Regional climate risk assessment from climate models using probabilistic machine learning

Problem

This work addresses the gap in capability between coarse global climate models and the fine-scale regional information necessary for effective decision-making in climate risk assessment. The authors highlight the limitations of existing methods, particularly the reliance on temporally aligned source and target data, which GenFocal aims to overcome. The paper is a preprint and has not yet undergone peer review.

Method

The core technical contribution is the GenFocal framework, which employs a two-stage process consisting of bias correction and super-resolution.

• Architecture: GenFocal is a generative AI framework designed to enhance the resolution of climate data.
• Algorithm: The method involves a two-stage approach:
  • Bias Correction: Utilizes a rectified flow mechanism to adjust discrepancies between coarse and fine-scale data.
  • Super-Resolution: Implements a conditional diffusion model to generate high-resolution outputs from the corrected data.
• Data:
  • Training Data Size/Source: The model is trained on 20 years of ERA5 reanalysis data (1980–1999) as the high-resolution target, with CESM2 LENS2 serving as the coarse-resolution source.
  • Input Features: The model uses 10 daily averaged climate variables as input.
  • Output Variables: It predicts 4 output variables sampled at 2-hour intervals.
• Training Compute: Specific details regarding the compute resources used for training are not disclosed.
• Other Mechanisms: The framework incorporates a domain decomposition technique to ensure temporal consistency in the generated outputs.

Results

GenFocal demonstrates significant improvements over existing baselines in various climate metrics:

• TC Count and Morphology: Accurately reproduces tropical cyclone statistics in the North Atlantic for the period 2010–2019 when compared to LENS2.
• Heat Index 99th Percentile: Achieves over a 35% reduction in average bias compared to the Bias-Corrected Spatial Disaggregation (BCSD) and STAR-ESDM methods.
• Tail Dependence of Temperature and Humidity Extremes: Shows a 44% reduction in error compared to STAR-ESDM and BCSD.
• 5-Day Extreme Caution Heat Streaks: Provides largely unbiased estimates, with average bias reductions of 44% and 57% compared to BCSD and STAR-ESDM, respectively. The available text does not report quantitative results for other metrics or benchmarks.

Limitations

The authors acknowledge the reliance on temporally aligned source and target data as a limitation of existing approaches, which GenFocal seeks to address. However, the paper does not discuss other potential limitations or challenges in the implementation of the framework.

Why it matters

The implications of this work are significant for downstream applications in climate risk assessment, as GenFocal provides a more accurate and reliable method for translating coarse climate model outputs into fine-scale regional data. This advancement could enhance decision-making processes in climate adaptation and mitigation strategies, ultimately contributing to more effective responses to climate change impacts.