**A seminar by Associate Professor Jan Dettmer from University of Calgary**

**Title: **The role of model selection and noise estimation when combining different data types in geophysical uncertainty quantification

**Abstract: **Geophysical inverse problems suffer from non-uniqueness. That is, individual physical data types cannot uniquely determine earth properties and processes. However, the information contained in data from multiple physical processes interacting with the subsurface can provide complementary information to reduce uniqueness issues. Such multiphysics inference of geophysical observations can improve the resolution of Earth structure and processes but is plagued by many subjective choices that practitioners are commonly required to make. Two main requirements are the specification of Earth models and specification of weights for data types. We will present the method of probabilistic multiphysics inference that employs Bayesian statistics to provide model selection and appropriate data weights. The method will be illustrated for two cases: (1) estimation of earthquake source parameters and (2) quantifying elastic earth properties. For case 1, we consider multiple seismic data types that are combined to improve the resolution of earthquake moment tensors. The data types include seismic waveforms, polarities of the first motions, and amplitude spectra of waveforms. To combine the different data types, data covariance matrices of the noise on each data type are estimated during Markov chain Monte Carlo sampling. This approach permits earthquake source characterization for data with poor signal to noise ratios. For case 2, we consider combining seismic and electromagnetic data to constrain elastic Earth properties. Since electromagnetic and seismic data sense the subsurface on different scales, the layering structure of the subsurface is estimated with flexible coupling, where the number of homogeneous layers is treated as unknown and the number of geophysical parameters for each layer are unknown. Hence, parameters for different physical processes are not required to share the same layering structure, which avoids over-parametrization. We demonstrate advantages for the near-surface characterization of permafrost soils (10s of metres in depth) and for the imaging of the lithosphere (100s of kilometres in depth).

For further information, please contact RSFAS Seminars.

All information collected by the University is governed by the ANU Privacy Policy.