Annual Seminar 2018

Issues specific to data sparse systems

Marc Bocquet (CEREA Joint Laboratory École des Ponts ParisTech and EdF R&D, Université Paris-Est, France)

Abstract

In this talk, I will present my view on the issue of data assimilation systems that use sparse observations. This subject will be introduced by outlying the specificities of sparse, usually surface in situ, observations. One way to account for a sparse network is to adapt the control space to the monitoring network using reduction methods, which I will discuss as a first topic. The primary goal of such methods is to reduce numerical complexity as well as getting a better conditioning for variational problems.  I will then focus on a less known, but quite fundamental topic of ill-defined data assimilation problems when using in situ observations in atmospheric chemistry models.  The first two topics are mainly discussed in a variational context. By contrast, the third topic will be focused on  the use of sparse observations within ensemble Kalman filters, which might be seen as a dual take on ensemble sampling errors.

Non-linearity and non-Gaussianity in variational data assimilation

Massimo Bonavita (ECMWF)

Abstract

The standard data assimilation theoretical framework assumes that errors in the analysis cycle evolve linearly and can be adequately described by Gaussian statistics. While these assumptions have long be recognised as a potential significant sub-optimality, the consensus in the global NWP community has been that they are not the main performance-limiting factor of analysis accuracy. Aim of this seminar is to help change this viewpoint.

We will show how the non-linearity and non-Gaussianity of the NWP data assimilation problem has increased over the years driven by the increases in model resolution and by the growing prominence of nonlinear observations in modern NWP data assimilation systems; and that the ability to deal effectively with non-linearity and non-Gaussianity is crucial to current analysis and forecast skill and will become an even more important factor in the future. Between the limited efficacy of purely linear solutions and the computational intractability of fully non-linear methods, a golden mean of perturbative approaches to non-linearity and careful treatment of deviations from Gaussianity appears the most promising path for further development.     

Ocean Data Assimilation for Numerical Weather Prediction

Phil Browne (ECMWF)

Abstract

As of June 2018, ECMWF has made seamless its forecasting systems, with all forecasts including a three-dimensional ocean and sea-ice model. This complements the existing atmosphere-land coupling present in the forecasts. With such an Earth-system approach being employed, initial conditions for all components need to be supplied. In particular, the ocean analysis is now a critical component of the forecasting system. In this talk we will describe the current and future strategies for ocean data assimilation at ECMWF, and how this fits within a coupled assimilation framework.

The differences in the ocean analysis system compared to the atmosphere are numerous and emerge from a number of intrinsic, practical, and technical reasons. One intrinsic difference is from the fact that the ocean has much longer timescales than the atmosphere. Practical differences come from the observation network in the ocean and the time it takes for deep ocean observations to be ingested into the NWP system. Technical differences stem from the ocean analysis system being developed separately from the atmospheric analysis and therefore using different tools which have to be combined.

In this talk we will introduce the ECMWF ocean analysis system. We will cover both the assimilation method and configuration, as well as details on the current observation network. This consists of in situ profiles of temperature and salinity, sea level anomaly, sea surface temperature, and sea ice concentration observations. We will look at the plans and challenges in using other available observations, such as lower level sea-surface temperature and sea ice concentration observations, observations of sea ice thickness, sea surface salinity, or ocean colour data etc.

Motivation and Methods in Earth System Data Assimilation

Mark Buehner (Environment and Climate Change Canada)

Abstract

Operational weather forecast models are increasingly being coupled with other components of the Earth system. For example, the ECCC operational global weather forecasts have been coupled with an ice-ocean model since November 2017. Benefits from coupled forecasts have been demonstrated even at shorter time-scales relevant for medium-range NWP. Such improvements can be attributed to coupled interactions associated with: tropical convection, hurricanes, extra-tropical storms, coastal upwelling, and rapid changes in sea ice cover. However, initialization of coupled models from independent assimilation systems for each component can be a challenge. Consequently, even as data assimilation methods for atmosphere-only applications are becoming more complex (including the use of hybrid methods), there is a clear motivation to improve the initialization of coupled models by coupling the assimilation procedures of the individual component systems. Various coupled data assimilation strategies have been proposed. This presentation will provide an overview of the need for, and potential benefits from, coupled Earth system data assimilation. Then, preliminary work on so-called weakly coupled data assimilation applied to the ECCC global coupled atmosphere-ice-ocean system will be presented. In addition, plans for future research on more strongly coupled assimilation and the potential application of scale-dependent approaches for estimating coupled background-error covariances will be discussed.

