Machine learning enhanced coefficient of thermal expansion (CTE) of graphene incorporated aluminium based metal matrix composite

Graphene-reinforced aluminum metal matrix composites (Al–Gr MMCs) have attracted significant attention for high-temperature structural applications due to their potential to combine low density with improved dimensional stability. While aluminum exhibits favorable mechanical properties and processability, it undergoes considerable thermal expansion at elevated temperatures. In contrast, graphene possesses low or negative thermal expansion characteristics and can act as a thermal constraint within the matrix, making Al–Gr composites promising for applications requiring thermal stability. However, experimental characterization particularly at high temperatures and higher graphene concentrations is limited due to challenges such as dispersion difficulties, agglomeration beyond 1 wt% graphene, and restricted availability of comprehensive datasets. To address these limitations, this research work highlights a physics-guided hybrid machine learning framework to predict the coefficient of thermal expansion (CTE) and overall length change of Al–Gr composites. The approach integrates physics-aware feature engineering with synthetic data augmentation to construct an extended dataset spanning a wider range of temperatures and compositions. For accurate interpolation within the experimentally observed domain, a multi-model stacking ensemble is employed, combining the best-performing models from five algorithmic groups using a Ridge Regressor meta-learner. To improve reliability in extrapolation, the ensemble is dynamically fused with a Physics-Informed Neural Network (PINN) and a Bayesian Neural Network (BNN). The PINN introduces physics-guided behavior through a gradient-based monotonic regularization term, encouraging predictions that follow physically consistent trends with increasing graphene content. The BNN, implemented using Monte Carlo dropout, provides approximate estimates of predictive uncertainty through stochastic forward passes. A composition-dependent weighting strategy is used to combine these components, allowing the framework to balance data-driven accuracy, physical consistency, and uncertainty awareness across different regimes. The proposed framework demonstrates very high predictive performance within the observed data range (R 2 ≈ 0.999176 with low prediction error) and is further used to explore thermal behavior up to 2.4 wt% graphene and temperatures up to 500 °C. The results capture key trends, including the gradual stabilization of thermal expansion at elevated temperatures and the overall reduction in effective CTE with increasing graphene content. While these predictions extend beyond currently available experimental data, they provide useful insights into material behavior and highlight the potential of physics-guided hybrid machine learning approaches for supporting composite design in high-temperature environments.