Seminar: AI for High-Performance Climate in The Age of Computation by Prof. Torsten Hoefler

Abstract: In the Age of Computation, traditional machine learning models are beginning to encounter data limitations. To address this, we are leveraging synthetic data and interacting agents to develop next-generation systems. This talk will illustrate how these advancements can revolutionize climate science. We will present an example from the recent Earth Virtualization Engines meeting in Berlin, where 140 scientists set the ambitious goal of achieving global 1km resolution simulations. However, we will demonstrate that the computational demands for such high-resolution simulations are currently unfeasible. By exploring the entire pipeline of weather and climate research, we will outline innovative solutions at each stage. For data assimilation, we propose the use of diffusion models inspired by image infilling techniques. In the simulation phase, we recommend combining traditional dataflow optimizations from the original Fortran code with AI techniques, utilizing our Data Centric Parallel Computing framework (DaCe) to replace certain simulation components. Additionally, we will question whether AI could eventually fully replace traditional simulations, highlighting the limitations of current methods and inspiring future research to overcome these challenges. In the post-processing stage, we will showcase how neural compression can be employed to store and analyse petabytes of climate data efficiently. Through various use-cases and studies, we will also address potential pitfalls in lossy data compression and propose solutions. Ultimately, this presentation will demonstrate how AI, integrated with existing simulation methods, can surpass current computational limitations, driving forward the field of climate science.

Bio: Prof. Torsten Hoefler is a Professor of Computer Science at ETH Zurich and the Chief Architect for Machine Learning at the Swiss National Supercomputing Centre. Prof. Hoefler is one of the world’s leading scientists in the field of high-performance computing, who is recognised for his foundational contributions to HPC and the application of HPC techniques to machine learning. Prof Hoefler is an ACM Fellow, IEEE Fellow, and Member of Academia Europaea. He received the ACM Gordon Bell Prize in 2019, was the first recipient of the ISC Jack Dongarra Award in 2023, and was awarded Max Planck-Humboldt Medal (the most prestigious German award to internationally outstanding mid-career scientists) in 2024. He has more than 300 papers in peer-reviewed international conferences and journals, numerous awards for best paper at top conferences (ACM, IEEE), and has a h-index of 76 (on Google Scholar).

Prof Hoefler’s research aims at understanding the performance of parallel computing systems ranging from parallel computer architecture through parallel programming to parallel algorithms. He is also active in the application areas of Weather and Climate simulations as well as Machine Learning with a focus on Distributed Deep Learning. Prof Hoefler is making invaluable contributions towards efforts to harness AI and high-performance compute for scientific research, and applications in areas such as weather simulation, healthcare, and the optimisation of algorithms for green computing.