MACHINE LEARNING IN THERMO-ELECTROCHEMICAL CELL MATERIAL IDENTIFICATION: FEATURE SELECTION AND ANALYSIS

Abstract

This research shows the use of Machine Learning models to predict the seebeck coefficient, feature selection, and analysis of ionic thermoelectric material using different feature selection strategies. The approach comprises data collection of ionic thermoelectric material including (matrix + ion donor) combinations with features and target (seebeck coefficient) from different publish papers.  Applying different feature selection strategies with cross validation mean absolute error and mean square error to determine optimum feature subset which improve accuracy and generalization of model. Subsequently multiple models were trained on each respective feature subset. To evaluate their performance Decision Tree model was the best model exhibit high R² and low mean absolute error and root mean square error trained on univariate selected feature subset. It is revealed that seeebck coefficient is dominated over few strong predictors, adding more features reduce the accuracy of model and introduce noise or overfitting. This finding also expose that reduction of features significantly accelerates the discovery of matrix, ion donor combinations for thermoelectrochemical cell. Further to achieve additional superior robustness and generalization the top selected model was subjected to hyper parameter optimization process. SHapley Additive exPlanations (SHAP) and correlation analysis was performed to interpret model behavior, determined most influential features, and relationship between features and target seebeck coefficient. FractionCSP3 of the matrix and NumRotatableBonds of the ion donor were identified the most important features using SHAP analysis. It is also found that FractionCSP3 of the matrix show positive correlation, while NumRotatableBonds of the ion donor exhibit negative correlations with seebeck coefficient. The Decision Tree models trained on univariate selected features   predicted many promising combinations, especially polyurethane-based, cellulose-based, and PVA-based along with gelatin ionogel, and PAM hydrogel systems with predicted Seebeck coefficients up to 42.8 mV/K.