Computer Guidance for Tropical Forecasting

Read...

Although we covered an "old-school" approach for short-term tropical cyclone track forecasts on the previous page, we have many sophisticated tools for predicting the track and intensity of tropical cyclones. Indeed, the advent of computer model guidance revolutionized weather forecasting, and tropical forecasting is no exception. Thanks to developments in computer guidance, reasonably accurate forecasts for tracks of tropical cyclones are now the norm several days in advance.

You're already familiar with how computer models work and what some of their main flaws are from your previous studies, and you should be familiar with some commonly used computer models and forecast variables used for forecasting in the middle latitudes. That basic knowledge is still applicable to the tropics, but tropical forecasters have some other computer guidance tools to work with, too. I'll break the discussion down into three parts -- global models, specialized tropical cyclone models, and ensembles.

Global Models

Since tropical cyclones are a global phenomena, forecasters often turn to the "global models" (that is, models that have a domain covering the entire globe that you should already be familiar with from your previous studies) to keep tabs on tropical cyclones in any basin. This includes both the "traditional" numerical weather prediction models run by the major modeling centers around the world (such as the GFS in the United States, ECMWF from Europe, CMC GDPS from Canada, the UKMET from the United Kingdom, and the JMA from Japan, among others), as well as artificial intelligence-based models that these centers have developed, such as the AI-GFS or the EC-AIFS (which stands for the European Centre Artificial Intelligence/Integrated Forecast System). 

As a quick reminder, traditional numerical weather prediction models and artificial intelligence (AI) models produce their predictions in different ways. Traditional numerical weather prediction models start with a representation of the current state of the atmosphere and then solve numerous mathematical equations that describe the dynamics and thermodynamics of atmospheric behavior to predict future states of the atmosphere. AI models are trained on decades of past weather reanalysis data, and using what they learn about how atmospheric patterns evolve from their training dataset, they can predict future states of the atmosphere after starting with a representation of its current state (often without incorporating physics at all). As you may recall from your previous studies, both modeling approaches have their strengths and limitations.

For an example of what a tropical cyclone looks like in the broad domain of a global model, check out the GFS forecast below. The "footprints" of four tropical cyclones (circled) are apparent as regions of relatively low sea-level pressure. As we've already discussed in this lesson, we can quickly get the idea that tropical cyclones are relatively small features in the scheme of things (certainly compared to the larger mid-latitude cyclones located at higher latitudes). Just a few decades ago, global models had resolutions that were so coarse that they weren't of much use in providing detailed looks at the core and wind field of a tropical cyclone, but resolution has increased so that global models can provide these details to some degree. Still, other models have been developed specifically to provide more detailed guidance for existing tropical cyclones.

Specialized Tropical Cyclone Models

Because global models aren't always the best at simulating the finer details of tropical cyclones, forecasters also turn to models specifically designed to forecast tropical cyclones. These models generally do not have a global domain, and only cover specific tropical basins. NOAA's flagship model developed specifically for tropical-cyclone forecasting is the Hurricane Analysis and Forecast System (HAFS), which became operational in 2023. Some benefits of the HAFS include the fact that it is "ocean coupled," which means that changes in the ocean and atmosphere respond to each other in the model, which is not the case in some global models. Ocean coupling in a model can be a big advantage because as you'll learn later, strong hurricanes can dramatically alter the characteristics of the ocean beneath them, which can then in turn alter the intensity of the storm.

The HAFS is also run at a relatively high resolution, with "nests" that follow individual storms along in time. Its high resolution means that it is capable of predicting small-scale structures within a storm. Of course, there's no guarantee that these small-scale details will be accurate for any given storm, but the ability to realistically simulate deep convective cells can be very helpful in simulating processes in the cores of tropical cyclones, which can improve intensity prediction, on average. As an example of the detail provided by these forecasts, check out the 6-hour forecast (below) of composite radar reflectivity and mean sea-level pressure for Super Typhoon Doksuri (2023), as it approached northern Luzon in the Philippines (opens in a new window).
 

The core of Doksuri was depicted with great detail as it approached northern Luzon, and the HAFS predicted a central pressure of 917 mb. But, as I just mentioned, while the HAFS can make highly-detailed predictions, there's no guarantee that they'll be accurate (the lowest estimated central pressure during Doksuri's life was 926 mb, so this was a pretty substantial error for a six-hour forecast).

The HAFS is actually run in two configurations -- HAFS-A and HAFS-B (note that the forecast prog above is from the HAFS-A). While the HAFS is not a global model, the HAFS-A configuration is run in all tropical basins. The HAFS-B configuration is only run on tropical basins under the responsibility of the National Hurricane Center and the Central Pacific Hurricane Center. The HAFS-A and HAFS-B also have some differences in their ocean coupling schemes and how they simulate some small-scale physical processes. Furthermore, tropical cyclones in the HAFS-B domain that have Doppler radar and other data collected during aircraft reconnaisance flights (opens in a new window) have some extra initialization data compared to storms in other basins.

