In the electronics industry, introducing Machine Learning (ML)-based techniques can enhance Technology Computer-Aided Design (TCAD) methods. However, the performance of ML models is highly dependent on their training datasets. Particularly in the semiconductor industry, given the fact that the fabrication process of semiconductor devices is complicated and expensive, it is of great difficulty to obtain datasets with sufficient size and good quality. In this paper, we propose a strategy for improving ML-based device modeling by data self-augmentation using variational autoencoder-based techniques, where initially only a few experimental data points are required and TCAD tools are not essential. Taking a deep neural network-based prediction task of the Ohmic resistance value in Gallium Nitride devices as an example, we apply our proposed strategy to augment data points and achieve a reduction in the mean absolute error of predicting the experimental results by up to 70%. The proposed method could be easily modified for different tasks, rendering it of high interest to the semiconductor industry in general.