Leveraging Self-Supervised Learning for Universal Active Learning in Materials Discovery

Active learning (AL) has been widely used to assist human materials scientists in achieving goals effectively by balancing exploration and exploitation. Many existing AL studies show that Bayesian optimization combined with feature engineering can conduct more efficient searches than human experts. However, preparing hand-crafted features is effort-intensive and suboptimal. Furthermore, since researchers apply AL for various purposes and scenarios, there is no consensus on benchmarks for comparing AL models in materials science. To address these challenges, we propose a neural network-based universal active learning framework that leverages self-supervised pretraining on large-scale unlabeled datasets. This approach eliminates the need for feature engineering and can be flexibly applied across various materials discovery scenarios. Additionally, we formulate several materials discovery scenarios as benchmarks for evaluating AL approaches. Our preliminary results indicate that the proposed method significantly enhances the sample efficiency of exploration in these scenarios.