Multiphysics inversion with variable complexity of receiver-function, surface-wave dispersion and magnetotelluric data reduces uncertainty for lithosphere structure

Abstract

We present a probabilistic multiphysics inversion based on Bayesian inference with trans-dimensional models. We jointly consider magnetotelluric, receiver function, and Rayleigh-wave dispersion data to infer one-dimensional lithospheric structure in the vicinity of Athabasca, Canada. The location is on the North American Craton with a cover of sediments from the Western Canada Sedimentary Basin. The trans-dimensional model uses layer nodes that include parameters that are activated or deactivated based on data information. Furthermore, the number of nodes is based on data information. Hence, the parameterization uncertainty is included in the uncertainty estimates. Furthermore, the layer nodes permit trans-dimensional decoupling such that some discontinuities may be represented by only some of the parameters. In probabilistic multiphysics inversion, it is important that the various data types are weighed objectively. Here, the weights are the data covariance matrices of the various data types. We apply empirical estimation of data covariance matrices and employ hierarchical scaling parameters to reduce the dependence on some assumptions required by the empirical approach. Hence, we account for noise variances and covariances, which is crucial for successful probabilistic multiphysics inversion. The parameter estimates and data covariance matrices are obtained with the reversible-jump Markov chain Monte Carlo algorithm with parallel tempering to enhance the efficiency. Since covariance matrix estimation changes data weights, the estimation process is carried out while samples are not recorded for inference. The results at the Athabasca site fit the data and produce plausible data covariance matrices for the data weights.