Welcome And Topic Preview

SPEAKER_00 0:00

Hi, this is Meteor Roger Steve Pelletier, and I am the weatherman. Here's a special discussion on weather and neural networks and artificial intelligence. Hope you enjoyed. A lot of great information here. For quite a long time now, I have been, or we actually have been very blessed to have Jeff Morrison send us his reports from uh at first it was Persiphone, New Jersey, and now Somerset. And uh Jeff has given us excellent detailed weather information that has been occurring in those places. Pretty much we will look at uh what happens in Morristown, in Newark, New York City, Philadelphia, and we compare all these numbers to come up with uh a pretty good picture of what the climate has been in the past month. We do this just about each and every month. And each and every month, Jeff shares that with us, Jeff Morrison. You've heard me talk uh about him uh quite some time, uh quite a few times in the past. And uh Jeff has shared some really excellent information about uh the uh weather prediction models that we use for weather forecasting and have been using for many, many years as well. So Jeff writes a few months ago we talked about how weather forecasts are prepared, and we discussed two more widely referenced numerical weather prediction models. One is called the GFS, the U.S. Global Forecast System, and the Euro, or the European Center for Medium Range Weather Forecast, the ECMWF. Now, both are large physics-based computer models which run on supercomputers and divide the atmosphere into three-dimensional grids. At each point, the models calculate variables like temperature, humidity, pressure, and wind, and then using complex physics equations, equations of motion, predict how these variables will evolve over time. Those time steps go from just minutes and go all the way out to hours and then days. Even with supercomputers, though, the models take hours to run, and small uncertainties in the initial atmospheric input, we call that the initialization, can quickly grow over time, garbage in, garbage out, impacting forecast's accuracy several days in the future. It starts out usually as some pretty good model uh output, but then if it starts off with some errors, the errors just become larger and larger over time. While these and other models remain the foundation of modern weather forecasting, artificial intelligence, AI systems are increasingly complementing these numerical models and rivaling them in their accuracy. Now, AI-based forecast models approach the problem differently. Instead of solving equations step by step, AI models learn patterns from historical data. Researchers train AI networks on decades of past weather observations and model outputs. Now, during this training, the AI system analyzes millions of examples of how atmospheric conditions evolve from one step to the next. The AI weather model learns to predict the next atmospheric state given the current one. So it does that comparison. And over time, the AI network begins to identify its own version of the complex relationships among these variables, like the jet stream position, ocean temperatures, and storm development. Now, once trained, the AI model will have the ability to generate weather forecasts very quickly in minutes versus hours, applying what it has learned during that training process. There are an increasing number of AI weather models being developed right now. Three of the most talked about are the GraphCast from the Google DeepMine, Forecast Net3 from Nvidia, and the third one is Aurora from Microsoft. Google's GraphCast is being developed in collaboration with the ECMWF, what we call the Euro. It uses a machine learning architecture called a Graph Neural Network, which represents the atmosphere as interconnected nodes. In tests with the ECMWF or the Euro, GraphCast has matched or exceeded the accuracy of the numerical models and can generate a global 10-day forecast in mere minutes on specialized hardware. Similarly, Nvidia is working with leading universities like Caltech and UC Berkeley to develop its AI weather model. ForecastNet, which is now in version 3, ForecastNet can predict global weather patterns in a fraction of the time, making it potentially useful for tracking fast developing weather phenomena like hurricanes, tornadoes, and flash floods. Now, a newer model attraction attention in Microsoft's Aurora being designed as a foundation model for atmospheric prediction, meaning it can be adapted to multiple weather and climate tasks all at the same time. Instead of being optimized for a single forecast product, say the 10-day forecast, it can be fine-tuned for predicting specific extreme weather phenomena like atmospheric rivers or the rapid change in a dangerous storm. The flexible design reflects the growing trend in AI towards large general weather models that can perform many related tasks. AI weather forecast offers a number of key advantages over the traditional numeric models. Number one, speed is obvious, with forecasts generated in seconds or minutes versus hours. So speed is number one. Number two, higher efficiency in that there is no need for large supercommunity capacity. Three, pattern recognition because these AI networks excel in finding patterns across vast amounts of data that may be difficult to model with traditional physical equations. I look at numbers and I try to find these patterns all the time. Neural nets do that in seconds. On the other hand, the models are only as good as the historical data used for training, and as climate impacts may change, past weather may become less reliable guides for future prediction. In addition, traditional models are based on physical laws, allowing meteorologists and scientists to understand why they produce the result that they do. AI models are very much like a sort of black box, making it possibly harder to diagnose potential errors. In the future, AI will not replace traditional meteorology. Rather, it will enhance it. Human expertise is still essential in interpreting the AI outputs, understanding local weather patterns, and effectively communicating forecasts to the public. As AI continues to evolve, it is making weather prediction faster, more reliable, and more accessible than ever, helping the public to better prepare for everyday weather conditions and those more impactful and extreme weather events. Well, we thank Jeff Morrison for that uh discussion about AI. I actually have been working with uh AI neural networks since the late 1990s. It was archaic, it was difficult, it was frustrating, it continues to evolve, and then actually it hasn't really come to its uh fruition, or shall I say, its successes until over the last few years. And that's because of the uh ability to put these neural nets together in order to learn. Neural nets learn patterns, and patterns are the key to everything. But like I said a little bit earlier, and this is my opinion, Steve Pelletier, that garbage in, garbage out. Jeff alluded to that also. If the data that we're using is bad, then the data that's coming out is also going to be flawed. So that's something that we're going to be looking for. We thank Jeff Morrison for that excellent representation and discussion about uh neural networks and AI in uh numerical weather prediction. And we look forward to his latest reports. So have a great evening, have a great day. We'll talk to your first thing real soon.