論文ID: 24-00072
This paper presents a data augmentation method for generating a surrogate model of numerical analysis results. The proposed method focuses on the relative frequency of learning data for generating a learning model using deep learning techniques. Generally, data augmentation techniques are known to be useful for improving prediction accuracy. Adding noise and data duplication are commonly used for predicting numerical simulation results, but it is essential to carefully consider the amount of noise or choose a duplication target. However, these techniques are not appropriate for generating surrogate models. The reason is that the numerical analysis results mostly have high data imbalance, and no specific solution has been presented. The method proposed in this paper solves this problem and aims to be a simple and highly versatile data augmentation method. This paper describes the application of the proposed method to predict two-dimensional stress fields. It was confirmed that by increasing the number of data augmentations using the proposed method, the prediction errors were reduced for three different stress components stably. Additionally, it was confirmed that the prediction accuracy improved 5.81 to 27.0% compared to that of the data augmentation by simple duplication.
