Machine learning approach for mechanical property assessment of industrial waste-filled epoxy-jute composites

The growing demand for sustainable materials has stimulated the development of bio-based composites, yet the combination of natural fibers and industrial waste fillers in polymer matrices has not been exploited synergistically. In line with Sustainable Development Goal 12 (SDG-12), the paper focuses on reusing the steel industry by-product Linz-Donawitz (LD) sludge to enhance the mechanical properties of epoxy composites when used with the jute fiber. Six composite specimens with constant jute fiber loading (20 wt%) and a range of LD sludge content (0-25 wt%) were prepared using hand lay-up technique. The best composition (60 wt% epoxy, 20 wt% jute, 20 wt% LD sludge) resulted in tensile strength of 61.84 MPa (28.8% better than neat epoxy), flexural strength of 31.81 MPa (41.8% better) and impact strength of 18.026 kJ/m². Interfacial defects and agglomeration of particles led to a decrease in mechanical properties beyound 20 wt% sludge. This experimental data was used to train four machine learning models to forecast mechanical properties given compositional inputs. On training data, XGBoost achieved R² = 1.0000 with near-zero errors (MAE = 0.0005 MPa, RMSE = 0.0008 MPa). However, when trained on a small dataset of six specimens, this perfect fit is mostly due to memorization of the training data, as opposed to predictive power. The findings suggest the risk of overfitting, mainly in the cases of Decision Tree and Gradient Boosting models. More realistic estimates of model performance are given by cross-validation (R² = 0.94 ± 0.04 in the case of XGBoost). The ML models can thus be used to analyze exploratory composition-property trend analysis in this particular composition space, as opposed to extrapolative prediction. These results both validate the possibility of hybrid composites that use industrial waste to obtain mechanical performance equivalent to standard natural fiber composites and indicate that waste can be valorized, although any assertion of ML predictive capacity should be carefully hedged due to limitations in the datasets.