De-risking CO₂ sequestration in sandstone reservoirs using frequency-dependent AVO attributes for enhanced fluid detection and seismic monitoring

Abstract

Accurate seismic monitoring of subsurface CO₂ storage fundamentally depends on capturing frequency-dependent dispersion and attenuation phenomena introduced by squirt flow under partial CO₂ saturation. Traditional seismic forward-modelling approaches, which predominantly rely on low-frequency Gassmann–Biot theory, typically neglect these frequency-dependent squirtflow effects. This omission introduces uncertainties in quantifying velocity dispersion and attenuation, consequently impacting seismic amplitudes and AVO responses that are critical for reliable reservoir characterization and effective CO₂ plume monitoring.

To address these limitations, this thesis integrates a frequency-dependent rock-physics modelling approach with deep learning techniques applied to the Jurassic Sognefjord Formation at the Smeaheia β-structure, located within the Norwegian North Sea. Elastic and petrophysical log data from the Trow 32/2-1 well are initially conditioned and upscaled to seismic resolution. Subsequently, the Papageorgiou & Chapman (2017) hybrid squirt-flow model is systematically applied to simulate the CO₂-induced dispersion and attenuation effects across relaxed, dispersive and unrelaxed frequency regimes. These simulations explore the sensitivity of the seismic attributes to varying degrees of saturation (𝑆ᵥᵥ), capillary pressure (𝑞) and seismic attenuation (𝑄⁻¹), allowing for the evaluation of frequency-dependent reflectivity across a comprehensive range of synthetic scenarios and offset angles.

The squirt-flow modelling results demonstrate notable frequency-dependent compressional wave softening (ΔVₚ reaching up to –25%) and reductions in the compressional modulus (Δ𝑃Mod up to –46%) at intermediate to full CO₂ saturations under conditions of high attenuation (𝑄=10), in the dispersive state (𝑊꜀= 0), relative to the 100% brine baseline (𝑄=1000, 𝑊꜀= –2). To illustrate these combined effects, a frequency-dependent amplitude-versus-offset (FAVO) rock-physics template was developed from the forward modelled scenarios. This template highlights increased amplitude sensitivity at lower frequencies (≤ 20 Hz) and larger offsets (30°– 40°), conditions that clearly discriminate higher CO₂ saturations, particularly under elevated attenuation levels (𝑄= 10). Furthermore, larger capillary pressure contrasts (𝑞= 0.054) moderate these amplitude responses, reducing magnitudes and flattening AVO gradients, yet still preserving distinct signatures indicative of CO₂ saturation and reservoir pressure variations. Nevertheless, the conducted feasibility study, integrating frequency-dependent reflectivity modelling, spectral decomposition and FAVO analysis, revealed that squirt-flow-related dispersive changes manifest as seismic-scale phase shifts and amplitude variations. However, the precise quantification of such CO₂-induced signatures is complicated by seismic interference from the combined effects of thin-bed tuning and frequency-dependent squirt-flow, which together mask the squirt-flow attenuation effects expected in spectral decomposition while exaggerating AVO responses in the higher-frequency range (60–90 Hz).

Given these interpretational challenges, frequency-dependent spectral amplitudes were subsequently integrated into a two-dimensional convolutional neural network (2D CNN) as a proposed mitigation strategy. This CNN was designed with multiple parallel classification outputs, enabling simultaneous inversion for CO₂ saturation, capillary pressure, and attenuation parameters. After hyperparameter optimization, the CNN demonstrated excellent predictive performance, achieving an average accuracy of 96.99%, a holistic classification accuracy of 92.6%, and an average F₁-score of 96.92% across the different 𝑆ᵥᵥ, 𝑞, and 𝑄⁻¹ attribute classifiers. The achieved performance indicates the model’s capability in effectively distinguishing between the CO₂-induced responses and interference-related complexities, emphasizing the potential use of frequency-dependent spectral amplitudes for accurate reservoir attribute predictions.

Hence, this thesis ultimately demonstrates that integrating frequency-dependent rock-physics modelling with advanced machine-learning inversion techniques enhances the seismic monitoring and characterization of CO₂ plumes within sandstone reservoirs.