Update README.md

README.md

@@ -18,11 +18,70 @@ size_categories:
 [![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)
 
+## Details
+
+The original dataset can be found in the [OptoGPT publication 📝](https://arxiv.org/abs/2304.10294) and here on [HuggingFace](https://huggingface.co/datasets/mataigao/optogpt_data).
+
+**Key Enhancements**
+
+- Inclusion of an **absorption** feature in the model ➕📈
+- Increased the maximum **wave length** to 2,000nm 💡
+- Additional **materials**, e.g. EVA 👍
+
+## Structure
+
+```
+├── materials/
+│   ├── Ag.csv
+│   ├── Al.csv
+│   ├── ...
+│   └── ZnSe.csv
+├── train/
+│   ├── train-00.safetensors
+│   ├── train-01.safetensors
+│   ├── ...
+│   └── train-09.safetensors
+├── test/
+│   └── test.safetensors
+└── tokens.json
+```
+
+Each `*.safetensors` file contains 1 million thin film structures (indexed by `tokens.json`) as well as their simulated absorption, reflection and transmission (RAT) spectrum. 
+
+The spectrum tensors have the shape *(n_samples, 3 [RAT], 171 [bins],)* and are stores in `bfloat16`. The thin film layers have the shape *(n_samples, 20 [max_depth],)* and are stored as long.
+
+The thin film layer sequence is ordered from top to bottom, i.e., the lowest index is the top of the thin film layers (touches the air), and the highest index is the bottom layer (touching the back substrate). 
+
+
+## Details
+
+- All spectra have been simulated with [tmm_fast](https://github.com/MLResearchAtOSRAM/tmm_fast)
+- The *n* (refractive index) and *k* (extinction coefficients) of all materials can be found in CSV files in the `/materials` folder
+- The training and test data are pre-split and can be found in the `/train` and `/test` folder respectively
+- A full list of the vocabular, i.e., all possible tokens, can be found in the `tokens.json`, with the the format of `<material>_<thickness_in_nm>`
+- There are additional tokens for end of sequences (EOS), padding (PAD) and masking (MASK)
 
 ## 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%);">