Artificial neural networks (ANN) for predicting the I-V characteristics of single photon avalanche diode (SPAD)

This paper presents an Artificial Neural Network (ANN) model for predicting the I-V characteristics of Single Photon Avalanche Diode (SPAD). As a data-driven model, ANN is a promising alternative to Technology Computer-Aided Design (TCAD) and can substantially reduce the computational cost. We investigate three ANN structures by tuning various hyperparameters and network configurations for I-V simulation of SPAD device. We train and evaluate each of the structures separately using a large amount of real-world I-V data generated through TCAD simulations and compare their performances. The optimized structure implemented with <i>LeakyReLU</i> activation function demonstrates smooth convergence through the learning period with a higher <i>R</i>² of 0.9975, outperforming the models employing <i>ReLU</i> and <i>Tanh</i> activations. Furthermore, the <i>LeakyReLU</i> activation function based model achieves a notably lower testing error in predicting the I–V characterization of SPAD device. Therefore, ANN model shows its potential as a robust and efficient alternative to traditional TCAD model for SPAD device simulation.