Predicting mechanical properties from co-axial melt pool monitoring signals in laser powder bed fusion

The critical issue of part quality repeatability within and across multiple builds during metal additive manufacturing techniques presents a major roadblock to their adoption for large-scale manufacturing, particu-larly for sensitive applications like aerospace and nuclear. The property variations are a consequence of process fluctuations, which can potentially be measured using in-situ sensors. Thus, in-situ monitoring and control provide a pathway for alleviating this issue; however, a mapping between the in-situ signatures and the part properties is crucial to this endeavor. Most previous studies focus on employing in-situ signatures for predicting porosity and defects. The present work develops machine learning based surrogate models for predicting relative density, work hardening exponent, elongation to fracture, and uniform elongation using co-axial melt pool monitoring for IN718 tensile bars printed using GE Concept Laser M2. Autoencoders, along with 1st-order and 2nd-order statistics, are employed to extract representative features from the in-situ data. Non-parametric regression models are developed and optimized using active learning and grid search. The models provide ac-curate predictions for the ductility-related properties and are able to distinctly resolve the high-density and low -density samples. The models capture the property variations about a fixed set of printing parameters and can potentially be employed for monitoring and part qualification, reducing the reliance on post-build testing. Lastly, the model predictions show a high correlation to the true value of the properties and can thus be used for providing feedback signals for designing closed-loop control to ensure better part quality repeatability.

Automated summary

The critical issue of part quality repeatability in metal additive manufacturing techniques, particularly for sensitive applications like aerospace and nuclear, can be alleviated by in-situ monitoring and control. This study develops machine learning models based on in-situ monitoring to predict important properties of printed parts. The models accurately predict ductility-related properties and distinguish between high-density and low-density samples.