Process‐Based Climate Model Development Harnessing Machine Learning: II. Model Calibration From Single Column to Global

Abstract

We demonstrate a new approach for climate model tuning in a realistic situation. Our approach, the mathematical foundations and technical details of which are given in Part I, systematically uses a single‐column configuration of a global atmospheric model on test cases for which reference large‐eddy‐simulations are available. The space of free parameters is sampled running the single‐column model from which metrics are estimated in the full parameter space using emulators. The parameter space is then reduced by retaining only the values for which the emulated metrics match large eddy simulations within a given tolerance to error. The approach is applied to the 6A version of the LMDZ model which results from a long investment in the development of physics parameterizations and by‐hand tuning. The boundary layer is revisited by increasing the vertical resolution and varying parameters that were kept fixed so far, which improves the representation of clouds at process scale. The approach allows us to automatically reach a tuning of this modified configuration as good as that of the 6A version. We show how this approach helps accelerate the introduction of new parameterizations. It allows us to maintain the physical foundations of the model and to ensure that the improvement of global metrics is obtained for a reasonable behavior at process level, reducing the risk of error compensations that may arise from over‐fitting some climate metrics. That is, we get things right for the right reasons.

Calibration and Uncertainty Quantification of Convective Parameters in an Idealized GCM

“…The idea complements that of Bayesian uncertainty quantification, where instead of searching for a high probability region of parameter space with respect to data, one rules out regions of parameter space that are deemed inconsistent with the data. Couvreux et al (2020) and Hourdin et al (2020) recently constrained the parameter space of a parameterization scheme by approximating a plausibility function over the parameter space using a Gaussian process, and then removing "implausible" regions of parameter space where the plausibility function passes a threshold. This removal process is iterated until the uncertainty of the emulator is small enough, or the space becomes empty.…”

Process‐Based Climate Model Development Harnessing Machine Learning: III. The Representation of Cumulus Geometry and Their 3D Radiative Effects

“…Thirteen iterations were applied, reducing the NROY space from 11.7% of the original space after the first wave, to 8.40% after the twelfth wave, and 8.39% after the thirteenth wave, where the process was assumed to have reached convergence. It would have been possible to further reduce the NROY space by decreasing the rejection threshold or adding new constraints (new metrics), as was done in Couvreux et al (2020) and Hourdin et al (2020). However, the aim of this study is not to determine a unique set of acceptable parameters but rather to analyze the structure of the parameter space and compare various configurations that are acceptable given the arbitrarily chosen tolerance.…”

“…Does the best choice for cloud parameters match the LES-derived values of Section 2? The High-Tune: Explorer calibration tool (Couvreux et al, 2020;Hourdin et al, 2020) is used in the following to answer these questions. The tool is fully described in Part I.…”

“…Part I (Couvreux et al, 2020) describes this approach and the associated numerical tools. Part II (Hourdin et al, 2020) applies them to the calibration of a 3D climate model after prior calibration of the parameterization of shallow convection in the SCM/LES framework.…”

Calibration and Uncertainty Quantification of Convective Parameters in an Idealized GCM

“…Traditionally, parameters are calibrated ("tuned") by hand, in a process that exploits only a small subset of the available observational data and relies on the knowledge and intuition of climate modelers about plausible ranges of parameters and their effect on the simulated climate of a model (Randall & Wielicki, 1997;Mauritsen et al, 2012;Golaz et al, 2013;Hourdin et al, 2013;Flato et al, 2013;Hourdin et al, 2017;Schmidt et al, 2017;Zhao et al, 2018). More recently, some broader-scale automated approaches that more systematically quantify the plausible range of parameters have begun to be explored (Couvreux et al, 2020;Hourdin et al, 2020). However, to fully account for parametric uncertainty, we require a Bayesian view of the model-data relationship, where model parameters are treated as realizations sampled from an underlying probability distribution.…”