Characterising and Modelling Observation Errors

Sarah L. Dance (University of Reading)

Abstract

Observation uncertainty arises from a number of sources, including instrument noise; errors in the observation operator; issues in pre-processing; and mismatch between the scales observed and the scales resolved by the numerical model. In most cases, it is reasonable to assume that instrument errors are independent and uncorrelated. However, the other sources of error result in observation error correlations. 

Until recently, in operational data assimilation, the observation errors have been assumed uncorrelated and processes such as observation thinning and variance inflation have been used to either reduce the correlated error or attempt to account for the unknown correlations. To improve the accuracy of the analysis and the number of observations assimilated, it is necessary to understand and account for the full, potentially correlated, observation error statistics. These error statistics cannot be calculated directly so must be estimated statistically. We discuss a method for diagnosing observation uncertainty using observation-model departures, and demonstrate the results for different observation types including Doppler radar wind and satellite inter-channel error statistics. 

Mathematical studies on the incorporation of correlated observation errors in variational assimilation have established that the computational work needed to solve the assimilation problem increases as the observation error covariance matrix becomes ill-conditioned. In particular we show that the rate of convergence of the assimilation scheme depends on the minimum eigenvalue of the observation error correlation matrix. To reduce operational costs of the assimilation, the error correlation matrix is reconditioned by altering its eigenstructure.

In recent years the improved treatment of correlated inter-channel observation errors in data assimilation has been shown to improve the analysis accuracy and forecast skill scores at a number of operational centres.  New work assimilating dense sets of observations with spatial correlations also shows promise.

Towards coupled assimilation in operational systems

Patricia de Rosnay (ECMWF)

Abstract

NWP centres are moving toward using an Earth system approach with operational coupled data assimilation.  Pioneer work on coupled data assimilation started in the context of coupled ocean-atmosphere reanalyses.  At ECMWF century long and satellite era coupled reanalyses, CERA-20C and CERA-SAT, demonstrated the capability of the outer-loop coupling to simultaneously ingesting atmospheric and ocean observations in the coupled Earth system model.

Current ECMWF Integrated Forecasting System (IFS) cycles now include the modular possibility to use coupled assimilation following different approaches, ranging from weakly coupled assimilation (WCDA) to outer-loop coupling also called quasi strongly coupled assimilation (QSCDA).  WCDA was implemented in operations for sea ice in 2018 and will be considered for the ocean in 2019.  Evaluation the ocean-atmosphere QSCDA for NWP is being evaluated and demonstrated promise in tropical areas, but exposed some challenging issues in the extratropics.

Land-atmosphere assimilation systems used for NWP generally rely on WCDA approaches. Recent developments at ECMWF make use of the Ensemble Data Assimilation to compute the Jacobians in the extended Kalman filter (EKF) soil moisture analysis. This approach enables a dynamic link to the meteorological conditions, providing a new component to land-atmosphere coupling. It reduces the cost of the EKF, which opens the perspective of QSCDA testing for land-atmosphere and makes it possible to extend the land analysis to more variables and more satellite observations.

This presentation gives an overview of activities conducted to develop coupled assimilation across the ECMWF operational systems.  Coupled assimilation is still a relatively new field of research, with many open questions (e.g. error growth time scales) and remaining technical challenges.  Ongoing work focuses on modular DA development to explore different coupling methods and coupling through the observation operator (e.g. for snow covered surfaces). It includes modular system infrastructure developments, with consistent suite definition for the different Earth system components. Moving toward operational coupled assimilation also involves major work on ocean, sea ice and land observing systems at all levels (governance, acquisition, pre-processing, archiving, monitoring, etc…) to ensure that they satisfy operational NWP and reanalysis systems requirements and constraints. 

