GLACE-2 Participants

A) We would prefer that the boreal summer case were run first, for consistency with the contributions from other modeling groups. If a supplemental set of austral summer simulations were submitted, however, they would be processed using our standardized routines, so the submitting group would learn about their model's performance in this time period.

A) The resolution used is solely the choice of the modeling group, the idea being that each modeling group knows best how to tailor the experiment to maximize its relevance to their own institution.s needs. A resolution of 250km is not out of line with some other expected submissions.

A) Free access will be the policy.

A) Step (1) on p. 9 of the experimental plan lists 3 options for initialization. For options 2 and 3, GLACE-2 will still require that the scaling be performed. For option 1, for which the scaling is used only to correct for forecast model biases, we accept the existence of a reasonable counter-argument: that the soil moistures generated offline, while biased relative to the model's climatology, are still more "realistic" and thus may be preferable.

This is a philosophical argument, and different groups will have different views on the usefulness of scaling in this situation. Because scaling for option 1 will probably not have a significant impact on the results, and because we want groups to perform the experiment in a way they feel would be most beneficial to their own needs, scaling is only *encouraged* for option 1. Again, though, scaling is required for options 2 and 3. For those using option 1, we just need to know for the write-up whether or not scaling was performed.

A) GSWP2 (http://www.iges.org/gswp2/sensitivity.html) offers a number of 10-year forcing datasets. The baseline (B0) has been criticized for having an overcorrection for precipitation gauge undercatch and for having poor winds. These are reasonable criticisms. However, given that some groups have already performed the baseline runs and may be depending on using those data for GLACE-2, we will not dismiss the use of the B0 forcing. Groups can use the B0 forcing if they so wish. If groups prefer to use another set of GSWP forcing data because they feel that would better suit the needs of their institutions, that's fine, *as long as the precipitation and radiation data are suitably scaled to GPCP and SRB observations*, as in B0. By having all groups rely on the GPCP and SRB observations, we should maintain the necessary consistency between participants, with differences stemming from the different choices being second order. Again, we will need to know what people use so that it can be documented.

A) For the experiments that do not utilize soil moisture initialization via GSWP, the initial soil moisture conditions should (optimally) still be consistent, if possible, with concurrent SSTs; this can be achieved by extracting them from parallel AMIP-type simulations using realistic SSTs. If your group has multiple parallel AMIP runs archived, that would be perfect. Most groups have at least one set of AMIP-based restart files archived; if only one is available, initial conditions for the different ensemble members could be taken from different years (same date) of this run. Initial conditions could even be taken from the different years of GSWP output. See the GLACE-2 plan for details.

A) No. All of us are probably traveling way too much these days, so for the present, we will move the project forward and communicate our results via the internet. This approach worked quite well for GLACE-1.

A) By far the best way to initialize the Series 2 simulations is to go into the GCM"s archives and choose archived restart files for the dates in question, but for random years. The model you are running must have been run for decades (centuries?) by other researchers, and they would have archived the land states they generated over the course of these runs -- maybe once a month, something like that. For a June 1 experiment, for example, choose ten random years and extract the June 1 restarts for those years. The restarts will contain the soil moisture distributions (and surface temperature distributions, and so on) that you can use for the initialization of Series 2.

A) It's very easy to obtain the multi-model estimates for any day of the ten-year period, but only at 1 degree resolution. [See the "data sites" link on the GLACE2 webpage]. Perhaps the data can be sensibly interpolated to your own grid. For you to make use of these data, however, you need to have information regarding the mean and standard deviation of soil moisture in your model as a function of day-of-year, as (easily) determined from a multi-decadal simulation with your model. This is critical. If you have these data, then the GSWP2 multi-model soil moistures can be transformed to soil moistures for your model using the equation on p. 11 of the experimental plan.

No, you don't need 100 different sets of soil moisture conditions. You only need 10. Here's why. Let SM19XX represent the soil moisture conditions you've obtained for the year 19XX, say from the GSWP exercise.

The key thing to remember is that because you're using different SSTs and different atmospheric initial conditions for the different start dates, the ten simulations for 1986 will be different from the ten simulations for 1987, and so on. You are not repeating simulations here. You are simply trying to capture the fact that a specific start date (day of year) is associated in Series 2 with a certain amount of soil moisture uncertainty, which we are capturing here with the set (SM1986, SM1987, ..., SM1995).

