Physics-Aware Graph Neural Network for FeFET Variability Analysis: Enabling 10 7 × Speedup, R 2 > 0.98 Accuracy, and Efficient Transfer Learning to New Architectures
Journal
Technical Digest - International Electron Devices Meeting, IEDM
Start Page
1
End Page
4
ISBN (of the container)
979-833156785-9
Date Issued
2025-12-06
Author(s)
Abstract
We propose and experimentally validate a Physics-aware Graph Neural Network (GNN) for FeFET variability analysis, addressing the limitations of traditional TCAD and DNN approaches. By representing the ferroelectric (FE) layer as a graph and incorporating multi-level physical features, our method achieves a 107× speedup compared to TCAD while maintaining high predictive accuracy (R2 > 0.98). We further demonstrate superior data efficiency and adaptability over conventional DNNs through transfer learning, successfully extending the model to an UTBSOI architecture with only 500 samples (R2 > 0.95) and, critically, from simulation to experimental data with 200 samples (Vth R2 > 0.985). This work presents a robust methodology that bridges the simulation-to-experiment gap, enabling rapid and reliable variability analysis for advanced FeFET technologies.
Event(s)
2025 IEEE International Electron Devices Meeting, IEDM 2025
Publisher
IEEE
Type
conference paper