The cyclones that hit Mozambique at the beginning of each year are among the natural causes of social and economic damage. But new forecasting models are making a difference and can help mitigate the impact.
Cyclone Chido, which hit Mozambique on 15 December 2024, was the latest to join the long list of atmospheric depressions that batter the country every year between December and May, causing deaths, injuries and a great deal of damage. The way the earth’s atmosphere works means that Mozambique is always one of the arrival points for cyclones from the Indian Ocean.
In recent years, scientists have been warning of signs that these natural phenomena are intensifying due to climate change. In the case of Cyclone Chido, the figures speak for themselves: at least 120 people died and another 868 were injured during the storm’s passage through northern and central Mozambique (Cabo Delgado, Niassa and Nampula in the north, and Tete and Sofala in the centre). Numerous infrastructures and companies were affected. The impact on the economy is severe. But new technologies are revealing new ways to help mitigate human suffering and damage to the economy. Artificial Intelligence (AI) is at the forefront.
Google DeepMind launches GenCast
At the end of 2024, scientists at Google DeepMind developed a meteorological model that surpasses the world’s most accurate systems: the new form of forecasting is able to cover a greater number of different initial conditions. The researchers announced that the GenCast model has improved computers’ ability to anticipate extreme events (such as intense cyclones) that are outside the limits of the data they were trained on. In other words, it will be possible to accurately predict unprecedented events, but which climate change is making more likely and more serious. Tests have shown that the model has greater capacity than the extremely reliable ensemble run by the European Centre for Medium-Range Weather Forecasts (ECMWF) in 97.2% of the 1320 metrics evaluated – including the prediction of extreme weather events, tropical cyclone trajectories and wind energy production.
GenCast’s machine learning approach used meteorological data from 1979 to 2018. As the processing capacity of computers continues to increase, improvements in forecasts in recent years have been iterative (in terms of quantity) rather than revolutionary. The new generation of AI weather models being launched by technology companies are trained on previous meteorological data and rely on machine learning techniques. They can be run on remote processing systems (in the cloud) in just a few minutes.
In the case of Mozambique, the ambition to create an early warning network may involve investing in this type of new generation system, but collecting accurate data remains a problem
Nvidia shows StormCast
Amid the peaks of the latest Atlantic hurricane season, Nvidia has announced a new generative AI model, called StormCast, to simulate atmospheric dynamics with high precision. In other words, the model makes more reliable weather forecasts at the mesoscale – a scale larger than storms but smaller than cyclones – which is fundamental for disaster planning and mitigation.
Integrated into Nvidia’s Earth-2 platform, which combines AI, physical simulations and computer graphics, StormCast offers auto-regressive hourly forecasts, anticipating future events based on past data. Complementing CorrDiff, another Nvidia AI model, StormCast improves high-resolution regional forecasts, which are essential for identifying physical risks related to weather and climate.
The company has indicated that CorrDiff has already proven to be a thousand times more efficient than traditional methods, with thousands of times less energy consumption. For example, Taiwan’s National Science and Technology Centre for Disaster Reduction plans to use CorrDiff to predict the path of typhoons, ‘work that previously cost almost three million dollars in CPU, but can now be done for around 60,000 dollars on a single system with an Nvidia H100 Tensor Core GPU.’ In other words, in addition to the capacity of the weather forecasting software, the company is also drawing on its silicon chip production.
In the end, human validation counts
Despite the advances, the final validation of weather forecasts will continue to be done by humans and it is unlikely that the new models, such as GenCast or StormCast, will replace all those currently in use. What is more plausible is that they will be used as yet another forecasting tool in the meteorologists’ toolbox. Their experience in observing results and forecasts on a daily basis gives them an advantage: they know when there are good and bad predictions and can use this to their advantage to make more accurate forecasts, say experts in the sector.
These systems are important tools because they use mathematical equations that describe how the atmosphere works, with calculations for everything from how moisture is transported to how jetstream winds can feed and guide storm systems. A huge amount of observations are fed into these models to produce projections for the next few hours or days.
In the case of Mozambique, the ambition to create an early warning network may involve investing in this type of new generation system, but collecting accurate and detailed data remains a problem. In September 2023, the country announced that it had meteorological stations in operation in half of the country’s districts (88), but without detailed information on the state of operation.
With only one weather radar installed in the centre, specifically in the city of Beira, the country needs another six to fully cover the territory.
