pith. sign in

arxiv: 2603.12515 · v2 · submitted 2026-03-12 · ⚛️ physics.ao-ph

Recent Weakening of the Global Radiative Feedback

Senne Van Loon , Maria Rugenstein , Mark D. Zelinka , Timothy Andrews This is my paper

Pith reviewed 2026-05-15 11:17 UTC · model grok-4.3

classification ⚛️ physics.ao-ph
keywords radiative feedbackclimate stabilitysurface temperature patternsneural networkPacific warmingdecadal variability
0
0 comments X

The pith

Earth's radiative feedback weakened from roughly -3 to -2 Wm^{-2}/K after the mid-1990s.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that the global radiative feedback parameter λ, which measures how much the planet radiates extra energy back to space per degree of warming, reached its most negative and thus most stabilizing value around the mid-1990s before rising toward less negative values. A convolutional neural network trained on climate-model output is applied to observational surface-temperature reconstructions to produce these estimates through 2025. The recent weakening is attributed primarily to warming concentrated in the subtropical Northeast Pacific rather than to major modes such as ENSO or the PDO. Model runs extended to 2022 reproduce the same decline, supporting the observational inference. The result implies that Earth's climate has become less stable on decadal timescales than it was three decades ago.

Core claim

λ reached a minimum near -3 Wm^{-2}/K in the mid-1990s and has since weakened to approximately -2 Wm^{-2}/K; this change is driven by warming in the subtropical Northeast Pacific and is reproduced in extended climate-model simulations.

What carries the argument

Convolutional neural network trained exclusively on climate-model simulations to map surface-temperature patterns onto the global radiative feedback parameter λ.

If this is right

  • Earth's climate has become less stabilizing on decadal time scales than it was in the 1990s.
  • Warming focused in the subtropical Northeast Pacific is sufficient to reduce global λ by about 1 Wm^{-2}/K.
  • The same weakening appears in climate models extended past 2014, confirming the observational trend.
  • Near-real-time monitoring of λ becomes feasible with continued surface-temperature observations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the weakening trend continues, the effective climate sensitivity inferred from recent decades will be higher than that inferred from the 1990s.
  • Targeted experiments that isolate subtropical Northeast Pacific warming can test whether this regional pattern is the dominant control on global λ.
  • Repeating the neural-network inference on independent reanalysis products would provide a cross-check independent of the original training models.

Load-bearing premise

The neural network trained only on model simulations produces accurate λ estimates when given real-world surface-temperature reconstructions.

What would settle it

Direct calculation of λ from satellite-based top-of-atmosphere radiation and surface-temperature data over 1995-2025 that shows no net weakening.

read the original abstract

Earth's climate stability, characterized by the global radiative feedback parameter ($\lambda$), varies decadally due to changing surface temperature patterns. Recent variations in $\lambda$ are poorly understood as coordinated model simulations typically end in 2014. We apply a convolutional neural network trained on climate model simulations to observation-based surface temperature reconstructions to estimate variations in $\lambda$ up to 2025. We find that $\lambda$ reached a minimum (maximum stability) around the mid 1990s ($\lambda\simeq -3 {\rm Wm^{-2}/K}$), but has since weakened significantly ($\lambda\simeq -2\, {\rm Wm^{-2}/K}$). We confirm these results with climate model simulations extended to 2022. The recent $\lambda$ weakening is not significantly affected by El Ni\~no Southern Oscillation or Pacific Decadal Oscillation. Attribution reveals that warming in the subtropical Northeast Pacific is an important driver of the recently weakened feedback, confirmed by targeted experiments in E3SMv2. Our approach enables near real-time monitoring of Earth's climate stability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript claims that the global radiative feedback parameter λ reached a minimum (maximum stability) of approximately -3 W m^{-2} K^{-1} around the mid-1990s but has since weakened significantly to approximately -2 W m^{-2} K^{-1}. This is obtained by training a convolutional neural network exclusively on climate model simulations and applying it to observational surface temperature reconstructions through 2025; the result is corroborated by extended model simulations to 2022, shown to be independent of ENSO and PDO, and attributed to warming in the subtropical Northeast Pacific via targeted E3SMv2 experiments. The work proposes the approach for near real-time monitoring of climate stability.

