AI weather forecaster outperforms meteorologists

But first, many have expressed alarm that their jobs might be threatened in future by artificial intelligence. And now weather forecasters might be among them, because Google’s DeepMind have announced that a new type of AI, known as GraphCast, can predict medium term weather conditions, and they claim it ‘outperformed traditional systems in 90% of tested cases’. At the moment, weather supercomputers run simulations powered by physics equations to crunch variables like temperature and pressure to come up with a forecast. But this is slow and very energy hungry. GraphCast uses 40 years of historical data combined with machine learning to make the same predictions but in a matter of minutes…

Matthew Chantry - What we've seen with GraphCast is a significant step forward in the abilities of these models. They now have, in our eyes, we've done some assessments, and have some legitimate claims to have equalled and sometimes outperformed the physical models for deterministic forecasts. And they've shown this not only for scores, but also for some of the most impactful use cases of weather forecasts such as tropical cyclone track predictions. We at ECMWF view this as a hugely exciting technology to lower the energy costs of making forecasts, but also potentially improve them. There's probably more work to be done to create reliable operational products, but this is likely the beginning of a revolution. This is our assessment in how weather forecasts are created.

Matthew Chantry from the European Centre for Medium-Range Weather Forecasts on the impact of GraphCast.

But what does this mean in practice? And how does it affect those that actually use the data to create weather forecasts? I’ve been speaking to the founder of the British Weather Services and author of ‘Weather or Not?’, Jim Dale...

Jim - Let me take you back to about 40 years ago when I first started in this career. I was in the Royal Navy at the time, and meteorology at that time was fairly rudimentary. It was advancing, it always does, but prognosis charts as we call them were very limited. Often it was a bit of a push/shove from the forecaster more than anything, from a supercomputer. If it was a supercomputer, it wasn't so super at that time, if that makes sense. In other words, it was evolving, but it was very limited. It didn't go much beyond 48-72 hours. Over the period that then followed, bit by bit and little by little, yes, the supercomputers came in. They always seem to get a little bit more super as time goes on. Technology just gets that little bit better. The number crunching gets that little bit more. Any advancement in these things is always a welcome one, but the proof is in the pudding: the tasting of it. We'll wait and see.

Will - With current supercomputers, what information do they provide you with to inform your decisions?

Jim - You can look at any part of the atmosphere going from surface right up to the troposphere, the top lid of the weather, and they will suggest movements in pressure, air pressure, temperature profile, in wind and wind speed direction. Some of these computers now do go two/three weeks ahead. Sounds a bit silly, and I'll openly say to you, never rely on the third week or even sometimes the second week. You've got to compare and contrast these various computers. And there are various ones coming out of various countries, for example. There's the American model, the British model, there'll be the French, then the Japanese and the German. They don't always sing from the same hymn sheet. One of the skills of forecasters is to put that all together. So you interpret what you see and you convey that message. When you convey it, it is in the most simplistic terms in the English language or in numerals, suggesting how much snow or how much rain.

Will - AI is sweeping through the scientific community and meteorology is no exception. With these new AI that can take so much more data about precipitation and pressure and more accurately calculate future weather, have they changed how this process works?

Jim - They have, and they will continue to do so. Artificial intelligence is something that will continue to move in step with meteorology and climatology without a question. Nothing's perfect though, even super duper computers and the advancement of AI does not necessarily mean that you end up with the right result - what actually happens. Meteorology and forecasting is not an exact science. It never has been and probably will never be. Why? Because the atmosphere is extremely chaotic, unpredictable. Certainly you get global phenomena like El Ninos and La Ninas, mixing with climate change, mixing with differing seasons. I mention climate change for a good reason because that introduces that word chaos to a certain degree. There was hurricane Otis which occurred off Mexico a few weeks ago. 24 hours before the advent of the hurricane, it looked fairly innocuous. A tropical storm, but nothing special. And within 24/36 hours it was hurricane force five. This unpredictability will stay there and AI has to get older and help the humans on this side to provide the correct messages and to interpret properly to ensure we get as best possible accuracy. But there will always be a human element.

Will - There's been much concern in all sorts of job sectors and business sectors that AI is going to drive out the role of people. Is there a concern that the role of presenting the weather is being phased out by digital systems?

Jim - Like in every other walk of life, it's a possibility. I'm sure it will be utilised online and offline, on TV and various other places. So there is some concern, if that's the right word. The weather is one of those components of our lives that is very personal. We like to chat, we like to have anecdotal things that perhaps I haven't thought about, if that makes sense. In other words, you add things in little bits and pieces that actually make your life go round on that particular day. The weather affects us in so many different ways. I don't think AI is going to take that out of it. In terms of pulling that personal feeling into the weather around us and what that actually means to us, I think when you transpose forecasts, put them out there and say this is what's coming and, by the way, this is what it means to you, and particularly climate change in some of the extreme events, then I think that's the next step for humans to take. Nevermind AI. I'm sure AI will catch up, but I think humans have got a leader here and that personal approach is something that I would advocate for any meteorologist out there.