A Critical Examination of Active Learning Workflows in Materials Science

Abstract

Active learning (AL) is an increasingly important approach for data-efficient machine learning (ML) in materials science. It is widely used, from building training datasets to guiding autonomous materials discovery platforms. However, the performance of AL workflows depends on a number of often implicit design choices that are rarely examined systematically. Here, we critically analyze commonly used AL strategies in materials science, highlighting overlooked assumptions, hidden biases, and methodological limitations across different applications. Based on this, we provide practical guidelines to enhance the efficiency and reliability of AL workflows for materials science applications.