Study on One-dimensional Bayesian Inversion of Magnetotelluric Data

Abstract:

Bayesian inversion is a non-linear inversion method, which is based on the probability estimation theory of statistics. The model parameters such as observation data and measurement errors are regarded as random variables for inversion, and the posterior probability distribution of the model parameters is obtained. In this way, not only the optimal model (MAP, the maximum of posterior probability distribution) can be obtained, but also the uncertainty information of the inversion results can be obtained, so as to understand the reliability of the inversion results.

Firstly, the paper expounds the principle and inversion process of linear inversion theory and Bayesian inversion theory, and then studies the classical Markov Chain-Monte Carlo sampling method (MCMC) in Bayesian inversion. In the research process, we first elaborated the principles of Monte Carlo method and Markov Chain, and then combed the flow of MCMC sampling and Metropolis-Hasting (MH) sampling methods. After that, we do Bayesian inversion and Occam inversion respectively for the same data, and compare the advantages and disadvantages of the two algorithms. Finally, aiming at the problem of low sampling efficiency in high-dimensional model space in MCMC algorithm, we design MCMC algorithm to improve sampling efficiency by reducing the dimension of model space. We further use this method to invert real MT data collected from the northern Tibetan Plateau, the inversion results will be presented and discussed during the presentation on the fall meeting.

* This study is funded by National Natural Science Foundation of China (41774087).