Abstract
The electromigration (EM)-induced reliability issues in very large scale integration (VLSI) circuits have attracted increased attention due to the continuous technology scaling. Traditional EM models often lead to overly pessimistic prediction incompatible with the shrinking design margin in future technology nodes. Motivated by the latest success of neural networks in solving differential equations in physical problems, we propose a novel mesh-free model to compute EM-induced stress evolution in VLSI circuits. The model utilizes a specifically crafted space-time physics-informed neural network (STPINN) as the solver for EM analysis. By coupling the physics-based EM analysis with dynamic temperature incorporating Joule heating and via effect, we can observe stress evolution along multi-segment interconnect trees under constant, time-dependent and space-time-dependent temperature during the void nucleation phase. The proposed STPINN method obviates the time discretization and meshing required in conventional numerical stress evolution analysis and offers significant computational savings. Numerical comparison with competing schemes demonstrates a 2x ~ 52x speedup with a satisfactory accuracy.
Introduction
As semiconductor technology continues to scale into the nanometer regime, the current densities carried by metal interconnects have reached unprecedented levels. This intensification exacerbates electromigration (EM), a physical phenomenon where the momentum transfer from flowing electrons causes the gradual displacement of metal atoms. This atomic migration results in the accumulation of mechanical stress within the wire: tensile stress at the cathode can lead to void formation, while compressive stress at the anode can cause hillocks. Both outcomes pose significant threats to the long-term reliability of integrated circuits.
Accurately predicting stress evolution is critical for modern VLSI design. However, traditional empirical models, such as Black’s Equation, are often overly conservative, leading to inefficient designs that waste valuable area and power. Conversely, high-fidelity numerical methods like the Finite Element Method (FEM) are computationally prohibitive for large-scale interconnect trees. FEM requires complex spatial meshing and temporal discretization, where the computational cost scales poorly with the geometric complexity of modern multi-segment wire structures.
The STPINN Framework
To overcome these limitations, we introduce the Space-Time Physics-Informed Neural Network (STPINN). Unlike conventional deep learning models that rely solely on large datasets, STPINN integrates the underlying physical laws—specifically Korhonen’s Equation for stress diffusion—directly into the neural network's training process.
The STPINN architecture is mesh-free, utilizing automatic differentiation to compute the necessary spatial and temporal derivatives. This approach eliminates the need for rigid grids, allowing for a continuous representation of stress across the space-time domain. The objective function is constructed as a multi-term loss that includes the residuals of the governing partial differential equations (PDEs), initial conditions, and boundary conditions at the wire terminals and junctions.
Technical Innovations for Convergence
A significant challenge in applying neural networks to EM analysis is the extreme disparity in the magnitudes of physical constants. For instance, values related to Boltzmann’s constant and current density can differ by over 30 orders of magnitude, leading to vanishing or exploding gradients that prevent model convergence.
To address this, we implemented a linear preprocessing and scaling scheme. By transforming the input space and normalizing the physical parameters, we stabilized the training process. As shown in the gradient analysis, this preprocessing ensures that the gradient magnitudes remain within a range conducive to optimization, allowing the network to learn the complex stress evolution accurately.
Furthermore, we addressed the non-differentiability of stress at wire junctions through a "virtual distance" technique. By introducing a small virtual separation at intersections, we smoothed the stress gradients, enabling the neural network to effectively model multi-segment interconnect trees. The architecture also employs a multi-channel design to handle non-constant diffusivity, which occurs when temperature varies across the wire over time.
Results and Performance
The STPINN model was rigorously evaluated against COMSOL (a standard FEM solver) and analytical benchmarks across various geometries, ranging from single segments to complex 19-segment trees and cross-shaped terminals.
The results demonstrate that STPINN achieves a remarkable speedup of 2x to 52x compared to traditional FEM methods. This efficiency gain is particularly pronounced in complex topologies where FEM's meshing requirements become a bottleneck. Despite the significant reduction in computation time, accuracy remains high, with a mean relative error of less than 1.22% compared to FEM and less than 0.44% compared to analytical models.
A key finding of our study is the impact of dynamic temperature. By incorporating Joule heating and the heat-sinking "via effect," STPINN reveals that fluctuating temperatures significantly accelerate the rate of stress buildup compared to steady-state assumptions. This highlights the importance of physics-informed models in capturing real-world operating conditions.
Conclusion
STPINN represents a significant advancement in EM reliability analysis, offering a fast, accurate, and mesh-free alternative to traditional numerical solvers. By embedding physical constraints into a neural network, we provide a tool that can keep pace with the shrinking design margins of future technology nodes. Future work will focus on scaling this approach to full-chip power grids and incorporating the stochastic nature of metal grain structures to further enhance predictive capabilities.
Summary updated