Reliable Climate Projections: The Final Frontier for Stochastic Parametrisation

Lead Research Organisation: University of Oxford
Department Name: Oxford Physics

Abstract

Reliable climate projections are needed by governments and industry to make decisions in response to climate change. For example, the governmental National Flood Resilience Review, published in September 2016, used risks calculated from climate projections to evaluate how to best protect the country from extreme weather and flood events. By weighing up the cost of investment against the predicted risks, the government decided to invest in temporary protective flood barriers.

However, it is likely that our current climate projections do not accurately represent the possibility of different climate outcomes. On seasonal timescales, where we can test the skilfulness of our predictions, we can see that our current models are unable to reliably predict the likelihood of different events. The shortcomings observed in seasonal forecasts will also infect climate prediction. This limits the usefulness of current climate models for decision-making.

I propose to develop a new technique for representing uncertainty in climate models. This will improve our ability to predict the likelihood of different climate outcomes. The key aim is to improve the reliability of climate projections, and therefore their usefulness to decision-makers and end users.

In particular, I will target uncertainty in climate prediction due to the limitations of the climate model. The process of representing the atmosphere, oceans and land-surface and their many interactions in a piece of computer code is a large source of uncertainty. Some processes in the atmosphere take place on too small a scale to be explicitly represented on the grid used by a computer model, where grid points are typically 10-100 km apart. Instead, these processes are accounted for using "parametrisation schemes".

In the case of weather and seasonal forecasting, random numbers are widely used as a tool to help represent these small-scale processes in the model. These so called "stochastic" parametrisation schemes improve the consistency of the model with the underlying physical processes. The introduction of stochastic parametrisation schemes produced a step-change improvement in weather and seasonal forecasts.

However, current stochastic schemes do not directly target the physical source of uncertainty. Instead they simply target the impact of the uncertainty on the forecast. While this is sufficient for short-range applications, these schemes are not mathematically rigorous, and do not necessarily conserve moisture, mass or energy over longer time periods. This is very problematic for climate prediction.

My research will provide a set of new, physically motivated stochastic parametrisation schemes. By targeting uncertainty at the source, these schemes will be useful for climate projections as well as on weather and seasonal timescales.

I will develop this set of new stochastic schemes by comparing a typical forecast model to a high-resolution atmospheric simulation. My project partners across Europe will provide these state-of-the-art simulations in which small-scale processes are explicitly modelled. In the forecast model, these small-scales must be represented through parametrisation schemes. By comparing the two, I will identify processes that are better represented in a probabilistic manner instead of in the traditional deterministic manner.

I will initially evaluate my new schemes in weather and seasonal forecasts. This provides a way to validate the representation of fast timescale processes in the forecast model. Accurate representation of these processes is necessary for realistic climate simulation, due to the non-linear nature of the atmosphere. Having validated my new approach in this way, I will produce a probabilistic climate change projection where, for the first time, model uncertainty is accounted for in a complete and physically consistent manner.

Planned Impact

The following groups will benefit from this research

Operational forecasting centres

The new stochastic schemes I produce will be included into two operational models during the project: the European Centre for Medium Range Weather Forecasts and the UK Met Office. The new suite of stochastic parametrisation schemes will improve the skill of forecasts on a range of timescales. In particular, the reliability of weather and seasonal forecasts will improve, which will make them more useful for end users. This will directly benefit the modelling centres, as it increases the economic value of their forecasts. The schemes will also be suitable for use in climate models, where they have the potential to improve probabilistic climate projection. It is anticipated that other modelling groups worldwide will readily adopt the schemes.

Policy makers

Using a complete set of physically motivated stochastic parametrisation schemes to represent model uncertainty has the potential to improve probabilistic climate projection compared to the perturbed parameter or multi-model approaches currently used. The new schemes will improve the reliability of the projection for a given forcing scenario. This will directly benefit policy makers, who can make better-informed decisions. This will increase the effectiveness of policy, and facilitate planning to mitigate and adapt to climate change.

