Time-Lapse Seismic Inversion Using a 3D Neural Operator

Time-lapse (4D) seismic inversion is a key tool for monitoring subsurface changes caused by fluid injection and production, but model-based workflows require repeated forward and adjoint evaluations, a wavelet estimate for every vintage, and survey-specific regularization tuning. We propose a 3D neural operator framework for post-stack time-lapse impedance inversion. A U-shaped neural operator is trained on synthetic impedance models and their convolutional post-stack responses and serves as a mesh-independent, differentiable surrogate of the post-stack forward operator. During inversion, the operator is frozen, and the impedance model is updated by gradient descent with automatic differentiation using one set of settings for all vintages of a survey. On the Kimberlina synthetic CO${}_2$ model, the inversion reduces the error of the initial model for every vintage (relative error from 3.1–3.2% to 2.8–2.9%). On the Sleipner field data, the recovered impedance changes are concentrated inside an independently mapped CO${}_2$ plume (median change 4.7–7.3 times larger than outside), the predicted 4D data difference correlates with the observed one at $r=0.92$–$0.93$ inside the plume, and the operator is applied without retraining to the full $1001\times 468\times 249$ survey grid. We also quantify the training and fine-tuning costs and discuss the limits of a learned convolutional surrogate.