Publications

Article Preview

With scaling, physics-based analytical MOSFET compact models are becoming more complex. Parameter extraction based on measured or simulated data consumes a significant time in the compact model generation process. To tackle this problem, ANN-based approaches have shown promising performance improvements in terms of accuracy and speed. However, most previous studies used a multilayer perceptron (MLP) architecture which commonly requires a large number of parameters and train data to guarantee accuracy. In this article, we present a Mixture-of-Experts approach to neural compact modeling. It is 78.43% more parameter-efficient and achieves higher accuracy using fewer data when compared to a conventional neural compact modeling approach. It also uses 43.8% less time to train, thus, demonstrating its computational efficiency.

Article Preview

This work presents a novel approach of using knowledge transfer to increase the accuracy of artificial neural network (ANN)-based device compact models, or neural compact models. This is useful when the amount of data available for training an ANN is limited. By utilizing relatively abundant data of a previous technology node, physical phenomena that are not evident in the limited data of the target technology node (e.g. gate-induced drain leakage) are accurately predicted. When meta learning algorithms are used, the accuracy of the model significantly increases, with relative linear error 10 times lower compared to the case when prior knowledge is not incorporated. The proposed methodology can be used to model future generation devices with limited data, utilizing data from well-characterized past technology node devices.