README.md

---
language:
- en
tags:
- chemistry
- fuel
- engines
- YSI
---

# YSI Predictor — Yield Sooting Index Model

## 📌 Overview

This repository contains a machine learning model for predicting the **Yield Sooting Index (YSI)** of single-component fuel molecules directly from their **SMILES** representation.

**YSI is a soot formation metric** used in combustion science.  
- **Lower YSI → cleaner combustion**
- Highly relevant for **diesel replacement fuels**, **bio-fuels**, and **oxygenated fuels**.

This model supports:
- molecular design and optimization,
- genetic algorithms (e.g., CREM),
- Pareto optimization (CN vs YSI),
- rapid candidate screening.

---

## 🧠 How It Works

The prediction pipeline uses:
- **RDKit** — molecule parsing
- **Mordred** — 2D/3D molecular descriptors
- **FeatureSelector** — dimensionality reduction
- **Tree-based regression model** trained on experimental YSI values

**Prediction flow:**
1. Input SMILES → RDKit Molecule
2. Mordred descriptors generated
3. Feature selection applied
4. YSI predicted using trained regressor

Two model artifacts are included:

model.joblib # trained regressor
selector.joblib # feature selector used during training



---

## 🧬 Training Data

The model was trained using a curated dataset of **experimentally measured YSI values**, covering a diverse set of fuel molecule structures:

Includes:
- linear alkanes
- branched alkanes
- cyclic hydrocarbons
- aromatics
- oxygenated species (ethers, esters)

YSI range in dataset: **≈ 3 → 80**

---

## 📊 Performance

Performance was evaluated on both training and **held-out test** sets.

### ⭐ Training Performance

| Metric | Score |
|--------|--------|
| RMSE   | **6.9661** |
| MAE    | **4.0581** |
| R²     | **0.9309** |

---

### 🧭 Test Performance

| Metric | Score |
|--------|--------|
| RMSE   | **5.9667** |
| MAE    | **3.8324** |
| R²     | **0.9440** |
| MAPE   | **18.38%** |

The **test R² = 0.9440** shows strong predictive accuracy.

---

### 📉 Generalization Check

| Metric       | Value |
|--------------|--------|
| Train RMSE   | **6.9661** |
| Test RMSE    | **5.9667** |
| Δ (Test − Train) | **−0.9994** |

➡️ The negative Δ indicates **no overfitting**, and even **better test performance** due to more stable distribution.

---

## 🚀 Usage

Below is a minimal example showing how to use the model in Python.

> The feature calculation must match the training pipeline.

```python
import joblib
from rdkit import Chem
from shared_features import featurize_df, FeatureSelector

# Load model & selector
model = joblib.load("model.joblib")
selector = joblib.load("selector.joblib")

def predict_ysi(smiles: str):
    mol = Chem.MolFromSmiles(smiles)
    df = featurize_df([smiles])
    X = selector.transform(df)
    y = model.predict(X)
    return float(y[0])

print(predict_ysi("CCCCCCC"))