An accurate machine learning model to study the impact of realistic metal grain granularity on Nanosheet FETs
In this work, we present a machine learning neural network model to predict the impact of realistic metal grain granularity (MGG) variability on the threshold voltage and on the characteristics of a silicon-based 12 nm gate length nanosheet FET. This model is based on the multi-layer perceptron (MLP) machine learning architecture. As realistic MGG maps consist of the distribution of grains on the gate with different work-function values, it is relevant to apply algorithms such as the principal component analysis to reduce these features to the most representative ones. Once the realistic MGG features are correctly reduced without losing information, we train two different neural networks with the neurons in the output layer as the only difference, to predict the and the characteristics, respectively. The comparison between TCAD results and the model, shows excellent agreement for the mean and standard deviation of distributions for different average grain sizes values (from 3 nm to 10 nm) demonstrating the accuracy of the machine learning model. Also, we study the amount of data needed to accurately train the MLPs, leading to results that allow us to drastically reduce the computational time required to perform variability studies for state-of-art nano FET devices.
keywords: Machine learning, TCAD, nanosheet FET, metal grain granularity (MGG), Variability
Publication: Article
1693301452437
August 29, 2023
/research/publications/an-accurate-machine-learning-model-to-study-the-impact-of-realistic-metal-grain-granularity-on-nanosheet-fets
In this work, we present a machine learning neural network model to predict the impact of realistic metal grain granularity (MGG) variability on the threshold voltage and on the characteristics of a silicon-based 12 nm gate length nanosheet FET. This model is based on the multi-layer perceptron (MLP) machine learning architecture. As realistic MGG maps consist of the distribution of grains on the gate with different work-function values, it is relevant to apply algorithms such as the principal component analysis to reduce these features to the most representative ones. Once the realistic MGG features are correctly reduced without losing information, we train two different neural networks with the neurons in the output layer as the only difference, to predict the and the characteristics, respectively. The comparison between TCAD results and the model, shows excellent agreement for the mean and standard deviation of distributions for different average grain sizes values (from 3 nm to 10 nm) demonstrating the accuracy of the machine learning model. Also, we study the amount of data needed to accurately train the MLPs, leading to results that allow us to drastically reduce the computational time required to perform variability studies for state-of-art nano FET devices. - Julian G. Fernandez, Natalia Seoane, Enrique Comesaña, Juan C. Pichel, Antonio Garcia-Loureiro - 10.1016/j.sse.2023.108710
publications_en