Artificial intelligence-driven precipitation downscaling and projections over Thailand using CMIP6 climate models

Global warming has intensified the hydrological cycle, increased the frequency and severity of extreme precipitation events, and necessitated the collection of accurate future precipitation data for effective disaster mitigation and informed decision-making. The research evaluates the performance of artificial intelligence (AI)-driven downscaling techniques (Dynamic Neural Network with Memory (DyNN-Mem) and Hybrid Long Short-Term Memory Convolutional Neural Network (LSTM-CNN)) for scaling down CMIP6 Global Climate Models (GCMs) daily precipitation outputs across Thailand. Model performance evaluation for the historical period (2014–2022) relied on statistical indicators, including R2, MAE, and RMSE, by comparing simulated data with precipitation records from the Thai Meteorological Department (TMD). The results demonstrate that DyNN-Mem outperforms the Hybrid LSTM-CNN architecture in terms of R2 (0.55–- 0.78), MAE (0.22–0.29), and RMSE (1.28–3.30) for multiple GCMs. Using the MPI-ESM1-2-LR and CAMS-CSM1-0 GCM outputs, DyNNMem showed better spatial and temporal characteristics of precipitation, translating into better capturing the precipitation’s spatial and temporal characteristics. MPI-ESM1-2-LR and CanESM5 were found to be the most reliable CMIP6 GCMs for reproducing historical precipitation patterns in the Thai region. AI-based downscaling enhances regional climate forecasts; DyNN-Mem with MPI-ESM1-2-LR improves daily precipitation forecasts in Thailand.