Fundamental study for Structural Analysis using Machine Learning

Abstract

In design process, recent digitizing technologies make total cost reduced, however CAE computation relatively becomes larger than before. In the situation 1D-CAE or model-based simulation techniques become popular in the early stage of design process. Machine learning has an ability to reduce more cost, however accurate regression using machine learning should be established. The important point is data design and amount of data to be trained for machine learning. Material design and numerical fluid dynamics should handle a large amount of data and those data is the effective learning data. Many successful research works using machine learning are reported recently. On the other hand, in the fields of stress analysis or structural analysis, several research works are only presented. In this work, we propose an effective input data augmentation technique for convolutional neural network to be applied to accurate regression prediction. In engineering phenomena, several important parameters make a formula with polynomials using the parameters. Possible polynomials are computed in advance and the polynomials are arranged in grid like a digital image. The arranged polynomials are an augmented input data of convolutional neural network for regression prediction. All of input values are normalized between 0 to 1 or -1 to 1. The corresponding output data with input data are also normalized. Furthermore, usual data augmentation is conducted to prevent overfitting.