Weather/Climate Session Details
Monday, August 9, 1:30-4:30
(All times for the workshop are listed in Central Time, UTC -5)
(All times for the workshop are listed in Central Time, UTC -5)
Quicklinks to session details: Opening Session (ML1) Weather and Climate Aquatic Sciences Hydrology Translational Biology Closing Session (ML2)
Other workshop links: Workshop Home Workshop Logistics Poster Sessions Workshop Booklet
YOUTUBE LINKS: Please go to the KGML YouTube Channel for all available recorded presentations.
Session Organizers: Elizabeth Barnes, Ben Cash, Tim Delsole, Imme Ebert-Uphoff
SPEAKERS
1:30-1:50 Timothy Delsole, George Mason University: Overview of Knowledge-Guided Machine Learning for Weather and Climate (Presentation Slides) (Presentation Video)
1:50-2:15 Pierre Gentine, Columbia University: Hybrid modeling (physics plus machine learning) to improve prediction of the hydrological cycle (Presentation Slides) (Presentation Video)
2:15-2:40 Katie Dagon, National Center for Atmospheric Research: Machine learning-based feature detection to associate precipitation extremes with synoptic weather events (Presentation Slides) (Presentation Video)
2:40-3:05 Peter Dueben, European Centre for Medium-Range Weather Forecasts: Challenges when preparing machine learning tools for use in operational weather predictions (Presentation Slides) (Presentation Video)
3:05-3:10 BREAK
3:10-3:35 Antonios Mamalakis, Colorado State University: Assessing methods of explainable artificial intelligence (XAI) by using attribution benchmark datasets (Presentation Slides) (Presentation Video)
3:35-4:00 Laurie Trenary, George Mason University: Skillful statistical prediction of sub-seasonal temperature by training on dynamical model data (Presentation Slides) (Presentation Video)
4:00-4:25 Pedram Hassanzadeh, Rice University: Building physical consistencies into neural networks for weather/climate modeling (Presentation Slides) (Presentation Video)
Abstract: In this session we consider applications of Knowledge Guided Machine Learning (KGML) to weather and climate applications. Specifically, we consider any connection between scientific knowledge and machine learning, i.e. we include methods that add knowledge to machine learning and methods that extract knowledge from machine learning, as well as any combination of the two. Methods for adding knowledge include i) Incorporating physical constraints into machine learning methods; and ii) Transfer learning, which allows knowledge gained in one domain to be transferred to another domain. Methods for extracting knowledge include: i) Using explainable AI (XAI) to extract important spatial patterns from ML models and/or explore the physical processes leveragedpi by the ML method; ii) Using causal discovery to extract causal relationships from data; and iii) Using machine learning methods to learn a system’s PDEs from data. All of these topics are covered by invited speakers in this session, with the exception of PDE discovery being discussed in the Opening Session. This talk seeks to provide a general overview of these topics, incl. additional context, and examples from our own work. We thus hope to set the stage for the innovative approaches discussed in the invited talks to follow.
Bio: Timothy DelSole is a full professor in the Department of Atmospheric, Oceanic, and Earth Sciences at George Mason University, and a senior research scientist at the Center for Ocean-Land-Atmosphere Studies. His research focuses on the extent to which weather and climate changes can be predicted on time scales from weeks to year. He currently serves as co-Chief Editor of Journal of Climate.
Abstract: In recent years, we have witnessed an explosion in the applications of machine learning, especially for environmental problems.Yet for broader use, those algorithms may need to respect exactly some physical constraints such as the conservation of mass and energy. In addition, environmental applications (e.g. drought, heat waves) are typically focusing on extremes and on out-of-sample generalization rather than on interpolation. This can be a problem for typical algorithms, which interpolate well but have difficulties extrapolating. Finally, interpretation of machine learning algorithms can be difficult, limiting the trustworthiness in those algorithms. I will here show how a hybridization of machine learning algorithms, imposing physical knowledge within them, can help with those different issues and offer a promising avenue for climate applications and process understanding.
Bio: Pierre Gentine is the Maurice Ewing and J. Lamar Worzel professor of geophysics in the departments of Earth and Environmental Engineering and Earth and Environmental Sciences at Columbia University. His group is focusing on understanding the future continental hydrological and carbon cycles using a combination of modeling, observations and machine learning.
Abstract: Extreme precipitation events continue to have wide-ranging impacts across the world. Rainfall associated with atmospheric rivers, tropical cyclones, mesoscale convective systems, and fronts can cause devastation to communities and ecosystems. Machine learning-based detection algorithms can help with the automated classification of these synoptic weather features. Here we use new and existing machine learning algorithms to identify these types of systems in climate model output, and validate the results using observational and reanalysis products. We then associate the detected features with precipitation extremes to better understand the sources and mechanisms of extreme precipitation events. We further compare results using model simulations with present-day and future climate forcing, to study how extremes might change and evolve with climate change.
Bio: Dr. Katie Dagon is a project scientist at the National Center for Atmospheric Research (NCAR), working in the Climate and Global Dynamics Laboratory. Her research focuses on modeling the impacts of climate change on land-atmosphere interactions, climate variability, and extreme events. She is also interested in machine learning approaches to climate science, including quantifying uncertainty in model projections of climate change. From 2017-2019 she was an Advanced Study Program postdoctoral fellow at NCAR. Katie obtained her Ph.D. in Earth and Planetary Sciences from Harvard University in 2017 and her B.S. in Mathematics-Physics from Brown University in 2010.
