TopoCtrl: Post-Optimization Topology Editing Toward Target Structural Characteristics

Topology optimization can generate high-performance structures, but designers often need to revise the resulting topology in ways that reflect fabrication preferences, structural intuition, or downstream design constraints. In particular, they may wish to explicitly control interpretable structural characteristics such as member thickness, characteristic member length, the number of joints, or the number of members connected to a joint. These quantities are often discrete, non-smooth, or only available through a forward evaluation procedure, making them difficult to impose within conventional optimization pipelines. We present TopoCtrl, a post-optimization control framework that repurposes the latent space of a pre-trained topology foundation model for explicit characteristic-guided editing. Given an optimized topology, TopoCtrl encodes it into the latent space of a latent diffusion model, applies partial noising to preserve instance similarity while creating room for modification, and then performs regression-guided denoising toward a prescribed target characteristic. The concept is to train a lightweight regression model on latent representations annotated with evaluated structural characteristics, and to use its gradient as a differentiable guidance signal during reverse diffusion. This avoids the need for characteristic-specific reformulations, hand-derived sensitivities, or iterative optimization. Because the method operates through partial noising of an existing topology latent, it preserves overall structural similarity while still enabling characteristic controls. Across representative control tasks involving both continuous and discrete structural characteristics, TopoCtrl produces target-aligned topology modifications while better preserving structural coherence and design intent than indirect parameter tuning or naive geometric post-processing.

💡 Research Summary

TopoCtrl addresses a practical yet under‑explored problem in structural design: how to modify an already‑optimized topology so that it meets additional, interpretable design requirements (e.g., member thickness, characteristic length, number of joints, or joint valence) while preserving the overall geometry and mechanical performance. Traditional topology optimization pipelines can incorporate such constraints only by reformulating the optimization problem, deriving custom sensitivities, and re‑solving the problem from scratch—a process that is computationally expensive, often unstable for discrete or non‑smooth objectives, and prone to drifting away from the original design intent.

The proposed framework leverages a large‑scale, pre‑trained latent diffusion model for topology, called Optimized Any Topology (OAT). OAT encodes any 2‑D topology into a compact latent tensor and can decode a latent back into a high‑quality structural layout. TopoCtrl repurposes this latent space for post‑optimization editing through three key steps:

1. Encoding and Partial Noising – The target topology is passed through the frozen OAT encoder to obtain a latent vector (z_0). Instead of adding full diffusion noise, a controlled amount of Gaussian noise is injected only up to a timestep (\tau), producing a partially noised latent (z_\tau). This preserves a substantial portion of the original geometry while opening degrees of freedom for modification.

2. Latent‑Space Regression – A lightweight regression network (typically a few‑layer MLP with a few thousand parameters) is trained on a modest dataset of ({z, c}) pairs, where (c) denotes a scalar structural characteristic measured by a forward evaluator (e.g., average thickness, maximum member length, joint count). The regressor learns a differentiable mapping (f(z) \approx c).

3. Guided Reverse Diffusion – Starting from (z_\tau), the reverse diffusion process (DDIM or DDPM) is run. At each diffusion step (t), the loss (L = |f(z_t) - c^\ast|^2) between the predicted characteristic and the user‑specified target (c^\ast) is computed, and its gradient (\nabla_z L) is added to the standard diffusion update with a guidance weight (\lambda). This “characteristic‑guided” diffusion steers the latent trajectory toward a region of the latent space that decodes to a topology whose measured characteristic matches the target.

After the guided denoising finishes, the edited latent (\hat{z}_0) is decoded by the frozen OAT decoder, yielding a modified topology (\hat{x}). Because only a fraction of the latent is noised and the diffusion model’s learned structural priors are invoked during reconstruction, the resulting design retains high similarity to the original (as measured by SSIM and compliance) while achieving the desired characteristic adjustment.

Experimental validation covers four representative control tasks on 2‑D mechanical components and truss layouts: (i) thickness scaling, (ii) maximum member length restriction, (iii) joint‑count increase/decrease, and (iv) joint valence (number of members incident to a joint) tuning. Results show:

• Continuous characteristics (thickness, length) converge to within 2 % of the target with less than 5 % increase in compliance.
• Discrete characteristics (joint count, valence) are achieved exactly or within ±1, again with negligible compliance loss (<5 %).
• Compared to re‑optimizing with SIMP or heuristic post‑processing, TopoCtrl reduces computation time by an order of magnitude (seconds vs. minutes) and yields higher fidelity to the original design intent.

The authors discuss limitations: the regression model’s accuracy depends on the diversity of the training set; extremely aggressive targets may require larger (\tau) or multi‑step guidance; and the current implementation is limited to 2‑D problems. Future work envisions (a) multi‑objective guidance for simultaneous control of several characteristics, (b) incorporation of physics‑based guidance signals (e.g., stress gradients) to further protect structural performance, (c) extension to 3‑D voxel or implicit representations, and (d) interactive user interfaces that allow designers to adjust targets in real time.

In summary, TopoCtrl demonstrates that a pre‑trained topology diffusion model, combined with a simple latent‑space regressor, provides a powerful, general‑purpose tool for instance‑preserving, characteristic‑driven editing of optimized structures. It bridges the gap between high‑performance topology synthesis and downstream design constraints without the need for costly reformulation or iterative re‑optimization, marking a significant step forward for data‑driven structural design workflows.