A new detection and attribution method is presented and applied to the global mean surface air temperature (GSAT) from 1900 to 2014. The method aims at attributing the climate changes to the variations of greenhouse gases, anthropogenic aerosols, and natural forcings. A convolutional neural network (CNN) is trained using the simulated GSAT from historical and single-forcing simulations of twelve climate models. Then, we perform a backward optimization with the CNN to estimate the attributable GSAT changes. Such a method does not assume additivity in the effects of the forcings. The uncertainty in the attributable GSAT is estimated by sampling different starting points from single-forcing simulations and repeating the backward optimization. To evaluate this new method, the attributable GSAT changes are also calculated using the regularized optimal fingerprinting (ROF) method. Using synthetic non-additive data, we first find that the neural network-based method estimates attribuable changes better than ROF. When using GSAT data from climate model, the attribuable anomalies are similar for both methods, which might reflect that the influence of forcing is mainy additive for the GSAT. However, we found that the uncertainties given both methods are different. The new method presented here can be adapted and extended in future work, to investigate the non-additive changes found at the local scale or on other physical variables.

The internal variability pertains to fluctuations originating from processes inherent to the climate component and their mutual interactions. On the other hand, forced variability delineates the influence of external boundary conditions on the physical climate system. A methodology is formulated to distinguish between internal and forced variability within the surface air temperature. The noise-to-noise approach is employed for training a neural network, drawing an analogy between internal variability and image noise. A large training dataset is compiled using surface air temperature data spanning from 1901 to 2020, obtained from an ensemble of Atmosphere-Ocean General Circulation Model (AOGCM) simulations. The neural network utilized for training is a U-Net, a widely adopted convolutional network primarily designed for image segmentation. To assess performance, comparisons are made between outputs from two single-model initial-condition large ensembles (SMILEs), the ensemble mean, and the U-Net’s predictions. The U-Net reduces internal variability by a factor of four, although notable discrepancies are observed at the regional scale. While demonstrating effective filtering of the El Niño Southern Oscillation, the U-Net encounters challenges in areas dominated by forced variability, such as the Arctic sea ice retreat region. This methodology holds potential for extension to other physical variables, facilitating insights into the enduring changes triggered by external forcings over the long term.