Skip to content

FAQs and commentary related to:

T. R. Knutson, J. J. Sirutis, S.T. Garner, G. A,. Vecchi, and I. M. Held, 2008: Simulated reduction in Atlantic hurricane frequency under twenty-first century warming conditions, Nature Geosciences, doi:10.1038/ngeo202.

Prepared by Isaac Held in collaboration with the other authors of this paper — last substantive revision 06/15/09. Please contact  with any questions concerning this FAQ sheet.

For further information on the global warming/hurricane issue, see this page maintained by Tom Knutson.

1. What is our motivation in pursuing this research?
We feel that it is potentially very misleading to rely solely on empirical statistical relationships for predicting how global warming will affect hurricanes. The value of these statistical relationships is limited by the relatively small number of degrees of freedom in the observations, the necessity of choosing among a large number of possible predictors, and the lower quality of the observations as one moves back in time. Our work is focused on creating “dynamical” model alternatives to these statistical models, in which we attempt to simulate the physical system in question as realistically as we can, given our understanding of the underlying fluid dynamics and thermodynamics and given the available computational power. Others around the world are also pursuing this dynamical approach, and we are hopeful that as it matures it will greatly clarify the issue of global warming’s influence on hurricanes.
2. Why use a dynamical regional model?
Historically, attempts to study the relationships between global warming and hurricanes with dynamical rather than statistical models have been undertaken with either global climate models that are easily criticized as having too coarse a resolution to convincingly resolve hurricanes, or very high resolution models of individual storms with which one cannot address questions of storm freqeuncy. Our work is distinctive in the high resolution (18km) that we are able to use by restricting the domain to that of the Atlantic only. (There is an important global model of comparable resolution described in the literature — Oouchi et al. (2006) , which we comment on below.) Balancing the advantage of more easily moving to finer resolution, this regional approach has disadvantages as well. In particular, we need to rely on global models to provide input into our regional model when we predict the future. We would prefer to work with high resolution global models, and are pursuing global simulations at 50 km and 25 km resolution at GFDL. Some of the 50 km results are described in Zhao, et al (submitted)
3. Simple statistical models can be made to fit observed storm frequency over the 1980-2006 as well as our dynamical model; why spend such a large amount of computer time generating the same result with this complex dynamical model?
This is a key point that one needs to understand to determine whether or not to give special weight to this model’s predictions. The rules governing the evolution of the model are not of the form “when the shear is larger than X or the ocean temperature is less than Y do not allow a hurricane to form” — in fact, there is no explicit reference to the concept “hurricanes” within the model formulation. The rules governing the model operate at a much lower level of granularity, governing how the winds, pressures, temperatures, and humidities at neighboring grid points interact with each other – solving equations based on the fundamental principles of fluid dynamics, thermodynamics and radiative transfer. The model may chose to behave according to a simple “macroscopic” rule or it may not. To the extent that the observed variations in storm frequency emerge from the model, we have some confidence that the model is simulating atmospheric dynamics adequately at this granular level. Hopefully the proper macroscopic rules of hurricane formation emerge automatically. By capturing the underlying dynamics, rather than putting in by hand the rules that we think connect hurricanes with climate, we have more confidence that the model results will be transferable to novel situations where our familiar rules might break down — even if the fit to an observed time series of yearly hurricane frequencies is no better than that obtained with a statistical model.
4. Why not combine both dynamical and statistical methodologies?
These methods are complementary, and we believe current assessments of the state of the science relating hurricanes to global warming need to take both sources of information into account. We need to learn how to best optimize the information coming from both approaches. In particular, dynamical models can help us choose among competing statistical models exhibiting similar skill over the observed record, but built using different sets of predictors. For example, while impressive fits to interannual variations in storm activity in the Atlantic can be obtained using local tropical Atlantic sea surface temperatures (SSTs) as a predictor, Emanuel (2005), equally impressive fits can be obtained using as a predictor the difference between the tropical Atlantic SST and the tropical mean SST — see Swanson (2008). Models intermediate in structure between the simplest statistical approaches and fully dynamical approaches are also possible. The recent work of Emanuel (2008) is a good example.
5. How do we decide that the quality of this dynamical model is adequate for meaningful projections of future changes in hurricane activity?
We start by simulating the hurricane season over the Atlantic for each of the years between 1980 and 2006. In particular we count the number of hurricanes generated by the model, using the standard meteorological definition of a hurricane based on the maximum surface wind in the storm. We were impressed by the results for the year-to-year variation and the well-known upward trend over this time period (see figure 1 in the paper). The quality of this result is what encouraged us to proceed with the global warming study. It indicates to us that meaningful “macroscopic” rules governing storm genesis do emerge naturally from this model – that it must be handling the effects of changes in circulation (“shear”) and moist stability in a fairly realistic manner in order to simulation the year-to-year variability and the longer term trend with this level of realism. The confidence that anyone should place in our results should depend in large part on how impressed they are with the quality of this control simulation.The model has no difficulty in producing category 1 or 2 storms. When we look a the model’s strongest storms, the category into which they fall depends on whether we use the standard surface pressure criterion or the wind criterion, since the model has weak winds for a given central pressure for its strongest storms. Using the wind criterion the model produces no category 3-5 storms. Using the pressure criterion, the model does produce category 3 storms, and a small number of category 4 storms. This has been a disappointing aspect of the model, in that experience with other models of comparable resolution suggests that such a model should be able to generate stronger storms and a more realistic pressure-wind relationship. We are investigating this problem, but it remains a serious limitation to this study. We focus on tropical storm and hurricane numbers rather than intensity because of this limitation.