Of course, a superior strategy would be to collect, for 1 April 1986, a set of ten April 1 soil moistures consistent with 1 April 1986 SSTs. This could be achieved by extracting the soil moisture states from 10 parallel AMIP simulations. Many groups, however, don't have access to such data, but many groups do have access to multi-decadal AMIP simulations. Given the existence of such multi-decadal data, it would be preferable (if possible) to choose, for a La Niña year such as 1988, the soil moistures for La Niña years in the multi-decadal collection.

A) The diagnostic data may be in terms of monthly means, but the restart files themselves, if they were stored (for example) during an archived AMIP run, should include instantaneous values for the prognostic variables. Those would be the data to work from, if at all possible.

A) Using AMIP runs to initialize the atmosphere would not be the end of the world, because the soil moisture imprint on the simulations will last longer than the atmosphere imprint. Even so, if the realistic atmospheric initial conditions give you rain on day 2, and that rain wets the soil, then that may have some impact on skill in the coming month. Something to keep in mind... In any case, we isolate the land impact by subtracting the series 2 results from the series 1 results, so the impact of the atmospheric initialization may cancel itself out. Our advice: do what you can with the atmospheric initialization, but if you're forced to use the AMIP approach to atmospheric initialization, that's acceptable. We just have to document it.

A) It makes a difference if the SSTs include information from the future because GLACE-2 is being looked at as a component of the "Task Force for Seasonal Prediction" effort, and one of their "prime directives" is that any forecast performed in the effort be a true forecast, using knowledge only of initial conditions and the historical statistics of the climate. And in a sense, we are trying to forecast the historical anomalies. If we initialize the model on June 1 of 1988, for example, we want to see if we can forecast the 1988 drought *without* benefit of knowing what the actual July 1988 sea surface temperatures were. Now we could argue that because the Series 1 and Series 2 forecasts use the same SSTs, and because we subtract the skill obtained in Series 2 from that obtained in Series 1, the effect of the SSTs "cancels out", but that still doesn"t go over very well with the TFSP people.

A) It should be. In fact, as long as the SSTs you use don't contain "information from the future", you're free to use whatever you want. Your approach would only preserve some very high frequency SST information "from the future", which has got to be negligible.

A) Yes, you can certainly start with the Sheffield data, and hope that you can do the GSWP data later; whatever happens, we just have to explain things well in any papers we write. Re computational resources, however, replacing the 1986-1995 Sheffield period with the equivalent period from GSWP wouldn't add to the computational expense, other than the cheap offline part, and it would be nice if all the groups could use the same land surface forcing for the ten-year base period. So I would encourage you strongly, if possible, to use the GSWP-2 data for that period - unless, for example, your own institutional needs desire a 50-year analysis based on a single dataset.

A) We feel that being flexible about 2nd-order issues is key to maintaining successful participation in the project. The participating groups are doing these experiments *not only* to contribute to the multi-model results, but also to satisfy their own institution's interests regarding quantifying the value of land initialization. After all, we"re not sending people money to do these experiments; the money has to come from within the institutions, and so the experiments have to satisfy, to some extent, the specific needs of the individual group. If one group thinks using the NCEP SSTs is more valid, and another group thinks taking advantage of the continuity of the monthly Hadley SSTs will tell them more about their system, and if it is a second order issue (which it should be, in this case), then it makes more sense to let the individual groups decide what is best for them. Certainly some groups already have the plumbing in place to run with certain SST climatologies, and they might find the use of our persisted anomalies on top of that climatology far more trivial and scientifically valid than being forced to incorporate a separate dataset that hasn't been tested with their model.

In fact, in discussing the design of the experiment early on with TFSP, we decided to allow coupled atmosphere-ocean models to participate, with the coupled models relying only on the initialization of the land and ocean. This is because some institutions will be relying on such coupled models for their forecasts and will want to see how land initialization affects things in their systems. Talk about inconsistent SSTs! Still, though, given our strategy of comparing results with and without land initialization, and given the slow timescales associated with SSTs, the differences should still be second-order.

Of course, for a valid intercomparison, the first-order issues *must*, by all means, be uniform, and for GLACE-2, they are. That's why GLACE-2 will indeed tell us something useful. We will, of course, document the 2nd order differences in the write-ups.

A) It probably wouldn't make much difference. Even so, the whole point of the scaling exercise is to correct for biases in the model relative to "obs" (the GSWP offline results). For that reason, and because you say the AMIP run is only a little longer anyway, so that its statistics won't be that much better if all years were included, I suggest using the same periods to calculate the means and standard deviations. We want to correct for GCM biases only, and not for biases that come from both GCM deficiencies *and* the use of a few additional years of AMIP run data, especially if those few extra years include the extreme late 90's El Niño event.