Abstract: Crop growth simulation models are the primary instrument used for predicting crop developmental responses to climate change. The understanding of the applicability of phenological models is critical for measuring how climate change affects crop yields, particularly under warm climate conditions. Machine learning algorithm (MLP), maize simulation model (MAIS), response and adaption model (RAM), and Beta model (Beta) in warm climate were compared in this study. Based on phenological observation data of over 20 years for spring maize in Northeast China and summer maize in North China from agricultural meteorological station records and daily weather data, we separated the data into two categories: cold and warm years. Data were obtained by calculating the abnormality based on the mean temperature during the maize growing season. Cold years data were used to calibrate the models, and warm years data were used to validate them and subsequently evaluate their performance. The normalized root mean square error (NRMSE), mean bias error (MBE), and systematic bias of the simulation error against the average temperature of the growth period were used to evaluate the deviation between the simulated and observed maize phenology of the four models. The results showed that MLP performed much better than the three mechanistic phenology models during cold years. RAM outperformed the other mechanistic phenology models, followed by Beta and MAIS models. Our results indicated that the dates of MAIS and Beta were earlier than the observations, whereas the simulations of RAM and MLP were later than the observations. The proportion of sites with a significant trend of simulation error against the average temperature of the growth period for the RAM, MAIS, and Beta models was lower than that for MLP. Compared to that for MAIS and RAM, the proportion of sites with a significant trend for Beta was the smallest. In warm years, Beta performed better than the other models, followed by MLP, RAM, and MAIS. The simulations of the three mechanistic phenology models were earlier than the observations, but the simulations of MLP were later than the observations. Beta showed the smallest proportion of sites with a significant trend, followed by MLP, MAIS, and RAM. Overall, the models did not benefit from both calibration and validation. MLP performed well during calibration in cold years, but poorly in warm years. The overall performance of the mechanistic phenology models was worse than that of MLP in cold years, but they performed better in warm years. Different models are appropriate in various contexts. The MLP can be recommended to precisely reverse the impact of historical climate change on growth period. However, mechanistic models should be used to precisely predict the impact of future climate change on growth period.