with a 31.22% mean per-field reduction from initial grids. Plain CNN and context-removal variants achieve lower local errors of 0.00468 and 0.00442 and slightly lower adjusted reconstruction errors of 4.04% and 4.03%, respectively. The ten-field combined transfer setting gives full-model local error 0.00421 and adjusted reconstruction error 3.22%, against 4.99% for its initial grids. Small-perturbation attribution cosine similarity is 0.999993, while seven-channel agreement with replacement responses is 0.9643. These diagnostics describe local sensitivity and ranking consistency under specified standardized interventions. The results support a traceable prediction-to-reconstruction workflow, while the tested simplified variants limit claims of an accuracy advantage for the complete architecture.

Details