1D convolutional neural network with stacked bidirectional long short-term memory for seismic impedance inversion

Shang Huang, Paulina Wozniakowska, Marcelo Guarido, David J. Emery, Daniel O. Trad

The seismic impedance inversion problem is ill-posed and nonlinear because of insufficient data, and is limited by wavelet estimation and frequency band-limited data. A machine learning long short-term memory algorithm (LSTM) can capture long-term dependencies so that it can work with long and densely sampled well log data to eliminate these limitations and take advantage of the known rock physics trend with depth. In this work, two models including the stacked bidirectional long short-term memory (SBDLSTM) recurrent neural network, and 1D convolutional neural network (CNN) with stacked BDLSTM have been applied to the inverse problem P-impedance and S-impedance calculation. Near, mid, far offset seismic data, migration velocity and well log data attributes are provided to generate the training set. Extreme gradient boosting (XGBoost) is used as the baseline model for comparison. Results show that SBDLSTM can predict impedance more accurately than the XGBoost method in some rapidly changing layers. 1D CNN with stacked BDLSTM can also calculate a high-frequency impedance prediction with fewer artifacts. The promising aspect is that both SBDLSTM and 1D CNN with SBDLSTM approaches can maintain a good fit when given a small number of training datasets.