README.md

---
library_name: diffusers
pipeline_tag: text-to-image
license: apache-2.0
datasets:
  - OpenTO/OpenTO
---

# Optimize Any Topology (OAT)

This repository contains the official implementation of the **Optimize Any Topology (OAT)** model, a foundation model framework for shape- and resolution-free structural topology optimization.

**Paper**: [Optimize Any Topology: A Foundation Model for Shape- and Resolution-Free Structural Topology Optimization](https://huggingface.co/papers/2510.23667)
**Code**: https://github.com/ahnobari/OptimizeAnyTopology

<p align="center">
  <img src="https://github.com/user-attachments/assets/6200fa2c-0cd5-49af-897c-67688f28c446" alt="Optimize Any Topology Image">
</p>

## Model Details

### Model Description

Structural topology optimization (TO) is central to engineering design but remains computationally intensive due to complex physics and hard constraints. We introduce Optimize Any Topology (OAT), a foundation-model framework that directly predicts minimum-compliance layouts for arbitrary aspect ratios, resolutions, volume fractions, loads, and fixtures.

OAT combines a resolution- and shape-agnostic autoencoder with an implicit neural-field decoder and a conditional latent-diffusion model. It is trained on OpenTO, a new corpus of 2.2 million optimized structures covering 2 million unique boundary-condition configurations. OAT lowers mean compliance up to 90% relative to the best prior models and delivers sub-1 second inference on a single GPU across resolutions from 64x64 to 256x256 and aspect ratios as high as 10:1. These results establish OAT as a general, fast, and resolution-free framework for physics-aware topology optimization.

**NEWS: Accepted to Neurips 2025!**

-   **Developed by:** The authors of the OAT paper.
-   **Model type:** Conditional Latent Diffusion Model for structural topology optimization.
-   **Language(s):** Not applicable (generates structural layouts).
-   **Finetuned from model [optional]:** The model is trained from scratch using a two-stage process (NFAE then LDM).

### Model Sources

-   **Repository:** https://github.com/ahnobari/OptimizeAnyTopology
-   **Paper:** [Optimize Any Topology: A Foundation Model for Shape- and Resolution-Free Structural Topology Optimization](https://huggingface.co/papers/2510.23667)

## Uses

### Direct Use

OAT is intended for direct use in structural topology optimization. It can generate minimum-compliance layouts for a wide range of engineering design problems with arbitrary aspect ratios, resolutions, volume fractions, loads, and fixtures. Its fast inference capabilities make it suitable for rapid prototyping and design exploration.

### Out-of-Scope Use

This model is specifically designed for structural topology optimization. Its use for general-purpose image generation, tasks unrelated to engineering design, or without understanding the physical constraints and domain limitations is considered out-of-scope.

## Bias, Risks, and Limitations

The model's performance is tied to the distribution of the OpenTO training data. While comprehensive, potential biases or limitations may arise when applied to highly novel or out-of-distribution boundary conditions or material properties not well-represented in the dataset. Users should be aware that generated designs may require further validation (e.g., via Finite Element Analysis) to ensure real-world structural integrity and performance.

### Recommendations

Users should refer to the paper and the GitHub repository for a complete understanding of the model's capabilities and limitations. Validation of generated designs against established engineering principles is recommended, especially for critical applications.

## How to Get Started with the Model

For detailed instructions on installation, environment setup (including MKL optimized packages), training, and inference (sample generation and evaluation), please refer to the official [GitHub repository](https://github.com/ahnobari/OptimizeAnyTopology). The repository provides scripts and guidelines to replicate results and use the pre-trained checkpoints.

## Training Details

The model is trained in two stages:

1.  **Neural Field Auto-Encoder (NFAE)**: Maps variable resolution and shapes into a common latent space.
2.  **Latent Diffusion Model (LDM)**: Trained to generate samples using a conditional diffusion process on the pre-computed latents from the NFAE.

### Training Data

The model is trained on **OpenTO**, a new corpus of 2.2 million optimized structures covering 2 million unique boundary-condition configurations. The dataset is publicly available on Hugging Face 🤗 at [OpenTO/OpenTO](https://huggingface.co/datasets/OpenTO/OpenTO).

### Pre-Trained Checkpoints

Pre-trained checkpoints for both the Auto Encoder (NFAE) and Latent Diffusion Model (LDM) are available on Hugging Face:

*   **Auto Encoder**: [OpenTO/NFAE](https://huggingface.co/OpenTO/NFAE)
*   **Latent Diffusion Model**: [OpenTO/LDM](https://huggingface.co/OpenTO/LDM)
*   **Auto Encoder Large Latent**: [OpenTO/NFAE_L](https://huggingface.co/OpenTO/NFAE_L)
*   **Latent Diffusion Large Latent**: [OpenTO/LDM_L](https://huggingface.co/OpenTO/LDM_L)

These checkpoints can be loaded using the `.from_pretrained` function from the `OAT.Models` module.

## Evaluation

OAT has been rigorously evaluated on four public benchmarks and two challenging unseen tests. The results demonstrate that OAT significantly lowers mean compliance (up to 90%) compared to previous state-of-the-art models. It also achieves impressive inference speeds, delivering sub-1 second results on a single GPU across various resolutions and aspect ratios.

## Citation

If you find this work useful or inspiring for your research, please consider citing our paper:

```bibtex
@misc{optimizeanytopology2025,
      title={Optimize Any Topology: A Foundation Model for Shape- and Resolution-Free Structural Topology Optimization},
      author={Ahnobari, [Authors Not Provided In Prompt]},
      year={2025},
      eprint={2510.23667},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.23667},
}
```