Microstructural characterization and data-driven wear modelling of LPBF AlSi10Mg based on pin-on-disc studies

Abstract AlSi10Mg’s superior strength-to-weight ratio has attracted significant interest from the defence, automotive, and aerospace industries. Optimizing material performance and enhancing component reliability require accurate prediction of wear behaviour under various operating conditions. This work models and predicts the wear loss of additively manufactured AlSi10Mg produced by Laser Powder Bed Fusion (LPBF) using a data-driven approach. A total of 16 experimental samples were evaluated based on an L16 experimental design, considering applied load, sliding distance, and sliding speed as the primary test variables. The measured tribological responses included wear rate, coefficient of friction (COF), and friction force. To assess the material’s tribological performance, pin-on-disc wear experiments were conducted at various loads, sliding speeds, and sliding distances. Microstructural characterization was performed before and after wear testing, including analyses of porosity and grain size, to investigate the relationship between material structures and wear behaviour. Machine learning models based on Linear Regression, Support Vector Regression (SVR), Random Forest, Gradient Boosting, Gaussian Process, Decision Tree, Extreme Gradient Boosting (XGBoost), Artificial Neural Network (ANN), and K-Nearest Neighbors (KNN) were then developed and compared using the experimentally recorded wear loss data. An 80:20 train–test split was adopted for machine-learning analysis, with the models developed specifically to predict wear loss/wear rate from the experimental input variables. The combined experimental, microstructural, and data-driven approach provides useful insights into the wear behaviour of LPBF-fabricated AlSi10Mg. Gradient Boosting and Gaussian Process demonstrated promising predictive performance, while the results from the 16-experiment L16 dataset highlight the potential of machine learning for preliminary wear prediction. Further experimental data can support the development of more robust and generalizable predictive models.