betaFeedback

WorksPhysics and Astronomy

Bayesian inversion of multilayer CO2\mathrm{CO}_2 migration from seismic plume observations using a graph-based finite-rate invasion-percolation model

Abstract

Vertical migration of CO2\mathrm{CO}_2 in layered sandstone reservoirs is controlled by thin shale barriers whose properties are often poorly known. We develop a Bayesian framework that uses time-lapse seismic plume observations to estimate effective parameters governing lateral and vertical CO2\mathrm{CO}_2 migration. The forward model is a fast graph-based invasion-percolation model that extends conventional IP by representing both capillary-controlled filling of structural traps beneath shale barriers and finite-rate transfer through them. Parameters are inferred with approximate Bayesian computation and sequential Monte Carlo sampling, and the posterior samples are propagated into forecasts. Applied to real data from Sleipner, posterior simulations reproduce the broad distribution of CO2\mathrm{CO}_2 across nine sand units and its redistribution between 2010 and 2023. In contrast, the quasi-static model fails to reproduce this temporal evolution. Complementary synthetic experiments assess parameter recovery, forecasting, monitoring duration, and model misspecification. These experiments show that the information gained from monitoring depends on the migration events captured, with breakthrough and post-breach redistribution providing particularly strong constraints. This combination of fast simulation, probabilistic updating, and interpretable effective parameters makes the framework well suited to repeated forecast revision during active injection, especially when full-physics inference is too computationally demanding.

Introduction

The continued use of fossil fuels has increased CO2_2 emissions and remains a major driver of climate change. Carbon capture and storage (CCS) offers a means of reducing industrial emissions by capturing CO2_2, transporting it to a storage site, and injecting it into deep geological formations. The effectiveness and safety of CCS depends on the injected CO2_2 remaining contained within the storage complex. However, inferring subsurface CO2_2 migration can be challenging because subsurface properties are often poorly characterized, while time-lapse seismic data may be noisy or have limited resolution [1, 2]. A probabilistic framework is therefore needed to quantify uncertainty in both the inferred subsurface properties and the resulting future plume evolution.

Thin intra-formational barriers, such as shales, are an important source of uncertainty in layered storage reservoirs. These barriers strongly influence the lateral and vertical migration of CO2\mathrm{CO}_2 between permeable sand units. Their capillary properties depend on pore structure, interfacial tension, and wettability, which are difficult to determine because direct subsurface measurements are sparse [3, 4]. Furthermore, they may contain localized migration pathways where the CO2\mathrm{CO}_2 flows more easily [5]. As a result, the onset and rate of CO2\mathrm{CO}_2 migration between reservoir layers may remain uncertain even when the larger-scale reservoir geometry is well characterized.

These uncertainties can be reduced by conditioning flow models on monitoring observations through history matching and data assimilation [6, 7, 8, 9]. However, such approaches typically require many forward simulations, making computational cost a central challenge. Full-physics multiphase simulators capture the relevant flow processes, but can be too expensive to use within Monte Carlo-based inference. Surrogate, reduced-order, and multifidelity models therefore provide attractive alternatives by trading some physical detail for substantially lower computational cost [10, 11, 12, 13].

For buoyancy- and capillary-dominated migration away from the injection well, invasion percolation (IP) provides an appropriate reduced-physics representation of CO2\mathrm{CO}_2 migration. It represents migration through capillary invasion and buoyancy-driven filling and can reproduce the main plume pathways at much lower computational cost than full-physics simulation [14, 15]. It has already been used to investigate the influence of multilayer shale properties at Sleipner [3, 16]. Yet, conventional IP models are quasi-static and treat migration across a breached shale as instantaneous. They therefore cannot represent the gradual inter-layer transfer needed to explain the continued evolution of lower accumulations after CO2\mathrm{CO}_2 has reached overlying units [17, 18].

To address the need for a realistic representation of CO2\mathrm{CO}_2 migration that can be used for inference and forecasting, this paper develops a fast graph-based, multilayer IP simulator. Spill-point analysis is used to reduce each reservoir surface to a network of connected traps and spill paths, which is computed once and subsequently reused throughout the parameter estimate. Finite-rate vertical migration across shales is controlled by effective threshold pressures and post-breakthrough mobility parameters, which govern the onset and rate of vertical transfer, respectively. These parameters are inferred from time-lapse seismic plume outlines using approximate Bayesian computation with sequential Monte Carlo sampling (ABC-SMC). Forecasts of future plume evolution are then generated by propagating the inferred parameter uncertainty through the simulator, giving a probabilistic representation of the plume distribution and its uncertainty.

The framework is evaluated using both real-world observations and a controlled synthetic model. The Sleipner CO2\mathrm{CO}_2 storage site serves as a test case for assessing whether the reduced model can be conditioned on monitoring data to reproduce the observed plume distribution. It also enables us to investigate how uncertainty in shale topography propagates through the model and affects the inferred plume evolution. Complementing the Sleipner case, we consider a synthetic four-layer reservoir for which the true parameters and topographies are known. This controlled setting allows us to systematically assess parameter recovery, forecast uncertainty, the influence of monitoring duration, and the consequences of model misspecification.

The paper makes three main contributions.

  • An efficient graph-based formulation of multilayer IP simulator with finite-rate migration across intra-formational barriers. In the quasi-static limit, it is about 40 times faster than a cell-by-cell implementation on the reported benchmark.

  • A likelihood-free Bayesian treatment of uncertain barrier properties using time-lapse seismic plume outlines as observations.

  • A systematic assessment of how well the inferred model predicts plume evolution under both real-world and synthetic conditions.

Invasion-percolation model

Physical assumptions and capillary breakthrough

IP theory is used as a reduced-order model of far-field CO2\mathrm{CO}_2 migration, and represents migration as a sequence of capillary invasion and structural filling events [14, 15, 19]. By avoiding the solution of large systems of coupled nonlinear equations required by full-physics models, it enables efficient simulation of plume evolution. The approximation is appropriate when viscous pressure gradients are small relative to capillary forces. The relative importance of viscous and capillary forces is characterized by the capillary number

Ca=μUγ,\mathrm{Ca} = \frac{\mu U}{\gamma},

where μ\mu is a representative fluid viscosity, UU is the fluid velocity, and γ\gamma is the CO2\mathrm{CO}_2-brine interfacial tension. A limit of approximately Ca≈10−4\mathrm{Ca} \approx10^{-4} is commonly used to identify capillary-dominated migration [3]. For the Sleipner reservoir, Cavanagh and Haszeldine [3] estimate a capillary number of about 10−710^{-7}, which is well below the required threshold.

