10–12 Nov 2026
Arcotel Wimberger
Europe/Vienna timezone

Physics-Guided Gaussian Process Classification of Cracks in LPBF with Ti-6Al-4V Parts

12 Nov 2026, 12:00
20m
Room 2

Room 2

Oral Presentation Digitalisation, Artificial Intelligence & Simulation Simulation

Speaker

Ms Emily Esmeralda Carvajal Camelo (Materialise, LMSD Lab, KULeuven)

Description

Laser powder bed fusion (LPBF) produces complex metal alloy geometries and ready-to-use prototypes for industrial applications; However, this geometric freedom still depends heavily on preparation expertise and process tuning to achieve a successful first print. Among the failures that are induced by the manufacturing process cracks are especially critical, as they can lead to part rejection and are difficult to identify during early build preparation. This work presents a physics-guided feature engineering classifier based on industrial LPBF production using geometrical and simulation derived features.
The curated dataset consists of 127 LPBF cases printed in Ti-6Al-4V, including 68 cracked and 59 uncracked parts. The cases are relabeled from industrial production records using manufacturing inspection records into uncrack and crack cases. Each part is represented by 26 input features combining AM specific geometrical descriptors with residual stresses and strains extracted from Multiphysics numerical simulations of the printing process. The geometrical descriptors capture design characteristics relevant to LPBF printability, while the simulation-derived features encode mechanical indicators associated with the development of cracks. These features are selected using expert knowledge to preserve the representation physically meaningful and suitable for classification.
A Gaussian Process classifier is trained to distinguish cracked from uncracked cases. Using a train-test split, the model achieved 90% accuracy and precision for crack classification with a recall of 92% for the crack cases. These results indicate that there is an identifiable pattern between complex industrial geometries when using the correct simulation derived fields and specialized geometrical features for a specific failure case, such as cracks.
This proposed method supports early defect assessment after the build preparation and before printing for metal LPBF, specifically Titanium alloy, potentially reducing material scrapping as well as overall production costs.

Speaker Country Belgium, KULeuven
Would you like to publish your paper in the special issue of BHM "Berg- und Hüttenmännische Monatshefte" Yes

Authors

Ms Emily Esmeralda Carvajal Camelo (Materialise, LMSD Lab, KULeuven) Dr Michele Pavan (Materialis) Prof. Iuri Rocha (SLIMM Lab, Faculty of Civil Engineering and Geosciences, Delft University of Technology) Prof. Frank Naets (Mecha(tro)nic System Dynamics (LMSD), Department of Mechanical Engineering, KULeuven)

Presentation materials

There are no materials yet.