Detection and attribution of climate change: A deep learning and variational approach

Twelve climate models and observations are used to attribute the global mean surface temperature (GMST) changes from 1900 to 2014 to external climate forcings. The external forcings are decomposed into the effects of the well-mixed greenhouse gas concentration variation, the effects of anthropogenic aerosol concentration changes, and the effects of natural forcings. First, a convolutional neural network (CNN) is trained to estimate the simulated historical GMST from single-forcing experiments using outputs from the multi-model ensemble. We then use this CNN to solve the attribution problem using an original variational inversion approach. The variational inversion is first validated using historical climate simulations as pseudo-observations. Then we perform an inversion from observations. This provides a distribution of the GMST resulting from the three forcings. For 2014, inversions estimate that the greenhouse gases changes are responsible for a GMST anomaly within [0.8 degrees C, 1.9 degrees C], while anthropogenic aerosols and natural forcings anomalies are within [similar to 0.7 degrees C, similar to 0.1 degrees C] and [similar to 0.1 degrees C, 0.3 degrees C], respectively. The method designed here can be adapted and extended to attribute the changes of other variables or to focus on the regional scale. Impact Statement To devise efficient adaptation policies, it is key to understand the causes of past climate changes. Here, we present a method based on neural networks to estimate the past global mean surface temperature (GMST) anomalies caused by the changes in the greenhouse gas concentration, the variation of anthropogenic aerosols, and the variation driven by naturally occurring phenomena. This method is based on the training of a convolutional neural network using the estimations from 12 state-of-the-art climate models. Then we infer the most likely causes for the observed GMST changes from 1900 to 2014. The methodology presented could be applied in future studies to other variables or at the regional scale.