Neuro-Symbolic AI Controls Laser Powder Bed Fusion for Quality.

A novel neuro-symbolic closed-loop control architecture has been developed for laser powder bed fusion (LPBF), a key additive manufacturing process. This system integrates an in-loop ontology, which couples symbolic reasoning with statistical learning, to precisely set targets for a constraint-aware predictive controller. The core idea is to link process objectives and constraints to observable signals, allowing a description-logic reasoner to convert these into references and bounds for each scan. The architecture demonstrates its capability by addressing overhang dross, a critical quality limit related to melt pool depth that is difficult to measure during the build. It maps this unmeasurable quality metric to an observable width bound using a geometry- and power-dependent depth-to-width ratio, calibrated by a Gaussian process. The reasoner dynamically classifies upcoming features and applies active constraints, such as preventing lack-of-fusion at overhangs or enforcing energy-density caps. Simulations using a calibrated surrogate model for IN625 show that this neuro-symbolic approach eliminates dross produced by geometry-blind controllers. It maintains zero dross, degrades gracefully under plant mismatch, and can be retargeted to new alloys or constraints by simply editing ontology data, rather than requiring code changes. This research establishes the architectural feasibility, with experimental calibration being the next crucial step.