Forecast users

Improving operational probabilistic weather and seasonal forecasts improves our ability as a society to prepare, both for severe weather events such as storms, and for seasonal warnings of drought or flooding. A better-calibrated early warning system allows both companies and members of the public to plan and react. This research will therefore benefit society and especially those sectors of society that use weather and climate models for quantitative decision-making. Examples include:

- Industry
Forecast users in industry will benefit economically from the improved forecasts on a range of timescales. For example, the UK rail network can decide whether to invest in new 'slab tracks' that are very costly, but more resistant to buckling in hot weather, given that the number of hot days is predicted to increase due to changes in climate. On a shorter timescale, forecasts of extreme wind could inform a wind farm manager of the need to shut down the wind turbines to prevent damage in a storm. In both cases, if the forecasts used to inform these decisions are unreliable, businesses may make decisions that are costly to them (e.g. shutting down a turbine), while the benefits of making that decision are never realised (e.g. the storm does not materialise). Similarly, improved forecasts will directly increase the economic performance of these industries, and thereby improve the economic competitiveness of the United Kingdom.

- Charity sector
Improved forecasts on seasonal timescales have many important humanitarian applications, including the likelihood of floods or droughts for agricultural planning, and the projections of the prevalence of infectious diseases in a given season. The charity sector is showing an interest in using seasonal forecast information to plan relief programs in regions at risk: improved reliability of seasonal forecasts will improve the efficacy of these projects.

- Public
Improved weather forecasts benefit the public, as the information can be integrated into personal plans for hours or days ahead. This increased confidence in forecasts will in return be beneficial to the forecasting centres. The public will also benefit through improved decision-making by industries that provide public services, such as water companies, transport providers, and local government. Longer term, the health and life of the general public will benefit from an increased resilience to climate change, enabled by good policy decisions made today.

Publications

10 25 50
 
Description The most significant achievements of the award so far are:

1. Demonstrating the use of a coarse graining techniques for assessing model error. I propose a new technique which combines high-resolution simulations with a low-resolution forecasting model (a 'single column model'). This enables the error in the forecast model to be measured. This is important for producing probabilistic forecasts, in which we account for the uncertainty in our prediction. In particular, stochastic parametrisations provide a way to account for errors in our model when producing the forecast. I have used these new measurements of model error to assess existing stochastic parametrisation schemes as well as motivate new approaches (Christensen et al, 2018; Christensen, submitted). I have simultaneously made use of more traditional coarse-graining approaches to characterise uncertainty in specific processes of interest in a low-resolution model (Bessac et al, 2019).

2. I have continued to demonstrate the potential for using stochastic parametrisation schemes in climate models, with a particular focus on the beneficial impact such schemes have on the El Nino Southern Oscillation (Berner et al, 2018; Yang et al, 2019). It is important that seasonal and climate forecasts can capture the behaviour of the El Nino Southern Oscillation, as it has widespread impacts on extreme weather events, including the likelihood of droughts and flooding.
Exploitation Route These findings will be taken forward by the PI over the remaining part of the award. In addition:

- the approach of Bessac et al (2019) forms the basis for a new proposal submitted to the Natural Sciences and Engineering Research Council of Canada (results to be announced in April/May 2019)
- the new approach of Christensen et al, (2018) and Christensen (submitted) forms the basis of a proposed project under the World Weather Research Programme and World Climate Research Programme for a co-ordinated 'Model Uncertainty Intercomparison Project'. Discussions are underway to confirm which international groups wish to be involved.
Sectors Aerospace, Defence and Marine,Agriculture, Food and Drink,Energy,Environment,Leisure Activities, including Sports, Recreation and Tourism,Government, Democracy and Justice,Retail,Other