Coupled Land/Atmosphere Data Assimilation

Clara Draper (CIRES/NOAA)

Abstract

The land surface can have a profound impact on the diurnal evolution of the boundary layer. Additionally, land surface states, and particularly soil moisture, have a longer memory than the atmosphere, so that the land surface provides an important source of atmospheric predictability, up to at least seasonal time scales.  Atmospheric forecasts, across a range of time scales, can then be improved by using land data assimilation (DA) to improve a model’s initial land surface states.  The characteristics of the land surface differ from the atmosphere in several important ways, and this presentation will review how these differences have affected the design of land DA systems. Land DA is used by both the NWP community (working with AGCMs, to improve the atmosphere), and the hydrology community (working with offline land surface models, to improve water storage and fluxes).  Since these communities are ultimately interested in different processes, the approaches to land DA that have evolved within each are different. These differing approaches will be presented and compared. In both cases, the land DA has been tailored to improve the model components of greatest interest to that community (e.g., weather forecasts), however this does not then necessarily improve other model components (e.g., the land states). Nonetheless, progress is being made towards better integrated land DA approaches, capable of improving both the land and the atmosphere. As one example, results will be presented from recent coupled land/atmosphere DA experiments, in which satellite soil moisture is assimilated into a re-run of NASA’s MERRA-2 reanalysis, resulting in improved soil moisture, land surface fluxes, and screen-level temperatures.

Evolution of global observing systems

John Eyre (Met Office UK)

Abstract

Through its Rolling Review of Requirements (RRR) process, the WMO developed a “Vision for global observing systems in 2025”, which provides a high-level specification for a system of observing systems to address the need for observations for all WMO programmes and co-sponsored programmes.  The Vision is intended to be challenging but achievable on the 2025 timescale, and it covers both space-based and surface-based observing systems.  This presentation summarises this Vision and then reviews the progress and plans of the world’s space agencies to implement observing systems that contribute to meeting the space-based component of this Vision, with a focus on observations of interest to numerical weather prediction.  Under the new WMO Integrated Global Observing System (WIGOS) programme, a new “Vision for WIGOS in 2025” is being developed.  Key developments in the 2040 Vision, compared with the 2025 Vision, are presented both for the space-based and surface-based components.  The importance of the improving the timely international dissemination of observational data is stressed.

Comparison of data assimilation coupling strategies for Earth system models

Sergey Frolov (U.S. Naval Research Laboratory)

Abstract

Over the last decade, the Earth science community developed several prototypes of data assimilation methods for initialization of coupled model systems. These methods range from so-called weakly coupled data assimilation (DA) methods (where only the model guess is coupled but not the data assimilation system) to strongly coupled methods (where both the first guess and the DA are coupled). In this talk, I will present and contrast several prominent strategies for coupling in DA, including weakly coupled, outer loop-coupled, interface coupled, and strongly coupled methods. I attempt to provide guidance on when each of the methods is expected to perform best and what practical limitations we still need to overcome to improve the fidelity of the initial conditions for coupled Earth system models.

Coupling through the observation operator

Alan Geer (ECMWF)

Abstract

Satellite radiances are sensitive to many parts of the earth system. In the atmosphere that includes clouds, temperature and water vapour. At the surface that includes temperature, sea-ice, snow and soil moisture. Inverting radiances to obtain geophysical information requires the optimal separation of the wanted and unwanted information; one application's 'noise' is another application's 'signal'. This is a major challenge, but better use of satellite radiances has helped improve forecast quality over the last few decades. The move from assimilating temperature and humidity retrievals to assimilating radiances has eliminated the possibility of contamination by the a-priori. More recently, the direct assimilation of radiances affected by cloud and precipitation, i.e. all-sky assimilation, has the philosophy that rather than attempting to discard 'unwanted' information (e.g. by applying cloud detection) it is better to assimilate it. However, although the extraction of atmospheric information is increasingly sophisticated, the treatment of the boundary with the land, sea-ice and ocean is less so. A lot more satellite data is being used in situations where there is sensitivity to both surface and atmosphere, but it relies on less than optimal techniques, such as the skin temperature sink variable or dynamic surface emissivity retrievals. It would be better to constrain the inferred surface properties using geophysical modelling and by using accurate background errors. Furthermore, the same radiances that are assimilated for the atmosphere are often ingested separately via SST or sea-ice retrievals; these retrievals will have involved a sub-optimal representation of the atmosphere and can be several days late by the time they get into the system. The existing atmospheric radiance assimilation could help infer surface properties directly. Beyond this, the optimal use of satellite data demands a coupled data assimilation system, and that will require better geophysical and radiative transfer modelling at the surface.

