Anomalous-diffusion synthesis of non-Gaussian reservoir anomalies for time-lapse seismic inversion
We develop a physics-informed framework for learned time-lapse (4D) seismic inversion that estimates production-induced velocity changes associated with subsurface fluid migration. Existing learned inversion methods are commonly trained on seismic data generated from Gaussian diffusion models, which fail to capture the anomalous transport characteristic of heterogeneous subsurface reservoirs. To overcome this limitation, we introduce a data generator based on space-time fractional diffusion, in which the temporal and spatial fractional orders independently control memory effects and nonlocal transport, while elliptical anisotropy represents preferential flow. By sampling this parameter space, we generate a broad ensemble of physically consistent velocity perturbations spanning normal, subdiffusive, superdiffusive, and intermediate transport regimes. We then train a convolutional neural network to map time-lapse seismic residuals directly to velocity updates. Through extensive Monte Carlo validation, we show that the network accurately reconstructs production-induced anomalies across all transport regimes, with reconstruction errors governed primarily by anomaly amplitude and interface sharpness rather than by the diffusion regime itself. Our results demonstrate that embedding anomalous-transport physics into the training distribution substantially extends the applicability of amortized 4D seismic inversion, enabling robust recovery of heavy-tailed and anisotropic reservoir signatures beyond the Gaussian assumptions underlying existing approaches.