A Complete Density Correction using Normalizing Flows (CDC-NF) for CMIP6 GCMs

Area of Science:

• Climate science
• Atmospheric science
• Data science

Background:

• Global Climate Models (GCMs) are crucial for climate projections but suffer from biases, limiting their use in impact assessments.
• Existing bias correction (BC) methods like quantile mapping fail to accurately represent joint extremes and cross-variable relationships.

Purpose of the Study:

• To introduce a novel bias correction method, Complete Density Correction using Normalizing Flows (CDC-NF), that addresses limitations of traditional approaches.
• To improve the representation of GCMs' full joint distribution, particularly for extreme events and multivariate dependencies.

Main Methods:

• Developed CDC-NF, a method employing invertible transformations to adjust the complete joint distribution of GCM outputs.
• Applied CDC-NF to daily precipitation and maximum temperature data from CMIP6 GCM projections using NOAA nClimGrid-daily observations.
• Evaluated CDC-NF against traditional BC methods using metrics like Wasserstein Distance, RMSE, and PBIAS.

Main Results:

• CDC-NF showed significant improvements over traditional BC methods, especially for the 90th percentile extremes.
• The method demonstrated enhanced accuracy in preserving cross-correlation structures between variables.
• Improved performance in modeling compound extreme events was observed.

Conclusions:

• CDC-NF offers a robust advancement in bias correction for climate model outputs.
• This method enhances the reliability of climate projections and improves the accuracy of climate impact studies in a changing climate.