Comparative evaluation of machine learning regressors for the layer geometry prediction in wire arc additive manufacturing

Abstract

In this paper, a set of the most employed machine learning (ML) algorithms were trained and tested to assess which ones present the highest accuracy in predicting the layer geometry of the Ti-6Al-4V processed by plasma transfer arc deposition. Wire and arc additive manufacturing brings about the possibility of manufacturing large and robust components based on metal wires. One of the critical aspects to take into account during the manufacturing process is the layer geometry. Bead geometry depends on several processing parameters, e.g., arc voltage, welding current, travel speed, wire feed speed, and gas flow rate. The algorithms that better adjusted the prediction were multi-layer perceptron with five hidden layers, linear support vector regression, and boosting regressors, which combine multiple models to reduce overfitting risk.