Application of Machine Learning for Defect Mitigation and Roughness Estimation in EHLA AM Process

Abstract This study presents a data-driven methodology for optimizing Extreme High-Speed Laser Cladding (EHLA) processing parameters with the aim of identifying defect-free conditions and controlling surface roughness. The proposed approach combines experimental analysis, machine learning and experimental validation. In the first stage, classification algorithms were evaluated to identify processing conditions associated with defect-free coatings. The Multilayer Perceptron (MLP) provided the best overall classification performance and was used to support the screening of feasible processing conditions. Correlation analysis identified process speed, overlap and mass flow rate as the parameters most associated with defect occurrence, while stand-off, overlap, process speed and laser power showed the strongest relationships with surface roughness. In the second stage, regression models were developed using a reduced parameter space comprising laser power, process speed and overlap. The MLP provided the best predictive performance for surface roughness among the evaluated models. Experimental validation confirmed the formation of defect-free coatings under the selected processing conditions, although prediction errors increased for some less-represented or extreme conditions. Multilayer deposition further demonstrated stable roughness and defect-free coating formation under the selected conditions.