Transfer learning for accelerating phase-field modeling of ferroelectric domain formation in large-scale 3D systems

High-throughput phase-field simulations emerge as a compelling technique to predict the evolution of domain structures in ferroelectric materials. Despite their potential, their widespread use is impeded by significant computational costs, particularly for large 3D systems, limiting applicability in real-size scenarios and extensive parameter exploration. Here, we present a machine-learning approach tailored for accelerating 3D ferroelectric phase-field modeling. Our model exhibits remarkable accuracy in forecasting domain formation trajectories, providing stable predictions throughout entire simulation paths. Furthermore, we introduce a transfer learning strategy that facilitates the model extension to larger systems in a computationally tractable way, broadening its applicability to realistic scenarios. With a relative error below 10% when compared to high-fidelity methods, our surrogate serves as an efficient replacement for ferroelectric phase-field. By achieving speed-ups of up to 3300-fold compared to direct numerical solvers, the presented approach contributes to significantly accelerate the discovery and design of ferroelectric materials.