Machine Learning Based Classification of Non-functionalized MXenes into Metals and Non-metals
Introduction
This project deals with training classification models on physical properties of non-functionalized MXene molecules to classify them into metals and non-metals. The data required was obtained from aNANt database by IISc, Bangalore.
Data Visualization & Preprocessing
List of Features
Distribution of individual features
Corelation Heatmap
Feature Engineering
For each of the 5 atoms in the molecule we add the following properties to the molecule’s data:
- Neutron number
- Proton number
- Number of electrons
- Period
- Group number
- Atomic radius
- Electronegativity
- First Ionization Energy
- Density
- Melting point
- Boiling point
- Number of isotopes
- Number of shells
- Specific heat
- Mass Magnetic Susceptibility
- Molar Magnetic Susceptibility
- Thermal Conductivity
So, now, in addition to the 29 features we had earlier, we now have 85 new features (17 properties x 5 atoms per molecule). Therefore, a total of 131 features.
Distribution of individual features (After adding new features)
Correlation Heatmap
Class Imbalance
Results
Model : Random Forest (On oversampled data)
ROC Curve
ROC-AUC = 0.949
Accuracy = 93.2%
Model : Neural Network (Cost-sensitive)
ROC Curve
ROC-AUC = 0.962
Accuracy = 93.67%