Generalization capability of data-driven deep learning models for seismic full waveform inversion: an example using the OpenFWI dataset

Such geophysical fields as reservoir modelling, oil and gas exploration, CO2 sequestration, require robust subsurface velocity models. The state-of-the art approach to velocity inversion is the full waveform inversion (FWI) method. While FWI can lead to high-resolution velocity inversion, there are several challenges associated with the method. A number of data-driven deep learning approaches have been developed in the recent past that seemingly address these limitations. Therefore, we study the generalization capability of such data-driven FWI methods and demonstrate that a purely data-driven supervised learning is not sufficient for generalization to more complex and realistic models. Through synthetic tests, we show that the supervised data-driven methods suffer from noticeable degradation when tested with unseen data. Based on our findings, we argue that addition of the underlying physics is necessary to improve generalization capabilities of deep learning-based FWI models.