📄 earth science

Reducing forced-signal uncertainty in the INM RAS climate models using a pattern-recognition based approach

This paper integrates emergent constraints with a pattern-recognition methodology to reduce intermodel uncertainty in near-surface air temperature projections across INM RAS climate models, successfully narrowing historical warming patterns but failing to constrain the CO2-congruent signal projected for the late 21st century.

Original authors: Sergey Kravtsov, Andrey Gritsun, Sergey Gulev

Published 2026-09-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sergey Kravtsov, Andrey Gritsun, Sergey Gulev

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Climate scientists rely on complex computer programs, known as Earth System Models, to simulate how our planet's atmosphere and oceans have behaved in the past and to predict how they will change in the future. These models are essential tools for understanding the consequences of rising greenhouse gas levels, but they often disagree with one another. When different models run the same scenario, they can produce widely varying predictions for how much the world will warm or how that warming will be distributed across the globe. This disagreement, or uncertainty, makes it difficult for societies to plan effective responses to climate change. A major source of this uncertainty lies in how the models handle invisible physical processes, such as the formation of clouds, which are too small to be seen directly by the computer's grid but have a huge impact on the climate. To make better predictions, researchers are constantly looking for ways to narrow the gap between these different model outputs and the reality we observe in the historical record.

A team of researchers led by Sergey Kravtsov, along with colleagues from the Russian Academy of Sciences, has developed a new method to tackle this problem of model disagreement. They focused on three specific versions of a climate model developed at the Institute of Numerical Mathematics in Russia, which vary in their resolution and how they handle cloud physics. The researchers wanted to see if they could use a technique called pattern recognition to adjust the models' future projections so they matched the observed history of global temperatures more closely. Their approach involved breaking down the temperature data into two parts: a steady warming trend driven by carbon dioxide and a more complex, fluctuating background signal caused by other factors. By analyzing how these patterns behaved during the historical period from 1880 to 2005, they created a set of rules to rescale the models' outputs. The goal was to force the models to align with the actual temperature records we have, effectively "tuning" them before looking into the future.

The results of this experiment showed that the method worked well for the immediate past but hit a wall when looking further ahead. When the researchers applied their rescaling rules to the models' projections for the year 2100, the spread of predictions between the different models shrank significantly. The three models, which originally predicted different amounts of warming, were brought much closer together, all converging on a global temperature increase of about 3 degrees Celsius by the end of the century. This suggests that the method successfully reduced the uncertainty in the models' short-term forecasts by correcting for biases that were visible in the historical data. The researchers found that the models tended to overestimate the speed of warming in the future, and their adjustment process dialed these projections down to a more consistent level.

However, the study also revealed a critical limitation. While the rescaling worked for the historical period and the near future, it failed to constrain a specific type of climate signal that emerges later in the 21st century. The researchers discovered that there is a slow-moving component of climate change, likely related to how the ocean and atmosphere adjust to warming over decades, that does not show up clearly in the historical data. Because this signal was not visible in the past, the pattern-recognition method could not identify or correct for it. Consequently, even after the models were rescaled, they still disagreed on the magnitude of this late-century warming. The study suggests that while we can improve our models by matching them to what we have already seen, there are still hidden, slow-acting processes in the climate system that remain difficult to predict with current methods.

The researchers tested their idea by using one version of their model as a stand-in for the "real world" to see if the rescaling could bring the other models closer to it. This experiment confirmed that the method could indeed reduce the differences between the models, proving that the technique is mathematically sound. Yet, the core finding remains that the method is only as good as the data it learns from. If a climate change signal is too slow to be detected in the historical record, the method cannot fix the model's error regarding that signal. This implies that while we can refine our climate projections by learning from the past, there is a fundamental limit to how much we can reduce uncertainty without understanding these deeper, slower mechanisms of the Earth's climate system. The work does not solve the problem of model uncertainty entirely, but it offers a clearer path for understanding where our current tools succeed and where they fall short.

Technical Summary: Reducing Forced-Signal Uncertainty in the INM RAS Climate Models Using a Pattern-Recognition Based Approach

Problem Statement
Earth System Models (ESMs) exhibit significant inter-model uncertainty in future climate projections, particularly regarding near-surface air temperature (SAT). This spread arises from differences in model-specific parameterizations (e.g., cloud physics) and internal variability. While previous approaches have attempted to reduce this uncertainty through model weighting (based on ancestry, performance, and independence) or "emergent constraints" (linear relationships between historical predictors and future predictands), these methods often suffer from limited scope (focusing on single scalar metrics) or ad hoc predictor selection. The authors address the challenge of reducing the spread in gridded climate projections for a specific set of models—the Institute of Numerical Mathematics of the Russian Academy of Sciences (INM RAS) models—by integrating emergent constraints with a pattern-recognition methodology.

