---
language: en
tags:
- glass
- materials-science
- inorganic-chemistry
- regression
- scikit-learn
- xgboost
- property-prediction
license: mit
datasets:
- epam/SciGlass
---

# Vitreos 🔬

**Glass property prediction from oxide composition using machine learning.**

Developed by [Doruk Doğular](https://github.com/dorukdogular) · [GitHub](https://github.com/dorukdogular/vitreos) · [Live Website](https://vitreos.streamlit.app)

## Models

| Property | Samples | R² | MAE |
|---|---|---|---|
| Tg | 76,377 | 0.85 | 44 K |
| Density | 31,173 | 0.88 | 0.26 g/cm³ |
| Refractive Index | 58,913 | 0.83 | 0.036 |
| GFA | 11,858 | — | 69% acc |

## Performance

![Predicted vs Actual](pred_vs_actual_all.png)

![Metrics Summary](metrics_summary.png)

![Feature Importance](feature_importance_all.png)

## Usage

```python
import joblib, json, numpy as np

model = joblib.load("tg_regressor.pkl")
with open("tg_features.json") as f:
    features = json.load(f)

comp = {f: 0.0 for f in features}
comp["SiO2"] = 72.0
comp["Na2O"] = 14.0
comp["CaO"] = 9.0

X = np.array([[comp[f] for f in features]])
print(f"Tg: {model.predict(X)[0]:.1f} K")
```

## Dataset

Trained on [SciGlass](https://github.com/epam/SciGlass) by EPAM Systems — 422,000+ inorganic glass compositions, ODbL license.

## Author

**Doruk Doğular** · [@dorukdogular](https://github.com/dorukdogular)

## License
MIT (model) · ODbL (dataset)