Data-driven and interpretable modeling frameworks for frost accumulation during natural convection on polymer composite fins

Frost formation on heat-transfer surfaces remains a critical challenge in refrigeration, cryogenic systems, and energy-efficient thermal management. Accurate prediction of frost layer thickness (FLT) is particularly important for carbon fiber reinforced polymer composite fins (CFRPCFs), where distinct thermophysical properties alter frost growth dynamics relative to conventional metallic substrates. Because existing empirical correlations struggle to capture the highly nonlinear heat and mass transfer mechanisms unique to composite substrates, this study develops interpretable machine learning frameworks to model FLT during natural convection. Using an experimental database of $$\:356$$ observations, four models, i.e., support vector machine (SVM), artificial neural network (ANN), radial basis function neural network (RBFNN), and adaptive neuro-fuzzy inference system (ANFIS), were constructed using physically informed dimensionless inputs and optimized via Bayesian optimization. The developed RBFNN model demonstrated superior predictive capability, yielding mean absolute percentage deviation (MAPD) values of $$\:1.07\text{\%}$$ and $$\:2.06\text{\%}$$ for learning and validation, respectively, representing a major improvement over literature correlations. A grouped nine-fold (condition-wise) cross-validation protocol, iteratively holding out entire operating conditions, confirmed robust model generalization to unseen operational states, while applicability domain analysis verified model reliability across the experimental parameter space. Crucially, Shapley additive explanations (SHAP) were integrated to resolve the black-box nature of these models. The resulting model-attributed feature-importance patterns were found to be qualitatively consistent with established frost-growth mechanisms. This work provides a highly accurate, stable, and physically interpretable framework for FLT prediction on composite fins, establishing a reliable modeling template for advanced thermal materials.

Hali tarjima qilinmagan