Machine learning (ML) has evolved as a revolutionary paradigm in additive manufacturing (AM). In metal AM, the final property of the material is governed by the complex process-structure-property (PSP) relationship. In the PSP relationship, processing conditions direct microstructure evolution, and microstructure further determines the final properties of printed components. Traditional physics-based simulation methods provide mechanistic insight but remain computationally expensive and difficult to scale across AM’s high-dimensional design space. Recent literature demonstrates that ML can learn nonlinear PSP relationships from simulation, in-situ monitoring, and experimental datasets, enabling rapid prediction of melt pool behavior, porosity, grain morphology, dendritic evolution, and resulting mechanical properties. The integration of microstructure as an intermediate state consistently improves predictive fidelity in structure-to-property models, where various image processing techniques significantly reduce data requirements and enhance generalizability. Further, ML driven inverse design has enabled bidirectional exploration of AM design space. ML demonstrated the ability to find optimal processing windows and propose new alloy compositions suitable for AM’s unique environments. The review further discusses how ML enabled surrogates can strengthen digital twin frameworks by enabling real-time monitoring and closed-loop control when integrated with in-situ sensor data. Overall, these improvements position ML as a framework that complements physics-based modeling and accelerates alloy discovery. The reviewed trends highlight key challenges, which include dataset heterogeneity, limited adaptation of advanced ML techniques, and insufficient integration of physics constraints. There is a need for standardized datasets and real-time in-situ prediction of microstructures.
Moses et al. (Wed,) studied this question.