Key points are not available for this paper at this time.
Additive manufacturing subjects every deposited layer to a thermal cycle of extraordinary severity. A laser or electron beam heats a melt pool past 2,000 °C and withdraws within milliseconds, leaving material to cool at rates reaching 10⁶ K s⁻¹. Multiplied across millions of such events in a single build, this generates microstructures, residual stresses, and porosity distributions unlike anything produced by casting or forging. Purely data-driven machine learning struggles to extrapolate reliably beyond its training conditions; physics-based finite element analysis is simply too slow for iterative design or real-time control. Theory-guided machine learning (TGML) resolves both problems simultaneously — embedding governing equations directly inside neural network architectures and achieving near-simulation accuracy from 4–10 times fewer experimental specimens at inference speeds orders of magnitude faster than FEM. This review presents a comprehensive synthesis to evaluate five TGML architecture families — physics-informed neural networks, thermodynamics-based neural networks, neural ordinary differential equations, hybrid physics–data surrogate models, and theory-embedded deep learning — under a single framework spanning all major AM process classes. For each architecture the review documents demonstrated successes alongside known failure modes, computational costs, and scalability constraints, correcting a tendency in prior surveys to emphasise successes exclusively. Accuracy is benchmarked across ten material–process–metric combinations; eight critical industrial-adoption gaps are identified with tiered priority ratings; and a three-horizon roadmap spanning 2025–2035 is presented with specific, verifiable deliverables. Drawing on 127 primary publications, the review provides a working framework for researchers, engineers, and funding bodies engaged with this rapidly evolving field.
Ignatius et al. (Wed,) studied this question.