Real-time generalizable structural topology generation via fourier and latent modulated neural networks informed by few prior designs

Topology optimization plays a crucial role in structural design, yet traditional methods remain computationally inefficient. Deep learning offers acceleration, yet most models require large training datasets and exhibit limited generalization. In contrast, human designers can infer new solutions from only a few prior cases, leveraging correlations between design intent and outcomes. Moreover, when design constraints vary smoothly, the resulting density fields exhibit strong spatial correlations, suggesting opportunities to exploit these patterns for efficient, data-driven topology generation. Building on this observation, we propose a neural network framework that encodes design intent into latent vectors and employs cross-attention with Fourier features to couple global constraints and local spatial information, enabling design-aware density field modeling. Specifically, the density at each mesh point is adaptively modulated by both local coordinate information and global design constraints, thereby enabling real-time structural inference for new design constraints without retraining. Beyond direct prediction, the optimized latent vectors and network parameters can be reused as initialization for unseen constraints, thereby providing high-quality starting points that accelerate subsequent optimization. The proposed method supports both single- and multi-scenario optimization, which allows the inference of high-resolution designs and the generation of new solutions for unseen design constraints using knowledge from only two prior samples. Across benchmark cases, it achieves substantial acceleration over traditional solvers and consistently outperforms existing neural approaches. These results demonstrate its potential as a real-time, scalable, and generalizable framework for intelligent topology optimization.