Physics informed neural network based prediction of hardness in FDM printed steel reinforced PLA composites

This study presents an integrated experimental, microstructural, statistical, and machine-learning framework for predicting the Vickers hardness of fused deposition modelling (FDM)-fabricated steel-reinforced polylactic acid (PLA) composites. Composite filaments containing 0, 10, and 20 wt% steel particulates were produced using twin-screw extrusion and subsequently printed by varying build orientation, infill density, and print speed. Scanning electron microscopy and EDAX analysis confirmed the incorporation of steel particles within the PLA matrix, with relatively uniform dispersion at lower reinforcement content and slight particle clustering at higher loading. Taguchi-based response analysis indicated that infill density was the most influential process parameter affecting hardness, followed by build orientation and print speed within the selected experimental range. Support Vector Regression (SVR), Random Forest (RF), and a Physics-Informed Neural Network (PINN) were developed to predict hardness. Among the evaluated models, the PINN achieved the best predictive performance, with RMSE = 1.177 and R² = 0.865, outperforming RF and SVR. The improved performance of the PINN is attributed to the incorporation of a physics-based monotonicity constraint into the learning process. The proposed experimental PINN framework provides a useful strategy for hardness prediction and process optimisation of metal-filled polymer composites fabricated by additive manufacturing.