Abstract:

To overcome the shortcomings like computational expensive and inability to generate structures with explicit geometry of traditional topology optimization algorithms represented by the Solid Isotropic Material with Penalization（SIMP） method and the level-set method， a deep-learning-based two-stage approach for real-time explicit topology optimization was proposed， which combined deep learning with the moving morphable components（MMC） method. In the first stage， a deep learning model was used to replace most of the time-consuming finite element analysis. The second stage performed a small number of iterative fine-tuning of the structure predicted by the deep learning model to generate the final optimized structure with explicit geometry. A relatively general data set was used to verify the feasibility and effectiveness of the framework quantitatively and qualitatively， and the relationship between the training degree of the deep learning model in the first stage and the quality of the structure generated as well as the total time consumed was studied. Experimental results show that this approach can save more than 90% of the computing time while maintaining the quality of the topology optimization structures generated.

Key words: computer application, topology optimization, deep learning, moving morphable components method, computer-aided design