You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Missing ablation: CISM melts less ice than CTSM does #713
For water conservation, CISM should melt the same amount of ice as CTSM does. The interpolation of accumulation and ablation from CTSM to the CISM grid is not conservative. The conservation fix is to apply two global renormalization factors, one for accumulation and one for ablation. This is done in module ../components/cmeps/mediator/med_phases_prep_glc_mod.F90. We obtain a global sum (based on qice_l_ec in each elevation class) on the CTSM side, and another (based on the downscaled qice_g) on the CISM side. The ratio of these sums determines the renormalization factor. This procedure would be conservative (or very close) if CISM were able to apply qice_g in every grid cell.
However, Gustavo's recent analysis shows that CISM's total Greenland ablation is 50–60 Gt/yr less than CTSM's ablation. Our explanation is as follows:
The next time CISM runs, ice advects into this grid cell. Since SMB = 0, ice is present at time n+1. Typically the thickness is small, H < 1 m.
CTSM, seeing that ice is present (ice_covered = 1.0), computes SMB < 0 based on the CISM topography.
The next time CISM runs, the SMB is negative enough to melt all the ice, leaving some unused negative SMB. The cell is ice-free at time n+2.
The cycle repeats.
This seems to be happening in many grid cells at once. As a result, CTSM adds more freshwater to the ocean than the ice mass change in CISM can balance.
@Katetc , @billsacks , @gunterl: Before we start rewriting code, we should confirm that this flickering could produce conservation errors of the size we're seeing. I've asked Gunter to look at the difference between CISM's SMB (received from the coupler) and acab_applied in coupled runs. If we find that this difference, when summed over the ice sheet, is large enough to explain ~50 Gt/yr of ice that doesn't ablate, then we'll want to think about changing the coupler.
To suppress the flickering, we would want the coupler to compute SMB < 0 in ice-free cells adjacent to ice-covered cells. This could be done by computing a new mask (call it smb_mask) in Glad. The mask would consist of ice-covered cells, plus one row of immediately adjacent ice-free land cells. We would pass smb_mask up to the coupler along with ice_covered. Then when SMB is computed, we would change some logic in subroutine glc_get_elevation_classes_with_bareland, such that cells with ice_covered = 0 but smb_mask = 1 can get their SMB from virtual elevation classes with SMB < 0.
We also might want to change the way the global ablation sum is computed for the ice sheet. The current sum (in subroutine med_phases_prep_glc_mod.F90) looks like this:
If there are many cells at the ice margin where the ice is thin and qice_g < 0, then this sum would overestimate how much ablation happens in CISM. In that case we would want to change how the sum is done. But let's first see what Gunter's analysis shows.
I've noticed that there are a few CTSM grid cells for which a large percent of the area flickers from year to year.
I've seen this as follows:
First, concatenating the Jan files for years 100-109:
cd /glade/derecho/scratch/hannay/archive/b.e30_alpha10c.B1850C_MTso.ne30_t233_wgx3.402/lnd/hist
ncrcat *h0a.010*-01.nc /glade/derecho/scratch/sacks/b.e30_alpha10c.B1850C_MTso.ne30_t233_wgx3.402.clm2.h0a.010_-01.nc
Then, in a python notebook: Plot diffs from year to year (first two sets of diffs shown here, others are similar):
This makes me wonder if the idea of just expanding the boundary by a single CISM cell will resolve these large flickerings. It's possible that it would, depending on the geometry of the underlying CISM cells in these CTSM cells with big flickerings, but I'm wondering if there's a bigger spreading than just a single CISM cell in some places.
Would it make sense to do away with the single grid cell expansion idea and instead make @whlipscomb 's proposed change for all bare land grid cells in the CISM domain? i.e.: For any non-ice-covered grid cell: If the bare land SMB is positive, send that positive SMB (as we currently do); if the bare land SMB is 0 but the appropriate GLC elevation class has negative SMB, then instead send that negative SMB. But as @whlipscomb noted, in the global integrals we'll treat the latter case as having a 0 SMB on the CISM side, with the assumption that CISM will use little or none of that negative SMB.
Hmmm, but actually, I'm realizing some other oddities. I'm going to post this as is and then follow up in another comment.
My previous comment made me realize another possible cause of the flickering: If there is a persistent positive SMB in CTSM's bare land column but a persistent negative SMB (at least as big as the bare land's positive SMB) in the appropriate glacier elevation class in that grid cell. This could give rise to flickerings encompassing a large fraction of a CLM grid cell with the current code. And I don't think the above proposal would help - at least as long as we're sending positive SMB from the vegetated column when there is such a positive SMB.
