AI-controlled wire arc 3D printing improves bead geometry consistency via adaptive learning

Wire Arc Additive Manufacturing (WAAM) is a metal 3D printing process that deposits weld beads layer by layer. Its complex, nonlinear dynamics—linking thermal fields to geometry—make consistent bead height and width difficult to achieve. In a new paper, researchers Chen-Lung Lu and John Wen treat WAAM as a multi-input/multi-output dynamical system (inputs: torch speed and wire feed rate; outputs: bead height and width). They use a recurrent neural network (RNN) trained on input/output data to predict bead geometry, then apply one-step-ahead predictive control for planning and real-time adjustment.

To handle changing thermal conditions as the part builds, the model is updated after each layer based on the prediction error from the previous layer. This adaptive step significantly improves prediction accuracy and control performance. Experiments on a robotic WAAM testbed with integrated line-scanner feedback demonstrated marked improvements in height and width consistency compared to both constant-input and static-model baselines. The framework offers a practical, data-driven pathway toward robust regulation of additive manufacturing processes without requiring complex physics-based models.