 
Title Forcing files for the ECMWF Integrated Forecasting System (IFS) Single Column Model (SCM) over Indian Ocean/Tropical Pacific derived from 10-day high resolution CASCADE simulation 
Description This data set consisting of initial conditions, boundary conditions and forcing profiles for the Single Column Model (SCM) version of the European Centre for Medium-range Weather Forecasts (ECMWF) model, the Integrated Forecasting System (IFS). The IFS SCM is freely available through the OpenIFS project, on application to ECMWF for a licence. The data were produced and tested for IFS CY40R1, but will be suitable for earlier model cycles, and also for future versions assuming no new boundary fields are required by a later model. The data are archived as single time-stamp maps in netCDF files. If the data are extracted at any lat-lon location and the desired timestamps concatenated (e.g. using netCDF operators), the resultant file is in the correct format for input into the IFS SCM. The data covers the Tropical Indian Ocean/Warm Pool domain spanning 20S-20N, 42-181E. The data are available every 15 minutes from 6 April 2009 0100 UTC for a period of ten days. The total number of grid points over which an SCM can be run is 480 in the longitudinal direction, and 142 latitudinally. With over 68,000 independent grid points available for evaluation of SCM simulations, robust statistics of bias can be estimated over a wide range of boundary and climatic conditions. The initial conditions and forcing profiles were derived by coarse-graining high resolution (4 km) simulations produced as part of the NERC Cascade project, dataset ID xfhfc (also available on CEDA). The Cascade dataset is archived once an hour. The dataset was linearly interpolated in time to produce the 15-minute resolution required by the SCM. The resolution of the coarse-grained data corresponds to the IFS T639 reduced gaussian grid (approx 32 km). The boundary conditions are as used in the operational IFS at resolution T639. The coarse graining procedure by which the data were produced is detailed in Christensen, H. M., Dawson, A. and Holloway, C. E., 'Forcing Single Column Models using High-resolution Model Simulations', in review, Journal of Advances in Modeling Earth Systems (JAMES). 
Type Of Material Database/Collection of data 
Year Produced 2018 
Provided To Others? Yes  
Impact I am using this dataset to characterise the errors in a low resolution forecast model. These measurements will be used to improve the representation of model uncertainty in weather and climate models. 
URL http://dx.doi.org/10.5285/bf4fb57ac7f9461db27dab77c8c97cf2
 
Description CCSP-Brazil 
Organisation Centre for Ecology & Hydrology (CEH)
Country United Kingdom 
Sector Public 
PI Contribution Development of software to assess key modes of variability relevant to Brazilian Climate. Identify whether 1) modes differ between models and observations and 2) modes in a model change in a changing climate.
Collaborator Contribution Use of software to bias correct modes of variability in future climate simulations, to allow climate simulations to be blended with observational data to create a seamless time series for forcing land surface models.
Impact Software production underway for producing the required data sets.
Start Year 2018
 
Description SAMSI Machine Learning 
Organisation NCAR National Center for Atmospheric Research
Country United States 
Sector Academic/University 
PI Contribution Expertise in development of stochastic parametrisation schemes, analysis of ensemble weather forecasts, climate diagnostics. Contributions to experiment design, coding, performing experiments, analysis of data, writing publications.
Collaborator Contribution Expertise in machine learning and other advanced statistical techniques. Contributions to experiment design, coding, performing experiments, analysis of data, writing publications. Presentation of results at conferences
Impact Paper in preparation: Gagne DJ, Christensen HM, Subramanian A, and Monahan A Machine Learning for Stochastic Parameterization: Generative Adversarial Networks in the Lorenz '96 Model Presentations at: American Geophysical Union Fall Meeting in Washington DC, 2018 American Meteorological Society Annual Meeting in Phoenix, 2019 European Geosciences Union annual meeting in Vienna, 2019 International Union of Geodesy and Geophysics meeting in Montreal, 2019 The collaboration is multi-disciplinary: Gagne: Data science Christensen, Subramanian: Atmospheric science Monahan: Statistics
Start Year 2017
 
Description SAMSI Machine Learning 
Organisation University of Colorado Boulder
Department College of Engineering & Applied Science
Country United States 
Sector Academic/University 
PI Contribution Expertise in development of stochastic parametrisation schemes, analysis of ensemble weather forecasts, climate diagnostics. Contributions to experiment design, coding, performing experiments, analysis of data, writing publications.
Collaborator Contribution Expertise in machine learning and other advanced statistical techniques. Contributions to experiment design, coding, performing experiments, analysis of data, writing publications. Presentation of results at conferences
Impact Paper in preparation: Gagne DJ, Christensen HM, Subramanian A, and Monahan A Machine Learning for Stochastic Parameterization: Generative Adversarial Networks in the Lorenz '96 Model Presentations at: American Geophysical Union Fall Meeting in Washington DC, 2018 American Meteorological Society Annual Meeting in Phoenix, 2019 European Geosciences Union annual meeting in Vienna, 2019 International Union of Geodesy and Geophysics meeting in Montreal, 2019 The collaboration is multi-disciplinary: Gagne: Data science Christensen, Subramanian: Atmospheric science Monahan: Statistics
Start Year 2017
 
