Designer preference-driven topology optimization using a human-in-the-loop neural network

Abstract

Topology optimization is a powerful computational methodology for generating high-performance structures. However, ignoring aesthetics limits the practical application. User interaction helps address this limitation, but current methods lack generality, needing separate design and sensitivity analyses for each interaction form. To overcome this challenge, we propose a Designer Preference-Driven Neural Topology Optimization (DPDNTO) method, which utilizes a unified interaction parameter to formulate a unified, mask-based loss function to control geometric features and material distribution. In addition, feature sizes are flexibly controlled by adjusting the number of neurons in the DPDNTO method. By utilizing the backpropagation mechanism of neural networks, the proposed method efficiently updates design variables and automatically balances these multi-objective tasks through a dynamic parameter strategy. To provide further intuitive visual feedback, a multimodal large model is employed to render optimized structures into conceptual visualization images. Numerical experiments demonstrate that the proposed method not only enhances the aesthetic quality of the final designs but also improves structural stress performance and linear buckling resistance. These findings establish DPDNTO as a versatile and computationally efficient paradigm, bridging the critical gap between algorithmic optimization and preference-driven aesthetics and paving the way for advancements in fields such as architecture, industrial design, and advanced manufacturing.