Unsupervised learning inversion of seismic velocity models based on a multi-scale strategy

Abstract

Deep learning-based methods have performed well in seismic waveform inversion tasks in recent years, while the need for velocity models as labels has somewhat limited their application. Unsupervised learning allows us to train the neural network without labels. When inverting seismic velocity models from observed data, labels are often unavailable for real data. To address this problem and improve network generalization, we introduce a multi-scale strategy to enhance the performance of unsupervised learning. The first ‘multi-scale’ is derived from the conventional full waveform inversion strategy, in which the low-, middle- and high-frequency inversion results are successively predicted during the network training. Another ‘multi-scale’ is to introduce multi-scale similarity as an additional data loss term to improve the inversion results. With 12,000 samples from the overthrust model, our method obtains comparable results with the supervised learning method and outperforms unsupervised methods that rely only on the mean square error as a loss function. We compare the performance of the proposed method with multi-scale full waveform inversion on the Marmousi model, and the proposed method achieves better results at low- and middle-frequencies, and, as a result, it provides good initial models for further full waveform inversion updates.