MUMIP meeting 3
22 June 2022
Attendees
Lisa Bengtsson (NOAA / CIRES)
Ligia Bernardet (NOAA / DTC)
Judith Berner (NCAR / WWRP PDEF)
Hannah Christensen (U. Oxford)
Mike Ek (NCAR / DTC)
Richard Forbes (ECMWF)
Vasilli Kitsios (CSIRO)
Hugo Lambert (U. Exeter)
John Methven (U. Reading / WWRP PDEF)
Kathryn Newman (NOAA / DTC)
Romain Roehrig (CNRM)
Xia Sun (NOAA / DTC)
Nils Wedi (ECMWF / WCRP WGNE)
Keith Williams (UK Met Office)
Agenda
1. Updates from participating groups (all)
- Hannah Christensen (Oxford)
Funding application submitted to the Leverhulme Trust. We were invited to submit a full application in March following our successful outline submission in September. Awaiting the final outcome (mid July). This will cover a PDRA at University of Oxford (running the IFS with ECMWF), University of Exeter (running the MetUM with the UK Met Office), and Meteo France (running ARPEGE-Climat). Judith Berner (NCAR) is involved in an advisory role and providing a link to the DTC efforts.
MU-MIP advertising, most recently at ECMWF Model Uncertainty workshop last month. Very positive response from the community. White papers with recommendations from working groups to be published soon.
Discussion:
JB – note that these recommendations will be most relevant for NWP as opposed to climate or even paleoclimate models.
- Xia Sun (NOAA/NCAR DTC)
Year one of funded DTC project completed in May. Developed software and produced simulations with an array of CCPP SCMs forced by coarse grained ICON high-resolution runs. Currently developing and applying diagnostic tools.
Reported on settings used in CCPP SCM so far (see slides). See that the SCM is able to reproduce the spatial ICON evolution well. Domain averages also look promising, but with small biases in moisture developing near the surface
Request feedback on vertical level interpolation and what to do at the top
Request additional input variables: TKE, theta_l, theta_l_adv, rv_adv. Currently setting TKE to zero and backing out the others.
Request discussion on what variables should be output.
Request discussion on what format to store the data in. Currently 89M for each variable (hourly data, 44,000 grid points, all levels - big!!)
General discussion:
Ligia B: What does the model spin-up look like over the first few timesteps?
HC: If the spin up is strong, may need to run the model for longer to discard the spin-up
Response to discussion points raised
JB: zarr format is state of the art and could be better than netcdf.
VK: chunking in zarr can be a strong determinant of performance in both netcdf and zarr. But the optimal chunking may be a function of the analysis that you want to do
NW: likely that level specification more important near boundary layer than at the top
KW: We have interpolated ICON to MetUM model levels and cut off at top
HC: In the past I have padded the top with ERA data. Suggest to interpolate to whatever the default SCM model levels are as the physics will have been tuned for those levels
RR: TKE will be hard to provide as an extra input, as TKE is a subgrid variable, and the TKE for subgrid processes in ICON may be inconsistent with TKE for subgrid processes at lower resolution. Setting TKE to zero as you are doing will just add an extra spin-up
Lisa B: GFS with FV3 dycore does not have theta_l and rv as state variables. Why are they needed in CCPP? Likely related to dynamics choices in SCM …
HC: confirm that we only have the ICON state variables
- Keith Williams (UK Met Office)
Performed SCM simulations using coarse grained ICON test fields with Met Office model. Prepared namelist files for full set. Worked out logistics of batch submission etc
Question on output format. DEPHY is an input format. Are we happy with CF compliant netCDF?
Discussion:
HC: all location points should be in one file as a map
JB: Yes we should decide on a protocol for output. netcdf file for each variable?
KW: question over one variable per file or many per file
HC: all variables in one file would be unwieldy (using DTC number of 89MB per variable)
JB: space on cloud to help sharing of data
HC: so how many people are happy with zarr – we do not want to introduce barriers to using the data
VK: zarr is easy if you are using python. Incompatible with nco.
