Using Deep Learning for Data-Driven Seismic Velocity and Source Wavelet Inversion

Full-waveform inversion (FWI) requires an accurate representation of the source wavelet to enhance the reliability of model reconstruction. Traditional wavelet estimation methods depend on the accuracy of the existing velocity model. We propose an approach for the joint estimation of velocity and source wavelet in a data-driven manner. To achieve this, a dataset with randomly generated pairs of velocity models and source wavelets is created. Each wavelet is injected into its respective velocity model to simulate seismic records. A convolutional neural network (CNN) is then trained to predict both velocity and wavelet features from the resulting shot gathers. The trained CNN effectively mitigates the bias in velocity estimation caused by an inaccurate source wavelet. By jointly estimating velocity and wavelet, the method exhibits improved seismic velocity reconstruction compared with a network that predicts velocity alone. The approach demonstrates potential for producing the input necessary to enable high-resolution FWI-based imaging on both synthetic and shallow seismic field datasets.