Despite its promise, additive manufacturing (AM) remains constrained by recurring challenges: low build rates, insufficient reproducibility and a high sensitivity to statistical process variations. Over the past decade, researchers have compiled extensive defect taxonomies, investigated their detectability across a range of sensing modalities [1] and pursued both avoidance and correction strategies [2] [3]. Yet in many industrial use cases, critical defects, such as cracks, lack of fusion or local compositional anomalies, remain detectable only through destructive testing. This is neither scalable nor sustainable.

This plenary talk explores the potential of hybrid intelligence - physics-based process models enhanced by machine learning - in Laser Powder Bed Fusion (L-PBF) to move beyond reactive quality control toward proactive process resilience. We examine how machine learning can extend the applicability of physics-based approaches beyond the geometric and thermal constraints where traditional simulation becomes computationally challenging. Central to our discussion are three core questions:

What physical interactions drive defect formation? We examine how energy input, part geometry, material and powder-bed properties interact to produce defects.
How can we reduce defect probability before a build even starts, and during execution? We present approaches for adaptive parameter control, combined with real-time feedback and tailored to specific geometric or thermal constraints. We assess the role of in-process sensing (thermal radiation, visible light, and acoustic signals) and demonstrate closed-loop control at process-relevant timescales.
Can errors be reliably detected and remediated in-situ? We examine multi-modal process emissions as data sources for machine learning models that operate within physics-defined boundaries, and discuss the conditions under which such hybrid approaches offer genuine predictive value.

Finally, we address a critical question often overlooked in modern manufacturing: how can we design systems where artificial intelligence serves human judgment, not replaces it? We propose a shift from fixed parameter optimization toward Resilience by Design: adaptive systems that anticipate known error mechanisms, respond dynamically to process variations and evolve through validation. This requires three pillars: scientific rigor in process understanding, operational transparency in decision-making and human oversight in critical decisions. Hybrid intelligence plays a crucial role: not replacing physics-based knowledge or human judgment, but complementing them by learning patterns where traditional models reach their limits. The goal is not zero-defect manufacturing - a costly ideal in inherently variable processes - but rather manufacturing that is predictable, scalable and reliable.

References
[1] E. Duong, Dissertation RWTH Aachen, 2023
[2] T. Pichler, Dissertation RWTH Aachen, 2022
[3] O. Illies, Dissertation Universität Bremen, 2020
[4] A. R. Ziefuss et al., Materials & Design, 2024, 244, 113211.

Prof. Dr.-Ing. Florian Eibl