Assimilation of Atmospheric Composition

Antje Inness (ECMWF)

Abstract

Atmospheric composition impacts air quality, weather and climate. Air pollution is one of the biggest environmental health risks and it is therefore important to provide air quality forecasts on global, regional and local scales. Ozone and aerosol interact with radiation and clouds microphysics, which has to be accounted for in weather forecasting and climate simulations. ECMWF has included atmospheric composition in the IFS in the last 13 years and is now running the Copernicus Atmosphere Monitoring Service (CAMS) to provide daily pollution forecasts and a range of datasets on global and regional atmospheric composition, e.g. near-real-time estimates of fire emissions, reanalyses of atmospheric composition and greenhouse gas forecasts and analyses (see atmosphere.copernicus.eu).

The assimilation of atmospheric composition data comes with a whole new range of challenges compared to the assimilation of meteorological data in NWP. Generally, atmospheric composition data are sparser, have larger biases and often do not resolve the scales of interest well, both spatially and temporally. A lot of processes take place in the boundary layer, which is not well monitored from space, and only a few species (out of over 100) can be observed.  While NWP forecasts depend predominantly on the initial state, parts of the chemical system are not sensitive to initial conditions, but dependent on model parameters or external fields (e.g. emissions, deposition, reaction rates). Data assimilation is challenging for short lived species (e.g. NO2) or for species that are not well constrained by the assimilated observations (e.g. individual aerosol components).

In this talk we highlight some of the special challenges faced when assimilating atmospheric composition data.  We will show how CAMS deals with the atmospheric composition retrievals we assimilate and will present examples of some of the CAMS datasets.

Linear model and adjoint coupling

Marta Janisková (ECMWF)

Abstract

The goal of data assimilation is to produce an accurate representation of the atmospheric state to initialize numerical weather prediction models. Using an assimilating model that describes atmospheric processes in a realistic way is key to provide better analyses and subsequent forecasts. Therefore, physical parameterizations have become essential components of data assimilation systems, and in particular of the four-dimensional variational method (4D-Var).

In variational data assimilation, physical parametrizations have two main applications. Firstly, they are used during the assimilation to link the model's prognostic variables (typically temperature, wind, humidity and surface pressure) to the observed quantities (e.g. radiances, reflectivities, backscatters). Secondly, particularly in 4D-Var, a good model with physical parametrizations is needed to evolve the model state from the beginning of the assimilation window to the time of each observation.

The benefits of including linearized physical parametrizations in the assimilating model was clearly demonstrated. However, their use also comes with some constrains and limitations that need to be considered. In particular, 4D-Var relies on the assumption that physical parametrizations and observation operators should be quasi-linear, to avoid convergence problems in the minimization of the cost function. But, at the same time, these should provide realistic enough sensitivities and model equivalent to observations, while remaining computationally affordable for operational purposes.

This talk will summarize the achievements and issues encountered when developing linearized physical parametrizations in an operational context. It will address such topics as computational efficiency, the validity of the tangent-linear approximation with increased resolution and longer time integrations, as well as parametrization complexity required to assimilate specific observation types. Finally, the future of the adjoint approach will be discussed.

Towards operational earth system assimilation : challenges and priorities

Jean-François Mahfouf (Météo-France/CNRS, Centre National de Recherches Météorologiques, France)

Abstract

The main purpose of this presentation will be to summarize the key challenges associated with the development of earth system assimilation for operational applications, together with the future directions that need to be explored by the scientific community. The challenges and priorities will be structured around the main components of the earth system : oceans, continental and oceanic surfaces (including sea-ice), atmospheric composition (gases and particules including hydrometeors), continental hydrology and the upper atmosphere. In order to complement the other presentations, a number of examples from ongoing developments undertaken at Météo-France in the context of limited area modelling and short-range forecasting will be presented. In particular, the level of complexity of the modelling system, that partly depends upon the spatial and temporal scales on interest, will be discussed since it has significant impacts on the building blocks of data assimilation. Indeed, one has to consider the number of variables to be analysed and their controllability by relevant observations, together with the estimation of their background error statistics with adequate filtering. The modularity of the various components of the earth system modelling offers an opportunity to reach a fully coupled data assimilation system in a progressive manner. On the other hand, the use of external modules may require to build collaborations outside the operational community for the code maintenance and its evolutions, with possible conflicts in terms of code structure, computing efficiency and data assimilation techniques. These specific threats and opportunities will be discussed.

