Latest / The Climate Translation / The Predictability Desert

Transcript

The Predictability Desert · The Climate Translation

0:00
All of this at one point have looked up a weather forecast on our phone or computer to plan
0:13
our days.
0:14
We can even get seasonal forecasts to determine the trends over the next few months, but if
0:20
you have a wedding to attend to three weeks from now and you want to know whether it will
0:24
rain, well suddenly the forecast becomes much less useful.
0:29
Meteorologists have a name for this awkward gap between short-term forecasts and seasonal
0:36
climate outlooks.
0:37
It's called the predictability desert.
0:40
Today we're going to explore why weather becomes so difficult to predict as we look farther
0:45
into the future and how scientists may be getting better at navigating one of meteorology's
0:51
most stubborn blind spots through the use of artificial intelligence.
0:57
Welcome back to the podcast, I'm Dr. Mac and this is the Climate Translation.
1:09
Let's start off this episode by looking at how a weather forecast is made.
1:14
To predict the weather, meteorologists begin by trying to describe what the atmosphere
1:20
looks like right now.
1:22
We use surface measurements like temperature and air pressure, humidity, wind speed and
1:27
direction.
1:28
We utilize satellites to observe clouds and moisture from space.
1:33
We use weather balloons to sample the atmosphere thousands of feet above the ground and aircraft,
1:41
ships, buoys, radar stations and other instruments all contribute additional information.
1:49
Everything is fed into enormous computer models that divide the atmosphere into millions
1:55
of small boxes and inside each box, the model uses the laws of physics to calculate
2:02
how the air, the heat, the moisture and the momentum will change over time.
2:08
And then it moves forward a few minutes, an hour and a day.
2:13
That becomes our forecast.
2:14
The problem is, we can never measure the atmosphere perfectly.
2:20
Imagine that the actual air temperature at one location is 68.27 degrees Fahrenheit.
2:27
But our analysis of the atmosphere estimates it at about 68.2.
2:32
Or perhaps there's no instrument at that exact location so the computer had to estimate
2:38
what the temperature was using nearby observations.
2:42
Those tiny differences may mean almost nothing in regards to the forecast or today or tomorrow.
2:49
But the atmosphere is what scientists call a chaotic system.
2:53
And in a chaotic system, very small differences at the beginning grow into larger ones as time
3:00
passes.
3:02
You may have heard this described as the butterfly effect.
3:05
The hypothetical scenario that if a butterfly flaps its wings in Brazil, it sets off a chain
3:11
of events that causes a tornado in Texas.
3:14
The idea is literally not that a butterfly creates the tornado.
3:19
It means that when a system is extremely complicated, tiny changes can eventually lead
3:25
to very different outcomes.
3:28
Imagine two weather models, starting with almost identical atmospheres.
3:33
And one, a small disturbance over the ocean is slightly stronger.
3:38
At first the forecast looked identical, but a few days later that tiny difference may
3:43
change where a storm develops.
3:46
That changes the winds around the storm, and those winds affect the weather system farther
3:51
downstream.
3:52
Eventually, the two forecasts may show completely different weather over your hometown.
3:59
Those small differences grow and interact, and the farther forward a model runs, the
4:04
more opportunities those differences have to grow.
4:08
This is why a forecast does not simply become bad on some magical day.
4:13
Uncertainty just gradually increases.
4:16
Tomorrow's forecast starts with an atmosphere that looks very much like the one we can observe
4:21
today.
4:22
A week from now, the model has already passed through thousands, maybe even millions of small interactions.
4:29
Several weeks from now, many different versions of the atmosphere may all be physically possible.
4:35
Modern forecasting has pushed that useful window pretty far.
4:40
Today's best models can provide valuable guidance at about one to two weeks.
4:46
But eventually you reach a point where knowing today's exact weather becomes less useful
4:51
for determining exactly what will happen on a particular day in the future.
4:56
It doesn't mean the atmosphere becomes unpredictable.
5:00
But it does mean we have to stop asking, will it rain at my house at 3 o'clock on Tuesday,
5:05
three weeks from now?
5:07
And instead ask, are the odds of a wetter than normal week increasing?
5:12
And that change from predicting one exact future to estimating the range of possible
5:18
futures is how meteorologists begin crossing into that predictability desert.
5:25
So if one exact forecast becomes unreliable after about a week or two, what do meteorologists
5:30
do?
5:32
Well, you stop relying on just one forecast.
5:35
Instead, you run the atmosphere over again and again and again.
5:40
It's called ensemble forecasting.
5:43
Imagine you're standing at the top of a hill with a bucket full of tennis balls,
5:47
and you pour them out at about the same place.
5:51
At first they all roll together, but as they move downhill,
5:54
each ball hits tiny bumps and rocks and dips and begin drifting left and right.
6:00
Eventually the group spreads out.
6:03
Whether models work the same way.
