Machine learning and finite element-based design of sustainable industrial waste-filled glass/epoxy composites with enhanced mechanical properties

The steel making industry generates significant amounts of waste, particularly in the form of Linz-Donawitz slag (LDS). This study evaluates the feasibility of recycling these wastes as raw materials for the manufacturing of polymer composite. Composite specimens were fabricated with 20 wt% short glass fiber (SGF) and different weight percentages of LDS particles from 0 to 22.5 wt%. A sharp increase in mechanical properties is observed with an addition of 20 wt% SGF. The tensile strength, tensile modulus, flexural strength, and impact strength rise to 221.37 MPa, 5.41 GPa, 204.23 MPa, and 28.7 kJ/m 2 , respectively. An increase in the LDS content was accompanied by a marginal decrease in strength values, however, because of the aggravation between the slag particles. This work uses three multi-output regression models Polynomial Regression (PR), Random Forest Regression (RF), and Neural Network (NN) to establish mathematical correlations between LDS content (0–22.5 wt%) and four mechanical properties tensile strength, tensile modulus, flexural strength and impact strength. Polynomial Regression offers continuous expressions of interpolation and optimization. Random Forest Regression is capable of capturing a threshold effect and a stage-wise degradation without any global mathematical forms. Neural Networks model complex non-linear interaction by using hidden layers. Findings indicate that there are significant inter-property correlations (r > 0.93) with 1.00 correlation existing between tensile modulus and impact strength. Negative interactions between short glass fiber (SGF) and other components result in amplified degradation of hybrid systems with 20 wt% short glass fiber. Each of the three models can predict untested compositions quantitatively and speeds up the development of composite materials. Finite element simulations performed in ANSYS were used to estimate peak stress of the specimens. The predicted results showed close agreement with experimental data, with 1.486%, and 5.333% errors for tensile and flexural strength, respectively, which confirm good correlation between numerical and experimental approaches.