Paradigm Evolution of Intelligent Discrete Processes in Specialty Metallurgy: From Experience-Driven to AI4Science

Specialty metallurgy, primarily including Vacuum Induction Melting (VIM), Electroslag Remelting (ESR), and Vacuum Arc Remelting (VAR) processes, constitutes the critical production route for superalloys and specialty steels deployed in high-performance applications such as aeroengine turbine disks and marine gas turbine blades. In contrast to conventional continuous steelmaking, specialty metallurgy is characterized by discrete production features—small batch sizes, high product mix, and extended production cycles—that pose unique challenges for intelligent manufacturing transformation. This review systematically traces the five-paradigm evolutionary trajectory of intelligent specialty metallurgy: from the Empirical Paradigm through the Theoretical Paradigm, Computational Paradigm, and Data Paradigm to the emerging AI4Science Paradigm, elucidating the technical foundations, applicable domains, and inherent limitations of each. A comprehensive literature analysis reveals that specialty metallurgy currently stands at the nascent stage of the Data Paradigm (Fourth Paradigm), yet structural constraints including sample scarcity, process discreteness, and prolonged feedback cycles preclude purely data-driven approaches from replicating their success in conventional steelmaking. Drawing upon frontier case studies in physics-constrained transfer learning and mechanism-data fusion, this review argues that specialty metallurgy should leapfrog the incremental development of the Fourth Paradigm and advance directly toward the AI4Science Paradigm—achieving reliable predictions under small-sample conditions through the ternary integration of physics, data, and causality. This review further proposes that AI4Science does not negate the preceding four paradigms but rather functions as a "meta-paradigm" that organically synthesizes empirical knowledge, theoretical equations, simulation data, and machine learning. Based on this perspective, future intelligent specialty metallurgy should promote synergistic development across all five paradigms, dynamically adjusting paradigm weights according to process maturity, ultimately forming an adaptive closed loop through multi-paradigm integration. Literature cases reviewed herein further indicate that AI4Science-type models can provide practical gains under data-scarce conditions, such as achieving high-accuracy prediction with only tens of target-domain samples and enabling second-level surrogate prediction for processes that would otherwise require hours or days of numerical simulation; more importantly, this paradigm helps open long-standing black-box problems involving unobservable high-temperature melt-pool states, inclusion evolution, and segregation formation.