Exploring active learning strategies for predictive models in mechanics of materials

Machine learning (ML) has found widespread applications in predicting material properties and mechanical behaviors across various scales in computational materials science. This data-driven approach typically relies on large datasets to train predictive models. However, labeling data samples through numerical simulations in materials science can be computationally intensive. In response to this challenge, this research delves into the utilization of active learning (AL) strategies to selectively label the most informative data samples for regression and classification models. Additionally, Several novel AL strategies were developed to enhance the development of probabilistic ML models. Through illustrative examples, this study demonstrated that AL could significantly boost ML training efficiency by labeling only a small subset of data samples while achieving exceptional model performance.