Predicting transient convective heat-transfer coefficients across different fin-and-tube heat exchangers is challenging due to domain shift in geometry and operating envelopes. We develop a two-step workflow. First, time-resolved air-side and water-side coefficients, αair (t)and αwater (t), are identified from air–water experiments using an energy-balance inverse model with a measurement-consistency filter. Second, multilayer perceptrons learn a surrogate mapping from measured operating variables to α(t)and are evaluated under a leave-one-exchanger-out protocol (LOHO). Geometry shift between exchangers is characterized using Mahalanobis and nearest-neighbour distances in the geometry-descriptor space. We report these distances for the held-out exchanger to contextualize out-of-distribution generalization. Across four input configurations, adding geometric descriptors reduces water-side error from ~4–5% to below 1% MAPE (mean absolute percentage error ) on the held-out exchanger. The best air-side models reach 5–7% MAPE (with R≈0.96"–" 0.97) and show partial attenuation of sharp peaks during fast transients. The results show that geometry is essential for transferable water-side prediction, while air-side accuracy is limited by stronger flow–fin interactions and effective label noise. The framework enables fast transient performance modelling and rigorous cross-exchanger validation.