Electromigration Analysis for Interconnects Using Improved Graph Convolutional Network with Edge Feature Aggregation.

Area of Science:

• Electrical Engineering
• Materials Science
• Computer Science

Background:

• Electromigration (EM) is a critical reliability concern in advanced integrated circuits.
• Traditional hydrostatic stress analysis using partial differential equations (PDEs) is computationally intensive and impractical for full-chip analysis.
• Accurate hydrostatic stress prediction is essential for understanding and mitigating EM failures.

Purpose of the Study:

• To develop a computationally efficient and accurate method for predicting hydrostatic stress in circuit interconnects.
• To leverage graph neural networks (GNNs) for modeling interconnect structures and their associated stress distributions.
• To improve the speed and accuracy of electromigration analysis in integrated circuits.

Main Methods:

• Conceptualized circuit interconnect trees as graphs within a GNN framework.
• Generated ground truth hydrostatic stress data using finite element solution software.
• Developed and trained an improved Graph Convolutional Network (GCN) with edge feature aggregation and an attention mechanism.

Main Results:

• The proposed GCN model achieved a 15% improvement in Root Mean Square Error (RMSE) compared to the original GCN.
• The model demonstrated a significant increase in solution speed compared to traditional finite element methods.
• Successfully predicted hydrostatic stress values for nodes within interconnect trees.

Conclusions:

• The GNN-based approach offers a highly efficient and accurate alternative for hydrostatic stress analysis in integrated circuits.
• This method significantly accelerates electromigration reliability assessments, enabling practical full-chip analysis.
• The developed model contributes to enhancing the long-term reliability of advanced semiconductor devices.