Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Sketch2Topo: Using Hand-Drawn Inputs for Diffusion-Based Topology Optimization

Shuyue Feng, Cedric Caremel, Yoshihiro Kawahara

Abstract

Topology optimization (TO) is employed in engineering to optimize structural performance while maximizing material efficiency. However, traditional TO methods incur significant computational and time costs. Although research has leveraged generative AI to predict TO outcomes and validated feasibility and accuracy, existing approaches still suffer from limited customizability and impose a high cognitive load on users. Furthermore, balancing structural performance with aesthetic attributes remains a persistent challenge. We developed Sketch2Topo, which augments a diffusion-based TO model with image-to-image generation and image editing capabilities. With Sketch2Topo, users can use sketching to customize geometries and specify physical constraints. The tool also supports mask input, enabling users to perform TO on selected regions only, thereby supporting higher levels of customization. We summarize the workflow and details of the tool and conduct a brief quantitative evaluation. Finally, we explore application scenarios and discuss how hand-drawn input improves usability while balancing functionality and aesthetics.

Sketch2Topo: Using Hand-Drawn Inputs for Diffusion-Based Topology Optimization

Abstract

Topology optimization (TO) is employed in engineering to optimize structural performance while maximizing material efficiency. However, traditional TO methods incur significant computational and time costs. Although research has leveraged generative AI to predict TO outcomes and validated feasibility and accuracy, existing approaches still suffer from limited customizability and impose a high cognitive load on users. Furthermore, balancing structural performance with aesthetic attributes remains a persistent challenge. We developed Sketch2Topo, which augments a diffusion-based TO model with image-to-image generation and image editing capabilities. With Sketch2Topo, users can use sketching to customize geometries and specify physical constraints. The tool also supports mask input, enabling users to perform TO on selected regions only, thereby supporting higher levels of customization. We summarize the workflow and details of the tool and conduct a brief quantitative evaluation. Finally, we explore application scenarios and discuss how hand-drawn input improves usability while balancing functionality and aesthetics.
Paper Structure (11 sections, 3 figures, 1 table)

This paper contains 11 sections, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Topology Optimization and Generative AI
  4. Controllable Editing in Diffusion Models
  5. system overview
  6. Functionality and User Interface
  7. Technical Details
  8. evaluation and APPLICATION
  9. Minimum Compliance and Volume Fraction
  10. Iterative Design of a Chair
  11. CONCLUSION

Figures (3)

  • Figure 1: The left panel illustrates the workflow in prior work, where conditions are specified via text-based parameters and users cannot provide a specific target shape as an input condition. The right panel shows the Sketch2Topo workflow: users can sketch the target shape and directly specify physical constraints through hand-drawn inputs. In addition, users can use a masking tool to restrict topology optimization to selected regions.
  • Figure 2: Sketch2Topo GUI. (a) Brush tools for drawing the material region and physical constraints. (b) Parameters that are less suitable for sketch input remain text-based, allowing users to set the volume fraction and load direction. (c) Canvas area: the left canvas is used to draw the material and physical constraints, while the right canvas displays the generated result.
  • Figure 3: Iterative chair design using Sketch2Topo: (a) without the mask function; (b) with the mask function for creative exploration; (c)–(e) topology optimization results under three different mask conditions; and (f) the final rendered design (rendered using Nano Banana from google).