An Interpretable AI Model Can Now Predict How Defects in Metal 3D Prints Will Actually Affect Performance

Metal additive manufacturing - 3D printing of structural metal components - has produced some of the most complex and high-value parts in aerospace, defense, and medical applications. It has also produced a persistent quality problem: microscopic internal defects that form during the build process and can cause components to fail under stress. The question is not whether defects exist, but what to do about them.

Until now, quality assessment in metal additive manufacturing has relied heavily on a simple metric: porosity fraction. What percentage of the component's volume consists of voids? High porosity means low quality; low porosity means acceptable quality. This is a reasonable first approximation, but it throws away most of the information that actually matters. Two components with identical porosity percentages can have dramatically different mechanical properties depending on whether the pores are small and spherical or large and elongated, whether they are clustered near stress concentration sites or distributed evenly, whether they are at the surface or buried deep.

Moving Beyond Porosity Percentage

A research team led by Dr. Jeong Min Park of the Nano Materials Research Division at the Korea Institute of Materials Science (KIMS), in collaboration with Dr. Jaemin Wang and Prof. Dierk Raabe of the Max Planck Institute in Germany, developed an explainable AI model that characterizes defects by their morphological features - pore size, shape (non-circularity), spatial distribution - and directly links those features to mechanical properties. The results were published January 1, 2026, in Acta Materialia, one of the most prestigious journals in metallurgy, with an impact factor of 9.3.

The system uses microstructural images as input, automatically analyzing pore geometry and spatial arrangement. It then produces quantitative predictions of how those specific defect characteristics will affect mechanical performance - yield strength, ultimate tensile strength, fatigue life - and explains which features drove the prediction. Unlike conventional machine learning approaches that produce only a numerical output without justification, this explainable AI design shows the reasoning, making its predictions auditable.

Multi-Material Training Across the Full Process Chain

The research team assembled training data across three major metal additive manufacturing material classes - steel, aluminum alloys, and titanium alloys - and across multiple laser powder bed fusion (LPBF) process conditions. This multi-material approach is significant: a model trained only on one alloy system would have limited practical value, since the defect formation mechanisms and their mechanical consequences differ between materials.

The AI framework operates in two stages. The first stage assesses how process variables and powder characteristics influence defect formation - predicting what kind of defects will form under specific printing parameters. The second stage predicts how the resulting defect morphology will affect mechanical properties. Together, these stages create an integrated pipeline from process design to predicted performance: an engineer can evaluate multiple printing parameter combinations and rank them by predicted component quality before committing to a physical build.

"This research goes beyond simply reducing defects in metal 3D-printed components; it establishes a scientific framework that explains how specific types of defects directly influence performance," said Dr. Park. "We expect this work to contribute to the broader industrial adoption of metal additive manufacturing, particularly in high-performance sectors such as aerospace, space, and defense."

The Road Toward Industrial Deployment

The practical value proposition is substantial. Metal additive manufacturing parts can take hours or days to build. If a batch fails quality inspection, the material cost, energy cost, and time cost are all lost. An AI system that can predict likely defect characteristics and mechanical performance at the process design stage could reduce these losses significantly while also reducing the conservatism that engineers currently build into designs to compensate for quality uncertainty.

The research team's planned follow-up work focuses on extending this capability toward a digital twin-based quality management system suitable for industrial production environments. The current study provides the scientific foundation; translating it into a system that runs in real-time alongside a production LPBF machine, incorporating in-situ monitoring data, is the next engineering challenge.

As with any AI model trained on historical data, the framework's predictions are only as reliable as the quality and comprehensiveness of the training dataset. The model covers three material classes, but metal additive manufacturing encompasses many more alloy systems. Whether the defect-property relationships captured for steel, aluminum, and titanium generalize to other materials will require additional validation work.