Comparative performance analysis of optimization algorithms for hyperparameter tuning in LSBoost modeling of mechanical properties in FDM-printed nanocomposites

Abstract

Hyperparameter tuning is essential for developing accurate machine learning models, yet its role in predicting the mechanical properties of 3D-printed nanocomposites remains underexplored. This study evaluates the performance of three optimization techniques—Bayesian Optimization (BO), Simulated Annealing (SA), and Genetic Algorithm (GA)—in tuning a Least Squares Boosting (LSBoost) model for predicting the mechanical properties of fused deposition modeling (FDM) 3D-printed polylactic acid (PLA)/silica (SiO₂) nanocomposites. The properties assessed include modulus of elasticity ($E$), yield strength ($S_y$), and toughness at ultimate strength ($K_u$), influenced by key process parameters: extrusion rate (ER), SiO₂ nanoparticle concentration (SC), deposition layer thickness (LT), infill density (ID), and infill geometry (IG). Tensile specimens were produced using a Taguchi L₂₇ orthogonal array and tested under uniaxial tension. Tuning of the LSBoost model minimized a composite objective function involving root mean square error (RMSE) and (1 − $R^2$) loss metrics. Results demonstrated that GA achieved the best performance for yield strength prediction, with an RMSE of 1.9526 MPa and R² of 0.9713, while BO yielded the highest $R^2$ (0.9776) for modulus of elasticity prediction with a test RMSE of 130.13 MPa. For toughness, GA produced the lowest test RMSE of 102.86 MPa and the highest R² of 0.7953 among the optimization methods. Generally, GA consistently outperformed BO and SA in optimizing the LSBoost model across most mechanical properties, highlighting its effectiveness for hyperparameter tuning in the context of FDM-fabricated nanocomposites.