AI for the future of climate prediction

Go back to programme

AI for the future of climate prediction

Climate change is one of the existential challenges facing humanity. Recent advances in artificial intelligence (AI) are already transforming many science areas, including weather prediction.

In the first part of the workshop (10:00 – 16:00), we discuss the direction that AI may play to predict the climate, given that that AI applications for climate prediction remain in early stage of development. We explore the frontier of AI applications for prediction of climate change and associated risks, bringing together leading experts from academia, industry and research labs. Topics covered will include harnessing AI for enabling and supporting km-scale global modeling and reduced physical uncertainties.

In the second part of the workshop (16:15 – 18:00), we will discuss AI’s role in climate action. From greening AI technologies and standards to disaster management, the session will explore a plan of action and activities to take steps toward a greener world.

This presentation explores the convergence of Artificial Intelligence (AI) and climate modeling for high-resolution (km-scale) climate prediction. We will discuss the use of accelerated km-scale simulations to generate synthetic climate data, enabling the training of even more sophisticated climate models. We'll delve into the integration of observational data from diverse sources – weather stations, satellites, and airborne platforms – through diffusion models, a technique commonly used for image generation. To address the exascale challenge posed by climate data volume, we will present an AI-based compression method utilizing deep neural networks for efficient data representation. This approach allows for near-lossless reconstruction of the original data for subsequent analysis. We posit that this synergistic approach, combining high-fidelity data generation with advanced data assimilation techniques, has the potential to significantly improve the accuracy of climate predictions. This, in turn, can inform policy decisions and guide societal responses to climate change.

Climate projections and modelling the resulting impacts on humans and societies is critical to help humanity adapt to global warming. For weather, large-scale, pure machine learning models now outperform the best conventional, equation-based models for a wide range of scores. First steps towards the use of these or large-scale hybrid models for climate projections have been made. However, their potential remains currently unclear. The talk will critically analyse the skill of existing machine learning-based models for weather for their use for climate and impact modelling. Based on this, a set of open questions will be proposed. Addressing these in a systematic and scientifically reliable way will help us to understand what role large-scale machine learning can play for better climate projections and to realise the full potential of the technology.

This talk focuses on ITU (International Telecommunication Union) standards designed to assess and guide sustainable AI applications, particularly those shaping the development and deployment of environmentally conscious AI solutions. One of ITU’s standards groups, Study Group 5 (ITU-T SG5), plays a key role in evaluating the impact of information and communication technologies (ICTs) on climate change. Study Group 5 develops methodologies for this assessment and publishes guidelines for eco-friendly ICT usage. This talk explores greening AI technologies, standards, and how AI’s computational demands contribute to the ICT sector’s energy use and greenhouse gas (GHG) emissions.

The Green Digital Action track was launched at the United Nations Climate Change Conference in November 2023 (COP28) by ITU, together with over 40 partners from UN agencies, businesses, civil society, and governments. Its mission is to mobilize collective initiatives within the digital industry to combat climate change. By developing practical solutions and positioning the sector as a climate leader, the Green Digital Action track aims to enhance sustainability while advancing digital transformation. The goal is to create a shared vision for using the power of digital technologies while mitigating their environmental impact. This talk explores both the potential and pitfalls of ICTs in the context of environmental sustainability and how the ICT sector is using technology such as AI to reduce emissions.

Over its three-year lifetime, the Focus Group on AI for Natural Disaster Management (FG-AI4NDM), convened by the International Telecommunication Union (ITU), World Meteorological Organization (WMO), and United Nations Environment Programme (UNEP), has laid the groundwork for standards (producing three technical reports, a roadmap, a glossary, and a large language model based tool) and supported capacity sharing. In the future, there are plans to build on these outcomes by exploring complementary technologies for natural hazard management across a range of different disaster types. A Horizon-funded 3-yr project called Mediterranean and pan-European forecast and early warning system against natural hazards (MedEWSa) will apply the best practices developed by FG-AI4NDM and the potential of AI will be ascertained for decision support, multi-hazard forecast and impact assessment, risk transfer, and outreach, in regions of different coping capacities. The outcomes of this project will provide important insights into the versatility of AI solutions.

