Semantic Embeddings of Chemical Elements - Code and Data Repository

This repository contains the complete implementation, datasets, and trained models for the paper "Semantic Embeddings of Chemical Elements for Enhanced Materials Inference and Discovery" (arXiv:2502.14912).

This repository includes:

1. Source Code (MIT License)   - ElementBERT pre-training scripts using DebertaV2ForMaskedLM   - Downstream task implementations (regression, classification, Bayesian optimization)   - External validation and visualization tools

2. Datasets (CC0-1.0 License)   - 2025 publications training and test datasets   - Elemental embeddings from ElementBERT and baseline models   - Experimental property datasets for shape memory alloys, high-entropy alloys, and Ti alloys

3. Pre-trained Model (MIT License)   - Available on HuggingFace: https://huggingface.co/Neuquar/ElementBERT   - Based on DeBERTa-v3-xsmall architecture with masked language modeling

4. Web of Science Corpus Metadata   - Training corpus metadata with proper Clarivate attribution   - Only displays permitted public fields (Title, Authors, Source)   - See WOS_statement.txt for full copyright notice

Licensing

- Code: MIT License- Data: CC0 1.0 Universal (Public Domain Dedication)- Model: MIT License

Installation and Usage

Detailed installation instructions and usage examples are provided in the README.md file.

Reproducibility

All code, data, and models are provided to ensure full reproducibility of the results reported in the paper.