Research map / Vision-language and multimodal / Human and embodied modeling
Generating Physically Sound Designs from Text and a Set of Physical Constraints
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces TIDES, which jointly optimizes a CLIP-based visual alignment loss and a differentiable SIMP compliance objective (plus material cost) to generate structurally sound designs shaped by text prompts, validated in simulation and by 3D-printed 3-point bending tests.
Problem
Text-to-image models produce visually appealing designs that ignore physics, while structural optimization struggles to incorporate complex visual or aesthetic design goals; the paper targets generating designs that are both physically sound and match a text description.
Approach
Combine a frozen CLIP visual loss with a differentiable SIMP compliance loss over a shared density grid, using a Hill-function sigmoid to push densities to 0/1 and a compliance-based mask to remove floating, non-structural material instead of connected-component labeling.
Key results
- TIDES designs have compliance within the same order of magnitude as physics-only designs while vision-only designs have orders of magnitude higher compliance (per the paper's tower problem).
- In simulation for the 3-point bending beam, the 'large arch' design reached 285.71 compliance vs 299.55 for 'honeycomb, many hexagons' and 280.0 for physics-only; experimental compliance values were 1.0e-3 (physics-only), 1.13e-3 (honeycomb), 1.19e-3 N/mm (large arch).
- Across 30 trials on suspended bridge and hoop problems, text prompts (triangle, hexagon, arch) yield diverse designs with competitive compliance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add text-guided TIDES design generation with a large arch prompt instead of physics-only optimization | 3-point bending beam design · 3-point bending, experimental | compliance (N/mm): 1.0e-3 N/mm -> 1.19e-3 N/mm | Hurt |
| Add text-guided TIDES design generation with a honeycomb prompt instead of physics-only optimization | 3-point bending beam design · 3-point bending, experimental | compliance (N/mm): 1.0e-3 N/mm -> 1.13e-3 N/mm | Hurt |
| Add text-guided TIDES design generation with a large arch prompt instead of physics-only optimization | 3-point bending beam design · 3-point bending, simulation | compliance: 280.0 c -> 285.71 c | Hurt |
| Add text-guided TIDES design generation with a honeycomb prompt instead of physics-only optimization | 3-point bending beam design · 3-point bending, simulation | compliance: 280.0 c -> 299.55 c | Hurt |
| Remove the compliance mask (ablation) | TIDES · Structural design problems | floating material | Hurt |
| Remove the Hill function (ablation) | TIDES · Structural design problems | binary material distribution | Hurt |
| Remove the physics loss and optimize only the vision loss (ablation) | TIDES · Tower design problem | compliance | Hurt |
| Remove the vision loss and optimize only the physics loss (ablation) | TIDES · Tower design problem | visual alignment | Hurt |
| Vary the text prompt to specify triangle, hexagon, or arch supports | TIDES · Suspended bridge | compliance | No clear effect |
| Vary the text prompt to specify arch or triangle supports | TIDES · Hoop design problem | compliance | No clear effect |
Surprising findings
- For the hoop problem, arch and triangle prompt designs overlapped in performance with physics-only results, suggesting the visual loss can help escape local minima.
Limitations
- 2D simulation only; restricts 3D designs to layered 2D force applications.
- Manual cleanup pass required to remove some disjoint artifacts after compliance masking.
- Hyperparameters (alpha, n, beta weights) heuristically chosen; 3D-printed validation limited to a few beam designs.
Experimental setup
Benchmarks: tower design problem suspended bridge and hoop design problems 3-point bending test of 3D-printed beams
Metrics: compliance CLIP score target density force-displacement
Baselines: physics-only structural optimization (SIMP) vision-only CLIP optimization Stable Diffusion 2.1 starting designs Atlas3D DSO CLIP stylization topology optimization
Builds on: SIMP CLIP CLIPDraw Deep Daze
Closest papers by meaning
- EngiBench: A Framework for Data-Driven Engineering Design Research (NeurIPS 2025 · similarity 0.83)
- Optimize Any Topology: A Foundation Model for Shape- and Resolution-Free Structural Topology Optimization (NeurIPS 2025 · similarity 0.82)
- LEIA: Learned Environment for Interactive Architected Materials (NeurIPS 2026 · similarity 0.82)
- BikeBench: A Bicycle Design Benchmark for Generative Models with Objectives and Constraints (NeurIPS 2025 · similarity 0.81)
- Deep Electromagnetic Structure Design Under Limited Evaluation Budgets (ICML 2025 · similarity 0.81)
- Atlas3D: Physically Constrained Self-Supporting Text-to-3D for Simulation and Fabrication (NeurIPS 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.