Improvement of TCAD Augmented Machine Learning using Autoencoder for Semiconductor Variation Identification and Inverse Design

• Published in: IEEE access 2020-01, Vol.8, p.1-1
• Main Authors: Mehta, Kashyap, Raju, Sophia Susan, Xiao, Ming, Wang, Boyan, Zhang, Yuhao, Wong, Hiu Yung
• Format: Article
• Language: English
• Subjects: Autoencoder; CAD; Computational modeling; Computer aided design; Data models; Defect Identification; Doping; Electrical measurement; Inverse Design; Linear regression; Machine Learning; Model accuracy; Physical properties; PIN diodes; Predictive models; Regression models; Semiconductor device modeling; Semiconductor process modeling; Simulation; TCAD Simulation; Thickness; Variability
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2020.3014470

A machine learning (ML) model by combing two autoencoders and one linear regression model is proposed to avoid overfitting and to improve the accuracy of Technology Computer-Aided Design (TCAD)-augmented ML for semiconductor structural variation identification and inverse design, without using domain expertise. TCAD-augmented ML utilizes TCAD simulations to generate sufficient data for ML model development when experimental data are inadequate. The ML model can then be used to identify semiconductor structural variation for given experimental electrical measurements. In this study, the variation of layer thicknesses in the p-i-n diode is used as a demonstration. An ML model is developed to predict the diode layer thicknesses based on a given Current-Voltage (IV) curve. Although the variations of interest can be incorporated easily in TCAD simulations to generate ML training data, the TCAD-augmented ML model generally is overfitted and cannot predict the variations in experiment well due to hidden variables which also alters the IV curves. We show that by using an autoencoder, this problem can be solved. To verify the effectiveness, another set of TCAD simulation data is generated with hidden variables (dopant concentration variation) to emulate experimental data. Testing on the second set of data shows that the proposed model can avoid overfitting and has up to 15 times improvement in accuracy in thickness prediction. Moreover, this model is used successfully to perform inverse design and can capture an underlying physics that cannot be described by a simple physical parameter.