Pohang University of Science and Technology develops AI framework that predicts metal 3D printed part strength in seconds with 4x accuracy improvement

Researchers at Pohang University of Science and Technology (POSTECH), in collaboration with the Korea Institute of Materials Science (KIMS), have developed a data-selective machine learning (DSML) framework that predicts the mechanical strength of metal 3D printed parts in seconds, even when internal defects are present. Led by Professor Kim Hyeong-seop from POSTECH’s Graduate Institute of Ferrous & Eco Materials Technology and Department of Materials Science and Engineering, and senior researcher Park Jung-min from KIMS, the team trained the AI on a dataset combining LPBF process parameters — laser power, scan speed — with microstructural data, pore size, and spatial distribution. Validation on AlSi10Mg alloy, widely used in aerospace and automotive, showed a mean prediction error of 9.51 MPa, which the team reports is four times more accurate than existing methods. The framework also generates human-readable equations rather than operating as a black box, enabling interpretable predictions. The research was published in Acta Materialia.

This work directly addresses a persistent bottleneck in metal LPBF qualification: the cost and time required to characterize how process-induced porosity affects mechanical performance. Rather than attempting to eliminate defects — an approach that has limited scalability in production — the DSML framework accepts stochastic porosity as a given and models its impact on yield strength. This aligns with the aerospace qualification grind pattern, where the path to certification often hinges on demonstrating predictable performance despite inherent process variability. The ability to generate defect-aware design maps could compress the iterative test-build-test cycle that currently dominates qualification for critical aerospace and automotive components. The framework’s interpretability is a meaningful differentiator versus black-box neural network approaches, as regulatory bodies and engineering teams require traceable reasoning for certification decisions.

From an industry standpoint, this is a software-layer innovation that does not require hardware changes, making it potentially deployable across existing LPBF systems. The practical next step is validation on additional alloys — such as Ti-6Al-4V or Inconel 718 — and on parts with more complex geometries and defect distributions. If the framework generalizes beyond AlSi10Mg, it could reduce the experimental burden for qualification in regulated verticals. For now, the key limitation is that the model’s accuracy depends on the quality and breadth of the training dataset; real-world adoption will require integration with in-situ monitoring data streams from production machines.