This work presents a physically structured latent subspace model for signed distance function (SDF) representations, extending the original implicit representation . Our goal is to enhance design exploration within Model-Based Design (MBD) workflows. While implicit neural shape models offer high-fidelity 3D reconstruction, their latent spaces typically lack interpretability, as geometric factors are often coupled in a non-intuitive manner. This "black box" nature limits their utility as independent design variables in engineering contexts. We introduce a framework that anchors latent directions to human-interpretable geometric attributes, such as wheelbase, height, and width, during the training phase. The formulation utilizes global and local alignment losses to correlate latent variations with standardized physical specifications. Furthermore, a rank-agnostic compactness regularizer is employed to concentrate residual shape variations into a low-dimensional subspace without pre-specifying its dimension. This approach effectively bridges the gap between deep-learning-based surrogate models and traditional parametric design. Numerical experiments conducted on the large-scale DrivAerNet++ vehicle benchmark demonstrate that the proposed method maintains high reconstruction accuracy while providing superior controllability. The learned latent space enables monotonic and predictable shape edits with minimal cross-attribute interference. These results indicate that embedding physical structure into SDF latent spaces substantially improves numerical stability and consistency for vehicle-centric 3D modeling. The proposed method provides a robust, reduced-order parameterization suitable for downstream engineering tasks, such as aerodynamic optimization and simulation-driven design exploration.