Generalization challenges in optimizing heat transfer predictions in plate fin and tube heat exchangers using artificial neural networks

This study addresses the challenge of predicting air and water outlet temperatures in compact heat exchangers under unseen operational regimes, a critical gap in thermal modeling where traditional methods struggle with extrapolation. We evaluate artificial neural networks (ANNs) for their ability to generalize to an intermediate water supply temperature “C” (40–50 °C), situated between two trained ranges: “A” (30–40 °C) and “B” (50–65 °C). Using the BFGS (Broyden–Fletcher–Goldfarb–Shanno) algorithm, ANN models were trained on datasets “A” and “B”, with inputs including inlet temperatures, water flow rates, and air velocity. Performance was quantified via Mean Absolute Percentage Error (MAPE) and Theil's inequality coefficient. The ANNs achieved high accuracy within training ranges (MAPE = 0.50 % for air outlet temperature in “A”), and crucially, demonstrated reliable generalization to the unseen intermediate range “C”, with only modest error increases (MAPE = 2.64 % for air outlet temperature). Theil's coefficient confirmed stable predictions, underscoring ANN suitability for real-world applications such as HVAC systems and industrial processes, where operating conditions deviate from historical data. While results highlight ANNs as promising tools for extrapolation, we identify strategies to further enhance reliability. This work advances predictive modeling in thermal engineering, offering insights for optimizing heat exchanger performance under variable and untested conditions.