What's the problem ?
Most notably you will see "bulls eyes" of excessive QPF over short time ranges (like 6 or 12 hours). They are also accompanied by small but intense mid level vort maxima.  These vort maxima usually seem to "just appear out of no where" - that is they are not traced back to some existing feature in the initial conditions.  Sometimes you will see a weak low in the surface pressure field in association with the mid level vorts.

The GFS treats these as if they are real synoptic scale features.  They can contaminate the model output anywhere along the path of these features as they are carried along in the model's synoptic scale flow.

This can wreak havoc on those initializing forecast fields of surface wind and QPF in IFPS.

When does it happen ?
It happens most in the warm season.  It can startle you the first time you see it in late winter or early spring, but by early summer you get used to it.

Where does it happen ?
This "feature" has been observed  just about EVERYWHERE on the GFS grid.  Not only over land (even over complex terrain), but also over water.  It has also been observed in the southern hemisphere (over South America).

How to tell there's a problem

• Look at a loop of QPF (in either 6, 12, or 24 hour increments)
• Look for QPF maxima that seem inordinately large given the time increment: ~1.5"/6hr,  2.5"/12hr or 3.5"/24hr
• If you have access to this grid, look at the convective component of the total precip.  If the majority of the QPF is NOT produced by the convective component, you've likely got a grid scale feedback problem

If a 500mb vort "appears out of nowhere" at the time the QPF bulls eye is produced AND you can't trace the vort back to an existing feature in the initial conditions, you've likely got a grid scale feedback problem

What's it look like ?
Here is a loop of 6 hour QPF from the 12Z June 15 2003 GFS run:

Focus on Wyoming - note the precip maxima of 2"+ that occurs between 36 and 42 hours.. and again between 60 and 66 hours.

"Big deal" you say, "Convection can produce WAY more than that in even an hour !"  Very true, but for a model that doesn't explicitly model convection (convection is parameterized in the GFS), that's a lot of QPF for the model to be dumping out in a 6 hour time period.

Here is the corresponding loop of 500mb heights:

Note the vort max that suddenly appears at fhour 36 in Wyoming which you can actually follow through almost the rest of the loop.  That shouldn't look right to you.

If we look at a loop of the convective component of the total precip (that is, the portion of the QPF that was produced by the convective scheme):

we see that the convective component was less than half of the total.  At fhour 36 the convective component was less than 25% of the total.  This means that most of the QPF was not produced by way of the convective scheme, but produced by the model on the grid scale.  "So what." you say. Well, read on...

Why does it happen ?
Although the convective parameterization scheme in the GFS (the Arakawa-Schubert) was triggered and released SOME of the instability sensed by the GFS, not enough instability was released to prevent the model to react on a grid scale.

This is because the convective parameterization schemes emulates effects of convective updrafts, which are much smaller than the grid box.

Depending on the amount of instability that is in the grid box, the convective scheme redistributes heat and moisture in a realistic way, but over a SMALLER part of the grid box.  Although the effects of the parameterization do impact conditions on the grid scale, if the paramerization does not alleviate enough instability, the model will still sense the instability and start acting on it at 55km scale.

The model therefore acts on the remaining instability in a synoptic scale sense which can result in the GFS dumping all the precip it can as quick as it can over a small area (the grid box).

"What's wrong with that ?" you say.

Conceptually, nothing is wrong with this, because a result of the real atmosphere responding to instability sometimes does result in a large amount of rain occurring over a small area.

However, a tremendous amount of latent heat is produced in the model as the GFS dumps all this precip over a short period of time.  Again, this happens in the real atmosphere too.  BUT, the GFS is releasing this latent heat at  55 km resolution.  The problem here is that  when you heat the column of air at this resolution, the model senses it at a synoptic scale resolution (as opposed to the real atmosphere which typically does so at a meso and sub meso scale).

The model will then try to make adjustments at 55 km resolution (synoptic scale).  The result is a chain reaction whereby the model will try to converge mass into the region where pressures have lowered under the region where the latent heat was released.  This results in enhanced vertical motion over this region and enhanced divergence at the top of the model atmosphere.

When the model can't keep up with the increasing instability (in a feedback situation), the lower level moisture convergence increases.  The resulting updraft and convergence/divergence couplet slowly increase in depth and intensity.   Voriticy is increased within the grid box (through vortex stretching below the updraft maximum) and the system slowly propagates with the steering flow as it continues to maintain itself.

Just for fun, look at the loop of 250mb divergence:

Note the bulls eye of divergence in Wyoming associated with these QPF
maxes (that also appear out of nowhere and then disappears just as quickly).

The same is true for the 700mb Vertical Motion:

Although the signal here a little more subtle, it is still evident.
 

Although this process does happen in the real atmosphere (leading sometimes to the production of MCCs), the GFS does this on a synoptic scale.  Further, once the GFS has "scaled up" such a feature to the synoptic scale, it is very hard to dampen the feature out as quickly as it would do so in the real atmosphere.

Adjacent grid boxes sometimes begin to have grid-scale convection as well.  Once the instability is finally removed over many hours or even as long as a couple of days, the convective system created by the grid-scale precipitation scheme can slowly spin down.
 

The next two loops provide a dramatic example of this process.

Here is a loop of 500mb heights and vorticity from what was the 12Z AVN run from May 17 1999.

Note the vort max that appears in NW Kansas at fhour 60.   Although not shown, this vort max was progged by the AVN to continue northeast through IA and into NY producing a swath of heavy QPF along a stationary front.  It didn't verify.

