When it comes to extreme weather, an errant forecast can have serious effects. While advance warning can give people time to prepare for the weather as it did with the polar vortex last year, the absence of accurate advance warning can cause severe damage and loss of life, as it did with the failure to accurately predict the intensity of Hurricane Harvey in 2017. Now, a team of researchers from Rice University have leveraged the power of three supercomputers to apply deep learning to extreme weather prediction and improve outcomes.

“Generally, the numerical weather models do a good job predicting weather, but they still have some difficulties with extreme weather,” said Pedram Hassanzadeh, an assistant professor of mechanical engineering and earth, environmental and planetary sciences at Rice University, in an interview with the Texas Advanced Computing Center (TACC). “We’re trying to do extreme weather prediction in a very different way.”

That “different way” constituted a framework, built by Hassanzadeh and two of his PhD students (Ashesh Chattopadhyay and Ebrahim Nabizadeh), that approached extreme weather prediction as a pattern recognition problem and used deep learning to try to solve it. They chose to use two state-of-the-art approaches for deep learning pattern recognition: ConvNet, a convolutional neural network that has been popular over the last decade, and CapsNet, a more advanced capsule neural network that accounted for the relative position and orientation of weather features – a property that turned out to be crucial.

But to tackle weather data with advanced deep learning, they needed serious computational firepower.

This led the researchers to TACC and Pittsburgh Supercomputing Center (PSC), where they utilized three supercomputers to analyze large datasets and run machine learning codes. These were TACC’s Stampede2, a Dell EMC system benchmarked at 10.7 Linpack petaflops; TACC’s Wrangler, a data-focused Dell EMC system equipped with flash storage; and PSC’s Bridges, an HPE system that recently received a $1.8 million award to improve its deep learning capabilities.

“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.”

Using these supercomputers, the research team tested the framework on heat waves and cold spells using a small subset of atmospheric circulation data at a fixed altitude, along with some basic data on preceding surface temperature. They found that their advanced deep learning techniques were successful in predicting the events and outperformed simpler methods.

“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,” said Hassanzadeh.

The method bears some resemblance to early analog forecasting, whereby forecasters would attempt to predict weather events by studying earlier events that they expected to recur on a regular basis. Now, of course, the technique has evolved. “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.

With the basics of the approach now proven, Hassanzadeh’s team is beginning to take aim at the real competition: the operational models that provide the official weather forecasts for governments around the world. 

“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.”

About the research

The research referenced in this article was published as “Analog Forecasting of Extreme-Causing Weather Patterns Using Deep Learning” in the January 2020 issue of the Journal of Advances in Modeling Earth Systems. The paper, which was written by Ashesh Chattopadhyay, Ebrahim Nabizadeh and Pedram Hassanzadeh, can be accessed here.

To read TACC’s article on the research, by Faith Singer-Villalobos, click here.

Header image caption (from the researchers’ paper): Cluster centers of T2m [temperature at 2m above ground] anomalies at the onsets and Z500 [geopotential height at 500mb] patterns of 3 days earlier. The top (bottom) two rows correspond to summers (winters). S0 (W0) shows the average of T2m and Z500 patterns from days with no heat wave (cold spell). S1-S4 and W1-W4 are obtained from K-means clustering the anomalous T2m patterns at onsets into four classes, which roughly separates the extreme events into four geographical regions: Northern Canada (S1), Western US-Canada (S2), Southern US (S3), and Eastern US-Canada (S4) in summers, and North-West US-Canada (W1), Alaska (W2), North-East US-Canada (W3), and Northern Canada (W4) in winters. Rows 1 and 3 show the cluster centers while rows 2 and 4 show the average of Z500 patterns 3 days before the onsets for each cluster.