While the past may be a fixed and unchangeable point, machine learning sometimes makes predicting the future easier. Researchers at The Ohio State University have recently discovered a new method to predict the behavior of spatiotemporal chaotic systems, such as changes in the Earth’s weather, that are particularly difficult for scientists to forecast. This method is known as “next generation reservoir computing.”
The research, which was just published in the multidisciplinary journal Chaos: An Interdisciplinary Journal of Nonlinear Science, uses a brand-new, highly effective algorithm that, when combined with next-generation reservoir computing, can learn spatiotemporal chaotic systems in a fraction of the time it takes traditional machine learning algorithms.
Researchers put their algorithm to the test by predicting the behavior of an atmospheric weather model, a challenging problem that has been extensively researched in the past. The Ohio State team’s algorithm is more accurate and uses 400 to 1,250 times less training data to make better predictions than its rivals, traditional machine learning algorithms that can complete the same tasks. They used a laptop running Windows 10 to make predictions in a fraction of a second, which is roughly 240,000 times faster than conventional machine learning algorithms. Their method is also less computationally expensive, whereas solving complex computing problems previously required a supercomputer.
Wendson De Sa Barbosa, the lead author and a postdoctoral researcher in physics at Ohio State, said, “This is very exciting because we believe it’s a substantial advance in terms of data processing efficiency and prediction accuracy in the field of machine learning.” He claimed that the ability to predict these highly chaotic systems is a “physics grand challenge,” and mastering it could lead to novel scientific insights and advancements.
According to De Sa Barbosa, modern machine learning algorithms are particularly effective at forecasting dynamical systems because they can learn the underlying physical principles from prior data. Any complex real-world system can be predicted using machine learning models once there is sufficient data and computing power. Any physical process, such as the bob of a clock’s pendulum or interruptions in power grids, can be a part of such systems.
According to De Sa Barbosa, even heart cells exhibit fractal spatial patterns when they oscillate at an abnormally higher frequency than a regular heartbeat. Therefore, this research may one day be applied to better understanding and managing heart disease as well as a variety of other “real-world” issues.
Its behavior “could be reproduced and predicted if one knows the equations that precisely describe how these particular processes for a system will evolve,” he said. The clock’s swing position, for example, can be accurately predicted using only its current position and velocity. More complicated systems, like the weather on Earth, are much more difficult to predict because so many factors actively shape their chaotic behavior.
According to De Sa Barbosa, it would be impossible for scientists to make accurate predictions of the entire system without precise knowledge of each and every one of these variables and the model equations that show how these numerous variables are related. However, the nearly 500,000 historical training data points used in earlier studies for the atmospheric weather example used in this study could be decreased to only 400 with their machine learning algorithm while still achieving the same or better accuracy.
De Sa Barbosa hopes to move forward with his research in the future by using their algorithm to possibly speed up spatiotemporal simulations.
Because we still know so little about the world we live in, it’s critical to identify these highly dynamical systems and learn how to predict them more accurately.
Daniel J. Gauthier, professor of physics at Ohio State, was a co-author of the study. The Air Force Office of Scientific Research provided funding for their work.
