二极管
人工神经网络
计算机科学
电子工程
雪崩击穿
图形
电气工程
拓扑(电路)
单光子雪崩二极管
物理
逻辑门
齐纳二极管
PIN二极管
晶体管
电压
工程类
计算机模拟
作者
Zhixiang Cao,Zixuan Wang,Yuxuan Yang,Yujie Liu,Haoxiang Li,Xiangliang Jin
标识
DOI:10.1109/ted.2026.3697279
摘要
Single-photon avalanche diodes (SPADs) are essential for modern optoelectronic detection. However, the pronounced nonlinearity of avalanche multiplication and electric-field singularities at doping-gradient interfaces poses major challenges for conventional technology computer-aided design (TCAD) simulations, leading to convergence issues and high computational costs that limit SPAD design optimization efficiency. To address these challenges, this article proposes a multilevel residual-connected graph neural network (MLRC-GNN) for fast and accurate SPAD simulation. This model represents nonuniform simulation meshes as graphs and employs residual connections in a deep architecture to effectively capture complex multiphysics couplings within SPADs. A dataset of 8250 SPAD structures is generated using calibrated TCAD simulations for model training and validation. The results show that the MLRC-GNN achieves up to a$480\times $speedup and a 10% improvement in convergence probability compared with TCAD. The model also attains high predictive accuracy (R${}^{{2}} \gt 0.9939$) and reduces the mean squared error (mse) by approximately 33% compared with standard graph convolutional networks (GCNs) while maintaining strong generalization across diverse device geometries and bias conditions. This approach accelerates simulations and enables rapid identification of key physical fields governing photon detection probability (PDP) and dark count rate (DCR), providing an efficient, scalable framework for SPAD design and optimization.
科研通智能强力驱动
Strongly Powered by AbleSci AI