13/08/2026
I’m excited to share a comprehensive review published in Materials Research Letters, titled "Applications and challenges of machine learning in metal additive manufacturing."
Led by Associate Professor Wenchao Ke (Wuhan University of Technology) and first author Xianzhe Peng, with collaboration from Huazhong University of Science and Technology, this work systematically maps how Machine Learning (ML) is transforming the entire lifecycle of Metal Additive Manufacturing (MAM).
The Core Challenge: MAM is a multi-physics, non-equilibrium process where parameters like laser power, scan speed, and layer thickness create highly nonlinear Process-Structure-Property (PSP) relationships. Traditional trial-and-error methods struggle to achieve consistent quality and scalability.
Key Applications Covered:
🔹 Process Optimization: Moving beyond Taguchi methods and Volumetric Energy Density (VED). Bayesian optimization and active learning now enable efficient multi-objective searches across high-dimensional parameter spaces.
🔹 Defect Detection: Shifting from post-build inspection to in-situ monitoring. Multi-modal sensor fusion (thermal imaging, acoustic emission, visible light) combined with CNNs allows real-time identification of pores, cracks, and spatter.
🔹 Performance Prediction: Linking thermal history, microstructure, and mechanical properties. Models now predict strength, fatigue life, and magnetic performance directly from process signatures and SEM images.
🔹 Intelligent Design: Accelerating topology optimization and lattice structure generation using conditional GANs and genetic algorithms—dramatically shortening design iteration cycles.
The Game Changer – Physics-Informed ML (PIML):
Pure data-driven models risk physical inconsistency; pure physics models are computationally heavy. PIML bridges this gap by embedding conservation laws, boundary conditions, and simulation data into neural networks (e.g., PINNs for 3D temperature prediction without labeled data).
Critical Challenges Ahead:
The authors highlight five bottlenecks: scarce high-fidelity data, model "black-box" nature, poor cross-machine generalization, real-time deployment latency, and industrial safety validation.
Future Outlook:
We need public benchmark datasets, explainable AI (XAI), digital twins with closed-loop control, and edge-deployable lightweight models. The ultimate goal? A fully predictable, interpretable, and adaptive smart manufacturing ecosystem.
This review is a must-read for anyone working at the intersection of materials science, AI, and advanced manufacturing.
