A Highly Efficient Ozone (O3) Parameterization for Climate Sensitivity Simulations
Project ID: O3PCSS
Ozone is an important radiatively active trace gas in the atmosphere. It is coupled in a complicated way with temperature and circulation changes under climate change, yet many climate models prescribe ozone in a static way, because it is computational expensive to model it correctly. However, we know that those models that prescribe ozone (“non-interactive ozone”) do not capture some essential feedbacks fully. Thus, we propose a computationally cheap way to include “interactive ozone” into state-of-the-art climate models. The computationally cheap ozone chemistry module we are proposing will be based on the proof of concept study by Nowack et al., 2018. This scheme will be implemented into a climate configuration of the ICON-ART model (e.g. Schröter et al., 2018) and the resulting model system will be characterised, e.g. in terms of climate sensitivity (CS). CS is a key parameter that explains how much near surface warming we should expect when greenhouse gases are rising and it is known that CS depends on ozone feedbacks (e.g. Nowack et al., 2015).
Figure: Indicative illustration of how different climate states might be sampled in the CCM data acquired for WP1. Given community assumptions for “likely climate scenarios” some climate states will be better represented in the data set than others. The aim would be to allow for a quasi-continuous representation of climate states. In addition, chlorine loading as a metric for the severity of ozone loss will have different magnitudes. Thus, determining the size of the ozone hole.
Ozone is an important radiatively active trace gas in the atmosphere. It is coupled in a complicated way with temperature and circulation changes under climate change, yet many climate models prescribe ozone in a static way, because it is computational expensive to model it correctly. Here, we will address the challenge of creating a computationally cheap, transferable and well tested ozone chemistry module. The Earth System Science (ESS) domain challenge is to understand feedbacks in the climate system between composition, temperature and circulation. The computational challenge is to train such a model efficiently on existing data (hundreds of terabytes) and to implement the resulting module efficiently into a state-of-the-art chemistry-climate model for multiple platforms. Even though a comprehensive atmospheric chemistry is computationally expensive (e.g. Esentürk et al., 2018) many long (decadal to centennial) integrations haven been performed in international collaborations over the past years - many of them preserved at institutions like, e.g., CEDA (UK) and DKRZ (Germany). This formidable data base of chemistry-climate integrations will be used to train a machine learning model (Nowack et al., 2018) to generate a computationally cheap “replacement chemistry” for climate models. Here, a number of domain-based and computational challenges will be tackled: The scheme is based on temperature as the input variable – however, to make it transferable, systematic biases have to be dealt with (Nowack et al., 2019). Thus, the proof of concept scheme includes an option for a PCA based dimension reduction. Here, the trade-off between the (very large) size of the data sets (on which we learn) and the additional cost of performing the PCA (to reduce the data amounts and make the bias correction easier) has to be explored. In addition, our implementation of the scheme will consider as a parameter the so-called “chlorine loading”. A measure for the amount of anthropogenic ozone destruction by halogens (including – as a phenomenon – the well-known ozone hole in the southern hemisphere). There are different possibilities to implement this parameter dependence, by e.g. classifying the input data and creating “different” models for different regimes, with a provision of validity boundaries (this is approach is known to work) or, alternatively, to use the chlorine loading parameter as an additional input feature supplemental to the temperature (this second possibility has yet to be tested in this project – however, with a direct link to HAICU/AIM we have the opportunity to create vouchers for additional support actions – a mechanism well tested in PBs institute).
Given the proof of concept results, there is no doubt that a well-balanced – cheap, accurate and portable – ozone chemistry module will be achieved. However, the implementation of such a scheme into an existing chemistry-climate model will add a novel aspect to the project. In the later part of the project this ozone chemistry module (classified as a parametrization module in the context of a comprehensive climate model, thus called O3P for short), will be implemented in ICON-ART. ICON-ART is a comprehensive (atmospheric) chemistry-climate model that is developed by a consortium of partners (the ICON consortium: DWD, MPI-M, DKRZ, KIT – with KIT being responsible for the comprehensive composition module ART). Schröter et al., 2018 provides an overview regarding aspects of the ICON-ART implementation for first climatic applications and Weimer et al., 2021 describes the use of nested grids for downscaling a comprehensive chemistry to regional scales. Here, we want to make sure that the O3P will run in both configurations – global and regional – in a seamless manner. Again, domain and computational knowledge are both indispensable to achieve a good implementation so that the scheme stays computationally cheap and accurate, even when varying grids are used. Challenges that will be faced during the implementation phase include: Making the system fit for running on hybrid architectures (classical CPU versus GPU based computations – this task is obviously part of a bigger picture, because the ICON model itself is currently ported on GPUs – so the challenge is to track the ICON developments such that the O3P can run in multiple environments (CPU and GPU based) – to support this process the ICON gatekeeper at IMK-ASF will support this work as well). Also dealing with large data sets for validating the results and judging the quality of the results (e.g. defining the appropriate tests and metrices) will be a rewarding task (here, initiatives like NFDI4Earth and experience from other projects, like EOSC Synergy and the Helmholtz Analytics Framework (HAF) will be available).
To finish of the project, the ESS relevant question regarding climate sensitivity and the role of interactive ozone in the climate system will be addressed in (at least) two longer integrations that compare the climatic model responses for interactive and non-interactive integrations – addressing the meteorological motivation of this proposal. Here, feedbacks between ozone, temperature and circulation will be analysed – and changes that rely on the direct coupling between ozone and temperature (via the radiation) will be identified and quantified. O3P will be openly available for other models as well.