The ocean is changing due to weather forecasts. Known for its complex, labor-intensive processes relying on satellites, sensors and supercomputers, the industry is being transformed to enable more accurate and timely predictions than ever before.
Met Office is at the forefront of this shift, blending AI with traditional physics-based models to provide richer insights, greater computational efficiency, and time lier predictions. Kirstine Dale, the organization's chief AI officer, describes the shift as a “noisy revolution,” with machine learning models running tens of thousands of times faster than traditional methods today.
But according to Dale, AI is not a silver bullet. And human input is still necessary to interpret unprecedented events, examine results, and maintain public trust that has been built over decades. In this interview, compiled for length and clarity, Dale explains how Met Office integrates AI into its operations, the opportunities it offers, the need for human opinion in the technological revolution.
How does Met Office work with AI and machine learning?
Met Office embeds AI into end-to-end capabilities, helping to realize the value of this groundbreaking technology in terms of weather and climate science and service.
Over the past few years, we have integrated AI, especially across the forecasting chain. For example, we worked with DeepMind to create a deep learning system that turns radar data into highly accurate short-term rainfall forecasts. It also uses machine learning to develop “emulators” to replace some of the most computationally expensive physics in weather models and improve site-specific temperature prediction for bias correction.
It wasn't called AI at the time, but for a very long time it has used Gaussian processes for downscaling, clustering, cloud classification, and more. That means data science is nothing new for Met Office. We haven't come to AI from scratch, so we were able to accelerate that development quite quickly.
Why is AI becoming such a game changer for weather forecasting?
Met Office is essentially a big data organization. Weather observations are collected around clocks around the world, combining measurements of important variables with satellite images to depict what is happening in the atmosphere.
Approximately 21.5 billion observations appear through the organization every day. It runs 3M line code and produces 18 terabytes of data every day. This is an incredible amount of data flowing through an organization. So we found that when the AI revolution began, and how we could leverage the tools and techniques that emerged from the AI revolution to extract value and meaning from the dataset.
What makes AI different is mainly in terms of accuracy, or does it have other benefits?
Accuracy is definitely a part of it. Machine learning models actually began to become more competitive at the end of 2022. Now we're matching (and sometimes surpassing) physics-based approaches when it comes to accuracy. But that's not the only advantage. Another real game changer is speed.
These models are not only a little faster, but they can be tens of thousands of times faster than traditional physics-based models. That's important for two reasons. First of all, they run so fast that they can generate predictions that are much closer to the moment you need them. Secondly, they are computationally light. Training models is still expensive, but once training is done, the inference phase (prediction execution) requires relatively little calculations. This reduces the environmental footprint of modeling and opens up the possibility of running predictions on small systems. In the future, you can imagine predictions generated by laptops and smartphones. This will greatly improve accessibility.
To put this in context, the accuracy of the predictions has been historically improved about one day per decade. This is often referred to as a “quiet revolution.” AI can dramatically accelerate that curve. It could lead to real step changes that you might call a “noisy revolution” compared to the steady profits of the past.
Where are the limits of AI in weather forecasting?
AI will not replace physics-based predictions. Physical models are essential to explain why Predictions are unfolded in a specific way, which is essential to trust, and AI models are limited by the data being trained. They may struggle with unprecedented events, such as volcanic eruptions, or extreme events that have not yet appeared in training data. And in climate change, that's a real concern. As the future climate doesn't look like the past climate, there are limitations to machine learning models trained using historical weather.
In fact, we can see that AI and physics-based models are running side by side. Physics for insight and trust, AI for speed and efficiency. When used together, they can enable stronger and more accessible predictions than ever before.
What other challenges have you seen in deploying AI?
Honestly, this is still relatively new and there are a lot of things because people are still deciding how to use it. But the main thing is trust. You need to be completely confident in the ability of machine learning models to function properly and accurately and be competitive with physics-based models.
For us, the big challenge is ensuring that we trust the model before we share it. Otherwise it undermines the public's trust that has spent the last 150 years of buildings.
How can you ensure that you have this trust?
Over time, we developed stringent procedures to assess the performance of new systems. That framework now applies to AI as well. We will not operate machine learning models until they are thoroughly tested against strict performance standards. The process does not start from scratch. We have built up decades of experience in testing, verifying and improving models, and integrating AI gives us confidence that we can trust our results.
Do you think we are just at the beginning of a paradigm shift around AI in weather forecasting in the future?
Yes, I think we do. Machine learning approaches for weather forecasting are still being developed and while there is real potential, the shift has not yet fully happened. I often think about this through the lens of Amara's Law. We tend to overestimate the impact of new technologies in the short term and underestimate its impact in the long term. That applies to AI.
Early on, some people imagined that Met Office avatars would make predictions, which sounded fun, but obviously an overestimation of what could be possible right away. What's more likely is that as a global community, we underestimate how transformative AI will change over the long term.
That said, our expert human predictors continue to be central. Their role may evolve, and although they have many tools at their freedom, their judgment and experience remains as important as ever.
