Results in Physics (Jul 2021)
Bridging the analytical and artificial neural network models for keyhole formation with experimental verification in laser melting deposition: A novel approach
Abstract
Recent scientific and technological developments demonstrated that laser melting deposition (LMD) could yield near-net-shape parts by additive manufacturing. Under context, it was noticed that the primary operating parameters (i.e., laser power, scanning speed, and powder flow rate) significantly influence the generated keyhole's dimensions, ultimately affecting the deposited clad characteristics. In this study, a novel approach is proposed to generate a pool of datasets to implement artificial neural networking (ANN) in manufacturing process automation. A mathematical model was developed to approximate the keyhole's top and bottom widths and penetration depth, depending on the primary operating parameters. Single-layers of AISI 304 stainless steel were deposited via LMD to verify the mathematical results. It was shown that the mathematical model had predictions close to the experimental results. The validated model was used in correlation with the ANN model based upon 3-10-3 (no. of inputs-no. of neurons in hidden layer-no. of outputs) architecture. The outputs based on the given inputs, via the verified mathematical model, were used for the ANN model training. The results of experiments were compared with mathematical and ANN models. It was found that an increase in the laser scanning speed decreases the deposited layer's width, and a direct correlation has been inferred between powder flow rate and laser power with layer width. An increment in the scanning speed reduces the layer height and keyhole's dimensions. A direct correlation has been observed between powder flow rate and layer height, irrespective of the keyhole's dimensions. Laser power was in a direct relationship with layer height and keyhole dimensions. The developed mathematical model can estimate the keyhole's dimensions with an accuracy of (5–9) %. However, the output predicted for keyhole's dimensions by ANN, in a range of (2–3.5) %, is much closer to the experimental results, which identifies the ANN as a potential tool for the 3D printing process automation.