6. Does this dynamical model have free parameters that have been chosen to improve the fit to observations?
As described in our first paper on this model, there is one parameter that we vary that controls the total number of storms generated. This parameter is the rate at which we restore the observed large scale state of the atmosphere in the model interior. If left to its own devices, the model generates too many storms. We believe that this is because we are running without a convective parameterization – we simply allow the model to convect at resolved scales. To first approximation, a convective parameterization scheme, based on assumptions concerning the sub-grid scale structure of meteorological fields, allows the model to convect more easily. A consequence of running without such a scheme is that the model develops a modest cold bias in the upper troposphere, because convection is too inhibited. This destabilizes the atmosphere and causes too many storms to be generated. By restoring to observations on large scales we stabilize the atmosphere and reduce the number of storms – the stronger the restoration the stronger the reduction in storm number. (This is our current hypothesis as to what is going on in the model at any rate.) Once we choose the value of this restoration parameter, we hold the value fixed for all of our simulations. We would prefer not to have to use this interior nudging, and we consider it a model imperfection.
7. Why choose to run the model without a convective parameterization?
We find that this model performs at least as well as, and possibly better than, a version with one particular parameterization scheme that we have worked with in the past, and it is certainly a simpler model without using this complex closure scheme. The choice of running without a convection scheme was thus an application of “Occam’s razor”, based on this limited set of simulations. We have not been able to compare the quality of our simulations using a larger set of alternative convective parameterizations, because of the computational expense of this model. Running without a convection parameterization scheme is admittedly unconventional at this resolution.
8. Is the interior large-scale nudging exerting a strong control over individual genesis events in the control simulation?
Some of our colleagues have questioned whether the technique by which we force the model towards the observed large-scale state of the atmosphere is somehow forcing the observed storms to form in the model during the control period. If this were the case, it would clearly reduce the value of our control simulations as a test of model quality. But this is not the case; the individual storms in the model do not form in the same location or times as observed storms. If we perturb our initial condition a bit at the start of the hurricane season, we change the places and times at which storms form, and the total number of storms within the season generally differs somewhat as well. We hope to make this point more clearly in future publications on this model.The boundary conditions in the model – especially the boundary conditions over Africa at the eastern boundary of the domain – do generate disturbances that provide some raw material for storm development. But ongoing research suggests to us that the interannual variability of genesis in the model has relatively little to do with interannual variability in the high frequency boundary forcing.
9. What information do we use from global model projections when we downscale to project changes in Atlantic tropical storm activity?
A point of confusion we have encountered among some colleagues regards the use that we make of the global climate model projections. We do not use the tropical “weather” generated by the global models. We only use the change in the seasonal mean state of the atmosphere, and the change in sea surface temperatures, from the global model projections for the 21st century. We believe that these long-term seasonal mean changes are likely to be the most reliable component of the global projections. We keep the same “weather” in the boundary and large-scale interior forcing as in our control simulations. The reduction in storm activity that we see in our model is solely a response to the changes in seasonal mean climate predicted by the global models.
10. What factors control the hurricane activity response in this model?
Some of the key features of the seasonal mean response projected by the global models are discussed by Vecchi and Soden (2007.a) and Vecchi and Soden (2007.b) — including the change in vertical shear and in the moist stability. With regard to the latter, the key point is that, in global warming simulations, models produce very little change in moist stability – the tropical atmosphere stays close to a moist adiabat, which requires the upper troposphere to warm more than the surface air. One way to get our regional model to produce more storms (as already indicated above) is by reducing the moist stability of the troposphere, that is, by reducing the strength of this upper tropospheric warming. This immediately raises the following centrally important question with regard to the model’s simulation of the increase in Atlantic storm activity over the last two decades:
11. To what extent is the model’s increase in activity in the past few decades due to a destabilization of the atmosphere over the Atlantic?
We are working on a follow-up paper in which we investigates how much of the model-generated increase from 1980-2006 is due to changes in atmospheric stability and how much is due to changes in circulation (“shear”). We will report on these results as soon as we have finalized them, but we already know that stability is a significant factor for the model’s increase in storm activity over this period.
12. Is the reduction in moist stability over the Atlantic over the past few decades, with which the model is forced, an artifact?
The data with which we force this model during the control simulations – both through the boundary conditions and through the large-scale interior nudging – comes from the NCEP/NCAR reanalysis. It contains a substantial trend towards reduced stability over the 1980-2006 period: the upper troposphere does not warm enough to maintain the moist stability. This could be a real trend or it could be an artifact. Trends in tropical and global mean stability have been the topic of extensive analysis and discussion in recent years, as summarized in CCSP report 1.1. Trends over the Atlantic share in the uncertainty in the tropics-wide stability trends, but there is an additional factor: the stability in the Atlantic is expected to be reduced (enhanced) when the Atlantic Ocean surface temperatures warm more (less) than the rest of the tropical oceans.The moist stability over the Atlantic is not determined by Atlantic surface temperatures alone. One can think of tropical tropospheric temperature, over the Atlantic and elsewhere, as being maintained by averaging over the convection throughout the tropics (much of this convection occurs in the Indo-Pacific region), because the tropical atmosphere is very efficient at flattening horizontal temperature variations. The Atlantic Ocean has indeed warmed faster than the tropical mean over the past 20 years, and on this basis alone one should expect some reduction in stability over the Atlantic during this period.