I thought I would investigate this by looking at QICE_FRZ and QICE_MELT from the same files and years I used to make the above plots. I expected to see a flickering in gridcell-average QICE_MELT in the grid cells with big area flickerings. To my surprise, though, QICE_MELT is 0 everywhere in Greenland for most years. I did find at least one grid cell and year where it's non-zero, so it's not being forced to 0 (it seems). I'm not sure what to make of this. This is really confusing to me. Maybe I'm doing something wrong here??? I do see negative qice values in some elevation classes in the relevant cpl hist files... I just can't understand why the gridcell-average QICE_MELT ends up 0 pretty much everywhere.
Update: This was just my doing something dumb: I was just looking at the January averages instead of the annual averages.
@billsacks, thanks for these plots and comments. I agree that stopping flickering in CISM alone won't fix the problem. I did a back-of-the-envelope calculation showing that ~1000 or more CISM cells need to be flickering to account for 50 Gt/yr, and I don't think that many cells are located adjacent to the ice margin.
It makes sense that a persistent positive SMB for bare land and negative SMB in the glacier ECs could cause this flickering (and I can't think of anything else that would explain it). For some flickering CTSM cell (like the one in the northwest that alternates between dark blue and dark red), I'd be interested to see a time series of SMB for bare land and the first few ECs. Also, I wonder if the bare land sits above the elevation of the CISM surface.
I don't know why you would have QICE_MELT = 0 in this situation when there's a substantial glacier area with SMB < 0. That's very odd.
I was wondering if you've thought about landunit conversions. For instance, when bare land turns to glacier, you would have pseudo fluxes for conservation. How large are the fluxes? Would they have the effect of warming the cell (which would promote melting) or cooling it? I'm concerned that the fluxes might be in a direction that would exacerbate the flickering instead of damping it.
@billsacks, I have a question about the accumulation sum for the land model. See the loop starting on l. 1168 of med_phases_prep_glc_mod.F90. The loop goes from ec = 1 to ungriddedCount, where I think ungriddedCount = 11.
Is qice_l_ec defined such that qice_l_ec(11,n) holds the bare land value? I wanted to confirm that bare land values are included in the sum.
@gunterl has posted new plots E1 and E2 , showing that ice disappears from flickering cells every other timestep, although SMB = 0. We should figure out why that ice is going away. A short run with (idiag, jdiag) set to that cell's coordinates, and maybe some verbose diagnostics turned on, could be helpful.
I expected to see a flickering in gridcell-average QICE_MELT in the grid cells with big area flickerings. To my surprise, though, QICE_MELT is 0 everywhere in Greenland for most years. I did find at least one grid cell and year where it's non-zero, so it's not being forced to 0 (it seems). I'm not sure what to make of this. This is really confusing to me. Maybe I'm doing something wrong here??? I do see negative qice values in some elevation classes in the relevant cpl hist files... I just can't understand why the gridcell-average QICE_MELT ends up 0 pretty much everywhere.
This was just my doing something dumb: I was just looking at the January averages instead of the annual averages.
From looking more carefully at the elevation class by elevation class QICE values in the grid cell in the northwest with big flickering, I see:
QICE is consistently positive in the bare land elevation class; this bare land elevation class has topographic heights that vary between about 315m (in years where the grid cell is mostly ice covered) and about 445m (in years where the grid cell is mostly ice free)
QICE is negative in the two lowest glacier elevation classes
QICE is positive in the higher glacier elevation classes
I also see that the flickering happens in these two lower glacier elevation classes. This suggests that the flickering could be driven by SMB rather than ice flow: When a CISM cell is ice free, it sees the positive SMB from the bare land elevation class; this causes it to become ice-covered. In the next year, it sees the negative SMB from its glacier elevation class; this causes it to become ice-free. And then the cycle repeats.
I have only looked at this one grid cell - one of the CTSM grid cells with the most flickering - but I think this analysis says that, at least for this grid cell, the original change sketched out above won't solve this problem.
@billsacks, I agree that the original change won't solve the problem. But I think the flickering is driven by switching back and forth between the bare land SMB and the EC2 SMB, as you say.
I've written out an example at the end of the document that @gunterl has been working on. If the numbers add up (that is, if there are enough cells leaving enough negative SMB unused), then this might explain a mass imbalance of ~50 Gt/yr. I've sketched out a possible solution in the document.
@gunterl has looked at the mass imbalance associated with the flickering cells; see Fig. F1 here. The unused mass is only ~1 or 2 m/yr, not nearly enough to explain an ablation deficit of ~50 Gt/yr. Although the flickering is unrealistic and I think we can fix it (through a more careful treatment of ice_covered; see my suggestions at the end of the document), it seems like the main problem lies elsewhere.
@Katetc and @billsacks, what do you suggest? I wonder if we should work with Claude some more, trying to understand exactly where the figure of 50 or 60 Gt/yr comes from.