We consider NN permeable sand units S1,…,SN\mathcal{S}_1,\ldots,\mathcal{S}_N, ordered from deepest to shallowest. Let Bi\mathcal{B}_i denote the shale barrier immediately above Si\mathcal{S}_i. For i<Ni < N, Bi\mathcal{B}_i separates Si\mathcal{S}_i from Si+1\mathcal{S}_{i+1}. At the pore scale, the pressure required for a non-wetting fluid to enter a pore throat is

Pth(r)=2γcos⁡ξr,P_{\mathrm{th}}(r) = \frac{2\gamma\cos\xi}{r},

where ξ\xi is the contact angle measured through the brine and rr is the pore-throat radius. For a water-wet system, the smaller pore throats of shale result in a higher entry pressure than in the surrounding sand [20, 21]. Migration between sand units is controlled by the capillary entry pressure of these barriers.

In the reduced model, each barrier Bi\mathcal{B}_i is assigned a single effective threshold pressure Pth,iP_{\mathrm{th},i}. CO2\mathrm{CO}_2 accumulates beneath the barrier until the buoyancy pressure at the top of the accumulation exceeds this threshold. Within each connected accumulation, we assume vertical pressure equilibrium and a horizontal gas-water contact. Let hi,j,th_{i,j,t} denote the local CO2\mathrm{CO}_2 column height in horizontal grid cell jj beneath barrier Bi\mathcal{B}_i at time tt. The corresponding buoyancy pressure at the top of the column is Δρighi,j,t\Delta\rho_i g h_{i,j,t}, where gg is gravitational acceleration and Δρi\Delta\rho_i is the density difference between brine and CO2\mathrm{CO}_2 in layer ii. Capillary breakthrough becomes possible when Δρighi,j,t≥Pth,i\Delta\rho_i g h_{i,j,t} \geq P_{\mathrm{th},i}, which defines the critical column height

hcrit,i=Pth,iΔρigh_{\mathrm{crit},i} = \frac{P_{\mathrm{th},i}}{\Delta\rho_i g}

As injection proceeds, CO2\mathrm{CO}_2 therefore accumulates within a sand unit according to its geometry until the critical column height is reached beneath the overlying barrier. Once breakthrough occurs, CO2\mathrm{CO}_2 migrates into the next sand unit, where a new accumulation develops until the next barrier is breached. Repetition of this process produces a sequence of CO2\mathrm{CO}_2 accumulations across the permeable units.

Finite-rate transfer across shale barriers

In conventional quasi-static IP, transfer through a breached barrier is assumed to be instantaneous. Subsequent CO2\mathrm{CO}_2 supply is therefore passed directly to the overlying unit, and the lower accumulation cannot grow beyond its critical column height. This approximation is appropriate when the timescale of vertical transfer is short relative to that of lateral migration [15]. However, at reservoirs such as Sleipner, lower accumulations continue to evolve after vertical migration has begun [17, 18]. This behavior is inconsistent with instantaneous transfer and suggests that the vertical transfer timescale is not negligible. We therefore retain quasi-static redistribution within each sand unit, but introduce a finite transfer rate across the shale barriers.

Let Li,jL_{i,j} be the thickness of BiB_i in cell jj. We treat a breached part of the barrier as an effective Darcy-like connection, with flow driven by the buoyancy pressure in excess of the entry pressure. The volumetric transfer rate through cell jj is then

Qi,j(t)=λAjLi,j[Δρighi,j(t)−Pth,i]+=λΔρigAjLi,j[hi,j(t)−hcrit,i]+,(1)Q_{i,j}(t) = \frac{\lambda A_j}{L_{i,j}} \left[\Delta\rho_i g h_{i,j}(t) - P_{\mathrm{th},i}\right]_+ = \frac{\lambda\Delta\rho_i g A_j}{L_{i,j}} \left[h_{i,j}(t) - h_{\mathrm{crit},i}\right]_+, \tag*{(1)}

where [x]+=max⁡(x,0)[x]_+ = \max(x,0) and AjA_j is the horizontal area of cell jj. We use one common effective mobility λ\lambda with units m2,Pa−1,s−1\mathrm{m}^2,\mathrm{Pa}^{-1},\mathrm{s}^{-1} for all intra-formational barriers, although the formulation could be extended to layer-specific mobilities. With this transfer law, the barrier threshold pressures and mobility together control the lateral and vertical distribution of CO2_2 in the reservoir. The Pth,iP_{\mathrm{th},i} determines when upward migration begins from each unit, whereas λ\lambda controls the subsequent rate. Both should be interpreted as effective model-scale parameters that incorporate the combined effects of the underlying pore-scale properties.

Equation (1) is a simplified representation of buoyancy-driven two-phase flow through a shale barrier. It neglects brine pressure, counter-current brine flow, and back-pressure from the receiving unit, all of which require more complete layered-aquifer models to represent explicitly [22]. Beyond these simplifications, the IP formulation describes only drainage, in which non-wetting CO2_2 displaces the wetting brine phase. Like other IP simulators such as Permedia [23, 24], it does not represent subsequent imbibition or residual trapping. Because the simulations considered here are restricted to the injection period, drainage is expected to dominate plume evolution, making imbibition and residual trapping of secondary importance. These processes could also be incorporated in future extensions of the model is desired.

Graph-based multilayer simulation

Repeated simulations of CO2_2 migration require a computationally efficient forward model. The geometric structure of the migration problem allows us to construct such a model by reducing each shale topography to a graph of connected structural traps. Because this graph depends only on the reservoir geometry, it can be constructed once and reused across simulations. Rather than updating individual cells in a three-dimensional grid, the simulator routes CO2_2 mass between these traps. For each surface, a spill-point analysis assigns cells to traps, identifies their outlets, and organizes nested and downstream traps into a directed hierarchy. For single units, such a construction is provided by SurfwaterIntegratedModeling.jl and described in detail by Andersen [25]. Its spill-point algorithms were originally in MRST-co2lab to analyse CO2_2 migration within a single sand unit [26]. Here, we extend the approach to a multilayer setting, giving a graph-based representation of the entire reservoir.

We construct directed graph for each sand unit, with structural traps represented as nodes and lateral spill connections as edges. Closed lateral boundaries are imposed, so CO2_2 cannot spill out of the model domain. Figure 1 illustrates the resulting hierarchy for a two-layer reservoir. Injected CO2_2 is routed to a trap, where it accumulates until either the trap reaches its lateral spill point or the pressure required to breach the overlying shale is exceeded. In the former case, excess mass follows the corresponding edge to a downstream trap within the same graph. In the latter, CO2_2 is transferred vertically to a directed graph for the above unit and introduced as a source at the horizontal location of the breach. By processing the layers from deepest to shallowest, the simulator propagates mass through the coupled graphs. In the quasi-static limit, the simulator can then be used to compute the entire time evolution of the plume in a single pass. For finite-rate transfer, the simulator is called repeatedly at adaptive time intervals to resolve the vertical transfer.

Layered reservoir and graph representation of a structural trap filling sequence

Figure 1. Cross-sectional (a) and Graph-representation (b) of a structural trap filling sequence. Nodes represent structural traps, and are labeled for clarity. Note that the directed edges do not represent the spill paths, but illustrate the hierarchy where traps combine into larger accumulations.

