Development of a mathematical model to enable optimal efficiency of the indabuko lithium-ion battery.

Abstract

Cathode materials are the foremost primary challenge for the vast scale application of lithium-ion batteries in electric vehicles and the stockpiles of power. Foreseeing the properties of cathode materials is one of the central issues in energy storage. In the recent past, density functional theory (DFT) calculations aimed at materials property predictions offered the best trade-off between computational cost and accuracy compared to experiments. However, these calculations are still excessive and costly, limiting the acceleration of new materials discovery. Now the results from different computational materials science codes are made available in databases, which permit quick inquiry and screening of various materials by their properties. Such gigantic materials databases allow a dominant data-driven methodology in materials discovery, which should quicken advancements in the field. This study was aimed at applying machine learning algorithms on existing computations to make precise predictions of physical properties. Thus, the dissertation primary goal was build best ML models that are capable of predicting DFT calculated properties such as, formation energy, energy band-gap and classify materials as stable or unstable based on their thermodynamic stability. It was established that the algorithms only require the chemical formula as input when predicting materials properties. The theoretical aspect of this work describes the current machine learning algorithms and presents "descriptors"-representations of materials in a dataset that plays a significant role in prediction accuracy. Also, the dissertation examined how various descriptors and algorithms influence learning model. The Catboost Regressor was found to be the best algorithm for determining all the properties that were selected in this study. Results indicated that with appropriate descriptors and ML algorithms it is feasible to foresee formation energy with coefficient of determination (R2) of 0.95, mean absolute error (MAE) of 0.11 eV and classify materials into stable and unstable with 86% of accuracy and area under the ROC Curve (AUC) of 89%. Lastly, we build a web application that allow users to predict material properties easily.