深層学習と線形最適化法を併用した最適トラス・トポロジーの生成法

抄録

The present paper deals with a prediction method for axial force distribution of optimal topology using neural network (U-Net) and truss topologies data. By reducing design variables based on following predicted data, we also propose a method which is more computational efficient than conventional optimization methods. In this paper, we use a truss topology optimization problem, which is to find the cross-sectional area of truss members, such that both the structural volume and compliance are minimized as examples. U-Net trained with the optimal solutions by optimization method can predict optimal truss topology even from untrained data. Moreover, optimal or quasi-optimal solutions are obtained by solving optimization problems with truss members exceeding a certain value as design variables. The applications of the proposed method are illustrated in numerical examples with discussion on effectiveness of the proposed method and computation efficiency.