Application of Noise to Avoid Overfitting in TCAD Augmented Machine Learning

Abstract

In this paper, we propose and study the use of noise to avoid the overfitting issue in Technology Computer-Aided Design-augmented machine learning (TCAD-ML). TCAD-ML uses TCAD to generate sufficient data to train ML models for defect detection and reverse engineering by taking electrical characteristics such as Current-Voltage, IV, and Capacitance-Voltage, CV, curves as inputs. For example, the model can be used to deduce the epitaxial thicknesses of a p-i-n diode or the ambient temperature of a Schottky diode being measured, based on a givenIV curve. The models developed by TCAD-ML usually have overfitting issues when it is applied to experimental IV curves or IV curves generated with different TCAD setup. To avoid this issue, white Gaussian noise is added to the TCAD generated curves before ML. We show that by choosing the noise level properly, overfitting can be avoided. This is demonstrated successfully by using the TCAD-ML model to predict 1) the epitaxial thicknesses of a set of TCAD silicon diode IV’s generated with different settings (extra doping variations) than the settings in the training data and 2) the ambient temperature of experimental IV’s of Ga2O3 Schottky diode. Moreover, domain expertise is not required in the ML process.