Ben Shipway and Caroline Bain are strategy officers who lead the Met Office’s model development team. In this blog, they outline the potential for future blending of AI and physically-based models, and the steps the Met Office is taking to achieve its ambitious goals.
Machine learning (ML) is a type of artificial intelligence that is transforming the way science is developed at the Japan Meteorological Agency. Working with existing physically-based models to improve speed, accuracy, and flexibility in weather and climate prediction opens up new possibilities.
We recently set clear goals. The best combination of physically-based and ML-based modeling provides world-leading weather and climate forecasting data.
But what does that blend look like? And how can we talk about it clearly?
To answer these questions, we developed a simple framework that describes the different ways ML and physics work together. It’s not just about technical choices, it’s also about building a common language to guide future decision-making and convey direction to a wide range of stakeholders.
5 ways to blend ML and physics
We categorized the approaches into five types. Each has its own strengths, opportunities, and challenges and can be used in different parts of a forecasting system.
Independent physics base
These are the models we’ve been using for decades. They use the laws of physics to simulate the Earth system. They are trusted and accountable. In addition to observations, the data generated by these models also forms the basis for training currently emerging machine learning models. However, ML models have drawbacks such as computational cost.
Hybrid (two subtypes)
The term “hybrid model” has been used to describe the blend of ML and physics in a model for some time, but it’s also used in other ways, so we’ve broken it down into two more distinct types.
- Hybrid integrated ML
- ML is an integral part of the model and has the potential to replace or enhance components or processes within established physics-based systems.
- Hybrid compound ML
- The ML model and the physical model are run separately, and the output of one influences the evolution of the other. For example, a local physics-based model may use boundary data generated from a global ML model.
Extended ML
As a post-processing step, ML improves predictions or predictions after they are made. It doesn’t change how the model runs, but it increases the value of the output. This has already been tested for blending and downscaling ensemble predictions.
Independent ML base
These are ML-only systems that make predictions without using physical models. They are fast, scalable, and often learn from physics-based data (such as reanalysis datasets) that convey decades of scientific knowledge.
Each of these categories can be applied to a single forecasting model (e.g., a hybrid integrated ML global climate model) or a modeling system (e.g., a hybrid composite ML ensemble forecasting system).
Why this framework is important
This is not just a technical classification, but a strategic tool. It helps us:
- Communicate clearly about how ML and physics are being used.
- Support innovation while identifying and maintaining reliable workflows.
- Design systems to suit different user needs and environments.
for example:
- Hybrid and augmented approaches enable incremental changes by adding ML without losing existing physics-based knowledge and credibility.
- Independent-ML could be a game-changer, but more work is needed to build trust and generalization
- Integrated and combined hybrids provide flexibility in how ML can be deployed within existing physics-based models or as part of a broader system.
Looking to the future
There is no one-size-fits-all solution. We expect to see a variety of approaches tailored to different prediction goals, computing configurations, and user needs.
This framework gives meteorologists and scientists a common starting point. This will help us speak clearly and confidently about the ‘best blend’ when shaping the future of weather and climate prediction.
Although this blog focuses on blending strategies, it sits within the broader context of ML innovation at the Met Office. This reflects our commitment to combining cutting-edge technology with deep scientific expertise. We continue to improve weather and climate predictions to help you make better decisions to stay safe and prosperous.
Learn more about the Met Office’s approach to leveraging the benefits of AI for weather and climate intelligence.
About this blog
This is the official blog of the Met Office News Team, dedicated to providing journalists and bloggers with the latest weather, climate science, business news and information from the Met Office.