Therefore, there are reasons to expect that there has been a downward trend in moist stability in the Atlantic over the past few decades, independent of the reliability of the tropical mean trend in the reanalysis. But we are still left with a concern that our model may be generating the observed trend in activity for the wrong reason. The implication would be that our model suffers from some cancellation of errors, if in fact the changes in stability over the Atlantic in the reanalysis are not accurate.

13. Are we confident in the projections by global models of future changes in moist stability and shear over the Atlantic?
There are several aspects to this question, firstly with respect to the tropics as a whole, and secondly with respect to the Atlantic relative to the rest of the tropics.All of the world’s global climate models (to our knowledge) predict that the tropics as a whole will stay rather close to a moist adiabat as the oceans warm. Is it possible that this prediction is in error? If there is a uniform destabilization of the tropics associated with global warming, this would have dramatic consequences for tropical storms throughout the world and for all of tropical meteorology. We are personally sceptical that the moist stability of the tropics as a whole is easily changed, but it is clearly vitally important to reconcile models and observations of recent trends in tropical stability. The results in this paper are entirely dependent on this prediction of small changes in moist stability.

As discussed above, it is plausible that the stability of the Atlantic atmosphere relative to the tropical mean is sensitive to warming of the Atlantic SSTs relative to the warming of the tropical oceans as a whole. The consensus of the global climate projections is that the recent warming of the tropical Atlantic relative to the tropical mean will not continue into the future – that is, it is not thought to be part of the response to increasing greenhouse gases. This aspect of the global projections could be in error, although we have no particular reason to question it. We hope that an effect of our paper, and other recent work along these lines, will be to focus attention on this issue of spatial structure in the forced trend, rather than on the absolute temperature change in the Atlantic in isolation. The issue of how much credibility the global models have in their simulations of the spatial structure of the trends in ocean surface temperatures is likely to be a central issue when assessing the credibility of projections of future tropical storm activity.

If the global models are correct in this respect, the implication is that much of the observed recent excess warming of the Atlantic is due either to internal mechanisms (variability in the Atlantic Overturning circulation, for example) or non-greenhouse gas external forcing (e.g. recovery from strong aerosol cooling).

We also believe that there is evidence that recent trends in shear over the Atlantic, which have favored storm activity in recent years, are also controlled by the spatial structure of the SST trends (e.g. Zhang and Delworth (2006)). Once again, model projections suggest that these changes in shear should not be simply extrapolated into the future, and the credibility of these projections depends once again on the credibility of the projections of the spatial structure of trends in tropical SSTs.

14. How sensitive is the hurricane response to different global model projections of large-scale changes over the Atlantic?
We are in the process of repeating the calculations in our paper using the changes in ocean and large-scale atmospheric conditions generated by individual global models from the CMIP3 PCMDI archive, rather than the ensemble mean of these changes across these models. We expect that those models that show larger Atlantic warming relative to their tropical mean warming (as compared to the average over all of the models) will generate less of a reduction than we describe in this paper, or even an increase in tropical Atlantic storms. But we do not anticipate any results comparable to the 200-300% increase over the next century that one obtains by a naive extrapolation of recent trends. Our guess is that this level of increase is only possible if the tropical temperature profile as a whole is strongly destabilized by greenhouse gas increases (in contrast to the projections from our global models) or if the Atlantic continues to warm more rapidly than the tropical oceans as a whole (also in contrast to model projections), so that the century-scale changes in the atmosphere over the Atlantic resemble those in the reanalysis over the past few decades.In passing, we note that the Oouchi et al global 20km simulation produces a substantial increase in Atlantic storm activity. However, the ocean temperature anomaly that they use for boundary conditions includes a larger increase in SSTs in the storm generation region in the Atlantic, relative to the tropical mean warming, than is present in the ensemble mean CMIP3 projection that we utilize here. This may be responsible for part of their simulated increase.