Accelerated Search for BaTiO₃-Based Ceramics with Large Energy Storage at Low Fields Using Machine Learning and...

Prior research has shown that BaTiO3-based ceramics are promising candidates for energy storage applications. However, achieving high energy storage density at low electric fields remains a challenge. It was already known that traditional methods for optimizing compositions are time-consuming and inefficient. No prior work had resolved the optimal balance between physical intuition and data-driven approaches in material discovery. That uncertainty drove the need for new strategies that integrate machine learning with experimental validation. This gap motivated the development of methods that reduce search spaces while maintaining discovery potential. No prior work had demonstrated how varying the level of physical insight affects the efficiency of finding high-performance materials. This uncertainty highlights the importance of combining domain knowledge with automated search techniques.

The aim is to accelerate the discovery of BaTiO3-based ceramics with high energy storage at low fields. The specific problem is the inefficiency of traditional methods in exploring large compositional spaces. The motivation comes from the need to reduce both time and cost in material development. This study addresses the challenge by comparing two machine learning strategies. The first strategy uses full compositional space for prediction. The second strategy narrows the search to a specific crossover region. The goal is to determine which approach leads to faster and more effective discovery. The study also seeks to understand how physical intuition influences machine learning outcomes. This approach is designed to optimize the balance between exploration and exploitation.

Two machine learning strategies are compared in this study. The first strategy applies property prediction across all compounds in the space. The second strategy preselects compounds in a specific crossover region. Both methods use active learning loops to guide experimental synthesis. The strategies differ in how much physical insight is integrated into the model. Strategy I relies on broad data without prior domain knowledge. Strategy II uses phase diagram information to narrow the search. Experimental validation is performed after each learning loop. The feedback loop between prediction and synthesis is key to the design.

Strategy II identified a high-performing compound after only two learning loops. The compound (Ba0.86Ca0.14)(Ti0.79Zr0.11Hf0.10)O3 achieved 73 mJ cm-3 at 20 kV cm-1. This value is the highest reported for low-field energy storage in this class. The performance was measured using dielectric testing at the specified field. Strategy II outperformed Strategy I in both speed and accuracy. The crossover region in the phase diagram was critical to the success. The compound lies between ferroelectric and relaxor phases. This result supports the value of integrating domain knowledge with machine learning.

The authors propose that integrating physical insights with machine learning improves discovery efficiency. Strategy II demonstrated faster identification of high-performance materials. The success of Strategy II suggests the importance of narrowing search spaces. The compound found using Strategy II is the best performer in the study. The results support the use of feedback loops between prediction and experimentation. The authors suggest that this approach can be applied to other material systems. The study does not claim that Strategy II is universally superior. The findings highlight the value of combining domain knowledge with data-driven methods.