README.md

---
tags:
- model_hub_mixin
- pytorch_model_hub_mixin
pipeline_tag: image-to-image
---

# Optimize Any Topology: A Foundation Model for Shape- and Resolution-Free Structural Topology Optimization

This repository contains the official implementation of the **Optimize Any Topology (OAT)** model, a novel foundation-model framework for structural topology optimization. OAT directly predicts minimum-compliance layouts for arbitrary aspect ratios, resolutions, volume fractions, loads, and fixtures, overcoming limitations of prior deep-learning methods. It combines a resolution- and shape-agnostic autoencoder with an implicit neural-field decoder and a conditional latent-diffusion model.

OAT was presented in the paper [Optimize Any Topology: A Foundation Model for Shape- and Resolution-Free Structural Topology Optimization](https://huggingface.co/papers/2510.23667) and was accepted to NeurIPS 2025.

**Code:** https://github.com/ahnobari/OptimizeAnyTopology

<div align="center">
  <img src="https://github.com/user-attachments/assets/6200fa2c-0cd5-49af-897c-67688f28c446" alt="OAT Model Demo Image" width="auto"/>
</div>

This model has been pushed to the Hub using the [PytorchModelHubMixin](https://huggingface.co/docs/huggingface_hub/package_reference/mixins#huggingface_hub.PyTorchModelHubMixin) integration.

## The Open-TO Dataset
The OpenTO dataset, which OAT was trained on, is publicly available on Hugging Face 🤗 at [OpenTO/OpenTO](https://huggingface.co/datasets/OpenTO/OpenTO). It is a new corpus of 2.2 million optimized structures covering 2 million unique boundary-condition configurations. When you run the code, the datasets module is used to automatically download this data.

## Pre-Trained Checkpoints
Pre-trained checkpoints for both the Neural Field Autoencoder (NFAE) and Latent Diffusion Model (LDM) components of OAT are available on Hugging Face.

To load checkpoints using these pretrained weights, you can use the `.from_pretrained` function on both the `NFAE` class and `CTOPUNET` class in the `OAT.Models` module.

The checkpoints are available at:
- **Auto Encoder:** [OpenTO/NFAE](https://huggingface.co/OpenTO/NFAE)
- **Latent Diffusion:** [OpenTO/LDM](https://huggingface.co/OpenTO/LDM)

Additionally, two checkpoints trained on a larger latent space (80x80) are available:
- **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)

## Sample Usage (Generating Solutions)
To generate samples, you can run the `GenerateSolutions.py` script provided in the GitHub repository. First, ensure your environment is set up as described in the [GitHub repository](https://github.com/ahnobari/OptimizeAnyTopology#environment).

For example, to run generation with 4 samples per problem on the OpenTO test set using 8 GPUs:

```bash
accelerate launch --config-file ./accelerate_configs/LDM.yml --num-processes=8 -m scripts.GenerateSolutions \
            --n_samples 4 \
            --splits test\
            --output_path GeneratedSamples.pkl \
            --guidance_scale 2.0 \
            --batch_size 64 \
            --num_sampling_steps 20 \
            --ddim
```

For more detailed instructions on training and evaluation, please refer to the [official GitHub repository](https://github.com/ahnobari/OptimizeAnyTopology).