Machine Learning Enabled Materials Design: From Empirical Trial-and-Error to Data-Driven Paradigm Shift

　　
　　Traditional materials research heavily relies on empirical trial-and-error approaches with long cycles and high costs. This review delineates the paradigm shift from "experience-driven" to "data-driven" propelled by machine learning (ML).​We commence with an overview of the principles and applications of three major categories of ML. Subsequently, using solid oxide fuel cell materials as cases, we elaborate on the implementation of ML across five scenarios: (1) data discovery—ML accelerates or even replaces traditional theoretical calculations and captures material data through text mining via large language models; (2) performance prediction—mapping "composition-structure-property" through feature engineering and ML models; (3) synthesis guidance—enabling closed-loop process optimization with self-driving laboratories; (4) characterization parsing—image recognition and reconstruction of material microstructures; (5) inverse design—generating candidate materials according to performance requirements. Finally, we address current challenges and outline future prospects for the development of ML on materials’ research.