Tracking ocean temperatures has long helped scientists measure Earth’s accelerating energy imbalance, but researchers at the University of California, Santa Barbara have now applied new machine learning techniques to extrapolate deep-ocean temperatures (below 2,000 meters) with surprising results.
Whereas past research has shown gradual warming that has accelerated over time, new research from Aaron Bagnell, a doctoral student, and Tim DeVries, an associate professor, has suggested that cool deep-ocean temperatures offset surface warming until roughly 1990. After 1990, the authors posited, more rapid warming caused a spike in surface ocean temperatures that could not be offset by prior deep-ocean cooling.
Cooling associated with past events like the Little Ice Age (which covered roughly the 14th–19th centuries) had a lingering effect; DeVries said it takes roughly 1,000 years for surface ocean waters to circulate to the deep ocean, and vice versa, meaning that those cooler waters had an impact even centuries later. After 1990, the cumulative impact of surface-level warming was enough to cause whole-ocean temperatures to rise. The study was published in Nature Communications.
Bagnell’s autoregressive artificial neural network (ARANN) is an adaptation of a machine learning method that he said improves on previous methods of estimating past deep-ocean temperatures. Historically, sparse data from the deep ocean have made it difficult, if not impossible, to accurately extrapolate whole-ocean temperatures. “Previous methods might have just taken data from 1 year or 1 month, and applied some mapping algorithm to fill in the gaps, whereas we’re considering the whole record all at once, and we’re letting this machine learning algorithm, or artificial intelligence algorithm, make connections between…what happens near the surface [and] what happens deeper down,” DeVries said.
With machine learning techniques, each data point can be influenced by every other data point in a way that accounts for the ocean’s dynamic properties. “So, in principle, the machine learning technique is learning how to adapt the relationships that it sees near the surface,” from around the present to prior time periods, as well as [in] deeper parts of the ocean, Bagnell said.
Accounting for Sea Level Rise
The result is a finding that not everyone agrees with. Greg Johnson, an oceanographer with NOAA’s Pacific Marine Environmental Laboratory, said the lack of data about deep-ocean temperatures prior to 1970 makes him skeptical about the potential accuracy of these findings. He also pointed to rising sea levels as evidence that global warming was proceeding steadily prior to 1990.
“The sea level has been going up fairly steadily.… From the 1990s to the 2000s and 2010s, it’s definitely accelerated considerably, but prior to that it was still rising globally, and that’s probably not all ice melt, which is what you’d have to conclude from this study,” he said, pointing to a global rise in sea surface temperatures over the study’s time frame.
Bagnell responded that data on sea level rise don’t necessarily contradict his findings: “Thermal expansion of the oceans due to warming accounts for roughly one third of the sea level rise observed today. The other two thirds is mainly due to the addition of freshwater from land sources like melting glaciers. Additionally, the historical record of tide gauges and satellites (for 1993 onwards) indicates that sea level rise has been accelerating from the 20th century into the 21st century, meaning sea levels likely weren’t rising as rapidly then as they are today. So the possibility that freshwater input accounted for a larger fraction of sea level rise during the 20th century isn’t out of the question.”
—Robin Donovan (@RobinKD), Science Writer