Enhancing global CMIP6 model temperature predictions using deep learning neural networks

Accurate prediction of maximum Surface Air Temperature (SAT) is essential for understanding regional climate dynamics and managing the impacts of extreme heat events. This study evaluates the performance of shallow and deep Artificial Neural Network models (sANN and dANN) in predicting mean maximum SAT over Andhra Pradesh (AP), India, during the March-April-May season for the period 1981–2022. The models sANN and dANN were designed with 2 and 4 hidden layers that will be trained using outputs from multiple grouped ensemble combinations of Coupled Model Intercomparison Project Phase 6 (CMIP6) models as inputs and observations from the India Meteorological Department (IMD) as targets, respectively. The predictive accuracy of these ensembles was assessed using statistical metrics such as bias, mean absolute error, correlation coefficient, index of agreement, root mean square error, and standard deviation. Among the CMIP6 ensemble groupings, dANN_Group-5 (GFDL-CM4_gr2, NorESM2-MM, INM-CM4-8, MPI-ESM1-2-HR, EC-Earth3-Veg-LR, and CMCC-CM2-SR5) performed best. The dANN_Group-5 model consistently outperformed MMM_Group-5 (CMIP6 ensemble) and other CMIP6 ensembles, demonstrating superior accuracy in capturing observed interannual variability, extreme temperatures, and spatial distributions. The dANN_Group-5 model (bias range: −0.25 °C to + 0.25 °C) effectively minimized biases and captured fine-scale regional temperature patterns, particularly in areas vulnerable to extreme heat. In contrast, the MMM_Group-5 model (bias exceeding − 4 °C) and the sANN_Group-5 model (bias range: −2 °C to + 1 °C) exhibited substantial deviations, characterized by overly smoothed trends that failed to reproduce observed extremes, especially the highest SATs (i.e., the maximum of maxima, or max of max SAT). Furthermore, the dANN_Group-5 model demonstrated the lowest uncertainty, with Confidence Interval (CI) widths generally below ± 0.25 °C across the AP. In comparison, the MMM_Group-5 model presented broader CIs, reaching up to ± 1.0 °C, underscoring the advantage of deep learning in reducing model uncertainty. These findings demonstrate the shortcomings of conventional climate models while highlighting the potential of deep learning methods such as dANN for regional climate modeling. The dANN model’s ability to integrate and optimize multi-model CMIP6 data establishes it as a robust tool for high-resolution temperature prediction, with significant implications for climate impact assessments, water resource management, and heat mitigation strategies. Future work should expand its application to other regions and incorporate additional climatic variables to further enhance its predictive capabilities.