The High-Tune Explorer : first tests and publications

High-Tune Explorer is a suite of programs developed in the frame of the High-Tune project, supported by the ANR french research agency. The data available on the DOI is linked to a 2-part publications in Journal of Advances in Modeling Earth Systems Process-based climate model development harnessing machine learning: I. a calibration tool for parameterization improvement A major task in the development of atmospheric models is the development of parameterizations to account for processes not resolved by the dynamical core. The improvement of models is slow partly due to the difficulty of encompassing key processes into parameterizations and because parameterizations contain `free' parameters that must be calibrated or `tuned'. Considering the number of parameters in a model, their calibration is a complicated task, generally done manually. Recently, machine learning has been proposed as a replacement for these parameterizations. However, when models are to be used for long-term projections, exploring states far from the training data, sole use of machine learning might be dangerous. It also seems counter-intuitive to replace our strong physical understanding with unconstrained systems. Our proposition consists in retaining parameterizations but adjoining new tools relying on machine learning to accelerate model development. In particular we use Gaussian process-based methods from uncertainty quantification to calibrate the free parameters at a process level. To achieve this, we focus on the comparison of single-column simulations and reference large-eddy simulations over multiple boundary-layer cases. This paper describes the tools and the philosophy of tuning in single-column mode. Part 2 emphasizes how this framework can help accelerate model development. II. model calibration from single column to global. In view of the importance of global numerical models for the anticipation of future climate changes, their improvement is often considered too slow. We present a new approach that we believe could boost model improvement significantly. This approach promotes the use of machine learning techniques developed by the "uncertainty quantification" community for the adjustment of model free parameters, or tuning. These techniques are applied to physics improvement at process scale, represented through parameterizations. In this approach, the tuning of the global atmospheric model is preconditioned by calibration of the model free parameters on a series of well documented cloud scenes for which explicit very high resolution simulations are available (see Part I). We demonstrate on a real example how the reduction of the parameter space with this approach allows us to save a large amount of computer resources and detract from the long and tedious by-hand phase of model tuning. By automating part of the tuning process, the approach enables climate modeler expertise to focus on understanding and improving the model physics through parameterization.