Geophysical inversion is challenging because realistic models of the Earth are nonlinear and their corresponding inverse problems are ill-posed due to non-existent or nonunique solutions that are sensitive to small changes in the data. There is now a significant body of work in geophysics that uses neural networks to represent an inverse operator. It can be difficult to obtain sufficient observational data sets to train a neural network because field data acquisition is both time consuming and costly. Therefore it has become common to create training data with forward models that capture physical laws, empirically validated rules or other domain expertise. In this work we discuss multiple parameter sample strategies based on Bayesian inference. Sampled parameters are propagated through a nonlinear forward model and the resulting set of inputs and outputs are used as labeled training data in supervised learning. Trained surrogate models are then used to invert a set of observations and obtain the desired geophysical parameters. Results will be shown on a one-dimensional nonlinear ill-posed inverse problem.
Training Data for Geophysical Inversion
Jodi Mead, Boise State University
2023 AWM Research Symposium
Computational Inverse Problems and Uncertainty Quantification [Organized by Julianne Chung, Rosemary Renaut, and Malena Sabate-Landman]