Semi-automatic tuning of coupled climate models with multiple intrinsic timescales: lessons learned from the Lorenz96 model

Abstract

The History Matching method is explained in detail then used for tuning a toy coupled model: the Lorenz 96 model • The importance of several design choices is demonstrated, especially when considering forced experiments such as AMIP and OMIP • We argue that this tuning method is semi-automatic and highlight the importance of the expert-in-the-loop paradigm when considering it for real coupled models

Bayesian History Matching Applied to the Calibration of a Gravity Wave Parameterization

“…Where σ z is the observational uncertainty and $${\sigma }_{\widehat{f}}$$ is the uncertainty in the emulator prediction. Typically for the univariate case (Andrianakis et al., 2015; Couvreux et al., 2021; Lguensat et al., 2023; Williamson et al., 2013), Pukelsheim's rule is invoked which states that for a continuous unimodal distribution 95% of the probability mass lies within 3 standard deviations of the mean value (Pukelsheim, 1994). For the multivariate case, as in this work with a 2‐dimensional z , there are various definitions of the implausibility.…”

“…This approach was initially developed to calibrate models for oil exploration (Craig et al., 1997) and has found wide utility in various disciplines. This includes in calibrating models of galactic formation (Williamson et al., 2013), HIV disease transmission (Andrianakis et al., 2015) and recently in calibrating multi timescale dynamical systems (Lguensat et al., 2023).…”

Bayesian History Matching applied to the calibration of a gravity wave parameterization

“…Where σ z is the observational uncertainty and σ f is the uncertainty in the emulator prediction. Typically for the univariate case (Andrianakis et al, 2015;Lguensat et al, 2023;Couvreux et al, 2021;Williamson et al, 2013), Pulksheim's rule is invoked which states that for a continuous unimodal distribution 95% of the probability mass lies within 3 standard deviations of the mean value (Pukelsheim, 1994). For the multivariate case, as in this work with a 2-dimensional z there are various definitions of the implausibility.…”

“…This approach was initially developed to calibrate models for oil exploration (Craig et al, 1997) and has found wide utility in various disciplines. This includes in calibrating models of galactic formation (Williamson et al, 2013), HIV disease transmission (Andrianakis et al, 2015) and recently in calibrating multi timescale dynamical systems (Lguensat et al, 2023).…”

“…By comparing these predictions to the true observables we calculate an implausibility statistic which is minimized in regions of space where the predictions agree with the observations or those with high uncertainties. By determining the regions where this implausibility is below a certain threshold we obtain the "Not Yet Ruled Out" (NROY) space (Lguensat et al, 2023), a uniform space of parameters that, relative to the uncertainties, simulate a QBO consistent with observations.…”