UNIVERSITY PARK, Pa. — A group researching machine learning and using it to predict weather was awarded the best paper prize at the “Tackling Climate Change with Machine Learning” workshop at the International Conference on Learning Representations (ICLR), held virtually in Vienna, Austria in May.
Romit Maulik, a collaborator at Penn State's Institute for Computational and Data Sciences (ICDS) and assistant professor in the School of Information Science and Technology, was co-author of the winning paper, “Scaling Transformer Neural Networks for Clever and Reliable Medium-Range Weather Forecasts.”
“This is a year-long collaborative effort between UCLA, Carnegie Mellon University, Argonne National Laboratory and Pennsylvania State University,” Maulik said.
The study investigated forecasting using modern artificial intelligence (AI) tools compared to traditional methods currently used for operational forecasting.
“This is a paradigm shift from looking at traditional forecasts provided by multiple agencies,” Maulik said. “These forecasts are typically obtained using very large computational resources, which can be computationally expensive. We asked ourselves, what if we did it differently?”
Morlick explained that the AI model, based on computer vision techniques, learns weather patterns by sourcing data from historical information such as archived forecasts and satellite imagery.
“The trained model can then make predictions in real time without access to very large computational resources,” Maulik says. “Once the neural network is trained and released, model deployment can be done efficiently on a laptop, and eventually on increasingly smaller resources such as a mobile phone.”
Maulik said ICLR hosts several workshops on AI subtopics where researchers can present their papers and get feedback.
“Being accepted into a highly competitive workshop gives the paper maximum visibility,” says Maulik. “We get good feedback from both the AI and domain science communities, which helps us improve our methods significantly. The award itself is great and is recognition of our efforts. However, our long-term goal remains the same: we want to improve our current models and find ways to compete with traditional weather forecasting approaches.”
One of the researchers' goals, Maulik said, is to be able to more effectively predict extreme weather events that current models may struggle to predict.
“We have our sights set on bigger challenges,” Maulik says, “but as computational scientists, we want to solve problems first, and think about tools afterwards. We try to balance classical and machine learning methods, not leaning toward one or the other.”
Authors and collaborators on this article include Maulik, Tung Nguyen, Rohan Shah, Hritik Bansal, Troy Arcomano, Veerabhadra Rao Kotamarthi, Ian Foster, Sandeep Madireddy, and Aditya Grover.