Description SAMSI Machine Learning 
Organisation University of Victoria
Department Department of Biology
Country Canada 
Sector Academic/University 
PI Contribution Expertise in development of stochastic parametrisation schemes, analysis of ensemble weather forecasts, climate diagnostics. Contributions to experiment design, coding, performing experiments, analysis of data, writing publications.
Collaborator Contribution Expertise in machine learning and other advanced statistical techniques. Contributions to experiment design, coding, performing experiments, analysis of data, writing publications. Presentation of results at conferences
Impact Paper in preparation: Gagne DJ, Christensen HM, Subramanian A, and Monahan A Machine Learning for Stochastic Parameterization: Generative Adversarial Networks in the Lorenz '96 Model Presentations at: American Geophysical Union Fall Meeting in Washington DC, 2018 American Meteorological Society Annual Meeting in Phoenix, 2019 European Geosciences Union annual meeting in Vienna, 2019 International Union of Geodesy and Geophysics meeting in Montreal, 2019 The collaboration is multi-disciplinary: Gagne: Data science Christensen, Subramanian: Atmospheric science Monahan: Statistics
Start Year 2017
 
Description SAMSI stochastic Wind Fluxes 
Organisation Argonne National Laboratory
Department Chemical Sciences and Engineering Division
Country United States 
Sector Public 
PI Contribution Expertise in stochastic parametrisations. Expertise in Earth-system modelling. Implementation of new stochastic parametrisation in European Centre for Medium Range Weather Forecasts Integrated Forecasting System.
Collaborator Contribution Expertise in statistical analysis, in particular, statistical modelling of space-time processes. Expertise in air-sea interactions and modelling of air-sea fluxes.
Impact Manuscript: "Stochastic parameterization of subgrid-scale velocity enhancement of sea surface fluxes", by Julie Bessac; Adam H. Monahan; Hannah M Christensen; Nils Weitzel Accepted in Monthly Weather Review (8 Feb 2019). Presentations at: Joint Statistical Meeting in Vancouver in 2018 International Union of Geodesy and Geophysics meeting in Montreal in 2019 This collaboration is multidisciplinary. Atmospheric science: Christensen Statistics: Bessac, Monahan and Weitzel
Start Year 2017
 
Description SAMSI stochastic Wind Fluxes 
Organisation University of Bonn
Department Department of Neurology
Country Germany 
Sector Hospitals 
PI Contribution Expertise in stochastic parametrisations. Expertise in Earth-system modelling. Implementation of new stochastic parametrisation in European Centre for Medium Range Weather Forecasts Integrated Forecasting System.
Collaborator Contribution Expertise in statistical analysis, in particular, statistical modelling of space-time processes. Expertise in air-sea interactions and modelling of air-sea fluxes.
Impact Manuscript: "Stochastic parameterization of subgrid-scale velocity enhancement of sea surface fluxes", by Julie Bessac; Adam H. Monahan; Hannah M Christensen; Nils Weitzel Accepted in Monthly Weather Review (8 Feb 2019). Presentations at: Joint Statistical Meeting in Vancouver in 2018 International Union of Geodesy and Geophysics meeting in Montreal in 2019 This collaboration is multidisciplinary. Atmospheric science: Christensen Statistics: Bessac, Monahan and Weitzel
Start Year 2017
 
Description SAMSI stochastic Wind Fluxes 
Organisation University of Victoria
Department Department of Biology
Country Canada 
Sector Academic/University 
PI Contribution Expertise in stochastic parametrisations. Expertise in Earth-system modelling. Implementation of new stochastic parametrisation in European Centre for Medium Range Weather Forecasts Integrated Forecasting System.
Collaborator Contribution Expertise in statistical analysis, in particular, statistical modelling of space-time processes. Expertise in air-sea interactions and modelling of air-sea fluxes.
Impact Manuscript: "Stochastic parameterization of subgrid-scale velocity enhancement of sea surface fluxes", by Julie Bessac; Adam H. Monahan; Hannah M Christensen; Nils Weitzel Accepted in Monthly Weather Review (8 Feb 2019). Presentations at: Joint Statistical Meeting in Vancouver in 2018 International Union of Geodesy and Geophysics meeting in Montreal in 2019 This collaboration is multidisciplinary. Atmospheric science: Christensen Statistics: Bessac, Monahan and Weitzel
Start Year 2017