The computational kernel is implemented in Rust [27], with Python [28] bindings used for the inference workflow. Its efficiency in the quasi-static limit is assessed against a cell-by-cell implementation. We use the Sleipner reservoir, represented by nine shale surfaces on a 65×11965 \times119 grid from the Sleipner 2019 Benchmark Model [29]. The cell-by-cell reference discretizes the depth axis into 324 vertical layers. Runtime measurements over a 27-year forward simulation were performed on a single core of an AMD Ryzen 7 5800X over 30 repetitions. The graph-based implementation achieved a median runtime of 7.09 ms, whereas the cell-by-cell reference required 288 ms, corresponding to an approximately 40 times speedup. We separately assess the computational cost introduced by finite-rate vertical migration within the graph-based model. Finite-rate migration is more expensive because vertical transfer must be resolved through adaptive time refinement. Across 17 values spanning −14≤log⁡10λ≤−4-14 \le\log_{10} \lambda\le-4, the median runtime was 64.2 ms, about 9.1 times that of the quasi-static graph-based model.

Bayesian inversion from plume observations

Observation model for plume outlines

To model the uncertainty in the CO2\mathrm{CO}_{2} migration, the parameters of the finite-rate IP model are inferred using Bayesian inversion. The inversion is informed by time-lapse seismic observations of the CO2\mathrm{CO}_{2} plume. These surveys rely on amplitude-difference, travel-time, and waveform information to monitor plume evolution over time. As a simplified observation model, we represent the interpreted plume outlines from the seismic surveys as two-dimensional binary masks indicating whether CO2\mathrm{CO}_{2} is detected in each horizontal cell. The relation between the simulated column heights and the binary outlines is illustrated in Figure 2.

Bird’s-eye view of the plume outline and cross-sections

Figure 2. Illustration of the relation between the simulated column heights hi,j,th_{i,j,t} and the binary plume outlines Yi,j,tY_{i,j,t}

Recalling that hi,j,th_{i,j,t} denotes the local CO2\mathrm{CO}_{2} column height in cell jj beneath barrier BiB_i at time tt, we define the binary mask for sand unit ii as

Yi,j,t=1{hi,j,t≥hdet},Y_{i,j,t} = 1\{h_{i,j,t} \ge h_{\mathrm{det}}\},

where 1{⋅}1\{\cdot\} is the indicator function and hdet≥0h_{\mathrm{det}} \ge0 is the minimum detectable column height. A non-zero hdeth_{\mathrm{det}} removes thin plume margins that might not be visible on limited-resolution seismic data. We mainly assume that hdet=0h_{\mathrm{det}} = 0 in this paper, corresponding to perfect detection of any non-zero column height.

We choose this observation model because it provides a simple and interpretable summary of plume evolution. However, it does not restrict the framework, as the simulator can be also combined with more complex observation models. An important limitation is that the column heights are derived from the topographic model, which may be uncertain or inaccurate in real applications. Errors in the topography can therefore introduce bias that is not accounted for in the current model.

Likelihood-free parameter inference with ABC-SMC

Using the simulator and observation model, we can relate the effective parameters of the finite-rate IP model to the observed plume outlines. We have NN parameters in total, N−1N - 1 threshold pressures and one shared mobility. Here the uppermost shale is assumed to be impermeable, although this could be relaxed. The inferred parameters are combined into a single vector

θ=(Pth,1,…,Pth,N−1,λ).\boldsymbol{\theta} = (P_{\mathrm{th},1}, \ldots, P_{\mathrm{th},N-1}, \lambda).

A prior distribution π(θ)\pi(\boldsymbol{\theta}) is assigned to θ\boldsymbol{\theta}, representing the beliefs before conditioning on observations. The prior is application-dependent, but suitable choice is log-normal distributions to ensure positivity. At each survey time tkt_k, the observations provide one binary plume mask Yi,j,kobsY^{\mathrm{obs}}_{i,j,k} for sand unit ii and cell jj. For a given parameter vector θ\boldsymbol{\theta}, the simulator and detection operator produce the corresponding simulated masks Yi,j,k(θ)Y_{i,j,k}(\theta). We denote the complete collections of observed and simulated masks across all sand units and survey times by YobsY^{\mathrm{obs}} and Y(θ)Y(\theta), respectively.

In a Bayesian formulation, the likelihood L(Yobs∣θ)L(Y^{\mathrm{obs}}\mid\theta) which describes the probability of the observations for a given parameter vector. Bayes’ rule gives the posterior

π(θ∣Yobs)∝L(Yobs∣θ)π(θ).\pi(\theta\mid Y^{\mathrm{obs}}) \propto L(Y^{\mathrm{obs}}\mid\theta)\pi(\theta).

However, for the interpreted plume outlines, no tractable likelihood is available. We therefore use ABC for likelihood-free inference [30, 31]. ABC samples parameter values θi\theta_i from the prior, passes them through the simulator, and compares simulated output with the observations. This comparison is performed summary statistics [32], denoted by s(θi)s(\theta_i) and sobss^{\mathrm{obs}}, respectively. By retaining parameter values that produce sufficiently close summaries, it approximates the posterior. However, directly sampling from the prior can be inefficient in high-dimensional parameter spaces. SMC improves this by evolving a population of parameter values through a sequence increasingly restrictive distributions toward the ABC posterior [33, 34].

ABC-SMC algorithm is implemented in PyMC [35, 36]. It has a Gaussian ABC kernel giving an approximate ABC posterior

πABC(θ∣sobs)∝exp⁡[−12d2(θ)]π(θ),(2)\pi_{\mathrm{ABC}}(\theta\mid s^{\mathrm{obs}}) \propto\exp\left[-\frac{1}{2}d^2(\theta)\right]\pi(\theta), \tag*{(2)}

where the discrepancy function d(θ)d(\theta) describes the distance between the observed and simulated summary vectors. For a set of VV summary statistics, the squared normalized discrepancy is defined as

d2(θ)=∑v=1V(sv(θ)−svobsϵv)2.d^2(\theta) = \sum_{v=1}^{V}\left(\frac{s_v(\theta)-s_v^{\mathrm{obs}}}{\epsilon_v}\right)^2.

Here, sv(θ)s_v(\theta) corresponds to the vv-th component of the simulated summary vector, while svobss_v^{\mathrm{obs}} is the corresponding observed summary. The ϵv\epsilon_v are fixed componentwise discrepancy scales that set the relative influence of the summary components. Starting from an initial population drawn from the prior, SMC moves toward the posterior through a sequence of adaptively chosen tempering stages. At stage ww, the population targets

πABC,w(θ∣sobs)∝π(θ)exp⁡[−βw2d2(θ)],0=β0<β1<⋯≤1.\pi_{\mathrm{ABC},w}(\theta\mid s^{\mathrm{obs}}) \propto\pi(\theta)\exp\left[-\frac{\beta_w}{2}d^2(\theta)\right], \qquad0 = \beta_0 < \beta_1 < \cdots\leq1.

