Addressing challenges inverse problem with convolutional neural networks and regulation techniques: Applications in extraction of physical parameters of semiconductors devices

Abstract

The instability of the inverse problem is caused by its nonlocal and non-causal nature. This study addresses the inverse problem of determining the physical parameters of semiconductor devices. Based on statistical inversion theory, the probability distribution (posterior distribution) of the SBHs has been estimated by convolutional neural networks. Regularization techniques were then applied to such a distribution to accurately determine the SBHs of semiconductor devices. The results reveal that the fluctuations in the predicted SBHs by convolutional neural networks are similar to the amplitude between the upper and lower envelopes of the free decay curve. The method achieves a maximum relative error below 3.4% when using theoretical diode current–voltage data as input and maintains a relative error of less than 7% when compared to traditional methods when using experimental current–voltage data. Furthermore, the proposed method offers a mathematical interpretation of the inverse problem and demonstrates the capability of the proposed method to extract the physical parameters of semiconductor devices with a small amount of data.

Keywords

• Inverse problem
• Schottky barrier height
• Convolutional neural network
• Metal/semiconductor contact
• Thermionic emission