Detection and attribution of climate change using a neural network.

Abstract

<p>A new detection and attribution method is presented using a neural network and applied here to the global surface air temperature (GSAT) from 1900 to 2014.<br />The changes of the GSAT are attributed to the variations of greenhouse gases, anthropogenic aerosols, and natural forcings. Using the outputs from CMIP6 (Coupled model intercomparison project phase 6) climate models simulations, we train a convolutional neural network (CNN) to reproduce the GSAT simulated by the historical simulations from that of single-forcing simulations. Then, we use a a interpretable AI method : the backward optimization of the trained CNN to estimate the attributable changes. Such a method does not imply any additivity hypothesis in the effects of the forcings. The uncertainty is estimated using different random single-forcings members as starting points for the backward optimization. For comparison, the attributable changes are also calculated using the regularized optimal fingerprinting (ROF) method. The skills of both methods are evaluated following a perfect model approach. The attributed changes are coherent for both methods in terms of amplitude and with some differences in term of uncertainties. The method presented can be adapted and extended in future work, to attribute the changes of other physical variables or to focus on the regional scale where the non-additivity of the effect of forcings could be more present.<br />Keywords : detection and attribution, climate models, non-additivity, neural networks, interpretable AI</p>