Machine Learning for Predicting Dielectric Energy-Storage Properties of Perovskite Oxides

Perovskite oxides with the ABO₃ structure are core candidates for dielectric energy-storage capacitors, whose recoverable energy density is governed by the dielectric constant, the breakdown field strength, and the nonlinear polarization response. This review surveys recent machine-learning studies on predicting these properties of oxide perovskites, covering three representative quantities—the dielectric constant, the breakdown field strength, and the ferroelectric Curie temperature—together with direct performance indicators such as recoverable energy density and efficiency. The works are compared along the evolution of methods, from manual descriptors and ensemble learning to graph neural networks, small-sample strategies, and generative inverse design, with dataset size, verification level, and accuracy documented. Three insights emerge: physical descriptor models remain competitive and interpretable under small-sample conditions; equivariant graph neural networks offer advantages for tensorial and finite-temperature tasks at higher data and computational cost; and generative inverse design is emerging, with preliminary success at the ferroelectric mesoscopic scale, but remains limited by scarce multimodal data and incomplete experimental-validation loops, which constitute the main challenges ahead.