Scientists from the National Center for Atmospheric Research (NCAR) are applying the same artificial intelligence (AI) techniques generally used with facial recognition technology to help improve predictions about the frequency of hail storms and their severity.

The team of scientists trained a convolutional neural network, which is a type of deep learning model often used to identify the features of an individual face for the purpose of facial recognition technology, to recognize the characteristics of a storm that could potentially signal the formation of hail as well as the size of hail — both of which are nearly impossible to predict.

According to the team, the storm’s structure is what influences the formation of hail. For instance, a supercell has a better chance of producing hail than a squall line. Yet, most current hail-forecasting methods only look at small segments of a storm and, consequently, cannot determine the overall structure of the storm.

Capable of inflicting considerable damage, large hail can significantly impact property and agriculture. As such, the team believes that the deep learning tools might offer further insight into the factors likely to produce hail, thereby enabling researchers to improve model predictions.

Meteorological factors that play a role in the formation of hail are wide ranging — air needs to be humid near the land surface, but dryer the higher up it is; cloud freezing levels need to be low to the ground; strong updrafts keep the hail above the ground long enough to increase in size; changes in wind duration and speed can also have an impact.

Even in the event that all of those criteria are met, hailstone size is still hard to predict and can vary significantly due to the path of the storm as well as the conditions along the path. The team identifies this as the point where storm structure becomes significant.

"The shape of the storm is really important," said NCAR scientist David John Gagne, who led the research team. "In the past we have tended to focus on single points in a storm or vertical profiles, but the horizontal structure is also really important."

Current models are limited in scope due to the mathematical complexity involved in representing the physical properties of the complete storm. However, machine learning reportedly bypasses the need for a model capable of solving the physics of the storm. Instead, the machine learning neural network consumed large amounts of data, looked for patterns and taught itself the features that suggest the possibility of hail. The team then trained the system using images of simulated storms in addition to inputting data concerning pressure, wind speed, temperature and direction. Hail simulations resulting from those conditions emerged as outputs.

The model then identified features of the storm associated with whether or not it hails as well as hailstone size. After successfully proving that it is capable of making predictions, the team examined which features of the storm the model’s neural network identified as most important. Following that, the team used a method that ran the model backwards to locate the storm characteristics that would need to combine with the highest probability of producing severe hail.

Based on this, the team determined that storms with lower than average pressure near the surface and with higher than average pressure at the storm top (resulting in updrafts) have a greater probability of producing severe hail. Storms with blowing winds from the southeast near the surface and from the west near the top will also produce severe hail. Likewise, storms more circular in shape also have a higher probability of producing hail, according to the team.

"I think this new method has a lot of promise to help forecasters better predict a weather phenomenon capable of causing severe damage," Gagne said. "We are excited to continue testing and refining the model with observations of real storms."

The research is published in the American Meteorological Society's Monthly Weather Review