Sorry, I just realized I missed this message. Yes, I think a good next step would be to figure out where Claude's diagnosis of the 50 - 60 Gt/yr discrepancy comes from, then try to reproduce that figure ourselves. Hopefully once we have a better handle on where that number comes from we'll be able to better determine whether it's a real problem, and if so, then be able to move forward with more analyses in a more focused way.
Please see CMEPS PR#720, which modifies the mediator in a way that brings the missing ablation to near zero. This PR is linked to upcoming PRs for CISM and CISM-wrapper.
@whlipscomb@Katetc - I propose closing this issue now that #720 and the associated CISM PRs are merged. I know that this issue hasn't been completely resolved (there are some remaining small known discrepancies, as well as some other potential smaller discrepancies that could warrant further investigation), but it seems like remaining issues would warrant their own issues, so we can start fresh without the confusion of the history of this issue. What do you think?
Activity
whlipscomb commentedon Oct 15, 2026
@Katetc , @billsacks , @gunterl: Before we start rewriting code, we should confirm that this flickering could produce conservation errors of the size we're seeing. I've asked Gunter to look at the difference between CISM's SMB (received from the coupler) and acab_applied in coupled runs. If we find that this difference, when summed over the ice sheet, is large enough to explain ~50 Gt/yr of ice that doesn't ablate, then we'll want to think about changing the coupler.
To suppress the flickering, we would want the coupler to compute SMB < 0 in ice-free cells adjacent to ice-covered cells. This could be done by computing a new mask (call it smb_mask) in Glad. The mask would consist of ice-covered cells, plus one row of immediately adjacent ice-free land cells. We would pass smb_mask up to the coupler along with ice_covered. Then when SMB is computed, we would change some logic in subroutine glc_get_elevation_classes_with_bareland, such that cells with ice_covered = 0 but smb_mask = 1 can get their SMB from virtual elevation classes with SMB < 0.
We also might want to change the way the global ablation sum is computed for the ice sheet. The current sum (in subroutine med_phases_prep_glc_mod.F90) looks like this:
If there are many cells at the ice margin where the ice is thin and qice_g < 0, then this sum would overestimate how much ablation happens in CISM. In that case we would want to change how the sum is done. But let's first see what Gunter's analysis shows.
billsacks commentedon Oct 16, 2026
Thanks, @Katetc and @whlipscomb !
I've noticed that there are a few CTSM grid cells for which a large percent of the area flickers from year to year.
I've seen this as follows:
First, concatenating the Jan files for years 100-109:
Then, in a python notebook: Plot diffs from year to year (first two sets of diffs shown here, others are similar):
This makes me wonder if the idea of just expanding the boundary by a single CISM cell will resolve these large flickerings. It's possible that it would, depending on the geometry of the underlying CISM cells in these CTSM cells with big flickerings, but I'm wondering if there's a bigger spreading than just a single CISM cell in some places.
Would it make sense to do away with the single grid cell expansion idea and instead make @whlipscomb 's proposed change for all bare land grid cells in the CISM domain? i.e.: For any non-ice-covered grid cell: If the bare land SMB is positive, send that positive SMB (as we currently do); if the bare land SMB is 0 but the appropriate GLC elevation class has negative SMB, then instead send that negative SMB. But as @whlipscomb noted, in the global integrals we'll treat the latter case as having a 0 SMB on the CISM side, with the assumption that CISM will use little or none of that negative SMB.
Hmmm, but actually, I'm realizing some other oddities. I'm going to post this as is and then follow up in another comment.
billsacks commentedon Oct 16, 2026
My previous comment made me realize another possible cause of the flickering: If there is a persistent positive SMB in CTSM's bare land column but a persistent negative SMB (at least as big as the bare land's positive SMB) in the appropriate glacier elevation class in that grid cell. This could give rise to flickerings encompassing a large fraction of a CLM grid cell with the current code. And I don't think the above proposal would help - at least as long as we're sending positive SMB from the vegetated column when there is such a positive SMB.
I thought I would investigate this by looking at QICE_FRZ and QICE_MELT from the same files and years I used to make the above plots. I expected to see a flickering in gridcell-average QICE_MELT in the grid cells with big area flickerings. To my surprise, though, QICE_MELT is 0 everywhere in Greenland for most years. I did find at least one grid cell and year where it's non-zero, so it's not being forced to 0 (it seems). I'm not sure what to make of this. This is really confusing to me. Maybe I'm doing something wrong here??? I do see negative qice values in some elevation classes in the relevant cpl hist files... I just can't understand why the gridcell-average QICE_MELT ends up 0 pretty much everywhere.
whlipscomb commentedon Oct 16, 2026
@billsacks, thanks for these plots and comments. I agree that stopping flickering in CISM alone won't fix the problem. I did a back-of-the-envelope calculation showing that ~1000 or more CISM cells need to be flickering to account for 50 Gt/yr, and I don't think that many cells are located adjacent to the ice margin.
