AI Isn’t Ready to Take Over Weather Prediction

A recent report by Wired reveals that while many weather prediction efforts are supported by intricate computer algorithms, humans are still doing a good amount of the legwork. We currently rely on the GOES-16 and -17 satellites (AKA the latest geostationary satellites) as well as the Global Forecasting System (GFS) and European Center for Medium-Range Weather Forecasts (ECMWF) models to forecast the weather with more precision than ever before. But despite several decades of computerized automatic forecast development, human-powered (or at least human-enhanced) predictions assisted by these technologies are more accessible and more accurate than their AI counterparts.

Like in plenty of other verticals, a fully-automated meteorological future faces many obstacles. Weather prediction produced solely by AI would require massive amounts of computing power, such as that supplied by exascale computers, a category of supercomputer capable of processing 1018 calculations per second. Three exascale computers are currently under development in the US, with the first—the Aurora supercomputer at Argonne National Laboratory—slated to go live this year, but meteorology isn’t the only research area in line to experience Aurora’s power. Accurate weather forecasting is also threatened by the inevitable full deployment of 5G, in which radio interference could negatively impact vital satellites’ ability to observe water vapor levels. Weather forecasting relies partly on monitoring the 23.8GHz signals emitted by water vapor.

One solution to this problem is to deploy more equipment in the lower-performance but longer-range C-band. The 24GHz signals used for mmWave 5G offer higher performance but weaker range, as discussed in this PCMag story from 2019. 5G signals in the 3GHz – 7GHz range will not interfere with future weather forecasting.

In the meantime, computer-generated forecasts lack the flavor necessary to effectively prepare for disaster. While algorithmic models are generally more accurate and efficient than humans at predicting mild weather, humans more consistently produce accurate predictions regarding bad weather (the latter of which is arguably more important to get right). An analysis of two decades of human, GFS, and North American Mesoscale Forecast System (NAM) predictions showed that humans beat the world’s two most popular models in the “bad weather” category anywhere from 20 to 40 percent of the time. In other cases, humans were able to add value to automated guidance, using the algorithm’s predictions as a foundation for more detailed forecasting.

None of this is to say that automated forecasts aren’t valuable. Instead, today’s meteorology students are taught to fight complacency by learning to defend their predictions using real-time and historical data. “There’s an old adage that ‘all models are wrong, some are useful,’” meteorologist Shawn Milrad, an instructor at Embry-Riddle Aeronautical College, told Wired. “Even if it’s a great forecast it’s going to be slightly wrong. It’s how you can add value to that model.”

Continue reading

Predictive Policing Software Shown to Entrench Bias, not Address It

A new analysis by Gizmodo has concluded that such software disproportionately targets poor communities and communities of color.

DeepMind AI Breakthrough Allows Prediction of More Than 200 Million Protein Structures

DeepMind has announced that has generated structures for all 200+ million proteins in the centralized UniProt database. This is a big deal for basic biological research as well as for efforts to tackle some of the most important scientific conundrums of our time.

Google Now Predictive Search Cards Resurface in Google Assistant

Google phased out Now as a brand a while back, and the revamped Feed doesn't include the same predictive features. Now, some of them are coming back in the Google Assistant.