Seismic poststack impedance inversion using geophysics-informed deep-learning neural network

Abstract Traditionally, seismic poststack impedance inversion is implemented using linear optimization algorithms. Recently, deep-learning neural networks have been successfully used to estimate the impedance from seismic data. First, we determine the general workflow of seismic poststack impedance inversion using supervised neural network (SNN). Next, we work to compute seismic impedance using geophysics-informed neural network (GINN). Similar to linear optimization algorithms, the inputs of GINN include real seismograms, a wavelet, and a low-frequency model. The loss function of GINN is designed to minimize the difference between real seismograms and synthetic seismic seismograms that are computed from the estimated impedance and input wavelet. To avoid lateral discontinuity of estimated impedance, the weights of GINN are trained using the seismograms of all seismic traces. We use synthetic and real seismic data to discuss the advantages and limitations of GINN, SNN, and traditional linear optimization algorithms. Not surprisingly, the signal-to-noise ratio (S/N) of seismic data and the phase of seismic wavelet are the most important factors that affect the accuracy of impedance estimated using GINN. The accuracy and resolution of impedance estimated using GINN is higher than linear optimization and SNN if the seismic data have a high S/N. The synthetic examples demonstrate that the accuracy of impedance calculated using SNN increases with the number of available training wells and linear optimization algorithms are more robust to noise than GINN and SNN.