NLP-ML

Natural Language Embeddings of Synthesis and Testing Conditions Enhance Glass Dissolution Prediction

This repository contains all the code and data for the paper "Natural Language Embeddings of Synthesis and Testing Conditions Enhance Glass Dissolution Prediction".

Repository Structure

NLP-ML/
├── Data/                                       # Datasets: exact train/test splits and unseen composition files
├── Descriptor_Training/                        # Training with structural descriptor features
├── MLPipeline/                                 # Core ML pipeline modules
├── Models/                                     # Trained models with normalisation statistics and loss values
├── NLP_Features/                               # NLP feature extraction and embeddings
├── Training_scripts/                           # Model training scripts (nn_Neural.py, xg_boost.py)
├── Model_post_processing-DIRECT_PIPELINES.ipynb  # Main notebook: reproduces all plots and results
├── jaccard_distance.ipynb                      # Jaccard distance plot code
├── text_to_embed.py                           # Python file to convert abstract to embed.py
├── requirements.txt                            # Python dependencies
└── vocab_mappings.txt                          # Vocabulary mappings for text normalisation

Reproducing results

Model_post_processing-DIRECT_PIPELINES.ipynb is the full pipeline to load the trained models and reproduce all plots and results reported in the paper. Run this notebook end-to-end after installing the dependencies:

conda create -n nlpml python=3.10
conda activate nlpml
pip install -r requirements.txt
jupyter notebook Model_post_processing-DIRECT_PIPELINES.ipynb

Training models

The Training_scripts/ folder contains the training code for the models:

• nn_Neural.py — trains the neural network model
• xg_boost.py — trains the XGBoost model

Data

The Data/ folder contains:

• The exact train/test splits of the datasets used in the paper
• CSV files of the unseen (out-of-distribution) glass compositions used to evaluate model generalizability

Trained models

The Models/ folder contains:

• The trained model files
• The corresponding mean/standard-deviation files used to normalise the data
• Loss value files from training

Additional analysis

• jaccard_distance.ipynb — code to generate the Jaccard distance plots

Citation

If you use this code or data, please cite:

@article{mannan2026natural,
  title={Natural Language Embeddings of Synthesis and Testing conditions Enhance Glass Dissolution Prediction},
  author={Mannan, Sajid and Nambudiripad, K Sidharth and Mandal, Indrajeet and Gosvami, Nitya Nand and Krishnan, NM},
  journal={arXiv preprint arXiv:2604.14078},
  year={2026}
}