A Mode-Aware Hybrid Machine-Learning Framework for Full-Field Warpage Prediction of Fan-Out Panel-Level Packaging After Debonding

Fan-Out Panel-Level Packaging (FO-PLP) enables high area utilization and manufacturing efficiency, but process-induced warpage caused by the coefficient of thermal expansion (CTE) mismatch and polymer shrinkage remains a major challenge. This study presents a classifier-gated hybrid machine-learning framework for the rapid and accurate FO-PLP warpage prediction using a database generated from a validated three-dimensional finite element process model. A Random Forest classifier first estimates the probability of each global warpage mode, while cluster analysis reduces the spatial training dataset. Two mode-specific artificial neural networks are then combined through probability-weighted fusion to predict the full warpage field and enable warpage prediction for previously unseen geometry layouts. The framework was evaluated on 16 independent finite element designs spanning both warpage modes. Compared with an equivalent single-network model, the proposed approach consistently achieved lower mean and maximum prediction errors across all designs, with the greatest improvements at the panel edges and corners where the prediction is most challenging. In addition, the clustering strategy reduced the training-set size and computational cost. These results demonstrate that integrating warpage-mode classification with mode-specific learning improves both the prediction accuracy and training efficiency, providing a practical tool for the fast warpage assessment of new FO-PLP layout designs.