IN033
Incorporating physics and domain knowledge to improve interpretability, explainability, reliability and generalization of Machine Learning Models (MLM) in the geosciences

Session ID#: 82358

Session Description:
This session focuses on approaches that incorporate physical principles and domain knowledge in a systematic manner to the design, training and inference of MLM. Such efforts can result in more interpretable and explainable models that help place new scientific findings in the context of existing knowledge. Further, MLM are more reliable and generalizable when they respect the principles governing the complex systems they model or analyze.

This session emphasizes the following themes applicable to the geosciences, including atmospheric and ocean sciences, as well as the intersections of numerical modeling and data assimilation with ML.

Methods and applications that use physical principles and domain knowledge

  1. in the design, optimization, training or inference of MLM
  2. to systematically identify weaknesses of MLM
  3. to improve reliability and generalizability, and reduce the cost of MLM
  4. to interpret, understand and visualize MLM output
  5. to represent new physical models and parameterizations from ML

Session co-organized by IS-GEO.

Co-Organized:
  • A - Atmospheric Sciences
  • OS - Ocean Sciences
Index Terms:

1942 Machine learning [INFORMATICS]
3365 Subgrid-scale (SGS) parameterization [ATMOSPHERIC PROCESSES]
4260 Ocean data assimilation and reanalysis [OCEANOGRAPHY: GENERAL]
4313 Extreme events [NATURAL HAZARDS]

Primary Convener:  Karthik Kashinath, Lawrence Berkeley National Laboratory, Berkeley, CA, United States
Conveners:  Savini Manthila Samarasinghe1, Benjamin A Toms1 and Pierre F J Lermusiaux2, (1)Colorado State University, Fort Collins, CO, United States(2)Massachusetts Institute of Technology, Cambridge, MA, United States
Primary Liaison:  Karthik Kashinath, Lawrence Berkeley National Laboratory, Berkeley, CA, United States
 
Using physics-based regularization in Artificial Neural Networks to predict thermochemical energy storage systems (486717)
Timothy Praditia, Thilo Walser, Sergey Oladyshkin and Wolfgang Nowak, University of Stuttgart, Stochastic Simulation and Safety Research for Hydrosystems (IWS/SC SimTech), Stuttgart, Germany
 
Peering into the black box of machine learning parameterization (Invited) (491067)
Noah Brenowitz, University of Washington, Atmospheric Sciences, Seattle, WA, United States and Christopher S. Bretherton, University of Washington, Seattle, WA, United States
 
Peering inside the black box: Using deep learning to improve understanding of weather phenomena (Invited) (499010)
Amy McGovern, University of Oklahoma, Computer Science, Norman, United States, Ryan Lagerquist, University of Oklahoma, Norman, United States and David John Gagne II, National Center for Atmospheric Research, Boulder, CO, United States
 
Trials and Tribulations of Supervised Learning for Computational Geosciences in Planetary Magnetospheres (516043)
Abigail Azari, University of Michigan Ann Arbor, Ann Arbor, MI, United States, Jeffrey W. Lockhart, University of Michigan Ann Arbor, Ann Arbor, United States, Michael Warren Liemohn, University of Michigan, Climate and Space Sciences and Engineering, Ann Arbor, MI, United States and Xianzhe Jia, University of Michigan, Ann Arbor, MI, United States
 
Machine Learning Parameterization of the Surface Layer: Bridging the Observation-Modeling Gap (520893)
David John Gagne II1, Tyler McCandless1, Branko Kosovic1, Amy DeCastro1, Richard Loft1, Sue Ellen Haupt1 and Bai Yang2, (1)National Center for Atmospheric Research, Boulder, CO, United States, (2)NOAA Air Resources Lab, Field Research Division, Idaho Falls, ID, United States
 