6:06
Scientists can run the same model many times, but make tiny changes to the starting conditions.
6:13
Maybe the temperature is adjusted, maybe the wind's a little stronger, maybe the humidity
6:17
is lower.
6:18
Each version represents a slightly different, but still realistic picture of what the atmosphere
6:24
might look like right now.
6:27
And then the computer lets all those versions move forward.
6:30
For the first few days, the forecast may stay tightly grouped, and that gives meteorologists
6:36
more confidence.
6:37
But farther into the future, the solutions might spread out.
6:41
One version may develop a storm over the eastern United States.
6:46
Another may keep the storm farther north.
6:48
Another may not develop it at all.
6:50
That spread is useful information.
6:52
It tells us how confident we should be.
6:56
If 45 out of 50 model runs show unusually cold weather three weeks from now, it doesn't
7:02
mean cold weather is guaranteed, but it does tell us the atmosphere may be leaning in
7:07
that direction.
7:09
And this is where that predictability desert becomes less empty.
7:13
Even if we lose the ability to predict the exact weather on a specific day, larger patterns
7:20
provide clues.
7:22
The atmosphere does not operate on its own.
7:25
It interacts with the ocean, with snow cover, with soil moisture, and sea ice, and with large
7:31
climate patterns like El Niño and La Niña.
7:34
Those features usually change much more slowly than the individual weather systems.
7:40
Think of all that as background conditions.
7:42
A single thunderstorm may be impossible to predict three weeks in advance.
7:47
But if the Pacific Ocean is unusually warm, or a large atmospheric circulation pattern
7:52
has become established, that can tilt the odds towards certain kinds of weather.
7:59
That's why sub-seasonal forecasts usually talk in probabilities, warmer than normal, cooler
8:05
than normal, weather than normal, drier than normal.
8:08
We don't say it's going to rain at 2:17 pm next Thursday.
8:13
That kind of detail matters.
8:15
A farmer doesn't really need to know whether all rain at noon on a particular day.
8:20
They may care more about the weather in the next three weeks.
8:24
Is it going to be unusually dry or unusually wet?
8:27
An energy company may not need to know the exact temperature four weeks from now, but knowing
8:32
the risk of an extended cold period will help them prepare.
8:37
So the predictability desert is not completely unpredictable.
8:41
It's more like trying to see through the fog.
8:44
You can't make out the details in the distance, but sometimes you can see the shape of what's
8:49
coming ahead.
8:51
And new research is taking those ideas one step farther.
8:55
Instead of only asking what the forecast says, researchers are asking what mistakes does
9:01
the forecast tend to make.
9:03
So how can artificial intelligence help us here?
9:07
It doesn't eliminate the natural chaos that defines our atmosphere, and it doesn't attempt
9:12
to calculate exactly what the atmosphere is going to be doing four weeks from now.
9:17
Instead, researchers are taking a different approach.
9:20
They're teaching the forecast to recognize its own mistakes.
9:25
In October of 2026, a team led by Hannah Guan at Microsoft Research and Harvard University,
9:33
and Soukayna Mouatadid at the University of Toronto published a study in Nature, introducing
9:40
a system they called probabilistic bias correction.
9:44
The phrase sounds complicated, but the basic idea isn't.
9:48
Imagine you have a friend who's almost always five minutes late.
9:52
If the friend tells you they're going to be there at 6 o'clock, your experience knows
9:57
it's going to be 6:05.
10:00
You're not predicting the future perfectly, you just simply learned your friend has a
10:04
pattern.
10:05
Weather models have patterns too.
10:08
A model may consistently make the same kind of errors in a particular place or during a
10:13
particular time of year, or several weeks into our forecast.
10:17
The new system looks at years of old forecasts and compares them to what actually happened.
10:24
Then it begins to learn.
10:26
It learns where the model tended to be too warm, too cold, too wet, or too dry.
10:31
But because these are probabilistic forecasts, the system does more than simply add or subtract
10:37
a few degrees.
10:39
Remember that ensemble model?
10:41
Maybe 50 different model runs produced a range of possible temperatures.
10:46
The AI can adjust the entire range of probabilities based on what it has learned from past forecasts.
10:54
The researchers tested this system using forecasts from the European Center for Medium-Range
11:00
Weather Forecasts.
11:01
One of the world's leading weather forecasting centers.
11:05
They looked at weekly average temperature, precipitation, and sea level pressure during
11:10
weeks three and four of a forecast, and the improvements were substantial.
11:16
Compared to the European model's existing correction methods, the machine learning system
11:21
improved forecast across more than 90% of the locations tested for temperature and pressure,
11:27
and about 98% for precipitation.
11:30
When applied to the European model's own artificial intelligence model, it roughly doubled
11:36
the forecasting skill at these long-term ranges.
11:41
And then the researchers tested it in the real world.
11:44
During an internatoinal forecasting competition run by the European model, in 2025, the system
11:50
using this technique over-performed forecast from six major government weather centers, several
11:57
other AI models, and dozens of competing teams.
