[PAOCQ] Maria Rugenstein (Colorado State)

“Connecting surface warming, global radiation, and radiative forcing with convolutional neural networks”

Recent research has highlighted that radiative feedbacks — and thus climate sensitivity — are not constant in time but depend sensitively on surface temperature patterns (“pattern effect”). First, I will review physical drivers of the effect and explain how it matters for near-term projections of climate change. Then, I will discuss three implications we have recently gained:

(1) I will introduce how a non-linear convolutional neural networks (CNN) can be trained on climate model output to learn the pattern effect and predict global-mean TOA radiation from surface warming patterns. We use explainable artificial intelligence methods to visualize and quantify where the CNN draws its predictive skill. Remarkably and different from traditional approaches, we can predict radiation under strong climate change from training the CNN on internal variability alone. This out-of-sample application works only when feedbacks are allowed to be non-linear or equivalent, changing in time.

(2) We train CNNs on internal variability of climate models alone, to avoid the need to rely on their erroneously simulated surface warming patterns and uncertain implementation of greenhouse gas and aerosol emissions. We then derive a new estimate of the Earth’s radiative forcing over the last decades (which is unobservable), by combining the radiative effect learned in this minimalistic way in combination with the observed energy imbalance. Our forcing estimate is substantially higher than traditional ones and points to a strong aerosol radiative effect.

(3) We employ linear regularized regression methods to quantify the spatial controls of global and regional lower tropospheric inversion strength (EIS). EIS is an important cloud controlling factor of shallow cumulus clouds, which are the most relevant player in the pattern effect. We show that for variations in near-global EIS, the tropical weak temperature gradient approximation is indeed a good explanation. However, spatial controls of EIS in the regions of shallow cumulus clouds are much more varied than suggested by the tropical weak temperature gradient assumption. We quantify to what degree internal variability and decadal-long trends in EIS in the Eastern subtropical Pacific are driven locally versus remotely.

PAOC Colloquium —

Interdisciplinary seminar series that brings together the whole PAOC (Program in Atmospheres, Oceans, and Climate) community. Seminar topics include all research concerning the physics, chemistry, and biology of the atmospheres, oceans and climate, as well as talks about societal impacts of climatic processes.

Contact: paoc-colloquium-comm@mit.edu