Generating Physically Sound Designs from Text and a Set of Physical Constraints

TL;DR

The paper tackles the challenge of generating physically sound load-bearing designs from textual descriptions. It introduces TIDES, a framework that jointly optimizes a differentiable structural solver and a CLIP-based visual judge, augmented by a Hill-function to push material densities toward binary values and a compliance mask to remove non-structural material. Key contributions include physics-constrained text-image generation, text-informed co-design that leverages visual prompts to shape structural features, and lab-validated demonstrations via 3D-printed beams showing acceptable alignment between simulated and experimental performance. This approach enables diverse, prompt-driven design exploration without requiring physics data to train new models, promising practical impact for rapid concept-to-prototype workflows while highlighting considerations around misuse and scalability to 3D geometries.

Abstract

We present TIDES, a text informed design approach for generating physically sound designs based on a textual description and a set of physical constraints. TIDES jointly optimizes structural (topology) and visual properties. A pre-trained text-image model is used to measure the design's visual alignment with a text prompt and a differentiable physics simulator is used to measure its physical performance. We evaluate TIDES on a series of structural optimization problems operating under different load and support conditions, at different resolutions, and experimentally in the lab by performing the 3-point bending test on 2D beam designs that are extruded and 3D printed. We find that it can jointly optimize the two objectives and return designs that satisfy engineering design requirements (compliance and density) while utilizing features specified by text.

Figures (20)

• Figure 1: Tower design problem. The red arrow indicates the location a force is applied, and the blue line indicates the support the design rests on. The visual performance (v) of each design is given by the CLIP score. The physical performance is measured by the compliance (c), low compliance indicates a greater resistance to the force. (A) Designs generated from our TIDES approach display features specified by text and resist the applied force. (B) Designs generated from a vision loss alone have no understanding of physics indicated by floating material and orders of magnitude higher compliance. (C) Designs generated from a physics loss alone are physically sound and display simple solid features.
• Figure 2: The proposed text informed design framework. The framework jointly optimizes a text to image visual loss, a physic-based loss, and a material cost.
• Figure 3: Generating designs at increasing resolutions for a tower structural optimization problem. A force is applied at the center of the design space indicated by the red arrow and the design rests on support given by the blue line. As the resolution increases feature complexity increases.
• Figure 4: Distribution of design performance across 30 trials. (A) and (B) results for a suspended bridge and hoop design problem, the 90% confidence ellipse is plotted for each trial. (C) and (D) examples of designs returned. All 30 designs for each trial can be found in the Appendix \ref{['sec:bridg30']} and \ref{['sec:hoop30']}.
• Figure 5: Generating designs for diverse initial conditions (force and support positions and text prompts). The designs resist deformation from the applied force while displaying features specified by text suggesting TIDES is robust to changes in the initial conditions. In Appendix: \ref{['sec:diverse_problems']} we provide the configurations for each design problem.