A new response surface methodology (RSM) based neural network (NN) modeling method is proposed for finned-tube evaporator performance evaluation under dry and wet conditions. Two RSM designs, Box–Behnken design (BBD) and central composite design (CCD), are applied to collect a small but well-designed dataset for NN training, respectively. Compared with additional 7000 sets of test data, for all the evaporator performance including total cooling capacity, sensible heat ratio and pressure drops on both refrigerant and air sides, the standard deviation (SD) and coefficient of determination of trained NNs are less than 2% and higher than 0.998, respectively, under dry conditions while those are less than 4% and greater than 0.974, respectively, under wet conditions. Classic quadratic polynomial response surface models were also included for reference. By comparison, the proposed model achieves higher accuracy. Finally, parametric study based on the trained NNs is conducted. This new method can remarkably downsize the training dataset and mitigate the over-fitting risk of NN.