A deep full-sequence convolutional neural network for 3D sedimentary microfacies prediction using time-frequency seismic data

To address the challenges of limited well control and the low accuracy of 3D sedimentary microfacies modeling using seismic data, this study proposes a prediction workflow based on a time-frequency deep full-sequence convolutional neural network (DFCNN). Building upon continuous wavelet transform for time-frequency decomposition, which expands the seismic information dimension and reduces interpretation uncertainties, a large, accurately labeled time-frequency spectrum dataset was constructed for sedimentary facies. The loss function, network architecture, and key parameters of the DFCNN were optimized, enabling accurate prediction of 3D microfacies. The method was applied to the Jurassic Sangonghe Formation in the MXZ area. The results show that the predicted 3D sedimentary microfacies agree well with the drilling data, achieving an accuracy of 82.6%, which demonstrates the applicability of the proposed workflow for sedimentary facies prediction.