Methodology
The study analyzes three versions of the INM RAS models (CM5, CM6M, and CM6P) under historical (1850–2014) and SSP3-7.0 scenario (2015–2100) simulations. The methodology proceeds in three main stages:

Forced Signal Decomposition:
- CO2-Congruent Signal: The authors isolate the CO2-congruent SAT variability by regressing the ensemble-mean SAT time series against the annual mean $\ln(CO_2)$ time series over specific training periods (ranging from 1880–2005 to 1880–2045).
- Residual Decomposition: The residual, CO2-detrended variability is decomposed using Signal-to-Noise Maximizing Patterns (S/NP). This technique identifies spatial patterns that maximize the ratio of ensemble-mean variance to total variance, effectively separating the forced signal from aliased internal variability. The study retains the top three S/NP modes.
Pattern-Recognition Based Rescaling (Emergent Constraints):
- The authors apply a least-squares procedure to rescale the model-derived forced signals (both the CO2-congruent component and the S/NP components of the residual) to match reanalysis data (NOAA 20CRv3 and blended ERA-20C) over the historical training period (1880–2005).
- This generates rescaling coefficients that map the model's historical evolution to the "ground truth" of reanalysis. These coefficients are then applied to the future projections (1880–2100) to produce a library of optimally constrained, gridded forced-signal time series.
Residual Analysis and Validation:
- The study investigates the "residual forced signal" that remains undetected by the pattern-based rescaling in the late 21st century.
- To validate the efficacy of the rescaling without future observations, the authors perform a "pseudo-observation" experiment where CM5 simulations are treated as the ground truth to rescale the CM6 models.
- An idealized Energy Balance Model (EBM) coupled with a Stommel-type ocean box model is used to hypothesize why certain slow-adjustment climate modes (e.g., related to meridional overturning circulation weakening) might only emerge in the late 21st century, evading detection by historical training.

Key Results

Signal Decomposition: The CO2-congruent signal patterns across the three models are spatially similar (correlations > 0.93) but differ in magnitude. The CM6 models exhibit stronger warming over the tropical belt and a more pronounced land-sea contrast compared to CM5.
Rescaling Efficacy: The pattern-recognition rescaling successfully reduces the inter-model spread in projected global warming.
- For the original unscaled models, the projected warming by 2100 ranges from ~3.5°C (CM5) to ~5°C (CM6 models).
- After rescaling using historical reanalysis data, the projected warming converges to approximately 3°C across all models, with the inter-model uncertainty reduced to roughly 0.5°C or less.
- The rescaling coefficients generally indicate that the CM5 model warms slightly slower than reanalysis (requiring an upward correction), while the CM6 models warm too fast (requiring downward corrections).
Limitations: Despite reducing the spread, the rescaling fails to fully constrain the simulated CO2-congruent climate signal emerging in the second half of the 21st century. A residual forced signal, characterized by a near-zero global mean but distinct spatial patterns, persists.
Validation: The "pseudo-observation" experiment confirms that the rescaling procedure reduces the differences between the CM6 models and the "true" CM5 simulation, though residual differences remain due to distinct model dynamics and the undetected residual signal.

Significance and Claims
The authors claim that their methodology offers a data-driven alternative to traditional emergent constraints by utilizing pattern recognition to remap gridded climate fields rather than relying on single scalar predictors. They posit that this approach acts as a "data-driven parameterization" for climate feedbacks that are misrepresented or missing in current ESMs, potentially elucidating the physical drivers of model uncertainty.

However, the paper maintains a modest stance regarding the ultimate resolution of uncertainty. The authors explicitly state that while the rescaling reduces the spread of warming patterns, it "fails to constrain a simulated CO2-congruent climate signal emerging in the second half of the 21st century." They attribute this to slow climate system adjustments (such as ocean circulation changes) that only become detectable above internal variability in the late 21st century, a phenomenon that historical training periods cannot reliably capture. Consequently, the study concludes that while pattern-based rescaling is effective for reducing historical-based uncertainty, inherent limitations remain in projecting late-century signals that emerge outside the training window.

More like this