Model error and parameter estimation

Chiara Piccolo (Met Office UK)

Abstract

An ensemble system requires both knowledge of the uncertainties in the initial conditions and the uncertainties in the forecast model. Data assimilation techniques are a natural way to estimate model error from observations because they allow for observation error to be taken into account in a systematic way. The statistic of the model error can only be estimated over a sufficiently long period under a stationarity assumption because the observations only represent a single realisation of the truth. Weak-constraint 4d-Var can therefore be used to fit a stochastic model to observations over a training period of time to generate an archive of analysis increments. This is not the same as estimating the real model error since analysis increments include the growth of the irreducible analysis error as well as the model error. Nevertheless, since variational methods provide a minimum variance estimation, this means that the analysis increments are the minimum required increments to allow the model to fit the observations within observation errors for a long period. If a reanalysis is performed using this system, and the analysis increment at each time can be regarded as a random draw from an archive of analysis increments with stationary statistics, then the reanalysis trajectory will be statistically indistinguishable from a realisation of a stochastic model forced by analysis increments randomly chosen from this archive.

Here we test how well the assumption of random increments is satisfied using the analysis increments from the Met Office data assimilation scheme, 4d-Var. We also test the performance of an ensemble of analyses and forecasts when using a model forced with increments randomly chosen from the archive. Finally, we compare this ensemble performance with the general approach of using stochastic schemes that simulate model error within the model itself.

Atmospheric composition coupled model developments and surface flux estimation

Saroja Polavarapu (Environment and Climate Change Canada)

Abstract

Many of society’s grand challenges, from identifying the impacts of air pollutants on human and ecosystem health to climate change, require knowledge of minor atmospheric constituents.  Thus the modeling and data assimilation of chemical constituents is integral to advancing scientific knowledge in these fields.  In addition, combining observations of atmospheric constituents with numerical models can provide improved estimates of constituent distributions as well as an environmental predictive capability.    While the techniques and tools of meteorological data assimilation can also be applied to atmospheric chemistry models, there are unique challenges associated with each scientific problem.  In this presentation, we use the context of the carbon cycle to showcase such challenges.  The greenhouse gas data assimilation problem pertains to the tracking of small exchanges of carbon masses between massive reservoirs in the terrestrial biosphere, the ocean and the atmosphere (with anthropogenic perturbations altering the balance).  The coupling between the systems occurs on both weather and climate timescales and the societal needs for greenhouse gas flux estimates ranges from global to urban scales.  Though greenhouse gases are primarily treated as inert tracers, the coupling of atmospheric motion and tracers in models and in data assimilation is intricate.  The assimilation of greenhouse gases can impact weather prediction quality, while the uncertainty in weather forecasts places limitations on the ability to estimate carbon fluxes.  This presentation will focus on the coupling between meteorology and tracer transport and its implications on the flux estimation problem.  The flux inversion problem is a type of data assimilation problem in which observations are used primarily to constrain fluxes rather than initial conditions, and an offline chemistry transport model is used to connect surface fluxes with observed atmospheric concentrations.  When the inverse problem is generalized, it leads to the coupled data assimilation problem for meteorology and chemistry.

Ensemble Data Assimilation and Particle Filters for NWP

Roland Potthast (Deutscher Wetterdienst)

Abstract

We discuss the setup of the ensemble data assimilation (EDA) and forecasting systems (EPS) which have been developed and are under development at the German Weather Service DWD and its COSMO partners.

This is first the ICON global+mesoscale model (two-way nested), 13km/6.5km resolution, with its hybrid ensemble variational data assimilation (LETKF+EnVAR) run on a 3h cycle, and the ensemble prediction system ICON EPS. Second, this system drives the high-resolution ensemble data assimilation system COSMO-KENDA (Kilometer Scale Ensemble Data Assimilation) with 2.2km operational resolution at DWD and up to 1km resolution at further members of the COSMO consortium (Germany, Switzerland, Italy, Russia, Poland, Romania, Greece and Israel) to provide initial conditions for the high-resolution ensemble forecasting systems, e.g. the operational COSMO-D2-EPS or experimental ICON-LAM EPS. The system is also successfully run on GPU based supercomputers.

Further, we show results on the tests of localized adaptive particle filter (LAPF) and a localized Markov chain particle filter (LMCPF), which are being tested for the global model setup, currently in the standard experimental global resolution of 40km. We discuss how to overcome filter collapse or divergence by adaptive rejuvenation, mapping into ensemble space based on spread estimators. We also discuss how to keep balances intact when drawing from probability distributions in combination with localization. Here, we also employ incremental analysis update (IAU) for the ICON model system.  Different further versions of particle-filters and hybrid ensemble-particle filters are under test both for ICON on the global scale as well as for COSMO or ICON-LAM on the convective scale in collaboration with colleagues from ETH, Reading and Potsdam.

