Presentation Information
[450202-02-01]Intelligent design of topology optimization considering physics-related information
Prof. Jun Yan (Dalian University of Technology)
Structural optimization design is an effective means to achieve innovative structural design configurations, widely applied in structural design fields such as aerospace, automotive, and marine engineering. However, with the increase of the problem complexity, the number of structural design variables has surged, traditional topology optimization methods face the significant computational challenges. Due to the powerful nonlinear learning and computational capabilities of deep learning algorithms, applying deep learning algorithms to topology optimization design to speed up the topology optimization process has emerged as the most promising new discipline in topology optimization. This paper significantly improves the efficiency of topology optimization by combining deep learning models with traditional topology optimization methods such as SIMP (Solid Isotropic Material Penalty) and MMC (Moving Morphable Component). By incorporating physics-related information (such as principal stress matrices and temperature gradient matrices) into model training, a neural network prediction model based on a small sample set is constructed, yielding highly accurate prediction results. A data preprocessing method and a new form of loss function PMSE (Penalty Mean Square Error) that conform to the data characteristics of the optimization algorithm are proposed to improve the model prediction accuracy for structural boundaries. At the same time, the effects of different input modes on the final prediction results are compared. The results show that the introduction of physical information related to the objective function can effectively improve the prediction accuracy of the model, which provides new ideas for topology optimization methods based on deep learning.