Application of Residual Neural Network Towards Semiconductor Device Modeling

In this paper, we explore the application of Artificial Intelligence (AI) in device modeling, focusing on the extraction of electrical characteristic parameters from current-voltage (I-V) curves of Silicon Carbide (SiC) MOSFETs. We trained neural networks with 500 I-V curve images captured from these devices. To the best of our knowledge, this is the first instance where the AI engine has been used to predict threshold voltages from I-V curves of device datasheets. Our study utilized a SiC MOSFET structure in Technology Computer-Aided Design (TCAD) software to generate I-V curves under various mobility and work function parameter variations, resulting in a dataset of 500 different devices’ I-V characteristics. Firstly, we modified the input and output layers of existing ResNet18 model to meet our application. Secondly, we developed a custom six-layer Convolutional Neural Network (CNN) for this application, and found improved accuracy and speed. This demonstrates the potential for further improvement and application of ResNet and CNNs in device modeling using images available in the datasheets. Our custom CNN achieved a MSE of 0.05 and an average training time of 22 seconds per epoch, demonstrating significant improvements over the modified ResNet18.