Neuro-Symbolic Closed-Loop Control of Laser Powder Bed Fusion with an In-Loop Ontology

Laser powder bed fusion (LPBF) is highly sensitive to local geometry, particularly at overhangs where heat dissipation is poor, leading to defects like dross. Existing control methods are either geometry-blind (reacting only to sensor data) or feedforward-only (lacking real-time feedback). This paper asks: can we create a closed-loop controller that is both geometry-aware and responsive to in-process sensing by using symbolic knowledge to bridge the gap between unobservable quality metrics and observable sensor data?

The researchers propose a neuro-symbolic architecture that integrates a standards-aligned ontology directly into the control loop. The system uses a description-logic reasoner to classify geometric features (e.g., overhangs) and dynamically select appropriate constraints. Because the critical quality metric—melt pool depth—cannot be measured in real-time, the ontology uses a geometry- and power-dependent depth-to-width ratio (calibrated via a Gaussian process) to map depth limits onto observable melt pool width bounds. This allows a model predictive controller (MPC) to enforce quality constraints based on the specific geometric context of each scan.

In simulations using an Eagar-Tsai surrogate calibrated to NIST AM-Bench data for IN625, the neuro-symbolic controller successfully eliminated dross defects that occurred under geometry-blind control. The system maintained dross at zero while keeping lack-of-fusion defects minimal. Furthermore, the architecture demonstrated robustness to plant mismatches and allowed for retargeting to different materials by updating ontology data rather than rewriting control code, confirming the feasibility of the neuro-symbolic approach.

This work bridges the divide between symbolic AI and real-time process control in additive manufacturing. By making geometric knowledge explicit and inspectable, the system provides a path toward more reliable, adaptable, and interpretable manufacturing processes that can handle complex geometries without requiring extensive retraining for every new part or material.