Calibrated squirt-flow analysis of Mississippian dolomites and deep-learning CO₂ plume segmentation from frequency-decomposed seismic data

Abstract

Carbonate reservoirs exhibit complex pore architectures, including moldic, vuggy, intercrystalline, and fracture porosity, that produce wide-ranging elastic-modulus-porosity relationships and frequency-dependent velocity dispersion. Standard Gassmann fluid-substitution workflows, which assume pore-pressure equilibration in the low-frequency limit, are insufficient to describe the measured ultrasonic response of these rocks, where squirt flow through compliant microcracks can generate measurable stiffening of the saturated frame. This study presents a calibrated Mavko-Jizba squirt-flow analysis of five dolomite core plugs from the Mississippian carbonate interval of the Wellington Field, south-central Kansas. Compressional- and shear-wave velocities and quality factors were measured at eight ultrasonic frequencies (156.3 kHz to 20 MHz) under three fluid-saturation states: brine, partial brine, and oil-and-brine. A constrained inversion procedure employing differential-evolution global search followed by L-BFGS-B refinement recovers a hierarchical parameter set separating global mineral properties, sample-specific frame properties, and saturation-dependent squirt-flow parameters. The calibrated model reproduces both V_P and V_S across the analyzed frequency band with tight clustering around the 1:1 parity line and no systematic bias. Attenuation predictions capture the first-order behavior of the measured quality factors but display wider scatter, reflecting both the intrinsic uncertainty of laboratory Q estimation and unresolved pore-scale heterogeneity. Attenuation contrasts between saturation states exceed the corresponding velocity contrasts, indicating that Q-based attributes provide greater fluid sensitivity than velocity in this low-porosity dolomite dataset. Porosity values recovered from the calibration are broadly consistent with independent laboratory estimates. Leave-one-sample-out cross-validation, performed with global parameters held fixed, indicates internal stability within the five-sample dataset and no obvious overfitting across the sampled heterogeneity. The results provide a calibrated laboratory workflow for applying the Mavko-Jizba framework to tight carbonate data and establish new experimental constraints on effective squirt-flow parameters in dolomitic lithologies. In a complementary field-scale investigation, a deep-learning pipeline is developed for automated CO₂ plume segmentation from frequency-decomposed seismic amplitude-difference maps at the Sleipner storage site, North Sea. An ensemble of encoder-decoder architectures (UNet++, MAnet, DeepLabV3+, FPN) trained on eleven spectral bands achieves a mean Dice coefficient of 0.937 and mean intersection-over-union of 0.882 under leave-one-out cross-validation, demonstrating that frequency decomposition improves plume delineation relative to full-bandwidth data alone. Together, the two studies illustrate the diagnostic value of frequency-dependent seismic attributes from laboratory ultrasonic measurements through field-scale monitoring.