At the confluence of scientific simulation and modern machine learning there exists an opportunity to develop a “middle path” that leverages the strengths of both approaches to build machine-learning emulators of numerical simulation models. From the perspective of machine learning, incorporating simulation data may significantly reduce the need of expensive observational data for training, as well as conditions the model on real-time observations for inference; from the perspective of scientific simulation, a streamlined, end-to-end process naturally leveraging observational data may reduce the reliance on closures and parameterizations to resolve finer scales that models purely based on solving partial differential equation (PDE) systems cannot address. Running on modern machine learning frameworks and acceleration hardware such as graphical processing units (GPUs), such a hybrid models would allow fast inference, easy sensitivity/what-if analysis of simulation output variables with respect to input observational data, and scenario analysis. We present initial work developing such models for two problems of relevance to climate and environmental science: turbulent flows and land-use modeling.
Lunch will be available for attendees outside the auditorium at 12pm.