The next value of βw\beta_w is chosen adaptively based on the effective sample size of the incremental importance weights, and the procedure continues until βw=1\beta_w = 1. At each stage, the particles are reweighted and resampled, after which a Metropolis-Hastings mutation kernel is applied to restore diversity and explore the current target distribution. The PyMC implementation uses a multivariate normal proposal fitted to the particle population, with the number of mutation steps determined adaptively from the reduction in particle correlation.

Plume-outline summary statistics

To ensure the ABC posterior is informative about the parameters of interest, the summary statistics should capture plume features related to both vertical migration between sand units and lateral migration within each unit. We use three summary statistics for each layer and observation time, with the summary vector expanded as additional surveys become available. The first statistic characterizes the distribution of CO2_2 between the sand units, while the second and third statistics characterize the lateral position of the plume.

If the reservoir topography and plume outlines could be resolved with sufficient accuracy, the stored mass could be estimated by integrating plume thickness over the detected footprint. In practice, seismic imaging provides only an uncertain representation of the reservoir topography, which can bias such an estimate. We therefore use footprint area as a simpler proxy for the relative amount of CO2\mathrm{CO}_2 in each unit. Let Ai,k=∑jYi,j,kA_{i,k} = \sum_j Y_{i,j,k} denote the number of occupied cells in unit ii at observation time kk, and let ak=∑i=1nAi,ka_k = \sum_{i=1}^{n} A_{i,k} be the total number of occupied cells across all units. For ease of notation, we omit denoting whether the mask YY is observed or simulated. The first summary statistic is then

s1,i,k=Ai,kmax⁡{ak,1}s_{1,i,k} = \frac{A_{i,k}}{\max\{a_k,1\}}

and measures the fraction of the total detected footprint contained in unit ii. For the second and third statistics, let (xj,yj)(x_j,y_j) denote the normalized coordinates of cell jj. We define the area-weighted first moments

s2,i,k=1max⁡{ak,1}∑jxjYi,j,k,s3,i,k=1max⁡{ak,1}∑jyjYi,j,k.(3)s_{2,i,k} = \frac{1}{\max\{a_k,1\}}\sum_j x_jY_{i,j,k}, \qquad s_{3,i,k} = \frac{1}{\max\{a_k,1\}}\sum_j y_jY_{i,j,k}. \tag*{(3)}

Because these moments are normalized by the total detected footprint rather than the footprint within each unit, they capture both the lateral location of the plume and the relative amount present in that unit. If no plume is detected at time kk, all three summaries are zero.

Restricting the summary vector to these three quantities balances retaining informative plume features against keeping the ABC comparison low-dimensional. Moreover, because the selected summaries are not known to be sufficient, the resulting ABC posterior depends on their choice and may differ from the posterior based on the full plume observations. Additional outline-based measures, such as intersection-over-union, could be included, but more summaries do not necessarily improve the inference. Redundant or weakly informative components can make simultaneous agreement with the observations more difficult and increase sensitivity to discrepancy scaling.

Application to the Sleipner storage site

Dataset

We now apply the framework to seismic plume outlines data from the Sleipner storage site in the North Sea. This provides a real-world test of the conditioning approach in a setting with a well-documented multilayer plume. Since the shale properties are unknown, the aim here is not to assess parameter recovery against a known truth. Instead, we examine whether the model reproduces the broad plume redistribution, and see which effective parameter combinations are informed by the data.

The storage site contains nine main sand units separated by eight intra-formational shale barriers. Seven barriers are approximately 1-2 m thick, while the uppermost has a mean modelled thickness of about 7.6 m. A much thicker shale succession seals the formation above [3], and is treated as an impenetrable top barrier. Figure 3 shows a cross-section through the shales at the injection location. Other thinner shale layers may occur within the main sand layers, but were not found to affect the CO2\mathrm{CO}_2 migration.

Figure 3. Cross-section through the Sleipner layers at a UTM easting of 438.5 km. The cross marks the injection location.

Injection of CO2\mathrm{CO}_2 separated from natural-gas production began in 1996 through a well near the base of the Utsira Formation. Callioli Santi et al. [16] summarizes the annual injection volumes. A baseline seismic survey was acquired in 1994, followed by multiple repeat surveys between 1999 and 2020 [1, 37]. The plume reached the uppermost sand layer by the first repeat survey in 1999 [3]. This rapid ascent has been linked to localized feeder pathways through the intra-formational shales [38, 39].

This study uses seismic monitoring data from the 2010 and 2023 surveys. The 2010 outlines and shale surfaces come from the Sleipner 2019 Benchmark Model in CO2_2DataShare [29], where the shale surfaces are represented on a 65×11965 \times119 grid of 50×5050 \times50 m cells. The 2023 outlines are derived from the full-waveform inversion results of Martinez et al. [39]. An important consideration is that the observed outlines are laterally mismatched with the topography, meaning that the outlined plume is not located in the structural traps of the shale surfaces. This could be caused by uncertainties in seismic imaging.

Inference setup and posterior parameter estimates

Each barrier is assigned one effective entry pressure, giving nine inferred parameters in total. Localized feeders are not represented in this study. However, they could be included by setting the local capillary entry pressure to zero in traps corresponding to the feeder locations identified by Martinez et al. [39]. The per-layer CO2_2 densities are set according to Callioli Santi et al. [16], Appendix A, the brine density to 1000 kg/m3^3, the sand porosity to 36%, and the irreducible water saturation to 30% [40]. The observation threshold is hdet=0h_{\mathrm{det}} = 0. We select log-normal priors

log⁡10Pth,i∼N(1.824,0.3262),log⁡10λ∼N(−11,12),\log_{10} P_{\mathrm{th},i} \sim\mathcal{N}\left(1.824, 0.326^2\right), \qquad\log_{10} \lambda\sim\mathcal{N}\left(-11, 1^2\right),

where Pth,iP_{\mathrm{th},i} is expressed in kPa and λ\lambda in m2^2 Pa−1^{-1} s−1^{-1}. The PyMC ABC-SMC sampler uses threshold=0.75 and correlation_threshold=0.01. The discrepancy scales for s1,i,ks_{1,i,k} is set to 0.3 times the corresponding observed area fraction, while to 0.8 for s2,i,ks_{2,i,k} and s3,i,ks_{3,i,k}. The high value for the lateral summaries reflects the lateral mismatch, as we cannot expect the model to reproduce the observed lateral distribution of the plume. Five independent chains contribute 1000 final particles each, giving 5000 posterior samples in total.

Figure 4 compares the marginal posterior distributions with the prior distributions. The posterior for Pth,5P_{\mathrm{th},5} shifts towards higher pressures and contracts, while Pth,8P_{\mathrm{th},8} remains broad. The mobility posterior also contracts. This suggests the model is able to identify which aspects of the reduced model are identified and which remain unresolved by the available monitoring data. However, note that narrowing distributions do not imply independent parameter identification. We see in Appendix A.1 that the posterior correlations of Pth,5P_{\mathrm{th},5} and Pth,8P_{\mathrm{th},8} with log⁡10λ\log_{10} \lambda are 0.72 and 0.66, respectively.

