Gezani Impacts Madagascar: How Did AI Models Perform?

With the Southern Hemisphere TC season in full swing, right now activity is focused over the Indian Ocean as a new MJO pulse begins to develop in the area. The main system of interest right now is Cyclone Gezani, which is impacting Madagascar near the city of Toamasina, which is a major port city, as a Category-3 equivalent with winds of 110 knots.

AI models performance for this system has been mixed. 5 days ago, EC-AIFS forecast the system to dissipate as it got picked up by a trough, before reaching Madagascar (Figure 1).

On the other hand, AI-GFS kept the system intact, though too weak, and brought it in close to the correct location along the coast (Figure 2).

One question is how Google DeepMind’s ensemble, which has become one of the leading models for TC forecasting in the Atlantic and East Pacific, performed for this system. Figure 3 shows the DeepMind forecasts from 12 UTC February 5, out to 7 days. A cluster of members showed the system dissipating, like EC-AIFS for this cycle. Another cluster of members kept the system alive and moved it towards Madagascar, though they were almost all too weak.

Two days later, the model had correctly locked on to the fact that Gezani was going to survive and move towards Madagascar, and more of them showed intensification into a strong cyclone (Figure 4).

The TC is fairly small, which may be part of why the intensity was a struggle for AI models, even with the corrections based on historical data that DeepMind does, which should have helped the forecast some, given there have been several systems with similar track/intensity in February in the historical dataset (Figure 5).

Overall, this case shows that there are still cases that prove somewhat challenging for AI models, though it is encouraging that the ensemble still suggested the possibility of extreme intensification even when most of the members did not.