12:01
One example is especially interesting.
12:04
In November of 2025, the system identified an increasing risk of an extreme cold outbreak
12:11
across the Eastern United States about four weeks before it happened.
12:16
The traditionally corrected European forecast missed that signal.
12:20
That doesn't mean that AI knew how cold Atlanta or New York would be on a particular afternoon,
12:27
but it did recognize that the probability of an unusually cold week was becoming important
12:33
much earlier than the traditional model did.
12:36
In effect, artificial intelligence didn't discover a way around the butterfly effect.
12:42
Instead, it's getting better at separating useful signals from some of the predictable
12:46
errors that are buried inside our forecast.
12:51
So what would actually change if we became better at seeing three to four weeks into the
12:55
future?
12:57
It turns out quite a lot.
13:00
Think about agriculture.
13:02
Farmers would have the ability to know when hot or dry conditions are coming and plan
13:06
their irrigation, planting, or harvesting decisions.
13:10
It would also benefit water managers.
13:13
Reservoirs have to balance storing water for later with leaving enough room for heavy
13:17
rainfall and flooding.
13:19
A more reliable signal several weeks ahead could give managers more time to prepare.
13:25
Energy companies care too.
13:27
If the probability of a major cold outbreak begins increasing a month in advance, utilities
13:32
can prepare for higher heating demand.
13:35
During the summer, an elevated risk of extreme heat can provide similar warnings about electricity
13:40
demand.
13:42
And emergency managers may gain additional time before periods favorable for flooding,
13:47
wildfire, or other weather hazards.
13:49
That's why extending useful forecasts doesn't require predicting the exact weather.
13:55
Sometimes just knowing where the odds are shifting is enough to change a decision.
14:00
But there's another reason why this research is so important.
14:04
Every time scientists improve a weather model, they learn something about how well that
14:08
model represents the Earth's system.
14:11
Where does it consistently run to warm?
14:14
Where is it producing too much rain?
14:16
Where is it struggling with interactions between our oceans and our atmosphere?
14:21
And how is it handling the soil moisture and the snow or sea ice?
14:26
Those questions sound a lot like what climate scientists are asking about our climate models
14:32
to predict how climate change is going to affect our Earth.
14:36
Weather models and climate models are not the same thing.
14:39
A weather forecast is trying to predict the actual sequence of events beginning from today's
14:44
atmosphere.
14:45
A climate projection asks a different question.
14:48
If you change the greenhouse gases, the land use, or the conditions, what kind of weather
14:53
becomes more or less common over decades?
14:57
You can't take a weather model that works well at four weeks and keep running it until
15:01
the year 2100.
15:02
Chaos would make that impossible.
15:05
But the two fields share many of the same challenges.
15:08
They both have systematic biases.
15:11
Both have to represent complicated interactions among the atmosphere, the ocean, the land,
15:17
and the ice.
15:18
And both rely heavily on ensembles rather than predicting one model run that represents
15:24
the only possible future.
15:27
In fact, scientists are already using some of the same broad ideas in climate research.
15:32
A study published just days after this new sub-seasonal forecasting paper created an ensemble
15:39
of more than 1,400 climate projections.
15:43
The researchers used bias correction and multiple model runs to better understand the range
15:49
of possible futures for temperature and the uncertainty around them.
15:54
So improvements and weather prediction can become another testing ground.
15:59
If a model repeatedly fails to reproduce something we can observe in the next few weeks
16:03
or years, scientists can investigate why.
16:07
Better observation improves the starting data.
16:10
Better physics improves the models.
16:13
And machine learning, or AI, can help identify biases that humans did not recognize easily.
16:21
None of that eliminates uncertainty from our climate projections, but it can help us
16:25
understand where that uncertainty comes from and improve the tools that we use to estimate
16:31
future risk.
16:32
And that might be the larger lesson of this predictability desert.
16:36
Our goal is not to make uncertainty disappear.
16:39
It's to understand it well enough so that our forecast becomes useful.
16:44
Whether we're looking at four weeks ahead or several decades into an uncertain future.
16:58
Weather will probably never be perfectly predictable weeks in advance.
17:03
The atmosphere is too complicated and the differences grow too quickly.
17:07
But it doesn't mean that the predictability desert has to remain empty.
17:11
With better models, larger ensembles and artificial intelligence, scientists are finding
17:16
useful signals farther into the future.
17:20
And in a complex world, those few extra weeks of warning can be enough to make a better
17:25
decision.
17:26
I'm Dr. Mac.
17:28
And this has been the climate translation.
17:31
If you have a question about the climate you've been too afraid to ask or if you have
17:35
a differing opinion, I want to hear from you.
17:37
I can use your viewpoints and future episodes.
17:41
You can reach me at theclimatetranslation at gmail.com.
17:45
I'll see you next time.