Figure 4. ABC posterior distributions estimated from the Sleipner data using a finite-rate model. Results are shown for the eight shale entry pressures and the shared mobility. Solid vertical lines indicate posterior medians, while dashed lines indicate the log-normal priors.

Posterior reconstruction and model adequacy

We pick 500 of the 5000 posterior samples, and use them in forward simulations. Figure 5 compares the observed footprint-area fractions with their posterior reconstruction distributions for survey times 2010 and 2023. The displays indicate how well the model captures the vertical distributions of detected area. Most observations lie within the 90% reconstruction intervals. The only exception is S9 in 2023, where the observed fraction lies just below the interval. Importantly, the main temporal shift is also reproduced. Storage unit S5 has the largest observed footprint fraction in 2010, while a larger share of the detected area occupies S9 by 2023. Appendix A.2 compares how the forecast layer-wise mass fractions change when using only the 2010 observation or both 2010 and 2023.

Figure 5. Observed and posterior fractions of total detected plume area in each sand unit of Sleipner. The fit uses a finite-rate model. Error bars show 90% posterior reconstruction intervals.

Using the observation rule from Section 3.1, each posterior simulation is converted to a binary plume-detection mask. Cell-wise recurrence is the fraction of these simulations in which CO2_2 is detected in a given cell, with values near one indicating consistent plume presence across the posterior ensemble. Figure 6 shows that the simulations generally place CO2_2 in all nine sand units by 2010, reproduce the large accumulation in S5S_5, and concentrate later footprint growth in the upper units. Mismatch between the assumed topography and observed plume geometry leads to weaker lateral agreement. In several middle and upper units, the simulated plume lies south of the observed outline or splits between competing structural traps. This is consistent with Callioli Santi et al. [16], who also found southward migration in continuous-shale models.

Figure 6. Posterior recurrence of the detected plume at the two survey times used for the Sleipner fit. Units S1,…,S9S_1,\ldots,S_9 are ordered from deepest to shallowest. Colour shows the fraction of posterior simulations containing detected CO2_2 in each cell, and black lines show the observed outlines.

Comparison with quasi-static transfer

The role of finite-rate transfer is assessed by refitting the site with a quasi-static IP model. We use the same priors and ABC-parameters are earlier, but omit the rate-limitation. Figure 7 shows that this model cannot reproduce the layer-wise distribution of detected CO2_2 area at both survey times. Only the top layer evolves between 2010 and 2023, which leads to overestimation of the detected area in that layer. The quasi-static fit also compensates for its inability to retain CO2_2 beneath a breached barrier by inferring systematically higher effective entry pressures. Posteriors can be found in Appendix A.3, and medians range from 81 to 118 kPa, compared with 33-71 kPa for the finite-rate model. Note that these higher values represent compensation within the reduced model rather than direct estimates of physical shale entry pressures.

Figure 7. Observed and posterior fractions of total detected plume area in each sand unit of Sleipner. The fit uses a quasi-static model. Error bars show 90% posterior reconstruction intervals.

Application to a synthetic multilayer model

Experimental design

The Sleipner application tests whether the framework can be conditioned on an observed multilayer plume, but the unknown shale properties prevent direct assessment of parameter accuracy. Synthetic experiments provide a complementary setting in which the true parameters and future plume evolution are known by construction, allowing direct evaluation of parameter recovery, out-of-sample forecasting, and sensitivity to monitoring duration and model assumptions. These tests remain conditional on the assumed model and simulator and therefore do not assess robustness to broader model discrepancy.

The synthetic reservoir contains four sand units S1,…,S4S_1,\ldots,S_4 separated by shale barriers. It has a 5 km×5 km5\ \mathrm{km} \times5\ \mathrm{km} domain is discretised on a 51×5151 \times51 lateral grid. CO2\mathrm{CO}_2 is injected at a constant rate into the centre of the deepest unit. Each sand unit is 50 m thick and each shale is 6 m thick, with the top of the stack at 800 m depth. The four structural surfaces are independent realisations of a Gaussian random field (GRF) [41] using a Matérn covariance with smoothness ν=3/2\nu= 3/2, correlation length 1 km and standard deviation 10 m. Sand porosity is 36%, the irreducible water saturation is 30%, and the CO2\mathrm{CO}_2 and brine densities are 700 and 1000 kg m−31000\ \mathrm{kg}\ \mathrm{m}^{-3}, respectively. The true parameters are selected as (Pth,1,Pth,2,Pth,3)=(44,52,66) kPa(P_{\mathrm{th},1},P_{\mathrm{th},2},P_{\mathrm{th},3}) = (44, 52, 66)\ \mathrm{kPa} and λ=10−10 m2 Pa−1s−1\lambda= 10^{-10}\ \mathrm{m}^2\ \mathrm{Pa}^{-1}\mathrm{s}^{-1}.

We evaluate effect of monitoring duration by considering two different observation scenarios. Scenario 1 surveys the reservoir every two years from year 2 through year 16. Scenario 2 uses the same truth, but includes only the first five surveys and ends in year 10. Figure 8 shows plume evolution through year 16 and a later forecast target at year 30. All three barriers are crossed during scenario 1, with CO2\mathrm{CO}_2 first entering S4S_4 in year 13. In scenario 2, the shallowest barrier therefore remains unbreached during the observation period.

Figure 8. True plume evolution in the GRF reservoir. Units S1,…,S4S_1,\ldots,S_4 are ordered from deepest to shallowest. Coloured plume extents are shown from year 2 to year 16, together with the forecast target at year 30.

Unless stated otherwise, the detection threshold is set to hdet=0h_{\mathrm{det}} = 0. The entry-pressures are given priors log⁡10Pth,i∼N(1.732,0.3042)\log_{10} P_{\mathrm{th},i} \sim\mathcal{N}(1.732, 0.304^2), centred at the mean 54 kPa of the three true values. The mobility prior is log⁡10λ∼N(−9.523,1.02)\log_{10} \lambda\sim\mathcal{N}(-9.523, 1.0^2), with a median three times the true value. Recovery therefore cannot be attributed to priors centred on the truth. We set discrepancy scales to 0.1 times the observed area fractions and 0.3 times the observed spatial moments. The ABC-SMC sampler in PyMC uses the same threshold and correlation-threshold as in Section 4. Five independent chains of 1000 particles yield 5000 final particles.

Parameter recovery

Fitting the model, Figure 9 compares the marginal priors and ABC posteriors with the true parameters. In scenario 1, all three barriers are crossed during monitoring, so all four posteriors contract strongly around the truth. The entry-pressure 90% intervals have widths of 1.3–4.1 kPa. Scenario 2 gives similar estimates of Pth,1P_{\mathrm{th},1}, Pth,2P_{\mathrm{th},2} and λ\lambda, but the posterior for Pth,3P_{\mathrm{th},3} is broad and right-skewed. Its median lies 18% above the truth and its 90% interval spans 46–200 kPa. This asymmetry reflects that Pth,3P_{\mathrm{th},3} is informed primarily through a lower bound, since values low enough to trigger an earlier crossing are disfavoured. The results therefore suggest that information gain is event-dependent rather than only a function of the number of surveys. Observations before a barrier is crossed can rule out thresholds that would imply an earlier crossing, but do not strongly constrain larger values.

