Abstract
Generative models have shown promise for engineering design automation, yet applying them to multi-point multi-objective design (MPMOD) faces two critical challenges: inefficient encoding of relationships among multiple working conditions and performance targets, and error accumulation from conditional independence assumptions in existing diffusion models. This paper proposes Physics-guided Conditional Variational Autoencoder (PCVAE), a framework that achieves one-shot generation of MPMOD-compliant designs through three innovations. First, an attention-based fusion module holistically encodes geometric features, working conditions, and performance requirements across multiple points. Second, a physics-guided input transformation, rooted in Newtonian impact theory, maps raw inputs into a physically meaningful feature space, reducing learning complexity. Third, a performance-driven guidance mechanism employs a decoupled evaluator—pre-trained on high-fidelity simulations—to compute a performance loss that steers generation toward physically feasible regions, even beyond training distribution. Validated on high-fidelity rocket datasets, PCVAE achieves 62.8\% lower MAE than state-of-the-art methods in single-point design and 7.4× error reduction in multi-point scenarios, while maintaining superior diversity. The framework establishes a new paradigm where physical knowledge and performance feedback enable direct MPMOD without iterative refinement.