Turning the tables on weather forecasting

The basic equations used in meteorology have remained fundamentally unchanged since they were developed by Norwegian physicist Vilhelm Bjerknes in the early 1900s, based on calculations of atmospheric dynamics. Progress in weather prediction since then has principally been in the enormous expansion of weather monitoring resources around the world and in the supercomputers we use to solve the equations that use this influx of data.

We also now have the ability to combine newer observation technologies, including ground-based radar, satellite imagery and meteorological readings, to give an incredibly accurate snapshot of current conditions across most parts of the world. So the true challenge of weather prediction today is assimilating all of that data and running sufficient numbers of simulations quickly enough to enable timely prediction updates.

At the RIKEN Center for Computational Science, we are fortunate to have access to the Fugaku supercomputer, one of the most powerful computing facilities in the world. Using even a small fraction of Fugaku’s power has allowed us to demonstrate what is possible at the cutting edge of weather prediction science.

One of the problems with our current ensemble simulation approach for weather prediction is that it is computationally intensive and requires nationally-supported infrastructure to be available to the public. This type of investment is not possible in many countries.

Even in Japan, the current operational weather prediction run by the Japan Meteorological Agency is updated once every hour at a two-kilometer resolution and uses a relatively limited dataset. Upgrading this operational system to something like our olympic demonstration would require further investment in infrastructure and a permanent tenancy on Fugaku.

We are beginning to turn our attention to AI as a potentially more efficient and accessible platform for weather prediction. Could an AI model running on a single desktop graphics processing unit (GPU) predict the weather with comparable accuracy to what we currently achieve using supercomputers?

This is a very important question that is being explored. But we still need data—AI is only as good as the data it is trained on—but assuming we can train an AI model on a sufficiently large and complete dataset, we believe this revolution could be possible and would put sophisticated prediction in the hands of the people.

The key to prediction is uncertainty: those of us pushing the limits of weather prediction are consumed by it. In our ensemble simulations, each simulation is equally probable, so the more of these simulations we run for each point in time, the better out modeling is at representing uncertainty.

However, regardless of the number of ensemble simulations, all predictions become increasingly uncertain as they move further into the future, through an accumulation of more possible consequences.

What this ensemble approach also shows us is that the tiniest fluctuation in any of the starting conditions or model assumptions can result in very different weather evolutions over time, also known as the ‘butterfly effect’.

This is a bane for forecasting, but it also raises an intriguing and potentially profound question: If the tiniest change can affect weather development so significantly in a simulation, could we not make that small modification in the real world to change the evolution of weather itself?

With our new ability to run thousands of ensemble simulations using high-resolution data, it is now feasible to accurately model the likely consequences of such a deliberate change to the weather.

My team has shown that it is technically feasible to run very large ensembles quickly to ‘experiment’ with the weather virtually, such as the consequences of small wind changes on the severity of typhoons or heavy rains. We think these experiments could one day lead to targeted weather modifications that could, for example, prevent the development of a predicted extreme event.³ AI could help pinpoint the most efficient modifications.

This of course is not something that could, or should, be undertaken lightly. Great care needs to be taken that our control efforts don’t have unintended consequences. But perhaps this is a way to turn the tables, to mitigate some of the weather impacts that we have caused as a result of climate change.

Our simulation sandbox is a great place to perform such experiments safely.