Marginal priors and ABC posteriors for the two GRF scenarios, showing Shale 1, Shale 2, Shale 3 and $\log_{10}\lambda$ distributions for observations to years 16 and 10

Figure 9. (a) Scenario 1. GRF reservoir model observed to year 16. (b) Scenario 2. GRF reservoir model observed to year 10. Marginal priors and ABC posteriors for the two GRF scenarios. Dashed black curves show the priors, blue histograms show the posteriors, solid black lines show the posterior medians, and dashed orange lines show the true values.

Forecasting plume migration

For each scenario, 500 randomly selected posterior samples are propagated from the final survey and forwards in time. Figure 10 shows the resulting recurrence maps for year 30. Scenario 1 has high recurrence within the true outlines and sharp transitions at their boundaries in all units. Scenario 2 consistently reproduces the two deepest units, whereas S3S_3 and S4S_4 have broad transition zones.

Figure 10. (a) Scenario 1. GRF observed to year 16. (b) Scenario 2. GRF observed to year 10. Posterior plume recurrence at year 30, estimated from 500 posterior simulations. Black lines show the true plume outlines. Cells with zero recurrence are not coloured.

Figure 11 shows the fraction of the total stored mass in each sand unit, forecast annually from the end of the observation period to year 50. The truth lies within the 90% posterior forecast interval at every plotted time in both scenarios. At year 30, scenario 1 has an average interval width of 0.014 and a root-mean-square error of 0.0004 for the posterior median across the four units. The corresponding values in scenario 2 are 0.290 and 0.0849. Thus, ending observations six years earlier increases these metrics by factors of about 20 and 200, respectively. The additional uncertainty is concentrated in S3S_3 and S4S_4 because the weak constraint on P3P_3 leaves the timing of the third barrier breach uncertain. See that the uncertainty intervals in S4S_4 include the scenario where no CO2\mathrm{CO}_2 reaches the shallowest unit, which is consistent with the low recurrence from fig:10b.

Figure 11. Forecast fractions of total stored CO2_2 mass in each sand unit. Solid blue lines show posterior medians, blue bands show 90% posterior forecast intervals, and dashed black lines show the truth.

Accounting for the finite-rate transfer

The experiments above use the finite-rate transfer law of Equation (1) both to generate and fit the synthetic data. We now perform a controlled model-discrepancy test where we fit both the finite-rate and quasi-static models to a truth from a finite-rate scenario. The same experiment is also repeated with a quasi-static truth, and the results are included in Appendix B.1. Settings for the reservoir, true parameters and ABC inference are mostly unchanged. but the finite-rate mobility prior is widened to log⁡10λ∼N(−9.0,2.02)\log_{10} \lambda\sim N(-9.0, 2.0^2) to include a quasi-static limit.

Figure 12 compares posterior distributions and forecasts under the finite-rate truth. The finite-rate model recovers the entry pressures and reproduces the true layer-wise mass fractions. The quasi-static model cannot retain CO2_2 beneath a breached barrier and compensates by inferring higher entry pressures that delay upward migration. Yet, even with this compensation, the forecast underestimates mass retention in the lowest unit.

Figure 12. Finite-rate and quasi-static fits to the finite-rate synthetic truth. In (a), dashed black curves show the shared entry-pressure priors and dashed orange lines show the true parameters. In (b), the models are compared with the true layer-wise mass fractions.

Sensitivity to the plume-detection threshold

The preceding experiments assume hdet=0h_{\mathrm{det}} = 0, so the observed footprint equals the full simulated plume footprint. To assess sensitivity to uncertain detectability, observations are generated with a true threshold of 3 m and fitted using assumed thresholds of 0, 3 and 8 m. Appendix B.2 illustrates the geometric effect of these thresholds, where we see that increasing the threshold progressively removes thin plume margins and small accumulations.

Figure 13 shows that misspecifying hdeth_{\mathrm{det}} biases the inferred parameters. The entry-pressure posteriors are similar for assumed thresholds of 0 and 3 m, whereas the 8 m scenario shifts towards lower pressures. Relative to the correctly specified hdeth_{\mathrm{det}}, the mobility posterior shifts towards lower values for hdet=0h_{\mathrm{det}} = 0 and higher for hdet=8h_{\mathrm{det}} = 8 m. This indicates that misspecified detectability can be absorbed by both effective barrier strength and mobility. Posterior contraction should thus be interpreted together with the assumed observation model, since detection errors can appear as changes in the effective migration parameters.

Figure 13. Posteriors obtained with assumed detection thresholds of 0, 3 and 8 m when the true threshold is 3 m. Dashed black curves show the priors and dashed orange lines show the true parameters.

Closing remarks

We have presented a Bayesian framework for estimating effective parameters that control vertical CO2\mathrm{CO}_2 migration in layered reservoirs. The framework targets buoyancy- and capillary-dominated settings where full-physics inference is too costly or an interpretable reduced model is preferred. It develops a graph-based finite-rate IP simulator, and combines it with ABC-SMC for parameter estimation from time-lapse seismic plume outlines. For fixed reservoir geometry, the graph is constructed once and reused across parameter evaluations. On the reposted benchmark, a finite-rate forward simulation had a median runtime of 64.2 ms, making Monte Carlo-based inference computationally tractable.

The Sleipner experiment tests the framework under real data and model limitations, rather than against known parameter values. We find the posterior simulations to reproduce the broad vertical distribution of the detected plume across nine sand units and the shift in footprint area towards the upper units between 2010 and 2023. The surveys narrow several marginal parameter distributions, but posterior dependence indicates that effective barrier strength and mobility are not independently identified. Lateral agreement between observed and simulated outlines is weaker, reflecting both the topography-trap mismatch and the omission of localized feeders.

Synthetic experiments provide complementary tests of parameter recovery and out-of-sample forecasting, and show how uncertainty can be reduced as additional monitoring data are acquired. When monitoring continues until every barrier has been crossed, the posteriors concentrate around the true parameters and forecasts closely reproduce future plume states. Ending the observation period before the final crossing instead leaves the corresponding threshold pressure weakly constrained and increases forecast interval widths. Fitting quasi-static transfer to finite-rate observations inflates the inferred threshold pressures by more than 50% for the most affected barrier, whereas the finite-rate model recovers the truth and approaches the quasi-static limit when transfer is instantaneous. Misspecifying the plume-detection threshold biases both the inferred entry pressures and mobility, which motivates explicit treatment of seismic detectability in future applications.

