Abstract
Machine learning (ML) has emerged as a transformative approach to accelerate material discovery. A critical challenge in building predictive ML models is the selection of representative and informative training datasets from large databases. In this study, we present a framework that integrates inducing points and different acquisition functions within active learning (AL) campaigns to optimize the selection of training data. Specifically, we implement Gaussian process standard deviation (GP STD), alongside expected improvement (EI) and probability of improvement (PI) to guide the selection of high-impact data points that enhance model accuracy and reduce predictive uncertainty. We do so in the context of methane adsorption in metal-organic frameworks (MOFs). The selected MOFs are evaluated across key structural properties —void fraction (VF), largest cavity diameter (LCD), pore limiting diameter (PLD), and accessible surface area (SA), to ensure that the training data captures a broad spectrum of the design space. Through an intersection analysis of MOFs chosen by different methods, we identify a consensus set of 611 MOFs that appear across all strategies. This subset is used to train a Gaussian process regression (GPR) model for predicting CH₄ adsorption, resulting in a highly accurate model with an 𝑅² of 0.951 and a mean absolute error (MAE) of 5.11 cm³ (stp)/g framework. This multi-method approach demonstrates how ML-driven selection of diverse and informative MOFs can lead to robust predictive models, providing a pathway for efficient material screening and design.
Supplementary materials
Title
Supplementary Information for Multi-Method Material Selection for Adsorption using Inducing Points, Active Learning, and Bayesian Optimization Approaches
Description
Supplementary Information for Multi-Method Material Selection for Adsorption using Inducing Points, Active Learning, and Bayesian Optimization Approaches containing other results.
Actions