PINL

This project applies machine learning regression models to predict the boiling points of chemical compounds using molecular descriptors. It includes data cleaning, feature exploration, visualization, multiple regression techniques, and model hyperparameter tuning.

Team Members

• Mujtaba-4T4
• TurboGlitch
• Hydroxymeth

Dataset

The dataset contains 777 molecular descriptors for 5431 molecules, along with their boiling point values (BP). These descriptors capture structural, topological, physicochemical, and electronic properties of molecules.

Important Descriptors Used in Analysis

Descriptor ID	Descriptor Name
D004	Molecular Weight
D368	LogP (octanol-water partition coefficient)
D294	TPSA (Topological Polar Surface Area)
D555	Number of Hydrogen Bond Donors
D556	Number of Hydrogen Bond Acceptors
D557	Number of Rotatable Bonds
D553	Number of Ring Atoms
D559	Number of Aromatic Rings
D370	AlogP
D560	Formal Charge
D373	Fraction Csp3

The dataset file used: chem_data.csv

Workflow Summary

1. Data Preparation

• Dropped unnamed or null columns
• Split dataset by Subset column into train, validation, and test sets
• Standardized the feature values using StandardScaler

2. Feature Exploration

• Correlation matrix between features and boiling point
• Scatter plots between each key descriptor and boiling point

3. Base Models

• Linear Regression using selected features
• Evaluation using Mean Absolute Error (MAE)

4. Additional Regression Models

The following regressors were implemented and tested:

• Ridge Regression
• Lasso Regression
• ElasticNet Regression
• K-Nearest Neighbors Regressor
• Support Vector Regressor (SVR)
• Decision Tree Regressor
• Random Forest Regressor
• Gradient Boosting Regressor

Each model was trained and evaluated using the validation set.

5. Hyperparameter Tuning

Grid search was used with cross-validation to find optimal hyperparameters:

• Ridge Regression: {'alpha': 10}
• Lasso Regression: {'alpha': 0.1}
• ElasticNet Regression: {'alpha': 0.1, 'l1_ratio': 0.9}
• K-Nearest Neighbors Regressor: {'n_neighbors': 5, 'p': 1, 'weights': 'distance'}
• Support Vector Regressor: {'C': 1, 'gamma': 'scale', 'kernel': 'linear'}
• Decision Tree Regressor: {'max_depth': 10, 'min_samples_leaf': 4, 'min_samples_split': 2}
• Random Forest Regressor: {'max_depth': 20, 'min_samples_split': 2, 'n_estimators': 200}
• Gradient Boosting Regressor: {'n_estimators': 200, 'learning_rate': 0.1, 'max_depth': 5}

6. Results

Boiling Point (BP) prediction was evaluated using three key metrics:

• MAE: Mean Absolute Error
• MSE: Mean Squared Error
• R²: Coefficient of Determination

Note: BP is measured in degrees Celsius.

Model	MAE	MSE	R² Score
Linear Regression	25.60	1213.80	0.820
Ridge Regression	16.17	519.43	0.921
Lasso Regression	14.27	430.53	0.935
ElasticNet Regression	14.16	429.01	0.935
K-Nearest Neighbors	19.64	781.71	0.882
Support Vector Regressor	14.57	441.36	0.933
Decision Tree Regressor	20.68	923.34	0.860
Random Forest Regressor	14.60	470.16	0.929
Gradient Boosting Regressor	14.79	464.63	0.930

Best Model

ElasticNet Regression performed the best overall, achieving the lowest MAE (14.16) and highest R² score (0.935).

Libraries and Tools Used

• pandas, numpy, matplotlib, sklearn