A physics-corrected ML/DL framework for heat flux field prediction in porous structure of sintered Ag

Sintered Ag is a promising die-attach material for wide-bandgap power modules, but residual pores produce spatially non-uniform heat transport that is not captured by scalar effective thermal conductivity (ETC) alone. We develop a physics-corrected machine learning (ML)/deep learning (DL) framework for rapid prediction of vertical heat-flux fields from sintered-Ag cross-sections. A U-Net is trained on 691 of 867 real microstructures using a height-normalized target, q* = qH/ΔT, to remove the deterministic dependence of heat-flux magnitude on image height. An independent XGBoost model predicts ETC from normalized microstructural descriptors, and a row-wise correction enforces cross-sectional heat-flow conservation. The normalized U-Net retains high accuracy on 88 held-out real samples (field R² = 0.9898; ETC error = 1.20%). On 616 held-out aspect-ratio crops, normalization improves field R² from 0.0933 to 0.9852, reduces ETC error from 30.39% to 1.73%, and changes the geometry-dependent log-log slope from −0.9661 to −0.0077. On 320 unseen synthetic realizations, Raw / Normalized / + ML / + row-wise field R² values are −0.4291 / 0.7820 / 0.8117 / 0.8742, with ETC errors of 60.61 / 12.74 / 6.49 / 6.49%. The complete pipeline requires 176.4 ms per sample, approximately 130 times faster than finite element analysis (FEA). The resulting regime-level gating rule is explicit: use height normalization for geometry shift, apply ETC scaling only when the independent scalar predictor is more accurate than the ETC implied by the field, and use row-wise conservation to redistribute flux without altering the scalar ETC.