Agenda Day 1: AI for Climate Science

Abstract:
AI has become a critical component of Earth system modeling and climate change research, which has experienced rapid growth in data from model output of increasing resolution in weather and climate models, and Earth observation systems that produce orders of magnitude more data than their previous generations. Efforts are underway in the weather and climate modeling community towards further refinement in horizontal resolutions of numerical atmosphere GCMs towards km-scale in order to explicitly resolve certain small-scale convective cloud processes and provide more realistic local information on climate change. At the same time, exascale HPC systems have arrived and in most cases are designed with GPU accelerator technology that offers opportunities in reasonable numerical simulation and AI training turn-around times that balance hybrid physics and AI modeling procedures. Ultimately output from global km-scale storm-resolving models will become the essential driver behind the deployment of Earth digital twins for programs like the EC Destination Earth, NVIDIA Earth-2, and others. This talk will describe advances in the development of NVIDIA-based weather and climate AI models that provide the critical components towards the vision of Earth system digital twins.

Abstract:
This talk will outline three revolutions that happened in Earth system modelling. The quiet revolution has leveraged better observations and more compute power to allow for the constant improvements of prediction quality of the last decades, the digital revolution has enabled us to perform km-scale simulations on modern supercomputers that further increase the quality of our models, and the machine learning revolution has now shown that machine-learned models will change the way how we do numerical weather predictions and have a very strong (but still rather unknown) impact on climate modelling. This talk will summarize the past developments, explain current challenges and opportunities, and outline how the future of Earth system modelling will look like.

Abstract:
We will start the discussion with a landmark achievement: the first global simulation of the full Earth system at a 1.25 km grid spacing. Our talk will focus on how we used the Alps supercomputer to model the intricate flow of energy, water, and carbon across the atmosphere, ocean, and land. We will detail the heterogeneous setup and optimization techniques that enabled us to achieve an exceptional time compression of 82.5 simulated days per day, allowing for extensive studies of the Earth system. We will present our methodology for reducing code complexity by half while increasing both performance and portability. Finally, we will put this all into context of emerging AI methods and systems and provide an outlook into how to combine physics-based simulation and data-driven AI methods.

Abstract:
Resilience is beautiful, but it is not like beauty: You don't know it when you see it. Rather, you have to measure it. AlphaGeo takes a Darwinian approach to climate resilience, namely by emphasizing adaptation as the key to survival. Our journey includes milestones such as harmonizing climate hazard models at granular scale, building a comprehensive global adaptation layer, devising a scoring methodology to quantify resilience, and an econometric toolkit for investors and asset managers to calculate climate-adjusted cashflows and valuations. The diplomatic and economic discourse on climate has favored mitigation solutions as global public goods and demoted adaptation because it is viewed as a local, private good. However, local adaptation strategies are a critical component of national and systemic resilience even as they unlock a new geographical arbitrage among locations based upon degrees of adaptation. Investors and policy-makers must prioritize adaptation investments as an urgent contribution towards this systemic resilience.

Abstract:
Humanity faces an unprecedented challenge. Over the coming decades we must figure out how to mitigate and adapt to a rapidly changing climate, ensure resilient water supplies, sustainably feed a human population growing towards 10 billion people, all while stemming an ongoing and catastrophic loss of biodiversity. To succeed we will need to make transformational advances in our ability to query the Earth in order to understand what is where, how much is there, and how fast it is changing, all to inform our ability to ask and answer key questions of what could, and should, be there? This ability to monitor, model, and manage Earth’s natural resources at appropriate spatial and temporal scales is a grand challenge for AI. This talk will cover some of the opportunities, challenges, and advances in the deployment of AI for Earth, and present a research and deployment agenda commensurate with the risks and rewards the field presents.

