Learning to learn about the earth, using Bayesian inference

Title: Learning to learn about the earth, using Bayesian inference

Presenter: Anandaroop Ray (Geoscience Australia)

Abstract: To understand earth processes, geoscientists infer subsurface earth properties such as electromagnetic resistivity or seismic velocity from surface observations such as magnetotelluric data or seismograms. These properties are used to populate an earth model vector, and the spatial variation of properties sheds light on the underlying earth structure and phenomena, from groundwater aquifers to plate tectonics. I will show that in order to make accurate inferences about earth properties, inferences can first be made about the underlying length scales of these properties. From a mathematical point of view, the length scales can be conveniently thought of as “properties” of earth properties. This can be treated in an “infer to infer” paradigm analogous to the “learning to learn” paradigm which is now commonplace in the machine learning literature. A non-stationary trans-dimensional Gaussian Process (TDGP) is used to parameterise earth properties, and a multi-channel stationary TDGP is used to parameterise the length scales. Using non-stationary kernels, i.e., kernels with spatially variable length scales, earth models with sharp discontinuities can also be represented within this framework. As GPs are multi-dimensional interpolators, the same theory and computer code can be used to solve geophysical problems in 1D, 2D and 3D. This is demonstrated through a combination of 1D and 2D non-linear regression examples and a controlled source electromagnetic field example.

Biography: Anandaroop Ray (“Anand”) started his career as a non-seismic geophysicist with Shell Exploration and Production in 2007. In 2010 he joined the PhD programme in marine electromagnetics at the Scripps Institution of Oceanography in San Diego, California. In 2014 he completed his thesis focusing on uncertainty estimation in electromagnetic inversion for marine hydrocarbon exploration. From 2012-19, he worked for Chevron R&D on various problems – controlled source electromagnetics (CSEM), seismic full waveform inversion (FWI), reservoir properties from seismic (RPFS), airborne electromagnetics (AEM), statistical hydrocarbon exploration lookback analyses, and the role of machine learning in geophysics. The question most asked through his work is “how credibly can we interpret our inversion model(s),” the answering of which often requires the use of high-performance computing (HPC) techniques. He currently co-advises a PhD student at Columbia University on Bayesian geophysical inversion and has been active in convening and organizing the Uncertainty in Geophysical Inversion session at the American Geophysical Union’s Fall Meeting. In March 2019 he joined the Minerals, Energy and Groundwater Division at Geoscience Australia, where he continues to work on inverse uncertainty, model representation and geostatistics.

Register: https://us02web.zoom.us/webinar/register/WN_j_B7TWtMQvOSxy9zTLj5Iw