End-to-End Generative Design and Physically Interpretable Performance Evaluation of Casting Part Ribs Based on VAE-GAN and Task-Adaptive Graph Attention

Abstract

The design of stiffened thin-walled structures, commonly employed in die-cast automotive and aerospace components, is constrained by limited high-fidelity engineering data and the inherent difficulty of decoding mechanical mechanisms within complex, non-periodic topologies. Data are scarce and complex topologies are hard to interpret. We propose a framework that combines a VAE-GAN with an adaptive graph attention network (GAT). In this framework, the VAE acts as a structural prior. It maps irregular rib layouts into a smooth continuous latent manifold defined by geometric and topological features. This reduced mode collapse. Generated layouts remained valid and coherent, capturing implicit design rules without explicit rules. The augmented dataset also helped the GAT predictor capture spatial connectivity among rib nodes. The adaptive attention mechanism assigned weights to critical regions, improving interpretability across eight tasks. Experimental results showed that the GraphVAE-GAN framework generated novel stiffening rib layouts with 100% validity and structural uniqueness, achieving high-level coupling reconstruction among design variables. For maximum stress prediction under six loading conditions, the baseline model trained on 1184 real samples achieved R2 = 0.74 on a fixed test set. After expanding the training set to 3000 samples, R2 reached 0.91. Results were averaged over 5 runs with fixed seeds. Furthermore, the predictive model surpassed seven mainstream machine learning and deep learning benchmarks across all evaluated mechanical tasks, achieving a mean R2 value above 0.93 across the eight distinct prediction tasks while yielding substantially lower maximum absolute errors compared to the best-performing baseline configurations. These findings suggest a viable strategy for navigating high-dimensional topological design spaces in lightweight structural components and offer a physically grounded trajectory for automating the design exploration of complex engineering systems under multi-objective performance constraints.