Google has a pretty darned informative article on using AI for weather forecasting. (Wired has a similar article, but it’s behind a paywall)
https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/
Right, wrong, or indifferent, I’m concerned that if we start getting our weather forecast directly from an AI, who do we blame when they are wrong?
Brian
SV Serenity
Nonsuch Nereus #003
Pax River MD
I think it was in the 1980s that similar articles were being written about the introduction of super-computers to weather forecasting. I think the US NWS bought one of the first CRAY supercomputers. Atmospheric systems are remarkably complex and forecasting requires the acquisition and analysis of vast amounts of data. If AI can do this better than current methods that is good.
I am a bit old-fashioned about weather, especially for longer-range forecasts. I have used things like Windy, and for our offshore passages, Passageweather, but these are just uninterpreted GRIB files presented graphically. The NWS and other met agencies go beyond this by having a skilled forecaster take the raw data and interpret it. As to who to blame, the forecaster’s name is on the chart, in the case below, J Lewitsky. Also useful are various private weather companies (and individuals), We have used several. The Peri-peri net in South Africa is run by 4 or 5 guys who help you avoid weather issues along a potentially very dangerous coast. There are several in/for the Caribbean.
A final thought or two. It is worth studying how the weather works in the region where you sail. Also, a barometer is very useful in many circumstances. We crewed for a couple from Antigua to the Azores a few years ago. This passage is all about dealing with the Bermuda-Azores high. You can’t follow the rhumb line unless you have the fuel to motor for days, there just is not much wind at all in the high. You need to go north and then east to keep some wind. Go too far north and it is much further and you may get mid-latitude cyclonic storms, i.e. nasty cold fronts. Go too much east and the wind drops. I suggested to our captain that he keep an eye on the barometer. If it was going up, edge a bit left; if it was dropping head right. I think I suggested we try to follow the 1012 mb isobar. He had a nice brass barometer on the wall but said he never looked at it. He downloaded GRIBs every few hours and spent much of the passage on his laptop.
Hopefully AI should make weather forecasting more accurate and more local. In my area Long Island I find the forecasts are often off because the local weather office produce forecasts over too large an area. They don’t take into account the effects of the local maritime area and put an emphasis on land areas to the west.
Mike
BIANKA
1986 30U
Long Island
One of my most venerable sailor friends… just sailed the Atlantic again… followed Windy for weather and was amazed that the recent Hurricane in Western Mexico never appeared on their system!
J Jenkins
Captain Haddock
Nonsuch 36, #4
We have switched to Predict Wind from Windy and Wind Finder, The data is just easier to suss through. But off shore you better understand GRIB forecast files and how to use them. A fellow named Lee Chesneau, a retired NWF forecaster, gave an exellent lesson on the 500mb chart and the use of a barometre at a seminar at the Chicago Boat show one year. Learned lot in 2 hours!
He still does lectures and if you can well worth trying to attend one.
https://www.youtube.com/watch?v=lFclHHBEQeM
One of the benefits of a conventional numerical weather forecast is that we run “ensembles”. Not only are we not certain of the future weather, we don’t really even know the weather now or 6 hours ago. So we take that uncertainty and tweak our weather model’s initial conditions and re-run it. We look for differences. Small differences imply greater uncertainty and vice versa.
In the Canadian model there are 20 separate ensemble runs. If you’re interested you can see how the various models agree/disagree for e.g. the 500mb pressure level chart that @Thor mentioned: https://weather.gc.ca/ensemble/index_e.html
Currently, one of the limitations of the Google’s GraphCast is that it is “deterministic”; that is, it current emits a single simulation. Which is not as good at getting a sense of the probability of extreme events or what happens in a week+. But the Google researchers are up front with that limitation in the Science paper that describes the methodology and consider it an important next step in the process. Machine learning, in some form, either aiding or replacing conventional NWPs, is likely the future. Science paper here: https://www.science.org/doi/10.1126/science.adi2336
PredictWind is excellent as it provides forecasts from multiple different models. This is sort of the same idea as running with different initial conditions. So where the models agree, you can have greater confidence of it happening. The 1km models are good for showing the wind speedups over water vs land but that higher resolution doesn’t necessarily imply accuracy.
Another one I like is “Flowx” as it has a buttery smooth interface that allows you to swipe back and forth to see e.g. precipitation, wind. The free offering allows you to pick between three models: GFS (US), CMC(Canada), ECMWF (Europe). The ECMWF model is coarsest and doesn’t show wind gusts but is excellent for accuracy.
¯_(ツ)_/¯
Michael
“Grande Dame”
NS 36
I have certain fascination with the work of John van Neumannn the area of probability theory. But I did not know that his wife was a major player in modern weather forecasting…
