Purpose This paper aims to overcome the dataset dependency limitations in conventional machine learning isogeometric topology optimization (ITO) by developing a convolutional neural network (CNN)-integrated framework that eliminates the requirement for pre-optimized training dataset while achieving high-resolution topology generation. Design/methodology/approach A physics-informed CNN architecture is embedded within the ITO framework. Non-uniform rational B-splines (NURBS)-based parameterization synchronizes geometric modeling with structural response analysis through isogeometric analysis (IGA). The structural response obtained from IGA serves as physical information for constructing the loss function of the CNN. Compared with traditional sensitivity analysis, a sensitivity analysis strategy based on backpropagation is proposed. By utilizing the feedback of physical information, sensitivity analysis of the loss function is performed through backpropagation to optimize the CNN and generate the density distribution of the topology. A coordinate-to-density mapping mechanism via CNN nonlinear activation enables binarized density distribution with enhanced stiffness characteristics. Findings The feasibility and advantages of the proposed approach are validated through multi-dimensional numerical examples. Originality/value The proposed approach embeds a physics-informed CNN architecture in the ITO framework. Based on the physical information feedback obtained from IGA, the CNN is optimized to generate the optimal density distribution, achieving a topology with superior stiffness performance without pre-optimized training datasets. The proposed approach is demonstrated to be an effective ITO for applications in additive manufacturing.
Xia et al. (Tue,) studied this question.