Building a Hierarchy of Hybrid, Neural-network Parametrizations of Convection (Invited) (543245)
Tom Beucler, University of California Irvine, Earth System Science, Irvine, MA, United States; Columbia University, Earth and Environmental Engineering, New York, NY, United States, Pierre Gentine, Department of Earth and Environmental Engineering, Columbia University, New York, NY, United States, Mike S Pritchard, University California Irvine, Department of Earth System Science, Irvine, CA, United States, Stephan Rasp, Ludwig Maximilians University of Munich, Munich, Germany and Veronika Eyring, German Aerospace Center (DLR), Oberpfaffenhofen, Germany
 
Physics Guided Machine Learning: A New Paradigm for Modeling Dynamic Systems (545195)
Xiaowei Jia1, Jared Willard2, Anuj Karpatne3, Jordan Stuart Read4, Jacob A Zwart4, Alison Appling5, Paul C Hanson6 and Vipin Kumar1, (1)University of Minnesota Twin Cities, Minneapolis, MN, United States, (2)University of Minnesota Twin Cities, Minneapolis, United States, (3)Virginia Polytechnic Institute and State University, Department of Computer Science, Blacksburg, VA, United States, (4)USGS Wisconsin Water Science Center, Middleton, WI, United States, (5)USGS Integrated Information Dissemination Division, Data Science Branch, Middleton, WI, United States, (6)University of Wisconsin, Madison, WI, United States
 
Emulating Numeric Hydroclimate Models with Physics-Informed cGANs (559829)
Ashray Manepalli1, Adrian Albert2, Alan Rhoades2, Daniel Feldman2 and Andrew D Jones3, (1)Kensington, CA, United States, (2)Lawrence Berkeley National Laboratory, Berkeley, CA, United States, (3)Lawrence Berkeley Lab, Berkeley, CA, United States
 
Explaining State-of-the-Art Performance of Mixture Density Networks on Remotely Sensed Water Quality Parameters (567382)
Brandon Smith and Nima Pahlevan, NASA Goddard Space Flight Center, Greenbelt, MD, United States
 
Multi-resolution Cluster Analysis - Addressing Trust in Climate Classification (597779)
Derek DeSantis, Ventura, California, United States, Phillip J. Wolfram Jr, LANL, LOS Alamos, NM, United States and Boian Alexandrov, LANL, Santa Fe, NM, United States
 
Towards Physics-informed Deep Learning for Spatiotemporal Modeling of Turbulent Flows (617369)
Rui Wang1, Adrian Albert2, Karthik Kashinath2 and Mustafa Mustafa3, (1)Northeastern University, Khoury College of Computer Sciences, Boston, MA, United States, (2)Lawrence Berkeley National Laboratory, Berkeley, CA, United States, (3)Lawrence Berkeley National Laboratory, Berkeley, United States
 
Estimating model-form uncertainty for multi-scale systems (620819)
Jinlong Wu, Tapio Schneider and Andrew Stuart, California Institute of Technology, Pasadena, CA, United States
 
A Trainable Simulator: using unsupervised learning in conjunction with computational methods to rewrite our equations, applied to multiphase flow (621070)
Alejandro Francisco Queiruga, Lawrence Berkeley National Laboratory, Berkeley, CA, United States
 
Bayesian Learning Machines for Dynamical Ocean Models (629434)
Pierre F J Lermusiaux and Abhinav Gupta, Massachusetts Institute of Technology, Cambridge, MA, United States
 
Learning and Inference of Advective Fluid Transport in Marine Environments (630035)
Chinmay Kulkarni and Pierre F J Lermusiaux, Massachusetts Institute of Technology, Cambridge, MA, United States
 
Optimization of the K-Profile Parameterization Using Machine Learning (631446)
Justin Moshe Finkel1, Nathan M Urban2 and Luke Van Roekel2, (1)University of Chicago, Chicago, IL, United States, (2)Los Alamos National Laboratory, Los Alamos, NM, United States
See more of: Earth and Space Science Informatics
<< Previous Session Proposal | Next Session Proposal >>