Probing the impact of process variables in laser-welded aluminum alloys: A machine learning study

Abstract

This paper presents a pioneering approach, Bayesian machine learning (ML), for the estimation and characterization of critical laser welding features in Aluminum alloys, encompassing peak temperature, heat-affected zone width, and bead aspect ratio. The methodology involved constructing a laser welding database utilizing finite element simulations (FEM). The distinctive advantage of the Bayesian ML model lies in its capability to address challenges associated with excessive approximation and to account for uncertainties in parameters. This results in precise and resilient predictions of laser weld parameters across a spectrum of aluminum alloys. The findings underscored the model's efficacy in forecasting output targets, although regression analysis unveiled unique characteristics in data distribution and outliers specific to aluminum alloys. These outliers were primarily linked with the melting range of aluminum alloys, leading to the Al7075 alloy having the lowest prediction, and the Al1100 alloy the highest within the ML model. Additionally, the normalized average weight functions of input parameters were illustrated, clarifying their differing importance concerning diverse types of Al alloys in precisely forecasting output objectives. In light of these explanations, it remains consistent that laser power (LP) and welding speed (WS) inputs hold substantial sway across all alloys, while workpiece thickness (WT), beam diameter (BD), and initial temperature (IT) played comparatively lesser roles. Ultimately, this work contributes to a more profound comprehension of the relationships between input features and the geometrical and thermal behavior of laser-welded Al joints.