Artificial Neural Network Modeling of Thermal Conductivity Changes in Milk during Mechanized Khoa Production

An artificial neural network (ANN) approach was successfully deployed to model and predict the thermal conductivity of milk and concentrated milk systems during mechanized khoa production. Reliable data points (n=203) spanning wide operational boundaries of temperature (275.15–353.10 K), moisture content (48.82–92.00%), and fat content (0.00–11.17%) were compiled from established experimental studies to formulate and validate the model. A multi-layer feed-forward network optimized via the quasi-Newton algorithm using a 3:3:1 topology (three inputs, three hidden neurons with hyperbolic tangent activation functions, and a single linear output layer) demonstrated optimal predictive behaviour. The architecture yielded outstanding precision on independent testing subsets, demonstrating a strong correlation coefficient (R = 0.986), a minimal root mean squared error (RMSE = 0.0084 W/m•K), and a normalized squared error of 0.029 (normalized to the variance of the target data). Input sensitivity computations verified that product temperature (31.4% contribution) and moisture content (30.2% contribution) exert the highest thermodynamic control on thermal conductivity shifts, whereas fat content (4.4% contribution) exhibits a weaker but consistently inverse linear relationship. The resolved predictive equations were effectively embedded within a highly practical Microsoft Excel-based graphical user interface (GUI) to assist dairy process designers in real-time calculation, simulation, and industrial scaling of continuous scraped surface heat exchangers for indigenous milk confectionery manufacturing.