---
license: cc-by-4.0
datasets:
- HZBSolarOptics/MultiLayerThinFilms
metrics:
- mae
tags:
- science
- material
- inverse
- design
---

# OptoGPT++

[![License](https://img.shields.io/badge/Licenses-CC_BY_4.0-green)](./LICENSE)
[![Contact](https://img.shields.io/badge/Contact-SEAOPT-blue)](mailto:SE-AOPT-office@helmholtz-berlin.de)

Meet OptoGPT++ — an enhanced implementation of the OptoGPT, a decoder-only transformer that aims to solve inverse design of multi-layer thin film structures.

**Key Enhancements**

- Inclusion of an **absorption** feature in the model ➕📈
- Increased the maximum **wave length** to 2,000nm 💡
- Longer training time for **better predictive performance** 🤯

**Supporting Material**

**OptoGPT++**: https://github.com/jnitzz/OptoLlama \
**OptoGPT**: https://github.com/taigaoma1997/optogpt \
**ArXiV**: 📝 https://arxiv.org/abs/2304.10294

## Usage

### Install Dependencies

```bash
python -m pip install torch
python -m pip install safetensors
```

### Load Model Checkpoint

```python
from safetensors.torch import load_file

model = OptoGPT()

safetensors_path = "optogpt-model.safetensors"
state_dict = load_file(safetensors_path)
model.load_state_dict(state_dict)
```

## Useful Information

| Stat                | Value       |
| :------------------ | ----------: |
| #Parameters         | 108,381,113 |
| Best validation MAE |      0.0408 |
| Epochs trained      |       1,000 |
| Best epoch.         |         996 |
| Batch size          |         256 |
| n_blocks            |           6 |
| n_heads             |           8 |
| d_model             |       1,024 |
| max_seq_length      |          20 |

## Acknowledgements

This work is supported by the Helmholtz Association Initiative and Networking Fund through the Helmholtz AI platform, and the HAICORE@KIT grant.

## Citations

If you find our work helpful, please feel free to cite as following:

```
@article{ma2024optogpt,
  title={OptoGPT: a foundation model for inverse design in optical multilayer thin film structures},
  author={Ma, Taigao and Wang, Haozhu and Guo, L Jay},
  journal={Opto-Electronic Advances},
  volume={7},
  number={7},
  year={2024},
  publisher={Opto-Electronic Advance},
  doi={10.29026/oea.2024.240062}
}
```

----

<div style="float: left; position: relative; left: 50%; transform: translateX(-50%);">
  <a href="https://www.helmholtz-berlin.de/index_en.html"></a><img src="https://huggingface.co/datasets/HZBSolarOptics/shared-assets/raw/main/logo_hzb.svg" style="height: 80px; float: left; margin: 0 1em"></a>
  <a href="https://www.kit.edu/english/index.php"></a><img src="https://huggingface.co/datasets/HZBSolarOptics/shared-assets/raw/main/logo_kit.svg" style="height: 80px; float: left; margin: 0 1em"></a>
  <a href="https://www.zib.de/"></a><img src="https://huggingface.co/datasets/HZBSolarOptics/shared-assets/resolve/main/logo_zib.png" style="height: 80px; float: left; margin: 0 1em"></a>

  <a href="https://www.helmholtz.ai/"></a><img src="https://huggingface.co/datasets/HZBSolarOptics/shared-assets/raw/main/logo_hai.svg" style="height: 20px; float: left; margin: 2em 0; position: relative; left: 50%; transform: translateX(-50%);"></a>
</div>