Physics-guided self-supervised learning for low frequency data prediction in FWI

Abstract

In many geophysical machine learning applications, labeled data are either too expensive or impossible to collect. In this work, we developed a physics-guided self-supervised learning method to predict the absent low frequency (LF) components in acquired seismic data to overcome the cycle-skipping issue in full waveform inversion (FWI). Similar to self-supervised learning in Natural Language Processing (NLP), this physics-guided label-free learning method consists of two stages: the pretext task and the downstream task. For the pretext task, we proposed two algorithms, the pseudo-LF (PLF) method and the sparse inversion-based bandwidth-extension method (BWE), to automatically generate the pseudo-labels from the input high frequency (HF) data attributes. Once the pre-designed deep learning network (DNN) model, a variant of the InceptV4 network, is pretrained within the pretext task stage, the learned knowledge is transferred to the second stage - the downstream task, where the Progressive Transfer Learning strategy is employed for fine-tuning the model in an iterative manner. Once again, in the downstream task stage, labels are automatically generated and updated by a physics-guided inversion procedure, rather than being externally and manually labeled. In the beginning of the workflow, the pseudo-LF or the bandwidth-extension algorithm is implemented to approximately estimate the missing LF components, which are then injected into an FWI engine to reconstruct a velocity model. A forward modeling simulation is performed on this velocity model to generate a training set for the pretext training task. In the downstream task, the LF data predicted by the pretrained DNN are sent to the progressive transfer learning pipeline to be inverted by the FWI engine, producing an improved velocity model and a more representative training set to replace the previous one. The downstream task is performed multiple times iteratively to gradually fine-tune the DNN model by retrieving more and more subsurface information through the physics-based inversion. The numerical experiments demonstrate the high accuracy of the final predicted LF data and the effectiveness of cycle-skipping suppression. Presentation Date: Wednesday, October 14, 2020 Session Start Time: 9:20 AM Presentation Time: 11:25 AM Location: Poster Station 3 Presentation Type: Poster

Topics & Concepts