| Title |
Overview of Machine Learning-Driven Design and Manufacturing of Metallic Powder Materials |
| Authors |
(Trung Thanh Pham) ; (Woo-Hyeok Kim) ; (Jeoung Han Kim) |
| DOI |
https://doi.org/10.3365/KJMM.2026.64.9.836 |
| ISSN |
1738-8228(ISSN), 2288-8241(eISSN) |
| Keywords |
Powder metallurgy; Additive manufacturing; Machine learning; Data-driven; Physics-informed learning; Neural network |
| Abstract |
Powder-based manufacturing, including powder metallurgy (PM) and additive manufacturing (AM), links powder production, handling, shaping or consolidation, post-processing, and final property development through a history-dependent process chain. The same nominal processing parameters can yield entirely different outcomes when changes occur in powder morphology, chemistry, reuse state, machine condition, sensing setup, geometry, or the post-processing route. This inherent complexity makes conventional trial-and-error optimization prohibitively costly and severely limits the transferability of empirical process rules. To address these challenges, machine learning (ML) is increasingly deployed to extract underlying patterns from powder descriptors, powder-bed images, sensor streams, process logs, simulation outputs, and mechanical property data. In this review, we examine ML applications across the entire powder lifecycle rather than treating them as isolated algorithmic tasks. The reviewed studies cover feedstock characterization and optimization, in-situ monitoring, defect and process-state detection, property prediction, powder reuse assessment, post-processing-aware modeling, and process optimization. Across these domains, ML proves most effective when predictions are directly traceable to the powder state, process history, intermediate signatures, and validation labels. The literature also highlights recurring barriers, including small and uneven datasets, costly high-fidelity labels, incomplete traceability, weak generalization across different powder lots or machines, and critical trade-offs among accuracy, interpretability, transferability, and qualification relevance. Data-efficient strategies such as transfer learning, multi-fidelity modeling, and physics-informed or hybrid learning can mitigate some of these limitations; however, their efficacy depends heavily on specific assumptions, namely source-target similarity, a meaningful fidelity hierarchy, or valid physical priors. Ultimately, future progress will depend less on algorithm complexity alone than on the curation of traceable lifecycle datasets, uncertainty-aware validation protocols, and models with explicit applicability limits. Under these conditions, ML can serve as a practical decision-support tool for powder-based manufacturing while ensuring that experimental validation and process qualification remain central. |