Destination Earth: Using digital twins and AI to study climate change

Destination Earth DestinE digital twin Destination Earth DestinE digital twin

The 2021 Nobel Prize in Physics was awarded to researchers Syukuro Manabe and Klaus Hasselman for their work on "physical modelling" of Earth’s climate.

Their models – which use long-term historic data to understand how climate change impacts the planet – are now helping to give a clearer picture of ongoing planetary processes and how they could affect us all.

Destination Earth (DestinE), a new project from the European Commission, aims to build on existing climate change research and explore new dimensions with high-precision digital models of the Earth.

Set to be implemented over the next 7-10 years, the project is currently in early planning and workshopping stages.

At the heart of the DestinE project is the concept of the digital twin – a virtual Earth that simulates natural processes and human activity. Observing such replicas can help researchers understand change and help shape policies to mitigate extreme climate-related risks to society.

Artificial intelligence (AI) and machine learning will make this interactive framework more flexible and efficient, said Peter Bauer, Deputy Director, Research Department at the European Centre for Medium-Range Weather Forecasts (ECMWF), speaking at a recent AI for Good event organized by the International Telecommunication Union (ITU).

Launching in November, DestinE will be co-developed by ECMWF, the European Space Agency, and the European Agency for the Exploitation of Meteorological Data (EUMETSAT). The project aims to support policies that will help make Europe climate neutral by 2050.

Flexible replicas

Combining different simulations with unprecedented efficiency, digital twins can shed light on planetary cause-effect relationships while integrating basic needs for human sustenance, such as food, water, energy and health. DestinE also incorporates key European and global policy factors, like energy.

"The integration of policy sectors means we always have to think through entire chains of information and data," Bauer said, highlighting the need to “fully integrate different value chains in our digital twins."

To integrate multiple value chains and provide interactivity, digital twins demand massive processing power, with data flowing rapidly between high performing computing (HPC) systems in different countries, potentially woven together in “federated infrastructures.”

The project aims to launch its first two digital twins by December 2023: one to assess and predict weather-induced and geophysical extremes, and another to create analytical insights and test predictive scenarios for climate adaptation and mitigation policies.

Faster, improved results

Machine learning lets digital twins work faster, with reduced numerical precision further speeding up large-scale simulations.

"You can emulate model components and make them faster,” notes Peter Dueben, AI and Machine Learning Coordinator at ECMWF.

“Representing the same model component with the neural network is likely to make them faster than the original conventional scheme."

Machine learning improves the ability to simulate planetary activity, both in a default scenario or with varied conditions.

The system will reveal specific events, such as tropical cyclones, and provide tools to pursue hitherto impossible investigations. "You can think about information fusion more, bringing observations and the models closer together," Dueben says.

Co-designing solutions

But machine learning also presents challenges. For one thing, human researchers are interested in different angles.

"Domain scientists like to think about physics and physical reasoning, while machine-learning scientists like to think about data science problems with an input-output loss function," Dueben explains.

New findings, meanwhile, may exceed current computing power, he adds. "We are still only beginning to be able to tap into this resource, learning how to scale up our applications. Often, if you develop machine learning tools, it's difficult to integrate them back into the conventional modelling framework."

Reliance on precedent can raise questions about model predictions.

"If you build a machine learning tool for today's climate and then use it in a model to look into the future 100 years on, you will have weather situations that have never happened before,” Dueben pointed out.

“We often cannot trust machine learning tools to respond correctly to new situations because they can't extrapolate. There's a lot of research needed to build more trust in those methods."

Possibilities for the future

DestinE’s modelling approach could reveal a four-dimensional Earth system with an unprecedented level of detail and resolution. It could also produce maps to or from any point in time and space.

The framework, Dueben surmises, will eventually support AI-based user models within the digital twin.

For example, the project seeks to integrate additional digital twins, like a digital twin of the ocean, to serve sector specific use cases into the platform by 2027.

 

Image credit: starline via Freepik