Recent Weakening of the Global Radiative Feedback

Abstract

Earth's climate stability, characterized by the global radiative feedback parameter ($λ$), varies decadally due to changing surface temperature patterns. Recent variations in $λ$ 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 $λ$ up to 2025. We find that $λ$ reached a minimum (maximum stability) around the mid 1990s ($λ\simeq\SI{-3}{Wm^{-2}/K}$), but has since weakened significantly ($λ\simeq\SI{-2}{Wm^{-2}/K}$). We confirm these results with climate model simulations extended to 2022. The recent $λ$ weakening is not significantly affected by El Niño 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.

Figures (12)

• Figure 1: (a) The global feedback parameter $\lambda$ estimated from AGCMs that contributed amip-piForcing experiments to CMIP6 (CFMIP, blue, ending in 2014), an extended amip-piForcing experiment in HadGEM3-GC31-LL (green, ending in 2020), AMIP-F2010 experiments in E3SMv2 (orange, ending in 2022), the CNN applied to six historical surface temperature reconstructions (red, ending in 2025), and an estimate based on satellite observations (satellite, purple, ending in 2024). (b) The global radiative feedback parameter $\lambda$ estimated from the CNN applied to ${\bf T}_{\rm rc}$ (red, as in a); ${\bf T}^{{\neg I_{3.4}}}_{\rm rc}$ with the El Niño Southern Oscillation (ENSO) linearly removed (as defined by the Niño 3.4 region; pink); ${\bf T}^{{\rm \neg PDO}}_{\rm rc}$ with the Pacific Decadal Oscillation (PDO) linearly removed (brown). Thin lines are individual model simulations (for CFMIP), ensemble members (for E3SMv2), or temperature reconstructions (for CNN and satellite); thick lines are a mean across the thin lines per dataset.
• Figure 2: Attribution of changes in the global radiative feedback parameter. Panels a-c show the surface temperature patterns (linear regression slope of local temperature against global mean temperature). Panels d-f show the local contribution to the feedback parameter. The sum over all grid boxes equals $\lambda$ in the same period (indicated above the maps). Bottom panels show the difference between two periods for the surface temperature pattern (g-h) and feedback attribution maps (i-j). Maps are averaged over six reconstructions ${\bf T}_{rc}$, except for 1951-1980, with only 3 available datasets. All temperature patterns and attribution maps share their respective colorbars. Panels k-p show the contribution of six different regions (see panel a) to the global feedback parameter, considering only ocean area in each box. All values are anomalies with respect to the 1871-2025 mean. Thin lines are individual ${\bf T}_{\rm rc}$, thick lines are an ensemble mean (before 1940-1969, only one reconstruction is available). Values on the top right indicate the Pearson correlation coefficient with the global feedback (Fig. \ref{['historical_feedback']}).
• Figure 3: Decomposition of the net global feedback (a) into its shortwave (SW) clear-sky (b), longwave (LW) clear-sky (c), SW cloud (d), and LW cloud (e) components in the E3SMv2 ensemble (orange). Five members of E3SMv2 are run with HadISST-1.1 boundary conditions and fixed Northeast Pacific (NEP) SSTs (green). Thin lines are individual ensemble members, thick lines are an ensemble mean. We compare with bootstrapped 5-member averages from the 10-member E3SMv2 ensemble with HadISST-1.1 boundary conditions. Thick blue line shows the median, shading the two-sided 95% confidence bounds.
• Figure S1: Schematic of the CNN architecture used. The input maps (2m temperature) are passed through two convolutional layers, each with 32 kernels of size $3\times3$, followed by a max pooling layer and an ELU activation function. The result is flattened and passed through two fully connected layers with 32 and 16 neurons, with a ELU and linear activation function respectively. The final result is a single number estimating the global-mean radiative response $R$. Adapted from Rugenstein25 and VanLoon25.
• Figure S2: Gradient of the CNN, as the derivative of the output (global mean radiation $R$) to the input (near-surface temperature ${\bf T}$). The maps can be interpreted as a local radiative feedback. Panel a shows the gradient of the CNN pretrained on 7 large ensemble climate models. The gradient is averaged over all training years and models. Panel b shows the gradient of the CNN finetuned on CFMIP simulations, averaged over all CFMIP models and years in the training dataset.