Now here is a cross section through that vort max.
RH: Green shading (every 20 %)  Potential Temperature: white (5 deg K)  Wind: Blue (kts) Vertical Motion:  red ( - upward) Conv/Div: Yellow (- conv, + div)

Does the RH signature remind you at all of anything you've seen in real life ?  Kind of looks the profile of a CB with an anvil doesn't it !  Look at the convergence at the lower part of the "CB" and the divergence at the top.  Notice also the VV max at mid levels.  "Hey.. not too shabby !"  you may say.  But consider, this "CB"  produced by the model is on the order of  200 km across !  Not to mention that it didn't verify either.

The problem with this dramatic a response to instability by the GFS is that before you know it the model has produced a synoptic scale feature which is treated as real.

CONSIDER HOWEVER, that although this process leads to QPF bombs and spurious vorts over land, it serves quite well for predicting the development of tropical systems over water !!!
 

What's the remedy ?
We have been hinting that the GFS has been sensing instability properly, but responding improperly due to resolution.  This premise was tested by EMC's Global Modeling group by rerunning a case where a QPF bomb was observed in the operational GFS output.   The model was rerun at much finer resolution.
Remember, the goal is to improve how the GFS reacts to instability without degrading its performance with regards to development of tropical systems !
 
 

12Z May 04 2003 GFS run resulted in the following QPF forecast:

GFS 24 hour QPF forecast at ~80km res valid 12Z 05/05/03
Note the QPF bombs in MT, southern MN and the OH valley (75-100mm/24hr ~ = 3-4"/24hr)
 
 

Here is what was observed...

 
 

The GFS did pick up on the GENERAL areas that would see heavy precip, but was obviously lacking in detail.

Here is the GFS rerun at 15 km and 4 km resolution that day (it's actually the GFS run on a domain covering North America making it a Regional Spectral Model - or RSM)...

Note how the bombs are in MT and MN were reduced or rather better resolved.   But notice something else ?  The hi-res reruns don't produce enough areal coverage of .75-1.5" QPF (20-40 mm) amounts as compared to the observation.
 

Here is another rerun...

The 12Z GFS May 13, 2003 run resulted in the 24 hour QPF you see on the left (valid at 36 hours).  Note the QPF bomb in IA.  On the right is the observed 24hr precip via Stage IV valid at the same time.

Below is the 15 km RSM rerun.

Yes, yes... you are correct.. the placement of the "streaks" in OK, MO, and LA aren't that great.  But did you see the forecast up in the Dakotas and even IA and NE compared to the obs ???  That's a pretty darned good forecast... and light years better than the bomb forecast at 80 km resolution.

How do the reruns handle the vort maxima ?  Below are plots of absolute vorticity.  On the left you have the result from the operational GFS, and on the right the rerun at 15 km resolution (the RSM).  The scale here is actually 10x greater than what is typically plotted on 500mb charts.  So the "2" you see over eastern SD actually is a 20 unit vort max.   That QPF bomb help generate a 20 unit vort max whereas the RSM produced a very small scale 20-30 unit vort max (which would not run the risk of being scaled up
into the synoptic scale by the model).

These reruns have led EMC to some preliminary conclusions:

• The operational GFS IS properly sensing instability
• However, it is reacting to the instability at too large of a scale (because of the resolution of the operational model)
• Therefore, the mesoscale response is occurring on the model grid scale (which is synoptic scale)
• Running the model at a higher resolution will mitigate the problem (not available yet)
• These tests may hint that until the model resolution is sufficiently high enough, a possible partial solution would be to tweak horizontal diffusion.  This may selectively mitigate this QPF bomb problem without significantly impacting the GFS ability to spin up tropical systems (since the wave length of tropical systems is at least an order of magnitude larger than these vort maxima associated with the QPF bombs)
• Further testing at 4 km is needed

Does this happen in the Eta ?
Actually, the Eta responds in the same fashion as the GFS, but since its resolution is 12 km,  the problem doesn't manifest itself as dramatically.

How to still use the "contaminated" model output
Obviously you want to first look at another model for a second opinion.

Nonetheless, if you see one of these in the GFS over your area, you can still extract some useful information from the model.   Since the GFS senses instability properly, the model does well at hinting when the atmosphere is prime for producing an MCS like system "in the area" that the model has depicted.

Just like in the real atmosphere where an MCS can inhibit transport of moisture on the lee side of these features, so to does the GFS QPF bomb.

On the left is a 24 hour QPF valid on Day 5 from a 00Z MRF run.  You can see the QPF Bomb in the central plains.  On the right is the same forecast but from the following 12Z AVN run.  This run did not produce a bomb.

Note the difference in QPF amounts generated in eastern ND and northwest MN between the two panels.  The 12Z AVN run on the right is able to transport more moisture north as opposed to the run with the QPF Bomb.

Don't believe it ?? I didn't either when I first saw it.
But here is another example... 00Z MRF on the left,  12Z AVN on the right.

Note the difference in QPF generated over the Great Lakes when the QPF bomb is NOT produced.

Acknowledgments and References
Thanks to Dr. Hua Lu Pan and Dr. Pete Caplan (head of the Global Modeling Group EMC) for reviewing the output of the RSM runs with me.  And thanks to Dr. Bill Bua (COMET Scientist and NWP expert at EMC) for providing appropriate NWP PDS COMET web modules:

How Models Produce Precipitation and Clouds (see the CONV PARM link on the left)

Grid-scale Precipitation Bombs in the Global Forecasting System

(posted to the web 06/20/03)