Optimal machine learning feature selection for assessing the mechanical properties of a zeolite framework

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Zeolite mechanical properties like bulk and shear moduli are crucial for material design.
• Accurate prediction of these properties is essential for discovering new zeolites with desired functionalities.
• Existing methods for property prediction often lack sufficient accuracy and involve extensive computational resources.

Purpose of the Study:

• To develop a machine learning model for accurately predicting zeolite bulk and shear moduli.
• To identify critical features that govern the mechanical properties of zeolites.
• To accelerate the discovery of novel zeolites with enhanced mechanical characteristics.

Main Methods:

• Generated 896 zeolite descriptors using the matminer package.
• Utilized a database of 873 zeolite structures with density functional theory (DFT) calculated mechanical properties.
• Trained a LightGBM regression model using 45 and 249 critical features for shear and bulk moduli, respectively.
• Compared model performance against other regressors and feature sets.

Main Results:

• Identified critical features that significantly improve prediction accuracy for bulk modulus (17.3%) and shear modulus (10.6%).
• Reduced prediction uncertainty by one-third compared to previous methods.
• Demonstrated the robustness of the model through various training-test set ratios.
• Highlighted relationships between physical/chemical features and zeolite mechanical properties.

Conclusions:

• The developed machine learning model with critical features provides a highly accurate and efficient method for predicting zeolite mechanical properties.
• The identified critical features offer insights into structure-property relationships in zeolites.
• This approach can accelerate the exploration of vast hypothetical zeolite structures, leading to the discovery of advanced materials.