Comparative Analysis and Experimental Validation of Statistical and Machine Learning-Based Regressors for Modelling the Surface Roughness and Mechanical Properties of 316L Stainless Steel Specimens Produced by Selective Laser Melting.

Abstract

This article analyzes the surface roughness and mechanical properties of 316L samples produced by Selective Laser Melting (SLM) through the application of statistical regression and Machine Learning techniques. Response Surface Methodology, Multi-layer Perceptron, Support Vector Regression, Random Forests, Gaussian Process, and Adaptive Neuro-Fuzzy Inference System were fitted and compared to predict essential features in laser-printed metallic parts. Using the Box- Behnken design of experiment (DOE), a range of process parameters was selected to assess the impact of laser power, scanning speed, hatch spacing, and layer thickness on the surface roughness and mechanical properties. It is worth noticing that the simultaneous inclusion of hatch spacing and layer thickness as input variables hasn’t been widely investigated in the literature. The generalization capabilities of the fitted models were verified with additional experimental points not included in the training and testing procedures. The comparison conducted shows that there is no unique model capable of accurately predicting all the studied variables. It was necessary to consider more than one technique for reliable forecasting. Besides, combining the most accurate regressors into ensemble formulations increased, in some cases, the predictability of the stand-alone models themselves. Therefore, the two main contributions of this paper rely on the fitting of accurate modeling methods and characterizing the behavior of printing parameters on crucial features of the SLM process. Moreover, the presented modeling strategies could be used as references for further studies on the topic, particularly in optimization frameworks where these models could be implemented either as targets functions or constraints.