[Paper Review] Artificial intelligence and downscaling global climate model future projections

Rasmus Benestad|arXiv (Cornell University)|Jan 2, 2026

Climate variability and models0 citations

TL;DR

A critical review of AI/ML for downscaling global climate model projections, highlighting limitations, representativeness issues, and the value of traditional statistics and emulation approaches alongside AI/ML. It discusses proper benchmarking, data quality, and when AI/ML can be useful.

ABSTRACT

A critical review of artificial intelligence and deep machine learning (AI/ML) applied to downscaling of global climate model simulations provides some words of caution, based on past experiences and well-established principles. Recent papers tend to ignore more subtle successes with statistics and mathematical based downscaling, and there are examples of inappropriate evaluation strategies and incomplete accounts of the scientific progress when it comes to climate downscaling. An incomplete description state-of-the-art and a dogmatic approach to evaluation may give a deceiving impression that AI/ML is superior to more statistics and mathematics based methods.

Motivation & Objective

• Assess the use of AI/ML for downscaling global climate model future projections.
• Highlight limitations including data representativeness and out-of-distribution performance.
• Compare AI/ML methods with empirical-statistical downscaling (ESD) and traditional approaches.
• Discuss proper testing, benchmarking, and modeling strategies for climate downscaling.

Proposed method

• Provide a critical synthesis of AI/ML downscaling in climate science.
• Differentiate downscaling from interpolation and bias-adjustment.
• Discuss perfect prognosis, model output statistics, and hybrid PP-MOS strategies.
• Argue for emulation of regional climate models and use of common EOFs for transferability.

Figure 1: Clouds have many different scales, and different cloud types and structures illustrate different situations where links between large and small scales vary. They can also provide a metaphor for Earth’s climate going through transitions from a known state to new unknown one with different i

Experimental results

Research questions

• RQ1 What are the limitations of AI/ML methods when used for downscaling future climate projections?
• RQ2 How do AI/ML approaches compare to established statistical downscaling methods in terms of robustness and transferability?
• RQ3 Under what conditions can AI/ML provide reliable value in downscaling, such as emulation or pattern extraction?

Key findings

• AI/ML can emulate RCM simulations and study internal variability when trained on future-inclusive climate simulations.
• ESD and mathematics-based methods can be more robust with scarce or imperfect data and may outperform AI/ML for certain tasks.
• Downscaling requires leveraging links between large-scale and small-scale processes, not mere interpolation or bias-adjustment.
• Testing should focus on out-of-sample future climate performance across GCMs, not just historical statistics.
• There is a risk of dogmatic reliance on AI/ML results without transparency and replication.
• Proper benchmarking should compare against best alternatives, avoiding straw man methods.