Enhancing machine learning based seismic inversion with noise-augmented training datasets

SUMMARY An 18-layer, U-shaped convolutional neural network was trained to predict Vs models and identify near-surface void locations. To enhance seismic inversion accuracy for real world applications, the model is trained on synthetic data sets augmented with field noise. While models trained on noise-augmented data showed poorer performance on synthetic testing data sets, they achieved lower root-mean-square error values and the best results on field data. The velocity model resulting from full-waveform inversion based on noise-augmented model accurately detected a void-like low-velocity zone near the known void location. This approach shows that training with field-noise-augmented data allows machine learning models to generalize better to real-world conditions, increasing their reliability for velocity inversion in noisy environments. The results highlight the strong potential of this strategy, particularly if a diverse range of real noise samples is incorporated during training.