oa A Robust Self-Supervised Learning System for Low Frequency Seismic Data Extrapolation to Reduce Model Building Uncertainty

In order to suppress the cycle-skipping phenomenon in full waveform inversion (FWI) and reduce the uncertainty resulting from velocity model building practices, we developed a Deep Neural Network (DNN) based Self-Supervised Learning method to predict the low frequency (LF) seismic data by exploiting the implicit relationship connecting the LF data and the high frequency (HF). Employing the Progressive Transfer Learning strategy that seamlessly integrates a physics-based module and a DNN module, this Self-Supervised Learning system does not require access to labelled data, which are often impractical to collect in realistic projects. Instead, the training datasets are generated and automatically labelled within the learning system. Furthermore, this Self-Supervised Learning system is able to iteratively evolve the training set and update the DNN by gradually retrieving the subsurface information through the physics-based module to enhance the LF data prediction accuracy, thus propelling the FWI process out of the local minima. Throughout the workflow, the uncertainty of the LF extrapolation is rigorously monitored and quantified by the Self-Supervised Learning system.