Lest you think that NOAA didn't run tropical-cyclone specific models until the HAFS debuted in 2023, there's actually a history of such models going back to the early 1990s with the GHM (Geophysical Fluid Dynamics Lab Hurricane Model). Earlier generations of tropical-cyclone specific models also consisted of the HMON (Hurricanes in a Multi-scale Ocean-coupled Non-hydrostatic model), which became operational in 2017, and the HWRF (Hurricane Weather Research and Forecasting) model, which became operational in 2007. The HWRF in particular was ground breaking because it was first operational model to be able to assimilate Doppler radar data collected during aircraft reconnaissance flights in its initialization. The HMON and the HWRF are still being run, but are planned to be phased out.

Traditional numerical weather prediction models like the HAFS aren't the only specialized approach to tropical cyclone modeling, however. Specialized AI models focused on tropical cyclone forecasting also exist. These models are trained specifically on past tropical cyclone cases to produce forecasts for tropical cyclones (track, intensity, size, structure, etc.). But, these specialized models don't produce forecasts for the entire atmosphere as global AI models do. Some private sector companies like Google have been major developers of AI-based tropical cyclone models, which play a big role in our next modeling topic -- ensembles.

Ensembles

As you know, both traditional numerical weather prediction models and AI models are fallible, and often, various models have differing solutions. Indeed, check out the average cyclone forecast track errors (opens in a new window) of various computer models. Given that no models are perfect, and their solutions are often different, do forecasters have any tools at their disposal for helping them navigate the sea of uncertainty? Ensemble forecasts, to the rescue! Ensemble forecasting embraces the tendency toward differing forecast solutions by allowing forecasters to see a range of possible forecast outcomes, which allows forecasters to gauge uncertainty.

You've already been exposed to the basics of ensemble forecasting, but allow me to quickly review. Recall from your previous studies that the data used to initialize a computer model is always imperfect (we're nowhere close to being able to perfectly measure variables in the atmosphere everywhere at all times). So, the model initialization always contains errors. Ensemble forecasts are created by slightly altering the initial conditions fed into the model and / or altering the model physics (recall that a model's ability to mimic the atmosphere is not quite perfect). Each slight altering of the initial conditions or model physics generates an ensemble member. When there's very little spread in the solutions from all ensemble members, then the forecast isn't particularly sensitive to small errors in initialization or differences in model physics, and confidence in the operational model solution is high. But, when lots of spread exists among the individual member solutions, then the forecast is very sensitive to those differences, and confidence is lower.

Ensembles comprised of traditional numerical weather prediction models require a lot of computing power to run, and individual ensemble members are often run at reduced spatial resolution to conserve computing resources. But, ensembles for AI models exist, too, and while AI models require an immense amount of training data to develop, actually running the models on a daily basis is far less resource intensive than traditional numerical weather prediction models. So, AI ensembles can have many more members and can run much faster than ensembles from traditional numerical weather prediction models. To see an example of each in action, check out the image slider below, which shows track forecasts from the ECMWF ensemble for Super Typhoon Sinlaku (2026). The thick black line represents the actual storm track, while the multi-colored lines represent the ensemble member forecasts from the 00Z run on April 9.

It's clear that the ECMWF had a huge spread in track forecasts for Sinlaku, but if you toggle the image slider to see the corresponding ensemble run from Google DeepMind's AI model, it had a much tighter clustering of track solutions in this case, which could have helped forecasters narrow the scope of likely possibilities. Both sets of ensembles, however, carried the striking message that confidence steadily lowered with increasing forecast time as the spread in forecast tracks grew (it's simply the nature of the beast that errors associated with computer guidance grow with increasing time).

Ensembles can also give forecasters an idea of the range of possibilities for tropical cyclone intensity forecasts. For Sinlaku, this comparison of the ECMWF and Google DeepMind ensembles (opens in a new window) shows a wide range of possibilities. The black lines represent the storm's actual intensity, and we can quickly take away a few key messages. First, the vast majority of ensemble members underestimated Sinlaku's peak intensity (the black line peaks at the high end of the wind speed forecasts and "bottoms out" on the low end of the sea-level pressure forecasts). However, members of the Google DeepMind ensemble did a good job with timing its rapid intensification to a Category 5 storm. On the other hand, the ECMWF ensemble had a few members that better depicted Sinlaku's maximum intensity (below 900 mb), though they intensified the storm too slowly. Ultimately, forecasters utilize both "physics-based" and AI-based ensembles for key messages about forecast uncertainty. That's very helpful information, and it's much better to take into account the range of possibilities as opposed to locking in on a couple of operational model runs.

Other approaches to ensemble forecasting also exist (combining multiple ensembles into "super ensembles" or simply comparing many completely different models as an ensemble, for example). With many modeling options available (and I only covered the major ones on this page), it's important to remember from your previous studies that forecasters look for consensus among the models and diligently comb over real-time observations that might offer clues about which models have a better handle on a particular weather system. The same approach rings true for predicting tropical cyclones. If you're curious about where you can access model guidance from the global models and other models specifically created for predicting tropical cyclones, check out the list of resources in the Explore Further section below (the section also covers other models that I didn't touch on).

Yes, there's a wide variety of model guidance available to tropical forecasters. But, because tropical cyclones operate differently than mid-latitude cyclones, some unique forecast variables are of interest to tropical forecasters. We'll take a look at these variables next.