Joint Time-Lapse Seismic Impedance Inversion via Closed-Loop Deep Learning

Time-lapse seismic inversion estimates reservoir property changes from multi-vintage seismic data and thereby supports reservoir monitoring. Deep learning has been widely applied to seismic inversion due to its strong nonlinear modeling capability. However, supervised methods generally require large labeled datasets, and the cost of acquiring well-log data limits their use in time-lapse inversion. To reduce this dependence on labeled data, we extend existing closed-loop learning approaches to time-lapse elastic impedance inversion and develop a framework termed 4D-SeisInv Net. The framework comprises a seismic forward modeling network and two inversion networks that estimate the baseline elastic impedance and the time-lapse elastic impedance change. The estimated baseline impedance is used as conditioning information for predicting the time-lapse impedance change, and we jointly optimize the two inversion tasks under a unified objective function. The forward network further maps the inversion results back to the seismic domain, allowing unlabeled seismic data to participate in network training. Experiments with synthetic data show that, under the tested noise conditions, the proposed framework achieves higher inversion accuracy and greater robustness than conventional model-driven and supervised learning methods. For the field data, convolutional forward modeling of the inverted impedances yields seismic responses consistent with the observations, supporting the physical plausibility of the estimated time-lapse impedance changes.