Together, these results suggest a practical role for the framework during active injection. As new monitoring surveys become available, the inference can be updated to revise the effective migration parameters and corresponding plume forecasts. The value of each survey depends on the migration events it captures, with breakthrough and post-breach redistribution providing particularly strong constraints. Remaining posterior uncertainty then highlights aspects of the migration behaviour that are still unresolved by the available observations.

Future work should test robustness to uncertain shale topography and unknown localized shale breaks, and could infer additional observation-model parameters such as hdeth_{\mathrm{det}}. We only used plume outlines data in this study, and more informative seismic data and summary statistics may improve inference. Benchmarking the simulator against laboratory measurements and higher-fidelity flow simulations would clarify its fidelity to the underlying physics. This would also inform which additional physical processes should be included in the model. Furthermore, applying the framework to other storage sites can show how it performs under different geological settings and monitoring conditions. Lastly, as the fast simulator that was shown to match real-world Sleipner results rather well, we also foresee opportunities for using it as a surrogate model in for instance optimization of drilling plans, optimal survey design calculations, and other high-level tasks going beyond inversion.

Supporting Sleipner results

Posterior dependence

Faster post-breach transfer can be offset by later breakthrough, producing a trade-off between effective barrier strength and mobility. Figure A1 shows the correlations among the eight entry pressures and log⁡10λ\log_{10} \lambda in the estimated Sleipner ABC posterior from Section 4.2. See that the correlations between the capillary entry pressures and log⁡10λ\log_{10} \lambda are relatively small for the lower barriers. However, the association is stronger for Pth,5P_{\mathrm{th},5} and Pth,8P_{\mathrm{th},8}, whose correlations with log⁡10λ\log_{10} \lambda are 0.72 and 0.66, respectively. These dependencies show that narrow marginal distributions do not imply independent parameter identification.

Figure A1. Pearson correlations among the eight entry pressures and log⁡10λ\log_{10} \lambda in the Sleipner ABC posterior. Fit using a finite-rate model.

Effect of additional monitoring data

In the main Sleipner experiment, we used both observation times 2010 and 2023 for fitting the model. Figure A2 compares the forecast layer-wise mass fractions when using only the 2010 observation or both 2010 and 2023. As expected, the two-observation fit produces narrower forecast intervals, but the differences are small for the lower layers. The posterior medians mostly agree, with a notable difference in layer 5. The comparison therefore shows that the reduction in forecast uncertainty does not uniformly decrease when the monitoring record is extended. Interestingly, both fits forecast that the CO2\mathrm{CO}_{2} will not reach the top layer until after 2004, while in reality this has already occurred by 1999. This discrepancy indicates the time of monitoring matters. As the first observation is from 2010, we have little information about the early migration history, the reduced-order models choose parameters that delay upward migration.

Figure A2. Forecast layer-wise mass fractions for the finite-rate Sleipner model using only the 2010 observation or both 2010 and 2023. Solid lines show posterior medians, and shaded areas show 90% forecast intervals.

Quasi-static parameter estimates

Section 4.4 compared the finite-rate and quasi-static transfer models for the Sleipner data. We here provide the corresponding posterior distributions for the quasi-static model. See from Figure A3 that the quasi-static model produces higher posterior medians for every barrier than the finite-rate fit. This compensation is done to keep more CO2_2 in the lower units. Yet, from the forecast we found that this caused a misfit in the time-evolution of the layer-wise mass distribution. Also observe that the posteriors appear less concentrated around the medians. This could be due to the ABC-SMC algorithm struggling to converge, as it is unable to find parameter sets that match the observations.

Figure A3. ABC posterior distributions estimated from the Sleipner data using a quasi-static model. Results are shown for the eight shale entry pressures and the shared mobility. Solid vertical lines indicate posterior medians, while dashed lines indicate the log-normal priors.

Supporting synthetic results

Quasi-static limiting case

The main model-comparison experiment in Section 5.4 examines the consequences of fitting quasi-static transfer to finite-rate observations. We here provide a complementary test of whether the finite-rate model can fit a quasi-static truth. Recall that the mobility prior was deliberately chosen to be broad, so that the finite-rate model can approach a quasi-static limit where the transfer is effectively instantaneous.

Figure B4a shows the posterior distributions for the finite-rate and quasi-static models. The finite-rate model recovers the true entry pressures without apparent bias, while the mobility posterior shifts towards higher values but remains broad. Figure B4b shows that the two models produce nearly identical layer-wise mass forecasts, both closely matching the truth. These results indicate that the finite-rate formulation remains appropriate in the quasi-static limit and can therefore accommodate a broad range of transfer rates.

(a) Posterior distributions

(b) Forecast mass fractions

Figure B4. Finite-rate and quasi-static fits to the quasi-static synthetic truth. Panel (a) shows the shared entry-pressure priors as dashed black curves and the true entry pressures as dashed orange lines. The quasi-static truth has no true mobility to mark. Panel (b) compares both models with the true layer-wise mass fractions.

Effect of the plume-detection threshold

Section 5.5 compared the inferred parameters and forecasts for different plume-detection thresholds. Figure B5 illustrates how the detected plume extent changes with the threshold. Increasing hdeth_{\mathrm{det}} progressively removes thin plume margins and can eliminate small accumulations from the detected footprint. These geometric changes explain how an incorrect threshold can introduce bias in the inferred entry pressures, which is what was observed in fig:12a.

Figure B5. Detected plume extent at year 16 for detection thresholds of 0, 3, and 8 m.

Acknowledgements

We acknowledge support from the Centre for Geophysical Forecasting (grant no. 309960).

Statements and Declarations

Funding

This work was supported by the Centre for Geophysical Forecasting (grant no. 309960).

Competing interests

The authors declare no competing interests.

Data availability

The code and data is provided in a GitHub repository https://github.com/ellingsvee/CO2IPSimulator.

