🧪 Predicting Initial Dissolution Rates of Glasses Using AI

Welcome! This repository documents a complete AI-driven workflow developed during a one-year gap-year internship, focused on automating data extraction from scientific articles to create a database for modeling the initial dissolution rate of glasses—a critical property for materials used in nuclear waste containment.

🧠 Goal: Extract and structure glass composition and experimental parameters from PDF articles using LLMs, Langflow, OCR, and custom Python scripts, and use this data to train a model to predict the initial dissolution rate of glass under various conditions.

---

📌 Table of Contents

• 🔬 Project Background
• 🧰 Technologies Used
• 🧠 AI/ML Techniques
• ⚙️ Project Workflow
• 📊 Final Dataset & Modeling
• 📎 License & Citation
• 📚 References

---

🔬 Project Background

In the context of nuclear waste storage, glass is used to immobilize radioactive materials. Predicting how this glass alters (dissolves) over time in various conditions is essential for long-term safety.

The initial dissolution rate is one of the key metrics, but gathering experimental data on this is difficult due to:

• Scattered information across decades of publications
• Lack of structured datasets
• Variability in composition formats and testing conditions

---

🧰 Technologies Used

Category	Tools / Technologies
AI/LLM Workflow	Langflow (v1.2.0)
Language Models	Mistral Small 22B (internal API)
Development	Python, VS Code 1.96.2
Data Parsing	PDFMiner, PyMuPDF
OCR	Docling, Lab OCR Scanner
Database	SQLite + SQLAlchemy
ML Modeling	Scikit-learn
Visualization	matplotlib, seaborn

---

🧠 AI/ML Techniques

• Document Question Answering via Langflow
• LLM Prompt Engineering to validate articles
• OCR preprocessing for scanned papers
• Structured information extraction from PDFs
• Unit conversion and normalization of compositions
• Tabular data organization for ML
• Neural network modeling of the dissolution rate

---

⚙️ Project Workflow

Step 1: Article Collection & OCR

• Collected ~50+ articles from:

  • ScienceDirect
  • Google Scholar
  • Web of Science
  • Supervisor suggestions

• Used Docling and a lab OCR printer to digitize 3 scanned papers

Step 2: Relevance Check via LLM Prompting

Used Langflow + LLM with a strict specification sheet prompt to ensure each article includes:

• Full glass composition (~100%)
• Explicit initial dissolution rate (with unit)
• At least one experimental parameter (e.g., pH, temperature)

📄 Output: articles/valid_articles_metadata.json 🟢 20 valid articles (out of 50+ reviewed)

Step 3: Glass List Extraction

Extracted names of glasses with valid data using LLM, e.g.:

1. glass_1: NBS14/18
2. glass_2: NBSA
3. glass_3: NBSC
...

Step 4: Composition Extraction & Conversion

LLM prompts extract composition in various units → then converted using:

• convert_composition.py
• Output: mol% of elements (Si, O, B, Na, etc.)

Step 5: Full Parameter Extraction

For each test on each glass:

• Composition (mol%)
• Experimental conditions
• Initial rate (V₀ or r₀)
• Other metadata (irradiation, polishing, BET surface...)

🔢 Output: structured row for each glass-test combination (107 parameters)

Step 6: Populate the Database

• All cleaned data is inserted into an SQL database
• Converted to CSV: data/final_dataset.csv

Step 7: Model Training

• Used notebooks/03_model_training.ipynb
• Feature engineering on mol% + key conditions
• Trained regression model (e.g., RandomForestRegressor)
• Output: models/trained_model.pkl

---

📊 Final Dataset & Modeling

• ✅ 20 validated papers
• 🧪 70+ unique tests on glasses
• 🧾 107 parameters per test
• 📈 Model performance: R²=97% and RMSE=0.42

---

📎 License & Citation

📜 License

This project is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License.
You may not use this project for commercial purposes or publish derived works without attribution.

To view a copy of this license, visit:
http://creativecommons.org/licenses/by-nc/4.0/

📌 Citation Policy

If you use any part of this repository—including the workflow, prompts, dataset, or model—for research or publication, you must cite this work as:

@misc{ouanzougui2025glass,
  author       = {OUANZOUGUI Abdelhak},
  title        = {Predicting Glass Dissolution Rates with AI: A Langflow-Based Workflow},
  year         = 2025,
  url          = {https://github.com/OUANZOUGUIAbdelhak/GLASS-GPT}
}