Machine learning based prediction of thermal conductivity in graphene-reinforced aluminum metal matrix composites

Graphene-reinforced aluminum metal matrix composites (Al–Gr MMCs) are attractive candidates for thermal management applications because they pair the lightweight nature of aluminum with the high intrinsic thermal transport of graphene. However, experimental characterization is frequently constrained to narrow composition windows due to nanoparticle agglomeration, interfacial thermal resistance, and synthesis difficulties. To address this limitation, this study develops a data-driven modeling framework based on a 308-record dataset (spanning 0.0 to 10.0 wt.% graphene and thermal conductivities from 150.0 to 539.0 W/m·K) anchored on in-house experimental investigations and standardized via parametric curve discretization. A Dual-Track Prediction Engine is implemented to separate interpolation within the 0.0–10.0 wt.% domain from extrapolation at higher loadings (10.0–13.0 wt.%). Track A employs an Optuna-tuned weighted ensemble of Random Forest, LightGBM, and CatBoost regressors, whereas Track B combines fourth-degree Polynomial Regression (85% weight) with Gaussian Process Regression (GPR, 15% weight) to project continuous trends and quantify posterior predictive uncertainty ( ± 1 σ ). Evaluated on an 80:20 stratified train–test split across six concentration bins, the Track A weighted ensemble achieves a coefficient of determination ( R 2 ) of 0.9987, a root mean squared error (RMSE) of 3.40 W/m·K, and a mean absolute error (MAE) of 2.65 W/m·K, outperforming each individual base learner in reconstructing the composition–conductivity trajectory. Sensitivity analysis of the predicted curve shows that marginal thermal conductivity gains are highest below 1.0–1.5 wt.% (up to 10.4% per wt.%) before diminishing at elevated concentrations, while Track B provides smooth exploratory projections beyond 10.0 wt.% without boundary oscillations. These results demonstrate how combining tree-based ensemble learning with probabilistic Gaussian Process Regression supports preliminary composition screening in Al–Gr MMCs.