bcm4rcm: Bayesian committee machines for regional climate models
2023-2024 # climate # machine learning
This work applies an ensembling method to combine different regional climate model outputs and produces principled uncertainty estimates of precipitation under different climate scenarios. These estimates inform the likelihood of extreme events that could lead to flooding, landslides, or droughts. More specifically, Gaussian Processes are fit to the projections of Coordinated Regional Downscaling Experiment (CORDEX) members for South Asia. A Gaussian Process is a probabilistic and interpretable machine learning method that gives principled uncertainty estimates. The models are then combined using a Bayesian Committee Machine. This method allows modellers to overcome the computational complexity of Gaussian Processes, thus improving scalability while conserving the benefits of Gaussian Process models.