Enhancing Predictive Performance and Efficiency in CMOS with Optimized CNN Hyperparameters

Data-driven machine learning is valuable for analyzing transistor devices, but deep learning often requires large datasets and computational resources, limiting practical use. Traditional models applied to inline data tend to overfit due to high-dimensional parameters and data discontinuities. We propose a machine learning model that combines fully connected neural networks (DNN) and convolutional neural networks (CNN) to predict the parameters of 55nm 7V CMOS devices. This approach efficiently tunes hyperparameters, minimizes overfitting at lower computational cost, and achieves significant improvements in predictive performance.