PhysGen: Physically Grounded 3D Shape Generation for Industrial Design

TL;DR

The paper tackles the challenge of generating 3D shapes that are not only visually plausible but also physically feasible for industrial design, focusing on aerodynamic efficiency. It introduces PhysGen, a unified pipeline that marries a shape-and-physics variational autoencoder (SP-VAE) with a physics-guided flow-matching generator, using an alternating update strategy to co-optimize geometry and physics. The SP-VAE encodes both shape and aerodynamic properties (pressure field and drag) in a shared latent space, enabling end-to-end generation and physics-aware refinements. Experimental results on automotive and aircraft benchmarks show improved geometric plausibility and aerodynamic performance over prior methods, with additional benefits in single-view reconstruction and cross-domain generalization.

Abstract

Existing generative models for 3D shapes can synthesize high-fidelity and visually plausible shapes. For certain classes of shapes that have undergone an engineering design process, the realism of the shape is tightly coupled with the underlying physical properties, e.g., aerodynamic efficiency for automobiles. Since existing methods lack knowledge of such physics, they are unable to use this knowledge to enhance the realism of shape generation. Motivated by this, we propose a unified physics-based 3D shape generation pipeline, with a focus on industrial design applications. Specifically, we introduce a new flow matching model with explicit physical guidance, consisting of an alternating update process. We iteratively perform a velocity-based update and a physics-based refinement, progressively adjusting the latent code to align with the desired 3D shapes and physical properties. We further strengthen physical validity by incorporating a physics-aware regularization term into the velocity-based update step. To support such physics-guided updates, we build a shape-and-physics variational autoencoder (SP-VAE) that jointly encodes shape and physics information into a unified latent space. The experiments on three benchmarks show that this synergistic formulation improves shape realism beyond mere visual plausibility.

Figures (10)

• Figure 1: Physical knowledge is crucial for realistic and functionally efficient 3D shape generation. Without physics, (a) generated shapes may appear visually plausible yet violate basic physical feasibility, such as car wheels intersecting the body or chairs with broken or unstable legs, and (b) aerodynamic shapes produce wide turbulent wakes, indicating low aerodynamic efficiency. (c) In contrast, physics-guided generation produces shapes with coherent flow and reduced drag, achieving both aesthetic appeal and physical efficiency.
• Figure 2: Overview of the proposed framework. (a) The proposed SP-VAE learns a unified latent representation that jointly encodes geometric structure and physical properties. From this shared representation, three decoders reconstruct the 3D shape, surface pressure field, and drag coefficient, respectively. (b) The physics-guided shape generation iteratively bridges flow-matching updates and physical refinements, optionally conditioned on an image, such as a sketch. This alternating strategy updates the latent code to align with the desired 3D shape and physical properties, ensuring both visual plausibility and physical validity.
• Figure 3: Qualitative results of post-optimization and our unified generation. TripOptimizer produces distorted geometries and fails to recover the realistic shape, while our alternating method restores smooth and plausible surfaces that closely match the ground truth.
• Figure 4: Physical information improves shape generation accuracy. Starting from the physically unguided generation (gray), refining toward the target drag coefficient (blue) aligns the geometry more closely with the ground-truth shape (red).
• Figure 5: Physical information mitigates depth ambiguity in 3D generation from a real single-view image. Two shapes are first generated from the same image using different initial noises (red and orange cars). When physical guidance is applied during generation, the resulting shapes (blue and green cars) converge to similar front-view widths.