PREDICTING THE STABILITY OF CLAY MINERALS USING XGBOOST MACHINE LEARNING: EFFICIENCY AND ACCURACY THROUGH COMPOSITION DESCRIPTORS

• Main Author: M. N. F. Syamsul, Andi
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/81557
• Institution: Institut Teknologi Bandung

Clay minerals are materials that have various benefits, particularly in their swelling properties and cation exchange capacity. These materials have diverse compositional variations, each resulting in different properties. With the advancement of Ab initio computational methods such as Density Functional Theory (DFT), many studies have been conducted to measure the properties of these materials computationally. However, due to the numerous compositional variations of clay minerals, research using DFT methods often faces challenges. One important property that needs to be measured in determining the composition and structure of clay minerals is their thermodynamic stability. In this study, machine learning methods are used to predict the stability of clay minerals using composition descriptors. The machine learning method eXtreme Gradient Boosting (XGBoost) was chosen because it is known for its efficiency and high accuracy in performing regression. Three types of composition descriptors selected are Magpie, Fraction, and MatScholar. Each of these descriptors represents unique properties of the material’s composition: Magpie focuses on elemental stoichiometry, Fraction provides information about the fraction of elemental composition, and MatScholar uses vector embeddings based on scientific literature. Stability measurement was conducted by calculating the formation energy values using the XGBoost method. The validation process was done by comparing the formation energy values of clay mineral types, namely pyrophyllite and magnesium montmorillonite, obtained from DFT methods with the results from XGBoost predictions. The validation results show that XGBoost performs well in measuring the formation energy of clay minerals. Additionally, feature importance analysis of the descriptors provides insights into the physical properties that most affect the formation energy. These important features include maximum electronegativity, melting point, and maximum embedding from scientific literature, all of which have direct correlations with material stability. Thus, this study not only proves the effectiveness of the XGBoost method in predicting the stability of clay minerals but also helps in understanding the features that determine the stability of these materials.