Learning to Represent the Physics of Climate Transitions as an Inverse Problem

Ignacio Lopez-Gomez

Keywords: earth system models, climate transitions, inverse problems

Abstract:

Beyond the rise of global mean surface temperature, there is
growing evidence that anthropogenic climate change may
trigger rapid climate transitions if certain environmental
thresholds, known as tipping points, are surpassed. Formulating
effective policies to prevent and mitigate the effect of
these transitions requires Earth System Models (ESMs) that
can simulate them accurately across a range of possible socioeconomic
pathways. Advances in computational power now
enable simulating some of these transitions regionally at an
unprecedented resolution and scale, providing valuable information
about their underlying mechanisms. Physical parameterizations
within ESMs can be trained to represent such
transitions, by minimizing the mismatch between the ESM
statistics and those of the resolved regional simulations for a
wide range of conditions across the tipping point of interest.
This training task can be formulated as an inverse problem
and solved using ensemble Kalman methods. The methodology
is gradient-free, non-intrusive, and provides means to
quantify the parametric uncertainty of the trained model. We
demonstrate this approach by training a turbulence and convection
model to represent the stratocumulus to cumulus transition
in the eastern Pacific Ocean.