Efficient Microstructure-Property Exploration with Active Learning and Gaussian Process Regression

The advancement of materials design is often constrained by the computational expense of physics-based simulations, limiting the generation of comprehensive and diverse datasets. Addressing this limitation involves improving sampling strategies to more efficiently navigate the material design space and identify promising material configurations with the minimum number of simulations. This work introduces a data-driven active learning strategy that integrates probabilistic surrogate modeling methods, Gaussian process regression (GPR) model, and acquisition techniques guided by information theory, correlating underlying microstructural arrangements with resulting material behavior. We first expand the design space by including microstructures with diverse microstructural features. In each active learning iteration, samples with the most informative configurations are identified for detailed physical simulation. As the GPR model is iteratively updated, the model’s predictive capabilities are continuously enhanced, enabling both efficient exploration of the design space and a deeper understanding of structure-property relationships and capturing the dependencies for localized damage in diverse material systems including energetic materials and ceramics. Compared to conventional exhaustive sampling, our approach significantly minimizes the need for computationally intensive physics-based simulations while maintaining robust predictive power and uncertainty quantification. It also enables targeted material design and optimization, ultimately promoting the discovery of superior material systems.