Active learning-enabled first principles mapping of full-strain electromechanical responses in PbTiO3

Abstract

Deep elastic strain engineering has been recognized as a powerful route for tailoring electromechanical responses and for uncovering new functional properties in ferroelectrics enabled by extreme yet fully reversible elastic distortions. However, a systematic exploration of ferroelectric behavior across the full six-dimensional strain space—characterized by highly nonlinear and coupled evolutions of stress and polarization—has proved fundamentally challenging. Here, we develop a computational framework that integrates neural network regression, active learning, and high-throughput first principles calculations to map multi component electromechanical responses of PbTiO3 throughout the nonlinear large deformation regime. By exploiting crystallographic symmetry and entropy-driven sampling, the framework constructs an accurate predictive model using only 0.36% of the full strain space configurations. The model successfully captures both continuous and discontinuous electromechanical responses, with its accuracy validated against first principles calculations for strain conditions absent from training. This methodology provides a general strategy for efficiently navigating high dimensional property spaces, offering a powerful tool for deep elastic strain engineering in functional materials.