Bayesian-optimized deep neural network surrogate for orientation-driven anisotropic thermal conductivity prediction in hybrid polymer nanocomposites

• Proposes a physics-informed Bayesian-optimized DNN surrogate for anisotropic thermal conductivity prediction. • Introduces three engineered descriptors (OAI, ISD, SHF) capturing orientation and hybrid synergy effects. • Achieves perfect predictive accuracy with strong generalization. • Enables rapid orientation–composition mapping for hybrid BN–rGO nanocomposites. • Provides a scalable AI-driven alternative to experimental trial-and-error composite design. Efficient prediction of anisotropic thermal conductivity in hybrid polymer nanocomposites remains a major challenge due to the nonlinear coupling among filler morphology, interfacial physics, and processing-induced orientation. This study presents a physics-informed, data-driven methodology for constructing a high-fidelity surrogate model that accurately estimates in-plane and through-plane thermal conductivity of nanocomposites reinforced with boron nitride (BN), silane-treated reduced graphene oxide (rGO), and silver nanoparticles (AgNPs). A comprehensive secondary dataset was aggregated from published experimental studies, capturing filler characteristics, hybrid ratios, processing parameters, alignment strategies, and anisotropic conductivity responses. After extensive preprocessing including missing-value imputation, min–max normalization, and three composite-specific engineered descriptors were formulated: the Orientation Anisotropy Index to represent directional heat-flow pathways, the Interfacial Surface Density to quantify matrix–filler thermal coupling, and the Synergistic Hybrid Factor to describe cooperative multi-filler effects. A Deep Neural Network surrogate was implemented using Python and trained with stratified data splits. To achieve optimal architectural depth and generalization, Bayesian Optimization with a Gaussian Process surrogate and Expected Improvement acquisition function was employed. This approach enabled efficient exploration of the hyperparameter space, resulting a substantially enhanced predictive architecture. The optimized DNN achieved an R² of 0.99995, with RMSE and MAE values reduced to 0.00131 and 0.000998, demonstrating a notable improvement over existing baseline models such as Gradient Boosting Decision Tree, Multi-Layer Perceptron, FFANN, etc. (≈12–18% accuracy gain. Overall, the proposed Bayesian-optimized surrogate establishes a fast, reliable, and physics-consistent predictive engine, offering significant potential to accelerate composite design and guide orientation-controlled fabrication strategies in next-generation thermal interface materials.