Solid Rocket Propellant Performance Classification

Comparing Random Forest, SVM, and KNN for Propulsion Quality Prediction

Overview

Solid rocket propellants remain a critical technology in both defense and space exploration, offering lower production costs and simpler manufacturing compared to liquid-fueled alternatives. However, the challenge lies in producing high-quality propellant consistently.

This project applies machine learning classification techniques to a solid rocket propellant dataset to identify which chemical and physical parameters most strongly determine propellant performance quality. Three classifiers — Random Forest, Support Vector Machine (SVM), and K-Nearest Neighbors (KNN) — are trained, evaluated, and compared.

---

Dataset

• Source: Kaggle — Solid Rocket Propellant Burning Rate Data
• Observations: 137 samples
• Features: 18 variables

Feature Groups

Category	Features
High-energy oxidizers	NC+NG (Nitrocellulose + Nitroglycerin), RDX, HMX, CL-20
Metal additives	nAl (nano), mAl (micro), Al (raw), nAlN, Al₂O₃, nNi
Plasticizers & stabilizers	DEP (Diethyl Phthalate), nDPN (n-Diphenylamine)
Combustion regulators	Catalyst, Burning rate inhibitor
Measurements	Pressure (MPa), Burning rate (mm/s)

Custom Target Variable

A novel Propulsion Quality metric was derived to convert this regression dataset into a classification problem:

Propulsion Quality = Burning Rate / Pressure  →  min-max normalized to [0, 1]

This was then divided into three balanced classes using the 33rd and 67th percentiles:

Class	Propulsion Quality	Count
Low (0)	< 0.167	45
Med (1)	0.167 – 0.327	47
High (2)	> 0.327	45

---

📁 Repository Structure

├── Rocket_Propellant_Analysis.ipynb        # Full analysis notebook
├── images/                                 # Plots and visualizations
│   ├── correlation_heatmap.png
│   ├── feature_histograms.png
│   ├── boxplots.png
│   ├── class_distribution.png
│   ├── confusion_matrices.png
│   ├── model_comparison.png
│   ├── lime_importance.png
│   └── venn_diagram.png
└── README.md

---

Run Locally

1. Clone the repository

git clone https://github.com/YOUR_USERNAME/YOUR_REPO_NAME.git
cd YOUR_REPO_NAME

2. Install dependencies

pip install -r requirements.txt

3. Download the dataset

Download the dataset from Kaggle and place the .xlsx file in the project root. Then update the file path in the first cell of the notebook:

dosyaYolu = 'Solid_Rocket_Propellant_Burning_Rate_Data.xlsx'

4. Run the notebook

Open and run all cells in Rocket_Propellant_Analysis.ipynb using Jupyter or Google Colab.

jupyter notebook Rocket_Propellant_Analysis.ipynb

Google Colab: Upload the .xlsx file to your Colab session and update the path accordingly. The lime and matplotlib-venn packages will be installed via !pip install cells already included in the notebook.

---

⚙️ Methodology

1. Data Preprocessing

• Null value check (no missing values found)
• Label encoding: Low → 0, Med → 1, High → 2
• Propulsion Quality and Propulsion Class removed from feature inputs to prevent data leakage

2. Exploratory Data Analysis (EDA)

• Correlation heatmap
• Feature distribution histograms
• Per-class boxplots
• Class distribution bar chart

3. Train / Test Split & Scaling

• 80% training / 20% test split (stratified)
• StandardScaler fitted only on training data to prevent data leakage

4. Models

Model	Key Parameters
Random Forest	max_depth=5, min_samples_leaf=5, n_estimators=100
SVM	Default SVC with probability=True
KNN	k=7 (selected by cross-validating k=1..15; upper bound ≈ √137 ≈ 11.7)

5. Evaluation Metrics

• Accuracy, F1-Score (macro), Recall/Sensitivity, Precision, G-Score

6. Interpretability

• LIME (Local Interpretable Model-agnostic Explanations) — chosen over SHAP for better performance on small datasets and non-tree models
• Venn diagram — visualizing feature overlap across models' top-5 important features

---

Results

Model Performance (Test Set)

Model	Accuracy	F1-Score	Recall	G-Score
Random Forest	0.643	0.639	0.652	0.487
SVM	0.571	0.561	0.581	0.385
KNN (k=7)	0.536	0.511	0.548	0.314

Random Forest achieved the best test performance across all metrics.

Model Comparison

Confusion Matrices (Test Set)

• Low class (0) was the easiest to classify correctly across all models
• Med class (1) caused the most misclassifications — likely due to its overlapping characteristics with Low and High

Correlation Heatmap

Key correlations found:

• Burning rate inhibitor ↔ Burning rate: −0.85 (strong negative)
• Catalyst ↔ Burning rate: +0.60 (strong positive)
• Burning rate ↔ Propulsion Class: +0.68 (target variable's strongest predictor)

Feature Distributions

---

Feature Importance (LIME)

Top-5 features per model:

Rank	Random Forest	SVM	KNN
1	Pressure, MPa	nDPN, %	nDPN, %
2	Burning rate inhibitor	Pressure, MPa	Pressure, MPa
3	Burning rate, mm/sec	HMX, %	CL-20
4	NC+NG, %	Al₂O₃, %	Burning rate inhibitor
5	others, %	mAl, %	Al

Feature Overlap Across Models

Pressure (MPa) is the only feature in the top-5 of all three models — confirming it as the single most decisive variable. This aligns with combustion science literature where operating pressure is a fundamental driver of solid propellant burning behavior.

---

References

• Kaggle Dataset — Solid Rocket Propellant Burning Rate
• ML-based burning rate prediction (Wiley/Propellants Journal)
• Effect of aluminum content on burning rate (ScienceDirect)

---

📄 License

The dataset is sourced from Kaggle under its respective license. All analysis code is available for educational use.