From Processing to Programmable Properties: A Review of Artificial Intelligence in Metal Additive Manufacturing

Machine learning has accelerated process–structure–property (PSP) prediction in metal additive manufacturing (AM), but a predicted optimum is useful only if it can be realized through processing and verified experimentally. This review examines that transition by connecting physics-based simulation, data curation, microstructure representation and reconstruction, and inverse design. We assess the information each stage retains, the assumptions it introduces, and the uncertainty it passes to subsequent decisions. Three tests organize the synthesis: structural fidelity, independently evaluated property performance, and accessibility through a specified processing route. Using these tests, we compare forward and inverse methods in terms of their data requirements, physical constraints, and validation evidence, distinguishing metal-AM demonstrations from transferable approaches developed in adjacent fields. Recurring limitations include simulation–experiment mismatch, incomplete dataset provenance, representation-dependent ambiguity, and insufficient testing outside the training domain. We outline benchmarks and reporting practices that make these limitations visible and connect them to priorities in physics-informed, multifidelity, interpretable, and multimodal learning. The resulting roadmap places experimental feedback at the center of the design workflow and identifies the evidence needed to translate rapid predictions into reproducible control of microstructure and properties.