Whatever the outcome of COP 26 in Glasgow (AKA ‘the last chance saloon’), we will be living with wilder weather, often ravaging localized areas with little warning (e.g. that ice storm in sub-tropical Coffs Harbour several weeks ago). We need more reliable, real time weather forecasting tools to help mitigate the economic, social and human cost of these sudden, extreme swings in our weather.

Meteorologists for some time have used supercomputers to forecast weather, but they typically are reliable for forecasts over the next day or week. DeepMind, working with the UK Met Office, has now developed an AI-based ‘nowcasting’ program which pinpoints the timing, location and intensity of precipitation at high resolution up to two hours ahead.

This nowcasting program is one of the first practical applications of a new model of AI ‘thinking’ called a deep generative model (DGM).

What is DGM and why could it represent a step change for AI?

DGM is described as follows:

“Deep generative models (DGM) are neural networks with many hidden layers trained to approximate complicated, high-dimensional probability distributions using a large number of samples. When trained successfully, we can use the DGMs to estimate the likelihood of each observation and to create new samples from the underlying distribution. Developing DGMs has become one of the most hotly researched fields in artificial intelligence in recent years.”

None the wiser? Generative AI models are easier to understand when compared with the predominant way AI currently ‘thinks’ called ‘discriminative models’. A discriminative AI model can tell the difference between a dog and a cat, while a generative AI model can not only do that but also generate a highly realistic picture of a cat which does not exist: i.e. a ‘deep fake’.

As the vast lakes of training data are sucked in by a discriminative AI model, it will label the data (cat or dog). This is called supervised learning by AI. If you then show the AI a photo of a cat, it makes an assessment based on the probability that the image in the photo belongs to the category labelled cats than the other labelled category called dogs. It does this based on looking for a few telltale signs it has leant from the labelling process.

A generative AI model has to work harder in identifying the correlations in the learning data set, down to the pixel level. The goal is to build an AI model that can generate new sets of features that look as if they have been created using the same rules as the original data on which the AI was trained. The generative model does this by describing how a dataset is generated in terms of a probabilistic model and by sampling from this model, is able to generate new 'fake real' data.

Generative Adversial Networks (GANs) are the form of generative AI about which the AI world is most excited. A GAN consists of two engines: a generator and a discriminator. Essentially they play a game trying to checkmate each other. The generator produces an image which has to look as close as the generator can get to the real-world data it has studied. The discriminator then looks at the image and has to discriminate it from real-world data: i.e. say whether it is a fake. If the discriminator accurately identifies the fake, the generator knows it has to try harder in generating the next new image and if the discriminator cannot determine whether that next image is real or fact, the discriminator knows it has to sharpen its analysis…and so on. This is an illustration of why generative AI is called unsupervised learning by AI.

So in short, you can teach a discriminative AI to tell the difference between a real and a fake Van Gough but it can’t paint you a Van Gough of its own. A generative AI on the other hand can do just that – in fact so well that a discriminative AI may not be able to identify the Van Gough as a fake. 

A fun thing demonstrating the power of generative AIs is classic art portraits such as the Mona Lisa translated into ‘living portrait’ gifs.

Back to the weather

The problem the DeepMind developers set out to solve is that “smaller-scale weather phenomena are inherently difficult to predict due to underlying stochasticity” (i.e. haphazardness, irregularity). As generative AI models are fundamentally probabilistic at fine levels of correlations, they are well suit to grappling with this problem – or in the denser language of the DeepMind developers:

“By training these models on large corpora of radar observations rather than relying on in-built physical assumptions, deep learning methods aim to better model traditionally difficult non-linear precipitation phenomena, such as convective initiation and heavy precipitation. This class of methods directly predicts precipitation rates at each grid location, and models have been developed for both deterministic and probabilistic forecasts. As a result of their direct optimization and fewer inductive biases, the forecast quality of deep learning methods…has greatly improved.”

The reliability of the generative AI in predicting medium and heavy rain against the actual rainfall in part of Scotland was compared with the reliability of traditional short term forecasting methods. These traditional methods plug radar data into physical equations of the atmosphere, called numerical weather prediction (NWP) systems. The case study found that “deep learning systems produce forecasts that are e significantly more location-accurate than the [NWP systems].”

DeepMind also compared the ‘economic value’ of the generative AI models and the NWP systems. Economic value meant the value to “[o]operational meteorologists [who] seek utility in forecasts for critical events, safety and planning guidance.” For example, the BOM sending out urgent weather alerts on mobile phones or on the BOM app.

In a blind test between outputs generated by AI and NWP systems, the output from the generative nowcasting approach was preferred by 89% meteorologists when asked to make judgments of accuracy and value of the nowcast. Importantly, the AI output was identified as providing meteorologists with physical insight not provided by alternative methods. Meteorologists described DGMR as having the “best envelope”, “representing the risk best”, as having “much higher detail compared to what [expert meteorologists] are used to at the moment”, and as capturing “both the size of convection cells and intensity the best”.

But it’s still a forecast

DeepMind concluded that its generative AI model “provides improved forecast quality, forecast consistency and forecast value, providing fast and accurate short-term predictions at lead times where existing methods struggle.” But they also acknowledged that forecasting still has its shortcomings: "the prediction of heavy precipitation at long lead times remains difficult for all approaches.”

Beyond the weather, we are just beginning to understand the legitimate potential applications of generative AI. Another DeepMind project has used generative AI as a solution to the 50-year old protein folding problem, allowing high quality prediction for the shape of every single protein in the human body, as well as for the proteins of 20 additional organisms that scientists rely on for their research. 

Read more: Skillful Precipitation Nowcasting using Deep Generative Models of Radar