Coupled reanalysis at ECMWF

Dinand Schepers (ECMWF)

Abstract

The Centre’s ten-year Strategy to 2025 places great emphasis on the need to account for all relevant interactions between different components of the Earth system in ECMWF’s Integrated Forecasting System (IFS). One exponent of this is a growing research effort being devoted to coupled data assimilation, spearheaded by the production of various coupled reanalyses during the last few years. A major result of this of this effort is the production of CERA-20C, the century-long climate reanalysis assimilating a selection of conventional observations of ocean and atmosphere.

The latest coupled reanalysis produced at ECMWF is CERA-SAT a pilot for coupled Earth-system reanalysis in the satellite era. CERA-SAT was produced using the outerloop-coupled CERA assimilation system ingesting atmospheric, ocean, wave, sea ice and land surface observations provided by the full modern-day observing system. Additionally, weakly coupled wave and land surface analyses were produced based on a combination of optimal interpolation (OI) and simplified Extended Kalman Filter (EKF).

In this presentation we present the CERA-SAT coupled reanalysis, it’s approach to coupled data assimilation and a preliminary assessment of the benefits of coupled data assimilation in the context of the satellite era observation system. This will be done within the wider context of current reanalysis production at ECMWF (ERA5) and those envisioned for the future

Characterizing and modelling background error

Anthony Weaver (CERFACS)

Abstract

Data assimilation (DA) systems require quality background error covariances to make best use of information from Earth System observations. Developing techniques to estimate background error covariances and to represent them effectively and efficiently in DA algorithms have been active areas of research for many years. This presentation will start by reviewing the challenges involved in specifying background error covariances in DA algorithms, focusing on variational DA. In variational DA, the classical procedure to specify background error covariances is through a covariance model (often implemented as a control variable transform) that incorporates multivariate balance relationships and a spatial correlation operator. The correlation operator is the most computationally demanding component of the covariance model and different techniques have been proposed in atmospheric and ocean variational DA. Two techniques will be highlighted, namely the implicit diffusion operator approach with Chebyshev iteration solver and the direct convolution approach using compactly supported functions on sub-sampled grids. Both methods have attractive parallel properties (no global communications) and are capable of representing correlation functions on general (unstructured) meshes, making them suitable for next-generation Earth System models and high-performance computers.

Modern variational DA systems are based on a hybrid approach where an ensemble DA system is used to provide flow-dependent uncertainty information for a higher resolution deterministic variational analysis. The remainder of the presentation will focus on describing methods that use ensemble information for defining background error covariances in a variational analysis. First, the ensemble can be used to estimate parameters of the covariance model, such as the variances and shape parameters of the correlation functions. The local correlation tensor is a fundamental parameter of a diffusion- or function-based correlation model, and ways to estimate it from an ensemble will be described. Second, the ensemble can be used directly to build a low-rank sample covariance matrix from the ensemble. By construction, the sample covariance matrix contains general covariance information (multivariate, anisotropic...) but also a significant amount of sampling error due to its low rank nature. Localization is essential to render the sample covariance matrix useable for DA. Techniques to estimate localization functions and to apply them efficiently in variational DA will be discussed. The final part of the presentation will describe the hybrid B formulation, which provides a general framework for combining the advantages of the localized-sample covariance matrix and covariance modelling approaches.

Land, ocean, sea ice, wave coupled model developments

Nils Wedi (ECMWF)

Abstract

The coupling of atmosphere, land-surface, ocean, sea-ice, and waves in ECMWF’s Earth-system model requires a careful consideration of the highly varying temporal and spatial scales of the individual processes at the interfaces. Parameters such as skin temperature over land or sea surface temperature are well observed from satellite observations but are complex to model in dynamically coupled systems. This is modulated further by the exchange of information on different grids, masking of land, sea and lakes, as well as the exchange of temporally averaged information across the interface of different model components. Modelling advances are described that aim to improve the description of variables at these interfaces, together with challenges when moving forward towards a highly interactive atmospheric boundary layer embedded in a future Earth system model with explicitly simulated convection in the atmosphere and eddy resolving oceans.