Enhancing Climate Projections: A Comprehensive Density Correction Approach Using Normalizing Flows for Improved Bias Mitigation in Global Climate Models

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

Global Climate Models (GCMs) are crucial instruments for predicting and understanding our planet’s climate as they simulate the complex workings of the Earth’s climate system. Despite improvements brought on by increased computational capabilities and enhanced parametrizations, GCMs still present systematic biases, particularly when dealing with extremes and multivariate dependencies. These biases create challenges in both regional climate impact assessments and their application across various sectors.

The optimal application of GCM projections requires effective bias correction (BC). Typically, traditional BC approaches like quantile mapping only adjust the marginal distributions, which offers limited scope. These methods might correct single-variable biases but fall short when accounting for dependencies across different climate variables such as precipitation and temperature. It’s evident that ignoring such dependencies could misrepresent crucial interrelationships affecting climate insights and ultimately hinder accurate climate impact analyses.

Addressing these inadequacies, innovative techniques such as Multivariate Adaptive Constructed Analogs (MACA) and Asynchronous Canonical Correlation Analysis (CCA) attempt to preserve temporal covariability among climate variables. Similarly, deep learning-based methods like DeepSD utilize neural networks to learn intricate non-linear relationships between climate variables and geographical information. However, evaluations for preserving inter-variable spatial-temporal dependencies remain absent.

Extremely challenging to tackle within these models is the accurate replication of extreme event quantiles, such as severe rainfall or frequent heatwaves. Traditional methods excel at estimating central tendencies of climate distributions but falter considerably at the distribution’s extremities. The need for precise and reliable correction of these extremes is critical for informed climate projections, especially given the sparse historical data recording these events.

Given these complexities, our approach springs from the advent of deep learning innovations with a novel method called Complete Density Correction using Normalizing Flows (CDC-NF). This technique applies a class of deep learning models known as Normalizing Flows to accurately adjust the complete joint distribution of climate variables forecasted by GCMs. Normalizing Flows, renowned for their efficacy in transforming intricate distributions into simpler forms through invertible mappings, provide an ideal framework for this task. This approach adeptly addresses joint distribution biases and retains vital cross-correlation structures among climate variables, thereby improving the estimation of joint extreme quantiles.

The introduction of the CDC-NF method provides a significant leap forward as it leverages conditional normalizing flows to monitor dependencies between variables, such as precipitation conditioned by temperature. The method corrects biases strategically and theoretically, offering a transformative edge over traditional strategies. By preserving cross-correlations even at extreme quantiles, CDC-NF guarantees more accurate and trustworthy representations of climate conditions.

To address the gap in bias-corrected datasets for CMIP6 projections, the CDC-NF database was created. It uniquely offers a comprehensive approach with Complete Density Corrections for precipitation and maximum temperature projections. This state-of-the-art resource creates avenues for improved climate projections, particularly emphasizing joint extremes, thus enhancing the fidelity of sectoral climate impact studies.

As climate-related disasters increase globally, methods like CDC-NF equip researchers and policymakers with tools essential for comprehending and addressing compound extremes and cascading failures. By precisely replicating the joint and conditional densities of pertinent climate variables, this innovative approach strengthens predictive insights, vital for adaptive strategies and mitigation efforts amid looming climate challenges.