It makes sense that a persistent positive SMB for bare land and negative SMB in the glacier ECs could cause this flickering (and I can't think of anything else that would explain it). For some flickering CTSM cell (like the one in the northwest that alternates between dark blue and dark red), I'd be interested to see a time series of SMB for bare land and the first few ECs. Also, I wonder if the bare land sits above the elevation of the CISM surface.
I don't know why you would have QICE_MELT = 0 in this situation when there's a substantial glacier area with SMB < 0. That's very odd.
I was wondering if you've thought about landunit conversions. For instance, when bare land turns to glacier, you would have pseudo fluxes for conservation. How large are the fluxes? Would they have the effect of warming the cell (which would promote melting) or cooling it? I'm concerned that the fluxes might be in a direction that would exacerbate the flickering instead of damping it.
whlipscomb commentedon Oct 16, 2026
@billsacks, I have a question about the accumulation sum for the land model. See the loop starting on l. 1168 of med_phases_prep_glc_mod.F90. The loop goes from ec = 1 to ungriddedCount, where I think ungriddedCount = 11.
Is qice_l_ec defined such that qice_l_ec(11,n) holds the bare land value? I wanted to confirm that bare land values are included in the sum.
whlipscomb commentedon Oct 16, 2026
@gunterl has posted new plots E1 and E2 , showing that ice disappears from flickering cells every other timestep, although SMB = 0. We should figure out why that ice is going away. A short run with (idiag, jdiag) set to that cell's coordinates, and maybe some verbose diagnostics turned on, could be helpful.
billsacks commentedon Oct 16, 2026
This was just my doing something dumb: I was just looking at the January averages instead of the annual averages.
billsacks commentedon Oct 16, 2026
From looking more carefully at the elevation class by elevation class QICE values in the grid cell in the northwest with big flickering, I see:
I also see that the flickering happens in these two lower glacier elevation classes. This suggests that the flickering could be driven by SMB rather than ice flow: When a CISM cell is ice free, it sees the positive SMB from the bare land elevation class; this causes it to become ice-covered. In the next year, it sees the negative SMB from its glacier elevation class; this causes it to become ice-free. And then the cycle repeats.
I have only looked at this one grid cell - one of the CTSM grid cells with the most flickering - but I think this analysis says that, at least for this grid cell, the original change sketched out above won't solve this problem.
whlipscomb commentedon Oct 16, 2026
@billsacks, I agree that the original change won't solve the problem. But I think the flickering is driven by switching back and forth between the bare land SMB and the EC2 SMB, as you say.
I've written out an example at the end of the document that @gunterl has been working on. If the numbers add up (that is, if there are enough cells leaving enough negative SMB unused), then this might explain a mass imbalance of ~50 Gt/yr. I've sketched out a possible solution in the document.
whlipscomb commentedon Oct 19, 2026
@gunterl has looked at the mass imbalance associated with the flickering cells; see Fig. F1 here. The unused mass is only ~1 or 2 m/yr, not nearly enough to explain an ablation deficit of ~50 Gt/yr. Although the flickering is unrealistic and I think we can fix it (through a more careful treatment of ice_covered; see my suggestions at the end of the document), it seems like the main problem lies elsewhere.
@Katetc and @billsacks, what do you suggest? I wonder if we should work with Claude some more, trying to understand exactly where the figure of 50 or 60 Gt/yr comes from.
billsacks commentedon Oct 21, 2026
Sorry, I just realized I missed this message. Yes, I think a good next step would be to figure out where Claude's diagnosis of the 50 - 60 Gt/yr discrepancy comes from, then try to reproduce that figure ourselves. Hopefully once we have a better handle on where that number comes from we'll be able to better determine whether it's a real problem, and if so, then be able to move forward with more analyses in a more focused way.
whlipscomb commentedon Nov 9, 2026
Please see CMEPS PR#720, which modifies the mediator in a way that brings the missing ablation to near zero. This PR is linked to upcoming PRs for CISM and CISM-wrapper.
Merge pull request #720 from whlipscomb/lipscomb/fractional_ice_covered
billsacks commentedon Nov 13, 2026
@whlipscomb @Katetc - I propose closing this issue now that #720 and the associated CISM PRs are merged. I know that this issue hasn't been completely resolved (there are some remaining small known discrepancies, as well as some other potential smaller discrepancies that could warrant further investigation), but it seems like remaining issues would warrant their own issues, so we can start fresh without the confusion of the history of this issue. What do you think?