Machine-Learning-Guided Discovery of Thermally Anchored Polar Heterostructures for Harsh-Environment Capacitors

Abstract

The integration of dielectric capacitors into advanced power electronics requires reliable operation at elevated temperatures. However, the trade-off between recoverable energy density (Wrec) and efficiency () limits existing materials because thermal fluctuations can destabilize nanoscale polar structures and promote hysteretic ferroelectric states. Here, a machine-learning-guided strategy is used to identify a lead-free composition exhibiting a thermally anchored polar heterostructure (TAPH) that mitigates thermal degradation. A predictive model based on structural descriptors identifies 0.6Bi0.47Na0.47Ba0.06TiO3-0.4Bi0.20Sr0.70Hf0.20Ti0.80O3 as the optimal composition. Atomic-scale characterization reveals a heterogeneous polar state comprising tetragonal-, rhombohedral-, and cubic-like nanoregions. Local lattice distortion, Bi-containing bonding environments, and interphase electric and elastic fields are considered to cooperatively hinder the thermally driven coalescence of polar regions, while preserving the dynamic response of local dipoles. This heterogeneous polar state is associated with improved breakdown strength, low hysteresis, and stable high-temperature energy-storage performance.The resulting multilayer ceramic capacitors (MLCCs) deliver a Wrec of 24.12 J cm⁻³ with an η of 95% at 1160 kV cm⁻¹. From 20 to 200 °C, the devices maintain a Wrec of approximately 15.5 J cm⁻³ within ± 7.8% and an η of 95.6% within ± 6.8%. This thermally anchored structural framework provides a data-guided route to energy-storage dielectrics for high-temperature applications.