Extending Parametric Model Embedding with Physical Information for Design-space Dimensionality Reduction in Shape Optimization

Abstract: In this work, an extension of the parametric model embedding (PME) approach is presented, aiming to achieve more effective design-space dimensionality reduction for shape optimization in vehicle design. PME, rooted in principal component analysis (PCA), not only identifies a reduced set of critical modes but also re-parameterizes the original design space, enabling direct and interpretable manipulations of shape modifications within the reduced space. Alongside the "physics-informed" version (PI-PME), which enriches geometry with low-fidelity distributed and lumped physical quantities, a "physics-driven" variant (PD-PME) is introduced that focuses exclusively on physical parameters. Both formulations employ PCA to capture the principal modes of variability yet differ in their balance between geometric and physical information, through the ad-hoc definition of a weighted inner product. Through test cases involving the RAE-2822 airfoil, a bio-inspired underwater glider, a naval propeller, and the DTMB-5415 destroyer-type vessel, it is shown how the resulting frameworks provide a first-level assessment of design variability, offer interpretability regarding which original variables most strongly affect performance, and efficiently bridge geometric and physical parameters. Furthermore, lumped physical parameters can serve as a low-fidelity foundation for multi-fidelity optimization, directly leveraging the linear re-parameterization to drive the reduced design variables. Meanwhile, distributed physical parameters enable the construction of machine-learning-based reduced-order models to infer integral quantities of interest. By allowing the user to embed these insights early in the design process, PI-PME and PD-PME facilitate more robust, cost-effective exploration, paving the way for subsequent high-fidelity optimization.