Emulator Based Downscaling of Future Temperature by Evaluating Multiple Machine Learning Models in Peninsular Malaysia

This study presents the development of a hybrid, emulator‐based downscaling model over Peninsular Malaysia. Future dynamic downscaled precipitation from the Coordinated Regional Climate Downscaling Experiment (CORDEX), temperature outputs from the HadGEM2‐ES global climate model (GCM), and static predictors were used for the development of the models. Several ML algorithms were tested by training each on 10 representative years of CORDEX future climate projections from 2010 to 2099. The GCM and static data served as predictors, while the CORDEX data was the response variable. In the training phase, RF achieved the highest skill with an RMSE of 0.49°C, R 2 of 0.89 and KGEnp of 0.96, while Ada and ANN had larger errors and lower correlations (e.g., RMSE > 1.0°C) and were therefore excluded from the testing phase. In the testing phase, RF also showed the best performance among the models, with an overall performance of RMSE = 1.01°C, R 2 = 0.44 and KGEnp = 0.81 among the remaining four models. The final selected downscaled model based on the RF technique replicated mean, variability and extreme temperatures at the 95th and 99th percentiles. The monthly time series showed that the RF‐based model successfully reproduced the CORDEX data with a deviation of less than 0.5°C. Therefore, by emulating the outputs of physically based regional climate models, the proposed hybrid approach retains non‐stationary climate‐temperature relationships while significantly reducing computational overhead relative to full dynamic downscaling.