摘要
High-fidelity thermo-mechanical simulations for wafer-level packaging (WLP) reliability prediction, especially full 3D models, are time-consuming and expensive, making it difficult to obtain large training datasets. To reduce the simulation burden under a small dataset regime, this study develops an artificial neural network (ANN) predictor with an adaptive sampling workflow that integrates leave-one-out cross-validation (LOOCV) hotspot analysis and Sobol sequence sampling. Boundary points are selected as the initial design set to strengthen information near design limits. LOOCV results are used to localize regions where prediction errors concentrate, and Sobol low-discrepancy sequences generate new samples within the hotspot bounds for efficient data enrichment. The workflow is validated using 2D WLP simulation results, and is intended to support data-efficient learning to alleviate future 3D simulation costs.