JB: python more efficient if you have fewer, larger netcdf files
VK: suggest one variable per file, at all lat, lon, and lev
HC: main concern is to ensure data is provided in format which doesn’t turn people off. We would be a leader if we went for zarr, as for many other projects, data is provided in netcdf format. Tentative decision to go with netcdf – but can revisit later.
NW: do we have an estimate for total data size? Need this before we make a final decision
HC: should be discerning in terms of what variables we archive. Or we could only archive every third hour, as this is the time which we can compare to DYAMOND. Useful to have hourly data in first instance for internal spin up checks, but then exchange only three-hourly frequency data
Ligia B: We should iterate on list of variables offline
2. SCM protocol discussion
- Hannah Christensen
Choices made in forcing SCM can influence the results. Must at least report our choices. Ideally decide on common choice
First: forcing SCM at lower boundary. Either supply surface forcing fluxes or provide skin temperature and moisture. The two are inconsistent with each other
Second: Whether to use vertical velocity as part of dynamical forcing. Used by SCM for vertical advection
John Petch (UKMO) advised that skin temperature better than fluxes, but on the fence regarding vertical velocity.
Discussion:
KW: skin temperature more akin to atmosphere only GCM (e.g. provide SSTs, and let atmosphere compute fluxes)
JM: skin temperature better. How does atmosphere respond to a given forcing. Have to let surface fluxes be part of that response. E.g. https://rmets.onlinelibrary.wiley.com/doi/abs/10.1002/qj.49712455013
NW: Vertical velocity. Agree quite some sensitivity here to how vertical velocity is used. In IFS used within cloud microphysics. Could be specific to different models …
HC: conclusion is that there is sensitivity so we should be consistent, or that even if we all use it, it may enter in different ways, so we will not necessarily get consistency?
NW: agreed with the latter statement.
RR: surface forcing. ICON surface winds will show high variability, so the surface fluxes will be very different in ICON compared to low-res model with coarse resolution fluxes. So the input of energy etc could be substantially different in ICON to in SCMs, so we may drift to a different state. So RR would argue for fluxes. Also have u* to combine with fluxes to give momentum forcing
RR: horizontal winds. Geostrophic winds likely weak over the tropics. Should we use it, or just advected U, V. Or nudging and not care about wind
HC: agree higher resolution will lead to different fluxes in the low-res model, but would argue that there should be a parametrisation of this effect
RR: but other physics schemes may compensate for this lack of fluxes … a structural error
JM: variability in surface winds is important, but in danger of building in additional parametrisation that is not actually present in the SCM if we use fluxes. Feedbacks between flux variability and convective organisation – complex and would ideally be part of SCM parametrisations
Lisa B: most physics packages have a surface layer model which computes this …
RR: but different fluxes will lead to different profiles between ICON and SCM
HC: what have groups done to date?
XS: DTC CCPP work to date has used surface fluxes. Concern about lack of soil variables.
KW: UKMO has used skin temperature. Default is to use land surface model and, over ocean, this defaults to skin temperature. All our columns are over ocean.
Ligia B: suggests that CCPP could try some sensitivity experiments switching between surface fluxes and skin temperature
HC: tentative conclusion is force with skin temperature, but discussion to be continued offline
3. Next steps/ timeline
UOxf timeline: finalise variables required. produce all coarse graining input for ICON 40-day DYAMOND simulation by end of August
ACTION POINTS:
HC to add extra variables requested by DTC to SCM inputs where possible
HC to set up discussion document on what variables to output, including time frequency, levels, etc.
HC to follow up on SCM forcing with interested parties
XS to look at model spin-up – to help work out whether 3-hour forecasts are sufficient
XS to try switching between surface fluxes and skin temperature
All – final feedback on input files by 15 July
HC – to find new interim leader for project while HC on Maternity leave