Neural surrogate modeling of thermophysical properties in SiC-reinforced A356 aluminum composites for thermal management in brake applications

Aluminium metal matrix composites (AMMCs) reinforced with silicon carbide (SiC) are promising candidates for high-performance brake applications due to their superior strength-to-weight ratio and thermal stability. However, accurate prediction of thermophysical properties such as thermal conductivity and coefficient of thermal expansion remains challenging because experimental datasets are often limited and highly nonlinear. This study proposes a physics-informed shallow feed-forward artificial neural network framework to intelligently model and predict the thermophysical performance of SiC-reinforced A356 aluminium matrix composites. An experimental dataset, containing multi-composition samples of SiC-reinforced A356 composites, was used as the primary data source. To address data scarcity, a physics-guided data augmentation strategy combining interpolation between adjacent SiC compositions, uncertainty-aware noise injection (±2–5%), and strict thermodynamic and mechanical constraints was employed, expanding the dataset to nearly 120–150 physically consistent samples. The proposed Physics-Informed Shallow feed-forward ANN utilizes three key input features SiC reinforcement fraction, density, and elastic modulus and predicts thermal conductivity and thermal expansion coefficient as outputs. The model was implemented using Python. Results demonstrate excellent convergence behavior and strong agreement between predicted and experimental values. The proposed model achieved an R 2 of 0.985 with RMSE = 0.29 and MAE = 0.21, representing an improvement in predictive accuracy compared to conventional ANN and hybrid approaches. Unlike conventional data-driven models, the proposed framework explicitly embeds thermodynamic consistency, mixture-law constraints, and monotonic reinforcement-property relationships within the data augmentation and training pipeline. This transforms the neural model into a physically bounded surrogate predictor is suitable for brake-relevant thermophysical material assessment under limited experimental data conditions.