Improving efficiency in structural optimization using RBFNN and MMA-Adam hybrid method

Abstract

A significant challenge in traditional topology optimization (TO) methods lies in their low efficiency when handling large-scale design problems, primarily due to repetitive high-dimensional computations. Recently, the fusion of implicit neural representation with conventional structural topology optimization techniques has garnered considerable interest, owing to its various advantages, including the elimination of the need for filters. While this approach can yield design solutions comparable to those from traditional TO methods, it does not lead to clear efficiency gains; in some cases, the number of forward calculations is higher than in traditional methods. This study aims to enhance method efficiency by utilizing the radial basis function neural network (RBFNN) to implicitly represent the structure. Specifically, a set of trainable radial bases shapes and positions is employed to span the structure’s density field. Additionally, a hybrid approach is proposed, combining the method of moving asymptotes (MMA) with the Adam optimizer to update the neural network parameters. This updating technique accelerates convergence and enhances overall efficiency. Through adapting the bases via backpropagation and minimizing the loss function constructed based on traditional TO methods, our approach facilitates achieving design solutions with similar performance but with significantly fewer design variables and performance evaluations compared to traditional TO methods.