This talk launches a competition focused on optimizing energy usage for telecom sites. By maximizing solar energy utilization and minimizing reliance on mains and generators, the aim is to reduce the carbon footprint of telecom infrastructure. Participants of the competition will be tasked, using AI, to come up with smart energy scheduling policies that balance sustainability and reliability. Solutions need to be turned in by September 2024. The best solution will be showcased at the United Nations Climate Change Conference in November 2024 COP29.

The AI for Climate Action Innovation Factory, a start-up pitching competition, is a collaborative effort among United Nations agencies including the International Telecommunication Union (ITU), the International Atomic Energy Agency (IAEA), and the Food and Agriculture Organization (FAO). The 2024 edition aims to advance the utilization of AI in combating climate change, with solutions to be showcased at the United Nations Climate Change Conference in November 2024 (COP29). The Innovation Factory aims to accelerate the development of AI-driven solutions tailored for climate action by uncovering start-ups with the potential to produce AI solutions that can be implemented across different regions and industries. The focus is not only on technological development but also capacity building and knowledge dissemination to ensure that AI advancements are accessible and beneficial to all stakeholders involved in climate action. This approach helps bridge the gap between cutting-edge AI research and real-world environmental needs, promoting a more sustainable future.

AI Research for Climate Change and Environmental Sustainability

Claire Monteleoni

Despite the scientific consensus on climate change, drastic uncertainties remain. Crucial questions about regional climate trends, changes in extreme events, such as heat waves and mega-storms, and understanding how climate varied in the distant past, must be answered in order to improve predictions, assess impacts and vulnerability, and inform mitigation and sustainable adaptation strategies. Machine learning can help answer such questions and shed light on climate change. I will give an overview of our climate informatics research, focusing on challenges in learning from spatiotemporal data, along with semi- and unsupervised deep learning approaches to studying rare and extreme events, and precipitation and temperature downscaling.

 

AI and Earth Observation-Powered Insights for Enhanced Climate Prediction

Yifang Ban

Earth Observation (EO), with its synoptic view, large area coverage and frequent revisits, is essential for advancing climate prediction. Utilizing 50 years of EO time series data, it is possible to monitor the spatial patterns and temporal trends of our changing planet. EO provides comprehensive and continuous data on various environmental parameters, and has been used to derive essential climate variables (ECVs) such as fire, greenhouse gases, land cover, land surface temperature, galciers, and sea level. These ECVs are crucial for understanding and modeling climate dynamics. Deep learning, on the other hand, is one of the fastest-growing trends in big data analytics, proving to be a highly effective technique for large-scale image recognition, semantic segmentation, and change detection. This talk will showcase how AI and EO have been used to monitor urbanization and wildfires - significant contributors to climate change. By leverating AI and EO, we can swiftly and reliably map urbanization and monitor wildfires at a global scale, enabling the assessment of their environmental impact. These insights can be integrated into predictive models to forecast future climate scenarios. The integration of EO and AI not only enhances the accuracy of climate predictions but also supports the development of informed mitigation and adaptation strategies in the face of climate change.

 

Accelerating Responsible AI for Climate Action

Antonia Gawel

Climate change is one of humanity's most pressing collective challenges. Our approach to AI must be bold, responsible, and collaborative—that means developing AI in a way that maximizes the positive benefits to society, while responsibly managing its environmental footprint. This presentation explores how Google is accelerating climate solutions using AI, while responsibly managing the environmental impacts of this technology. Critically, partnership and collaboration will be central to harnessing AI's potential for positive climate action.

 

Climate Resilience through Earth System Foundation Models

Thomas Brunschwiler

This session explores the potential of Earth Observation and Weather Foundation Models (FM) operated in concert to evaluate climate impacts and facilitate nature-based adaptation strategies. We introduce the concept of self-supervised model pretraining and fine-tuning, encompassing sampling strategies and data streaming for inferencing services. This AI-as-a-platform approach enhances data efficiency, generalization, and few-shot learning, enabling economical scalability. Multi-modal and multi-domain FM architectures are presented which can deal with remote sensing and climate data, as well as translate between earth observations and text. We also present our vision of generic embedding sharing by FM-compressors to reduce data transfer latency, storage costs, and computational demands during inference time, as well as to improve data exploration. Case studies on climate and transition risks will illustrate the advantages of the Earth System Foundation Model approach.

Share this session