계층적 분류 및 회귀 그래프 신경망을 이용한 최적의 트러스 구조 예측 방법

Abstract

This paper proposes a graph neural network-based hierarchical classification and regression model to predict the optimal ground structure layout. A ground structure consists of a set of interconnected bars similar to a graph structure. The number of bars and joints depends on the size and configuration of the target design domains. Conventional neural networks, such as MLP and CNN, cannot handle such variable-dimensional data. Therefore, we adopted a graph neural network using the similarity between ground structure and graph. Only a few of the potential bar candidates remained after ground structure optimization converged. It causes the highly imbalanced distribution of bar areas, which is difficult to predict. Therefore, we construct the model to process a binary classification task to classify the presence or absence of remaining bars among potential bar candidates, and then sequentially process a regression task to predict the detailed value of the cross-sectional areas. To validate the proposed model, three types of ground structure optimization problems were defined: cantilever, simply supported, and L-shape beams. The model was evaluated with three evaluation metrics: classification, regression, and optimality. As a result, it predicts the presence or absence of the optimal ground structure with more than 95% accuracy and predicts the cross-sectional areas of the remaining bars with greater than 95% accuracy.