Purpose
The purpose of this work is to explore predictive model approaches for selecting laser cladding process settings for a desired bead geometry/overlap strategy. Complementing the modelling challenges is the development of a framework and methodologies to minimize data collection while maximizing the goodness of fit for the predictive models. This is essential for developing a foundation for metallic additive manufacturing process planning solutions.
Design/methodology/approach
Using the coaxial powder flow laser cladding method, 420 steel cladding powder is deposited on low carbon structural steel plates. A design of experiments (DOE) approach is taken using the response surface methodology (RSM) to establish the experimental configuration. The five process parameters such as laser power, travel speed, etc. are varied to explore their impact on the bead geometry. A total of three replicate experiments are performed and the collected data are assessed using a variety of methods to determine the process trends and the best modelling approaches.
Findings
There exist unpredictable, non-linear relationships between the process parameters and the bead geometry. The best fit for a predictive model is achieved with the artificial neural network (ANN) approach. Using the RSM, the experimental set is reduced by an order of magnitude; however, a model with R2 = 0.96 is generated with ANN. The predictive model goodness of fit for a single bead is similar to that for the overlapping bead geometry using ANN.
Originality/value
Developing a bead shape to process parameters model is challenging due to the non-linear coupling between the process parameters and the bead geometry and the number of parameters to be considered. The experimental design and modelling approaches presented in this work illustrate how designed experiments can minimize the data collection and produce a robust predictive model. The output of this work will provide a solid foundation for process planning operations.
Development of Supervised Learning Predictive Models for Highly Non-linear Biological, Biomedical, and General Datasets