AI is promising to transform weather forecasting

Well…..

I won’t hold my breath. I’ve seen some of this stuff and while it will “improve” certain aspects of automated forecasts (that is make them even more convincing) I doubt that they will actually improve the validity.

Currently many managers and bean counters are VERY hyped about AI being able for them to sack unwanted humans but they are opening a pandoras box which they are not qualified to quantify.

My kind of wet dream about AI is that someone is going to replace his HR department with some AI thingy which then is gonna tell him “I can’t let you do that Dave” when being told to fire department after department to be replaced by AI drones…(aka in 2001. Maybe some AI applications could be more immune to greed and the “Fawlty Towers Syndrome” than humans… but then again…

Airborne_Again wrote:

AI is promising to transform weather forecasting

Weather forecasting is already insanely good. What it isn’t good at is on the local level and shorter time spans, which in Norway with it’s geography is the main problem. Will AI be able to forecast burst of severe turbulence coming from 12:00 to 12:15 at ENVA the next day? Not even remotely possible due to the nature of chaotic systems. Will AI be able to at least better define the probability of burst of severe turbulence from 12:00 to 12:15 at ENVA? perhaps, but how is a human being for practical purposes going to interpret the difference from say prob 40 and prob 60 ?

I also think that AI is able to make weather forecast more convincing perhaps (in the same way Chat GPT is sort of convincing on the face of it), but not necessarily improve the underlying validity.

To think that AI is able to solve the underlying problem of complexity is far fetched when thinking how Chat GPT actually operates (but people keep on dreaming). Complexity is the least understood of everything in science and human knowledge. It’s a big black hole of knowledge and mathematics and philosophy, as is excellently explain here:

AI is a lot more than just ChatGPT. The latter is a pretty astounding technical achievement, but that doesn’t necessarily make it useful.

AI is all about pattern recognition, and has some serious achievements – the protein folding predictor for example. Meteorology seems like an excellent fit – there is a huge corpus of 100% accurate data. Feed the model every three consecutive days’ data for the last 20 years, training it to predict the third day given the previous two days.

In the US every weather office publishes four times a daily a human-written summary of how they arrived at the forecast. (I wish there was something like it in Europe). It makes for fascinating reading. It is full of subjective, experience-based judgements as to which of the various models to follow. That is the kind of thing that AI can do pretty well.

AI may be able to improve the forecasting model, but any model is dependent on input data. In order to achieve a major improvement in the output, one needs to improve the input as well – that is, create a bigger, denser observation network.

AI is all about pattern recognition, and has some serious achievements – the protein folding predictor for example.

AFAIK the protein folding has never been solved. It’s a complete and utter mystery. But perhaps you think of something else?

Pattern recognition only works so far, and is certainly no substitute for physics calculations. Sometimes the pattern itself is of the essence though, and in those cases AI certainly has merit. Billions and billions are put into AI, something ought to come out of it.

Just for the record. About 25 years ago I wrote a pattern recognition software based on Kohonen maps. It’s a general self learning algorithm useful where “pattern” is anything that can somehow be described using multiple vectors, like a computer representation of an image for instance. The technology has advanced multiple times since then though

Protein folding:

(I wish there was something like it in Europe). 

DWD does that. it is called “synoptic overview” and can be found here

the whole site for “hobby meteorologists” (which is used by professionals as well and the base for a lot of private suppliers) is here

[ long URLs changed to clickable links – see Posting Tips ]

DWD is charged for, mostly?

And I don’t think it is any better than ECMWF, which btw is available for free on windy.com which is now the standard wx briefing tool.

The “Threads possibly related to this one” below are worth a read, plus this one.

“AI” (neural learning) is increasingly used to examine stuff like MRI scans and it achieves good results, so it may well work for wx patterns. But that isn’t “AI” in the classic meaning. It is a heavily specialised pattern recognition task.