Abstract: The talk will provide a rough overview on the machine learning efforts that are currently investigated at the European Centre for Medium-Range Weather Forecasts (ECMWF). It will then discuss specific challenges and needs when using machine learning tools in the operational weather prediction workflow and outline why Knowledge Guided Machine Learning is important (and maybe essential) for some of the applications. The talk will close with an example for which a machine learning method without knowledge guidance failed.
Bio: Peter is the AI and Machine Learning Coordinator at European Centre for Medium-Range Weather Forecasts (ECMWF) and holds a University Research Fellowship of the Royal Society that enables him to perform research towards the use of machine learning, high-performance computing, and reduced numerical precision in weather and climate predictions. Peter has also a strong interest in the quantification of uncertainty of predictions for chaotic systems and is coordinator of the MAELSTROM EuroHPC-Join Undertaking project. Before moving to ECMWF, Peter has written his PhD thesis at the Max Planck Institute for Meteorology and has worked as PostDoc with Tim Palmer at the University of Oxford.
Abstract: Artificial neural networks (NNs) have shown great success in solving complex, nonlinear problems in the geosciences. Despite their success, NNs are difficult to interpret, which makes it hard for scientists to build trust for their predictions; highly important for the further use and exploitation of NNs’ potential. To address this, many methods have been recently suggested in the so-called eXplainable Artificial Intelligence (XAI) field, with the aim of attributing the NN predictions to specific features of the input and explaining their prediction strategy. Yet, the assessment of XAI methods in accurately explaining the NN strategy is typically subjective and no ground truth about how the attribution should look like is used. Moreover, benchmark datasets for regression problems in geoscience are rare. In this work, we provide a general framework to generate attribution benchmark datasets for regression problems, where the ground truth of the attribution is known a priori. Based on our framework, we generate a long dataset and train a fully-connected network to learn the underlying function that was used for simulation. We then use the ground truth of the attribution to assess the accuracy of different XAI methods and identify systematic advantages and pitfalls. Our work shows that there is room towards more objective assessment of XAI methods that may be of great importance for further application of NNs in geoscience.
Bio: Dr Mamalakis is a postdoctoral researcher at the Department of Atmospheric Science of Colorado State University. His research focuses on the application of machine learning (ML) and ML interpretability methods to climate problems, on climate predictability and teleconnections, climate change impacts, and hydrology. Dr Mamalakis holds a PhD in Civil and Environmental Engineering from the University of California, Irvine, and a MSc and a diploma from the University of Patras, Greece.
Abstract: In this study, we derive statistical models for predicting wintertime sub-seasonal temperature over the western United States. The statistical models are trained on two separate datasets, namely observations and dynamical model simulations, and are based on Least Absolute Shrinkage and Selection Operator (lasso). Surprisingly, statistical models trained on dynamical model simulations can predict observations better than observation-trained models. One reason for this is that simulations involve orders of magnitude more data than observational datasets. Nevertheless, the skill of sub-seasonal prediction is very low when measured by spatial average squared error. This result does not automatically mean there is no significant skill. For instance, certain large-scale patterns may be predictable, but this predictability may be obscured by local weather variability when local mean square error is used to measure skill. To identify large-scale predictable patterns, an optimization technique, called Skill Component Analysis (SCA), is applied. SCA finds the linear combination of variables that minimizes the normalized mean square error. Applying SCA to the lasso predictions reveals at least two patterns of large-scale temperature variations that are skillfully predicted. The predictability of these patterns is consistent between climate model simulations and observations. Not surprisingly, the predictability is determined largely by sea surface temperature variations in the Pacific, particularly the region associated with the El Nino-Southern Oscillation.
Bio: Laurie Trenary is an Assistant Research Professor in the Department of Atmosphere, Ocean, and Earth Science at George Mason University. Her work focuses broadly on topics related to climate change and prediction.
Abstract: Using deep neural networks (NNs) to represent the effects of subgrid-scale processes in multi-scale, nonlinear flows such as atmospheric and oceanic turbulence has received much attention in recent years. While even physics-agnostic NNs have shown promising results, they often need large training datasets, otherwise the data-driven subgrid-scale models could be inaccurate or cause numerical instabilities. Here we use two-dimensional turbulence as the testbed, and show how incorporating physics into the NNs can substantially reduce the amount of training data needed to obtain stable, accurate data-driven subgrid-scale models. Three specific ways of incorporating physics are discussed: data augmentation, building some of the symmetries of the flow into the NN’s architecture, and incorporating specific conservation laws into the loss function.
Bio: Dr. Hassanzadeh is an assistant professor at the departments of mechanical engineering and Earth, environmental and planetary sciences at Rice University. He received his MA in Mathematics and Ph.D. in Mechanical Engineering from UC Berkeley (2013). He was a Ziff Environmental Fellow at the Harvard University Center for the Environment and a Postdoctoral Fellow at the Harvard University Department of Earth and Planetary Science (2013-2016). He is currently leading or involved in a number of projects at the intersection of machine learning and weather/climate modeling supported by NSF, Office of Naval Research, Schmidt Futures, and C3ai.
Quicklinks to session details: Opening Session (ML1) Weather and Climate Aquatic Sciences Hydrology Translational Biology Closing Session (ML2)
Other workshop links: Workshop Home Workshop Logistics Poster Sessions