BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Anandaroop Ray (Geoscience Australia) DTSTART:20210628T053000Z DTEND:20210628T070000Z DTSTAMP:20260422T225722Z UID:RSES_Geophysics/1 DESCRIPTION:Title: Bayesian inference using trans-D Gaussian processes using trans-D Gaussian processes\nby Anandaroop Ray (Geoscience Australia) as part o f ANU Research School of Earth Sciences Geophysics Seminars\n\nLecture hel d in Jaeger 1 Seminar Room.\n\nAbstract\nTo understand earth processes\, g eoscientists infer subsurface earth properties such as electromagnetic res istivity or seismic velocity from surface observations such as magnetotell uric data or seismograms. These properties are used to populate an earth m odel vector\, and the spatial variation of properties sheds light on under lying earth structure and associated geodynamic phenomena. I will show tha t making accurate inferences about earth properties can require the use of a multilevel parameterisation\, where the same algorithm is used in a nes ted fashion to infer all levels of the parameterisation. This is an “inf er to infer” paradigm analogous to the “learning to learn” paradigm commonplace in the machine learning literature. A non-stationary trans-dim ensional Gaussian Process (TDGP) is used to parameterise earth properties (e.g.\, resistivity)\, and a multi-channel stationary TDGP is used to para meterise the length scales of the top-level parameterisation. Using non-st ationary GP kernels and efficient approximations to update GP kernel matri ces\, sharp discontinuities as well as smooth regions can be represented w ithin one earth model. As GPs are multi-dimensional interpolators\, the sa me theory and computer code can be used to solve geophysical problems in 1 D\, 2D and 3D.\n LOCATION:https://researchseminars.org/talk/RSES_Geophysics/1/ END:VEVENT END:VCALENDAR