Fast Electromigration Analysis of Multi-segment Interconnects Based on Recurrent Neural Networks

Predicting electromigration (EM) stress is a critical reliability challenge in power delivery networks. Existing semi-analytical Multilayer Perceptron (MLP) models neglect the Markovian sequential dependencies of EM diffusion and suffer from severe scalability issues due to global topological evaluations. To address these architectural and computational bottlenecks, this paper proposes a dynamic spatiotemporal stress forecasting framework. Specifically, a physics-aligned Recurrent Neural Network (RNN) is constructed to directly model the Markovian sequential evolution of EM stress. This is integrated with a tensor mapping-based dynamic local subgraph extraction strategy, effectively isolating spatial complexity from the global interconnect scale. At the numerical integration level, we develop a physics-driven adaptive hybrid convolution algorithm. Regulated by a stress relaxation time constant, this algorithm dynamically invokes discrete convolution during the early acceleration phase to strictly preserve the abrupt variations of the auxiliary function used for stress computation, and seamlessly switches to Gauss-Legendre quadrature during stable periods to maximize precision while eliminating smoothing artifacts. Furthermore, a simulation-time-based optimization strategy is incorporated to mitigate the prediction error gap between early and late simulation stages. Experimental results validate the efficiency of the proposed framework. Compared to the MLP baseline, our model achieves up to a 93.2% reduction in relative error and a 117.79× computational speedup. Against the Finite Element Method (FEM), it delivers up to 1335.6× acceleration while maintaining engineering-grade accuracy.