Paradigm evolution of intelligent discrete processes in specialty metallurgy: From experience-driven to AI4Science

Specialty metallurgy—including vacuum induction melting (VIM), electroslag remelting (ESR), and vacuum arc remelting (VAR)—is a critical production route for superalloys and special steels used in aeroengine turbine disks, marine gas turbine blades, and other high-performance components. Unlike continuous steelmaking, these processes feature small batches, high product diversity, long production cycles, and limited observability, creating distinctive barriers to intelligent manufacturing. This review traces the evolution of intelligent specialty metallurgy through five paradigms: empirical, theoretical, computational, data-driven, and AI4Science. Their technical foundations, application boundaries, and inherent limitations are systematically compared. The analysis shows that specialty metallurgy remains at an early stage of the data-driven paradigm, where sample scarcity, process discreteness, prolonged feedback cycles, and limited target quantification undermine the assumptions of conventional machine learning and restrict model transfer across grades, furnaces, and operating conditions. Drawing on advances in physics-informed learning, mechanism-data fusion, causal inference, and digital twins, this review argues that AI4Science provides a more suitable pathway than continued reliance on purely data-driven scaling. By integrating physics, data, and causality, AI4Science can constrain the hypothesis space, exploit simulation and expert knowledge, improve small-sample generalization, and support interpretable prediction and defect tracing. It should therefore be understood not as a replacement for the preceding paradigms, but as a meta-paradigm that integrates empirical knowledge, theoretical equations, numerical simulation, and machine learning. Future development should coordinate all five paradigms according to process maturity and progressively establish closed-loop workflows linking sensing, prediction, optimization, control, and quality feedback. Reported cases demonstrate high-accuracy prediction with only tens of target-domain samples and sub-second surrogate inference for otherwise time-consuming simulations, highlighting the potential of this framework to address unobservable melt-pool states, inclusion evolution, and segregation formation.