A critical review: Machine learning and advanced simulation in steel heat treatment

Steel heat treatment is governed by coupled heat transfer, diffusion, phase transformation, and stress development, yet industrial practice still relies on empirical diagrams and trial-and-error. This review critically examines machine learning and advanced simulation as design tools for that practice, spanning physics-based simulation, supervised learning, computer vision, physics-informed learning, surrogate modeling, inverse design, and digital twins. Quenching and tempering, thermochemical carburizing, localized induction and laser processing, press hardening, and additive-manufacturing post-processing are analyzed as separate mechanism-specific families rather than pooled as one process class. The evidence base is uneven: most studies rely on simulation or retrospective internal validation, only isolated reports include prospective component-scale tests, and none documents sustained closed-loop production control. Tree ensembles and kernel methods are strong baselines for the small, structured datasets typical of heat-treatment research, but no model family is universally superior. Validation design matters more than architecture: in one vision study, yield-strength R 2 rose from 0.43 under specimen-grouped validation to 0.97 under a leaky image split, and equation-generated targets merely emulate encoded correlations rather than validate physics. Hybrid models improve data efficiency and physical consistency, but evidence for extrapolation and plant transfer remains limited. A six-level validation hierarchy is proposed to separate random-holdout accuracy from grouped, external, temporal, prospective, and closed-loop evidence. Critical gaps persist in multimodal data integration, uncertainty quantification, causal process understanding, benchmarks, and safety-constrained control.