Tribological performance optimization of AZ91/TiB 2 magnesium matrix composites using stacked ensemble learning and Bayesian methods

Abstract

This work aims to develop and optimize lightweight, wear-resistant AZ91 magnesium matrix composites reinforced with titanium diboride (TiB₂) for demanding tribological applications. A hybrid ultrasonic-assisted mechanical stir casting technique was used to fabricate composites with 0–4 wt.% TiB₂, ensuring uniform reinforcement dispersion. The tribological behavior was systematically evaluated under dry sliding conditions using a Taguchi L18 orthogonal design with sliding speed, normal load, and TiB₂ content as variables. To address the nonlinear interdependencies among these factors and predict key tribological responses, wear rate (WR) and coefficient of friction (COF), a stacked ensemble learning (SEL) model was developed, integrating random forest, gradient boosting, and a polynomially expanded meta-regressor. The SEL model achieved high prediction accuracy (R² = 0.948 for COF, 0.908 for WR). Subsequently, Bayesian Optimization (BO) based on Gaussian Process Regression was employed for multi-objective process parameter optimization. The optimal combination—13.77 N normal load, 1.94 m/s sliding speed, and 3.98 wt.% TiB₂, yielded experimentally validated WR and COF values of 10.235 mg/km and 0.326, with <5% deviation from predictions. SEM analysis revealed mild abrasive wear, delamination and localized grain refinement as key mechanisms. This study conclusively demonstrates the efficacy of coupling advanced machine learning with experimental validation for intelligent design of Mg-based composites. The proposed SEL–BO framework reduces experimental workload while ensuring high-performance tribological outcomes, supporting its application in aerospace and automotive systems where both weight and surface durability are critical.