Science Success Story

Can Deep Learning Yield More Accurate Extreme Weather Forecasts?

XSEDE systems support pattern recognition-based extreme weather prediction

By Faith Singer-Villalobos, Texas Advanced Computing Center

 

Forecasting the weather patterns that cause extreme weather events is challenging despite decades of efforts and advances in numerical weather prediction (NWP). Modern forecasts use mathematical models of the atmosphere and oceans to predict the weather based on current weather conditions. Even with the increasing power of today's supercomputers, the forecasting skill of numerical weather models extends to only about six days, although there is some dependence on location, season, and type of weather pattern.

Rice University engineering researchers Ebrahim Nabizadeh (seated), Pedram Hassanzadeh, and Ashesh Chattopadhyay (not pictured) trained a deep learning neural network to predict extreme weather using World War II-ear "analog" weather forecasting. Credit: Jeff Fitlow, Rice University.

Persistent weather patterns that are often the drivers of extreme events are particularly hard to forecast. Improving the forecast of such events using NWP requires using higher-resolution models and running more simulations starting from almost the same weather conditions. The latter is needed to tackle the chaotic nature of the atmosphere, i.e., the famous butterfly effect. However, higher-resolution models and more simulations demand enormous computational resources.  

Pedram Hassanzadeh, an assistant professor in Mechanical Engineering and Earth, Environmental and Planetary Sciences at Rice University, and his PhD students Ashesh Chattopadhyay and Ebrahim Nabizadeh, recently introduced a data-driven framework that: 1) formulates extreme weather prediction as a pattern recognition problem, and 2) employs state-of-the-art deep learning techniques. Their findings were published in the February 2020 edition of the American Geophysical Union's Journal of Advances in Modeling Earth Systems.

To obtain their results, the researchers analyzed large data sets and employed machine learning codes on supercomputers at the Texas Advanced Computing Center (TACC) and the Pittsburgh Supercomputing Center with allocations through XSEDE. Each data set was several terabytes in size. In addition, they used data that had already been produced by supercomputers at the National Center for Atmospheric Research as input for the deep learning models.

"Our work would not have been possible without XSEDE's computing resources," Hassanzadeh said. "Stampede2, Wrangler, and Bridges enabled us to do this work. We have supplemental systems at Rice, but Stampede2 is the main supercomputing resource that my group uses, and Bridges enables us to efficiently work with very large datasets."

Deep learning is a form of artificial intelligence, in which computers are "trained" to make humanlike decisions without being explicitly programmed for them. The mainstay of deep learning, the convolutional neural network, excels at pattern recognition and is the key technology for self-driving cars, facial recognition, speech transcription, and dozens of other advances.

The advantage of a data-driven framework is that once trained on observational and/or high-resolution numerical model data, it can provide relatively accurate predictions at very little computational cost, which can augment and guide other NWP efforts by providing early warnings.

"Generally, the numerical weather models do a good job predicting weather, but they still have some difficulties with extreme weather," Hassanzadeh said. "We're trying to do extreme weather prediction in a very different way."

As a proof-of-concept demonstration, Hassanzadeh and team predicted heat waves and cold spells over North America using limited information about the atmospheric circulation at an altitude of around five kilometers, and in some cases, the surface temperature a few days earlier. The results of their demonstration suggest that extreme weather prediction can be done as a pattern recognition problem, particularly enabled by the recent advances in deep learning. In fact, the researchers found that more advanced deep learning methods outperformed simpler techniques, suggesting potential benefits in developing deep learning methods tailored for climate and weather data.

A schematic representations of the capsule neural network Rice University engineers created to forecast extreme weather events. Credit: Mario Norton, Rice University Digital Media Commons.

"We found that because the relative position of weather patterns play a key role in their evolution, using a more advanced deep learning method that tracks the relative position of features improves the accuracy and is also more robust when we don't have a large amount of data for training," Hassanzadeh said.  

Interestingly, pattern matching is the way people started doing weather prediction before and during the Second World War. In that era, people barely scratched the surface of what is possible today. And even integrating an equation into the weather system, a first step in a mathematical model, was not possible.

During that time, people did weather prediction by looking through catalogs of weather patterns and pattern matching — this is called analog forecasting. But meteorologists abandoned this approach after World War II once computers became more widely available.

The analog technique is a complex way of making a forecast, requiring the forecaster to remember a previous weather event that is expected to be mimicked by an upcoming event. What makes it a difficult technique to use is that there is rarely a perfect analog for an event in the future. It remains a useful method of observing rainfall over oceans, as well as forecasting precipitation amounts and distributions.

"In this paper, we show that with deep learning you can do analog forecasting with very complicated weather data — there's a lot of promise in this approach," Hassanzadeh said.

According to Hassanzadeh, a growing number of people in the weather and climate community are interested in how deep learning can help improve climate and weather modelling.

"I think we're showing people that this approach works," he said. "The next step for my group is to see if deep learning can be more accurate than the operational numerical weather models used for day-to-day weather forecasts. We may be able to train the neural networks using observational data, and it might work better and more accurately than what you get from the numerical weather models for predicting extreme events. We're going to focus on predictions with longer lead times, where the numerical models perform poorly. If it works, it will be a huge advance in weather prediction."

The study, "Analog Forecasting of Extreme-Causing Weather Patterns Using Deep Learning," was published in January 2020 in the Journal of Advances in Modeling Earth Systems (JAMES). The study co-authors are Ashesh Chattopadhyay, Ebrahim Nabizadeh, and Pedram Hassanzadeh of Rice University. This study was funded by NASA grant 80NSSC17K0266 and an Early-Career Research Fellowship from the Gulf Research Program of the National Academies of Sciences, Engineering, and Medicine. Computing resources were provided by TACC and PSC under the National Science Foundation-supported XSEDE project and Rice's Center for Research Computing in partnership with the Ken Kennedy Institute.

At A Glance

  • A growing number of people in the weather and climate community are interested in how deep learning can help improve climate and weather modelling.
  • Researchers from Rice University introduced a data-driven framework that: 1) formulates extreme weather prediction as a pattern recognition problem, and 2) employs state-of-the-art deep learning techniques.
  • Their findings were published in the February 2020 edition of the American Geophysical Union's Journal of Advances in Modeling Earth Systems.
  • To obtain their results, the researchers analyzed large data sets and employed machine learning codes on supercomputers at TACC and PSC. In addition, they used data that had already been produced by supercomputers at NCAR as input for the deep learning models.