Abstract:
Conventional damage functions, often linear or quadratic in temperature, fail to capture thresholds, tipping points, or sectoral and regional heterogeneity. This presentation introduces AI-driven damage functions—new approaches that leverage advances in climate science, economics, and machine learning to better represent risks and associated economic impacts related to climate. Artificial intelligence (AI) enables the estimation of damages by detecting nonlinearities, compound shocks, and context-specific dynamics in large climate and socio-economic datasets. Employing Spanish data, we show that these methods can enrich integrated assessment models with more realistic and flexible representations of climate damages. By rethinking how damages are quantified, this work strengthens the scientific basis for climate risk assessment and enhances the capacity of decision-makers to respond to escalating climate challenges.

Abstract:
Among the greatest challenges of the 21st century is to sustain the prosperity, health, and contentment of our global society while balancing its growing and expanding population and continued needs of plentiful energy, freshwater, clean air, and land resources. To meet this challenge, decisions and actions - at scale - must be informed via quantitative foresight and optimal solutions deciphered through advanced and rigorous prediction methods that identify the plausible and optimal development pathways. It is only through the continued advances of integrated model frameworks that represent the natural, managed, and built environments, the needed level of spatiotemporal details, and efficient computation enabled through AI-based methods that we can achieve informative projections of physical, transition, and termination risks, and what actions are needed to avoid and adapt to changing environments.  In view of this, prospects and needed progress in research methods, models, and predictions that effectively and efficiently represent physical, socioeconomic, and multi-sector coupling and feedbacks will be highlighted and demonstrated that: quantify implications of different decarbonization pathways under uncertainty; generate high-resolution, uncertainty-aware projections for extreme hazards and natural resource assessments as well as resilient building design; and enable use-inspired, open-science, FAIR data-visualization platforms to hyper-local scales. 

Abstract:
Climate change is profoundly affecting the global water, energy and carbon cycles and increasing the likelihood and severity of extreme events. Better decision support systems are essential to accurately predict and monitor environmental disasters and to optimally manage water and environmental resources. A Digital Twin Component (DTC) on Hydrology and Hydro-Hazards would provide breakthrough solutions for monitoring and simulation, but requires high-resolution (1 km, 1 hour/day) satellite Earth Observation (EO) data, fully integrated with advanced and spatially distributed modelling systems. The overall objective of the Digital Twin Earth (DTE) Hydrology Next project is to prototype a fully advanced end-to-end demonstrator of DTE at high spatial and temporal resolution (1 km and 1 hour target) by extending the successful implementation of the DTE Hydrology and DTE Hydrology Evolution projects carried out over the Po River Basin in Northern Italy and over the Mediterranean region, respectively. In particular, the previous DTE Hydrology (Evolution) projects have demonstrated the potential to progress towards an end-to-end reconstruction of the hydrological cycle at high spatial and temporal resolution through an effective combination of state-of-the-art EO data, in-situ observations, advanced hydrological and hydraulic modelling, AI and advanced digital platform capabilities. The DTE Hydrology Next project will provide further advances in this regard by implementing use cases for flood and landslide forecasting, drought monitoring and water resources management across Europe and targeted regions in Africa and Central America. The project will also further develop a cloud-based infrastructure for visualisation and full interaction with the simulations and results of the use cases.

Abstract:
High-performance computing (HPC) is experiencing a paradigm shift driven by the dual forces of architectural innovation and evolving algorithmic strategies. Traditional double-precision, CPU-centric systems are giving way to heterogeneous platforms that prioritize low-precision arithmetic and massive parallelism, spurred in large part by AI accelerator design. We examine the role of mixed-precision and “responsibly reckless” algorithms, which combine fast, low-precision operations with selective high-precision refinement to achieve significant performance and energy-efficiency gains without sacrificing numerical fidelity. The discussion extends to integer-emulated floating-point computation, such as the Ozaki scheme, which leverages AI-oriented integer tensor cores for scientific workloads, and the reproducibility challenges posed by these methods. Through this lens, we argue that the future of HPC lies in tight algorithm–hardware co-design, balancing performance, precision, and reproducibility to meet the demands of exascale and post-exascale systems.