Speevr logo

Climate | Extreme weather predictions with machine learning

Table of Contents

Can machine learning improve weather predictions?

I was discussing the potential for applying machine learning (ML) in climate prediction models yesterday with our colleague Adam Lund, a statistician by training. One common issue with ML models is that while they can excel in specific tasks, they often struggle to capture other important aspects of the system.

Sometimes for local weather and by violating the laws of physics

In more concrete terms, unlike the widely used General Circulation Models (GCMs) employed by meteorologists for weather predictions, ML-based models are typically limited to local forecasts and may sometimes violate the laws of physics. Only Superman could manage that. In statistical physics, random processes can be applied at micro-levels to explain broader variations and fluctuations of a system, but they must always adhere to the fundamental laws and principles of physics.

When we run out of people to infect with COVID-19

To provide a less abstract example from recent history, it wasn't uncommon for a couple of epidemiologists to extrapolate COVID-19 infection models beyond the total population sizes. This led to some individuals gaining overnight Twitter fame, which is likely to have polluted their objectivity and analysis.

Hurricanes are difficult to predict with standard weather prediction models

Another weather-related example: GCMs are highly effective at accurately predicting global weather patterns for up to 5-7 days, but they struggle with predicting hurricanes. The U.S. weather agency NOAA has a separate division dedicated to predicting tropical cyclones and hurricanes.

Predicting floods and landslides in Rio with ML

Leaving aside those nuisance laws of physics, a couple of years ago, a young atmospheric physicist at a university in Rio, Brazil, developed an ML early warning system to predict floods more accurately. It could alert local authorities to evacuate slum areas at high risk of landslides. She later received funding from the local government to advance her work, unlike most ML projects that fizzle out due to real-world impracticalities.

Rocket scientists not working at Morgan Stanley

In one of our recent updates, we wrote:

“One could easily mistake the folks at Morgan Stanley for running ensemble stochastic simulation models on non-linear complex systems for NASA. Even NASA scientists are forced to rethink their entire approach when their model predictions are so poor if they wish to remain employed.”

Coincidentally, today, a group of scientists at ETH (Einstein's university) in Switzerland published a paper in Nature on some research done along those lines (described above), which is beneficial to NOAA. The Washington Post wrote a simplified version of the paper.

Storylines for unprecedented heatwaves based on ensemble boosting
Nature | August 22, 2023

E. M. Fischer, U. Beyerle, L. Bloin-Wibe, C. Gessner, V. Humphrey, F. Lehner, A. G. Pendergrass, S. Sippel, J. Zeder & R. Knutti

Recent temperature extremes have shattered previously observed records, reaching intensities that were inconceivable before the events. Could the possibility of an event with such unprecedented intensity as the 2021 Pacific Northwest heatwave have been foreseen, based on climate model information available before the event? Could the scientific community have quantified its potential intensity based on the current generation of climate models? Here, we demonstrate how an ensemble boosting approach can be used to generate physically plausible storylines of a heatwave hotter than observed in the Pacific Northwest. We also show that heatwaves of much greater intensities than ever observed are possible in other locations like the Greater Chicago and Paris regions. In order to establish confidence in storylines of ‘black swan’-type events, different lines of evidence need to be combined along with process understanding to make this information robust and actionable for stakeholders.

[As a side note, Knutti is one of the top climate change commentators to follow on Twitter.]

What's ensemble & boosting?

The ETH group used a ML technique referred to as “boosting” (see reference on boosting technique—previously unfamiliar to me at least), which successfully predicted the 2021 heatwave in the US and Europe, unlike standard GCMs. In the context of this paper, “ensemble” refers to a group of stochastic (random) computer simulations where the output dynamics and statistical properties are analyzed.

What's special about the ETH study?

What's intriguing and distinctive about the ETH study is that the paper's authors adhered to physical laws and principles (i.e., the law of conservation of mass and energy that governs all GCMs) and still managed to predict extreme weather patterns. This opens the door for their methodology to be incorporated into existing weather prediction models.

Climate change deniers' interpretation

Climate change deniers may narrowly interpret the study to suggest that extreme weather patterns can be explained by endogenous factors. However, this fails to explain the initial conditions that lead to extreme heat waves in the first place. [Unfortunately, we have to shut them down early in every climate related update.]


Better predictions of extreme weather patterns within existing models 

The key takeaway from this study is the potential for improving local forecasts of extreme weather events worldwide within the current framework, as opposed to each region operating its own resource-intensive ML models, which is not only cumbersome but also environmentally unfriendly.

Subscribe to receive updates from Speevr Intelligence

Most recent by Speevr Intelligence

report

Share this page

Climate | Extreme weather predictions with machine learning

Researcher at ETH in Zurich find a novel approach to applying machine learning to improve local forecasts within existing weather prediction models which does