Fast TDDB Hotspot Detection of Large-Scale BEOL Circuits Using Physics-Informed Central Emphasized CNN

Time-Dependent Dielectric Breakdown (TDDB) has emerged as a critical reliability concern in Back-End-of-Line (BEOL) interconnects. In Electric Path Generation (EPG) based TDDB analysis, the evolution of metal ion concentration is governed by a Partial Differential Equation (PDE). Traditional numerical methods for solving this equation are time-consuming and demand considerable computational resources, making them impractical for large-scale circuits analysis. To address this challenge, we propose a physics-informed learning framework for fast prediction of ion concentration distributions in large BEOL layouts. The framework represents BEOL patterns as two-dimensional images, partitions large layouts into overlapping sub-graphs for encoding, and predicts transient concentration fields for each sub-graph. Based on the reconstructed full-layout concentration map, the method supports efficient TDDB hotspot analysis. The proposed Physics-Informed Central Emphasized Convolutional Neural Network (PICE-CNN) is applicable to diverse interconnect patterns and operating conditions, including different line widths, line spacings, and supply voltages. Experimental results show that the proposed method achieves up to 357.1 × speedup over conventional numerical solvers while maintaining high prediction accuracy, with the mean squared error below 1 × 10 − 4 in the emphasized central region.