Orenstein and team win global climate prediction challenge

IMPA researcher Paulo Orenstein is part of the winning team in the first stage of the AI Weather Quest, an international competition promoted by the European Centre for Medium-Range Weather Forecasts (ECMWF) and dedicated to advancing sub-seasonal climate forecasts. The group brings together researchers from Microsoft, Harvard University, the Massachusetts Institute of Technology (MIT), the University of Toronto and Rhiza Research.

The initiative aims to stimulate the development of models capable of anticipating temperature, precipitation and atmospheric pressure conditions two to six weeks in advance – an interval considered one of the most challenging in meteorology. The team, called MicroEnsemble, won first place in all the variables evaluated and beat both the operational dynamic models of six government forecasting agencies and the artificial intelligence models developed internally by ECMWF and the systems of 30 teams from around the world.

To meet the challenge, the researchers developed Duet, a model that combines two complementary strategies based on artificial intelligence (AI).

“Our model brings together two central ideas: a deep neural network based on transformers, capable of predicting the state of the weather several moments ahead, and a probabilistic bias correction technique, based on machine learning, which allows us to combine numerical and AI-based forecasts without the errors accumulating in a pronounced way. This integration means that Duet outperforms both traditional numerical models and purely AI-based models,” explains Orenstein.

With this hybrid approach, Duet achieved a more than 200% increase in predictive ability for temperature, precipitation and pressure at mean sea level. The winning model could even begin to be used operationally. According to Orenstein, ECMWF’s proposal is to integrate the winning models and test them in real time.

Sub-seasonal climate forecasts are strategic for areas such as water resource management, agricultural planning, forest fire prevention and the anticipation of droughts and extreme weather events, with a direct impact on the formulation of public policies and decision-making. In this context, the result also reinforces the role of applied mathematics. “The award highlights how mathematical ideas can have an impact on very concrete problems, such as weather and climate,” he said.

The winning team’s performance reinforces IMPA’s work in frontier research and highlights the contribution of its researchers to the development of scientific solutions to global challenges, integrating mathematics, data and artificial intelligence.

Check out the interview with Paulo Orenstein below.

IMPA: What does winning the AI Weather Quest in all categories mean for you and IMPA?

Orenstein: It’s great to be at the intersection of new algorithms, mathematical ideas and applications with the potential to generate concrete and broad benefits for society. Winning this stage of the AI Weather Quest also serves to show that several ideas we have proposed in the past, when put to the test, can in fact substantially improve sub-seasonal climate forecasts. This sub-seasonal horizon is considered one of the most challenging in the field and will probably continue to be an important source of new ideas in AI, statistics and climate. Participating in these advances in Brazil is particularly relevant, given the impact that climate change could have on the country. And, from an institutional point of view, doing so from IMPA highlights the fundamental role that mathematics can play far beyond the classroom.

IMPA: What differentiates the model you developed (Duet) from traditional weather forecasting approaches?

Orenstein: Classic weather forecasting approaches rely on physical models that simulate the functioning of the atmosphere. Although these models are very good in the short term (say, a few hours to a few days ahead), measurement and approximation errors begin to propagate and multiply, making forecasting many days ahead increasingly difficult. In addition, these models are computationally very expensive.

More recently, AI models have achieved good results in short-term climate forecasts, largely thanks to the abundance of atmospheric and climate data available. Nevertheless, these models also face difficulties in extending quality forecasts over many days.

Our model combines two ideas: (i) a deep neural network that uses *transformers* to predict the weather several moments ahead; and (ii) a probabilistic bias-correction technique, based on machine learning, which makes it possible to combine several forecasts – numerical and AI-based – without the errors accumulating in a pronounced way. This combination means that our model outperforms both forecasts based on numerical models and new models based purely on AI.

IMPA: What was it like working in a team with researchers from institutions such as Microsoft, Harvard, MIT and the University of Toronto?

Orenstein: A job like this is essentially collaborative. The volume of ideas, code and tasks needed to produce global forecasts week after week requires a coordinated effort from several people. In this sense, having top-level co-authors in different parts of the world was essential and very enriching for the success of the project.

IMPA: Can this model already be used by forecasting bodies or is it still in the experimental phase?

Orenstein: Yes. The idea of the European Center for Medium-Range Forecasting (ECMWF) is precisely that the winning models from each stage can be integrated and tested in real time. We are very excited about this possibility: over the last few months, our model has outperformed the main operational models used in the world.

IMPA: What are the next scientific challenges after this result?

Orenstein: The AIWQ competition awards prizes by season, and this was just the first. The next step is to adapt our ideas for the next season: winter (in the northern hemisphere) and summer (in the southern hemisphere). We know that the seasons have a strong impact on the quality of the sub-seasonal forecasts, so there’s a lot of work ahead. As well as preparing the model for the next stage, continuing in the competition is an opportunity to explore new ideas in statistics, mathematics and machine learning, advancing our understanding of sub-seasonal weather phenomena.

IMPA: What message does this award send to young researchers interested in mathematics, climate and artificial intelligence?

Orenstein: Mathematics is very broad, and there is room for all profiles and interests. In particular, I’m interested in mathematics that preserves a certain intrinsic beauty and, at the same time, has an impact on very concrete problems. In this sense, the award reinforces the potential of mathematical ideas in areas as diverse as weather and climate. It is worth remembering that the same mathematics behind recent meteorological advances is the mathematics that, decades ago, helped reveal fundamental limitations in our ability to predict chaotic systems, such as the weather. Finally, the result also highlights the reach that artificial intelligence is having in areas that are essential to society – and, luckily for us, it has also been a source of a lot of interesting mathematics.