Machine learning methods modeling waveform, multi-parameter full waveform inversion,and uncertainty quantification

Tianze Zhang

Full waveform inversion (FWI) is a potent technology capable of estimating subsurface parameters using seismic records. However, several challenges hinder its widespread application in large-scale field operations. Methods in machine learning, for instance, multiple layers perception(MLP), recurrent neural networks(RNN), Bayesian neural networks(BNN), and Fourier neural operators(FNO), have some appealing features. They can straightforwardly approximate a broad range of mathematical models and phenomena, making them easy to extend and adapt. Furthermore, machine learning algorithms tend to be straightforward to parallelize. These features might help us resolve the issues faced with FWI. In this thesis, I introduce strategies that leverage machine learning techniques to address the challenges faced with FWI, making FWI more feasible to implement in complex media and providing a confidence analysis for the inversion results. A primary concern is the necessity for accurate initial models in FWI; without these, FWI risks becoming ensnared in local minima. In this thesis, I propose recurrent neural network(RNN) isotropic elastic FWI. To verify the advantages of the RNN-based FWI, I calculate FWI using a list of objective functions, analyzing how the variation of the elastic parameters would influence the shape of the objective function (local minima) by using a toy model. Then, I proposed the elastic implicit full waveform inversion. Instead of directly updating the elastic parameters like in the conventional FWI, I use neural networks to generate elastic models and update the weights in the neural network to decrease data misfits. Numerical tests suggest that the implicit full waveform inversion can generate promising inversion results without using accurate initial models. Furthermore, how to develop a feasible uncertainty quantification method for FWI outcomes remains an open question. The estimation of the prior uncertainty and a method that effectively evaluates the inverse Hessian matrix are critical steps for the uncertainty quantification under the Bayesian inference. I introduce a method that uses the Bayesian neural network (BNN) to provide prior uncertainty for the elastic models and an algorithm that efficiently approximates inverse Hessian. I also consider the uncertainty under different convergence histories.During the inversion process, simplifications in wave propagation physics are often made to enhance computational efficiency. Yet, the extent to which such simplifications (modelling errors) impact the inversion results is seldom addressed. I proposed a method for the viscoelastic FWI that can quantify a specific type of modelling error, which can quantify the modelling error caused by the insufficient ability of the relaxation variables to model the constant Q model or a Q model that we desire, i.e., obtained from the field records. I also analyze the effect on the inversion results of such modelling errors.Our community is strongly motivated to find means of accurately computing wavefields with minimal computational expense. I propose the one-connection Fourier neural operator(OCFNO). I test the ability of such a network to “learn” to solve elastic wave equations. Computational speed-ups are significant: if the network is well trained, the computational speed for generating the wavefields is about 100 times that of traditional finite difference methods. Numerical validation of the networks for their ability to determine wavefields in both the time and frequency domains suggests that OCFNO outperforms the conventional Fourier neural operator structures in the accuracy of generating elastic wavefields.