The development of strategic materials such as spinels and perovskites is hampered by the long cycles and low efficiency of traditional trial-and-error methods, for which machine learning (ML) offers a disruptive data-driven paradigm. This review dissects the field's evolution from traditional machine learning (TML), reliant on manual feature engineering, to deep learning (DL) models capable of end-to-end autonomous feature extraction, examining their applications within forward screening and inverse design frameworks. We critically assess how DL overcomes TML's limitations by mapping atomic-level structures to macroscopic properties with greater precision, while also elucidating cutting-edge approaches to address the core challenge of high-quality data scarcity. However, we argue that fully realizing ML's potential requires shifting focus beyond predictive accuracy toward model reliability and robustness. In materials discovery and inverse design, the credibility of predictions depends on accurate uncertainty calibration. Integrating Bayesian learning and confidence-aware modeling can yield statistically reliable guidance for experiments and design decisions, enhancing the stability and trust of AI-assisted research. Future breakthroughs must pivot toward building standardized multi-modal databases, developing interpretable models that can unveil underlying physical mechanisms to overcome the "black-box" problem, and deeply integrating AI with automated experimental platforms to establish a closed-loop research ecosystem that truly accelerates scientific discovery.
Sun et al. (Wed,) studied this question.