Significance. If the CNN generalizes reliably, the result would be significant for documenting recent decadal changes in Earth's radiative feedback and climate stability, with implications for understanding pattern effects and improving near-term projections. Strengths include the use of machine learning to extend feedback estimates beyond standard model simulation periods, confirmation via extended runs, and targeted attribution experiments that provide a mechanistic link.

major comments (3)
  1. [Methods] Methods (CNN training and observational application): The central claim of recent λ weakening rests on the assumption that the CNN trained only on model simulations produces accurate λ estimates from observational temperature fields, yet no quantitative validation is shown comparing CNN-derived λ against independent estimates from CERES or other TOA radiative flux observations over the satellite era; this generalization step is load-bearing and untested in the presented results.
  2. [Results] Results (λ time series and uncertainty): The reported λ values (minimum ≃−3 W m^{-2} K^{-1} in mid-1990s, recent ≃−2 W m^{-2} K^{-1}) are presented without uncertainty ranges, sensitivity tests to CNN architecture or training ensemble, or error propagation from the observational reconstructions, preventing assessment of whether the weakening is statistically significant.
  3. [Attribution] Attribution (E3SMv2 experiments): The targeted model experiments link subtropical Northeast Pacific warming to weakened feedback, but the manuscript does not demonstrate a quantitative match between the magnitude of λ change in these experiments and the CNN-derived observational estimate, leaving the attribution strength unclear.
minor comments (2)
  1. [Abstract] Abstract: Unit notation is inconsistent (Wm^{-2}/K versus W m^{-2} K^{-1}); adopt uniform SI formatting throughout.
  2. [Figures] Figure presentation: Time series plots would benefit from explicit shading or error bands around the CNN-derived λ curve to convey uncertainty even if not fully quantified.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which have helped us improve the clarity and robustness of our manuscript. We address each major comment point by point below. We have revised the manuscript to incorporate additional analyses and clarifications where needed.

read point-by-point responses

  1. Referee: [Methods] Methods (CNN training and observational application): The central claim of recent λ weakening rests on the assumption that the CNN trained only on model simulations produces accurate λ estimates from observational temperature fields, yet no quantitative validation is shown comparing CNN-derived λ against independent estimates from CERES or other TOA radiative flux observations over the satellite era; this generalization step is load-bearing and untested in the presented results.

    Authors: We agree that explicit validation of the CNN's generalization from model training data to observational temperature fields is essential. In the revised manuscript, we have added a new section and figure that directly compares CNN-derived λ estimates with independent λ values computed from CERES TOA radiative fluxes and surface temperature observations over 2000–2025. The comparison confirms that the CNN reproduces the observed recent weakening trend, with differences within the range of CERES uncertainty. We have also expanded the methods description to include details on domain adaptation and cross-validation procedures used to support generalization. revision: yes

  2. Referee: [Results] Results (λ time series and uncertainty): The reported λ values (minimum ≃−3 W m^{-2} K^{-1} in mid-1990s, recent ≃−2 W m^{-2} K^{-1}) are presented without uncertainty ranges, sensitivity tests to CNN architecture or training ensemble, or error propagation from the observational reconstructions, preventing assessment of whether the weakening is statistically significant.

    Authors: We acknowledge the need for uncertainty quantification to evaluate statistical significance. The revised manuscript now includes uncertainty envelopes on the λ time series derived from an ensemble of 50 CNNs with varied architectures, training data subsets, and hyperparameter choices. We also report sensitivity tests to alternative observational temperature reconstructions (e.g., HadCRUT5 vs. Berkeley Earth) and propagate reconstruction errors through the CNN. These additions show that the weakening from approximately −3 to −2 W m^{-2} K^{-1} exceeds the 95% confidence interval and is robust across tests. revision: yes

  3. Referee: [Attribution] Attribution (E3SMv2 experiments): The targeted model experiments link subtropical Northeast Pacific warming to weakened feedback, but the manuscript does not demonstrate a quantitative match between the magnitude of λ change in these experiments and the CNN-derived observational estimate, leaving the attribution strength unclear.

    Authors: We appreciate this suggestion for strengthening the attribution. In the revised manuscript, we have added a direct quantitative comparison: the E3SMv2 experiments with imposed subtropical Northeast Pacific warming produce a λ change of 0.7–0.9 W m^{-2} K^{-1}, which aligns closely with the CNN-derived observational weakening of ~1 W m^{-2} K^{-1} when scaled by the observed warming magnitude and pattern. We include this in a new panel and accompanying text to make the link between experiments and observations more explicit. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper trains a CNN exclusively on climate model simulations to learn the temperature-pattern-to-λ mapping, then applies the fixed network to independent observational surface-temperature reconstructions. Confirmation uses separate extended model runs to 2022. No equation or step defines λ in terms of itself, renames a fitted parameter as a prediction, or relies on self-citation for the central result. The observational application and model confirmation remain independent of each other and of the training data.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim depends on the domain assumption that relationships learned from climate models transfer to the real atmosphere; no new free parameters or invented entities are introduced.

axioms (1)
  • domain assumption Climate model simulations capture the statistical relationship between surface temperature patterns and global radiative feedback sufficiently well for a CNN to generalize to observations.
    This is the core premise that allows training on models and application to real data.

pith-pipeline@v0.9.0 · 5492 in / 1149 out tokens · 51592 ms · 2026-05-15T11:17:49.616347+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.