5 January 2022 in Composites Science and Technology
10 August 2026 in Journal of Alloys and Compounds
12 May 2026 in Thin-Walled Structures
14 July 2026 in SSRN Electronic Journal
Sintered Ag is a leading die-attach material for wide-bandgap power modules, but residual pores induce pronounced spatial non-uniformity in the heat flux field, critical for assessing local hotspots and reliability. Finite element analysis (FEA) provides high-fidelity fields but is computationally prohibitive at scale, while deep learning (DL) surrogates suffer from severe magnitude miscalibration under distribution shift. We propose a physics-corrected hybrid machine learning (ML)/DL framework predicting the vertical heat flux field and effective thermal conductivity (ETC) of sintered Ag. A U-Net trained on 867 real cross-sections predicts the field; an independent XGBoost model predicts the scalar ETC, supplying a global magnitude-scaling correction; and a row-wise energy-conservation constraint is applied as post-processing. Raw DL predictions collapse to negative R² under out-of-distribution shift (aspect-ratio crops and synthetic microstructures), but the corrections recover R² to 0.979 and 0.880 without altering the integrated ETC. The pipeline runs ~130× faster than FEA.