References

References

  1. [1]Arts, R., Eiken, O., Chadwick, A., Zweigel, P., Meer, L., Zinszner, B.: Monitoring of CO2 injected at Sleipner using time-lapse seismic data. Energy 29(9), 1383–1392 (2004) https://doi.org/10.1016/j.energy.2004.03.072 . Accessed 2026-03-24
  2. [2]Chadwick, R.A., Arts, R., Eiken, O., Kirby, G.A., Lindeberg, E., Zweigel, P.: 4D Seismic Imaging of an Injected CO2 Plume at the Sleipner Field, Central North Sea. Geological Society, London, Memoirs 29(1), 311–320 (2004) https://doi.org/10.1144/GSL.MEM.2004.029.01.29 . Accessed 2026-06-22
  3. [3]Cavanagh, A.J., Haszeldine, R.S.: The Sleipner storage site: Capillary flow modeling of a layered CO2 plume requires fractured shale barriers within the Utsira Formation. International Journal of Greenhouse Gas Control 21, 101–112 (2014) https://doi.org/10.1016/j.ijggc.2013.11.017 . Accessed 2026-03-15
  4. [4]Iglauer, S., Pentland, C.H., Busch, A.: CO2 wettability of seal and reservoir rocks and the implications for carbon geo-sequestration. Water Resources Research 51(1), 729–774 (2015) https://doi.org/10.1002/2014WR015553 . e-print: https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1002/2014WR015553. Accessed 2026-06-22
  5. [5]Martinez, R., Vinje, V., Mispel, J., Ringrose, P., Stovas, A., Landrø, M.: Enhanced imaging of multi-layer CO2 plumes using full-waveform inversion: Insights from the Sleipner storage site. International Journal of Greenhouse Gas Control 148, 104517 (2025) https://doi.org/10.1016/j.ijggc.2025.104517 . ADS Bibcode: 2025IJGGC.14804517M. Accessed 2026-04-22
  6. [6]Oliver, D.S., Reynolds, A.C., Liu, N.: Inverse Theory for Petroleum Reservoir Characterization and History Matching. Cambridge University Press, Cambridge (2008) https://doi.org/10.1017/CBO9780511535642 . https://www.cambridge.org/core/books/inverse-theory-for-petroleum-reservoir-characterization-and-history-matching/5369D69A09D806AB52E1FA3F35E51EC3 Accessed 2026-06-22
  7. [7]Oliver, D.S., Chen, Y.: Recent progress on reservoir history matching: a review. Computational Geosciences 15(1), 185–221 (2011) https://doi.org/10.1007/s10596-010-9194-2 . Accessed 2026-06-22
  8. [8]Crain, D.M., Benson, S.M., Saltzer, S.D., Durlofsky, L.J.: An integrated framework for optimal monitoring and history matching in CO$_{2}$storage projects. Computational Geosciences 28(2), 211–225 (2024) https://doi.org/10.1007/s10596-023-10216-3 . Accessed 2026-08-31
  9. [9]Teng, W., Durlofsky, L.J.: Likelihood-free inference and hierarchical data assimilation for geological carbon storage. Advances in Water Resources 201, 104961 (2025) https://doi.org/10.1016/j.advwatres.2025.104961 . Accessed 2026-08-31
  10. [10]Wen, G., Li, Z., Azizzadenesheli, K., Anandkumar, A., Benson, S.M.: U-FNO – An enhanced Fourier neural operator-based deep-learning model for multiphase flow. arXiv. arXiv:2109.03697 [physics.geo-ph] (2022). https://doi.org/10.48550/arXiv.2109.03697 · http://arxiv.org/abs/2109.03697 Accessed 2026-06-22
  11. [11]Wen, G., Li, Z., Long, Q., Azizzadenesheli, K., Anandkumar, A., Benson, S.M.: Real-time high-resolution CO2 geological storage prediction using nested Fourier neural operators. In: AGU Fall Meeting 2023. American Geophysical Union, San Francisco (2023). https://agu.confex.com/agu/fm23/meetingapp.cgi/Paper/1248100 Accessed 2026-06-22
  12. [12]Han, Y., Hamon, F.P., Jiang, S., Durlofsky, L.J.: Surrogate model for geological CO2 storage and its use in hierarchical MCMC history matching. Advances in Water Resources 187, 104678 (2024) https://doi.org/10.1016/j.advwatres.2024.104678 . Accessed 2026-06-22
  13. [13]Watson, F., Klemetsdal, Ø., Lye, K.O.: Multifidelity Monte Carlo method for rapid uncertainty quantification in CO$_{2}$storage applications. Computational Geosciences 29(4), 32 (2025) https://doi.org/10.1007/s10596-025-10372-8 . Accessed 2026-08-27
  14. [14]Carruthers, D., Ringrose, P.: Secondary oil migration: oil-rock contact volumes, flow behaviour and rates. Geological Society, London, Special Publications 144(1), 205–220 (1998) https://doi.org/10.1144/GSL.SP.1998.144.01.16 . Publisher: The Geological Society of London. Accessed 2026-01-07
  15. [15]Carruthers, D.J.: AAPG/Datapages Discovery Series No. 7, Multidimensional Basin Modeling, Chapter 3: Modeling of Secondary Petroleum Migration Using Invasion Percolation Techniques, 21–37 (2003). Publisher: AAPG Special Volumes. Accessed 2025-11-07
  16. [16]Callioli Santi, A., Ringrose, P., Eidsvik, J., Haugdahl, T.A.: Invasion percolation Markov Chains – A probabilistic framework for assessing vertical CO2 migration. International Journal of Greenhouse Gas Control 142, 104338 (2025) https://doi.org/10.1016/j.ijggc.2025.104338 . Accessed 2025-10-07
  17. [17]Neufeld, J.A., Huppert, H.E.: Modelling carbon dioxide sequestration in layered strata. Journal of Fluid Mechanics 625, 353–370 (2009) https://doi.org/10.1017/S0022112008005703 . Accessed 2026-08-05
  18. [18]Boait, F.C., White, N.J., Bickle, M.J., Chadwick, R.A., Neufeld, J.A., Huppert, H.E.: Spatial and temporal evolution of injected CO2 at the Sleipner Field, North Sea. Journal of Geophysical Research 117 (2012) https://doi.org/10.1029/2011JB008603 . Num Pages: 21. Accessed 2026-06-22
  19. [19]Wilkinson, D., Willemsen, J.F.: Invasion percolation: a new form of percolation theory. Journal of Physics A: Mathematical and General 16(14), 3365 (1983) https://doi.org/10.1088/0305-4470/16/14/028 . Accessed 2026-03-24
  20. [20]Berg, R.R.: Capillary Pressures in Stratigraphic Traps. AAPG Bulletin 59(5), 939–956 (1975) https://doi.org/10.1306/83D91EF7-16C7-11D7-8645000102C1865D . Accessed 2026-06-22
  21. [21]Espinoza, D.N., Santamarina, J.C.: CO2 breakthrough—Caprock sealing efficiency and integrity for carbon geological storage. International Journal of Greenhouse Gas Control 66, 218–229 (2017) https://doi.org/10.1016/j.ijggc.2017.09.019 . Accessed 2026-06-22
  22. [22]Jenkins, L.T., Foschi, M., MacMinn, C.W.: Gas injection and leakage in layered aquifers (2019). https://arxiv.org/abs/1908.08137
  23. [23]Halliburton: Permedia CO2 Modeling Tools Software Suite (2026). https://www.halliburton.com/en/products/co2-storage/permedia-co2 Accessed 2026-08-27
  24. [24]Ni, H., Møyn­er, O., Kurtev, K.D., Benson, S.M.: Quantifying CO2 capillary heterogeneity trapping through macroscopic percolation simulation. Advances in Water Resources 155, 103990 (2021) https://doi.org/10.1016/j.advwatres.2021.103990 . Accessed 2026-08-24
  25. [25]Andersen, O.A.: Topography-based surface water modeling in Julia, with support for infiltration and temporal developments. Journal of Open Source Software 10(109), 7785 (2025) https://doi.org/10.21105/joss.07785 . Accessed 2025-10-09
  26. [26]Møll Nilsen, H., Lie, K.-A., Møyn­er, O., Andersen, O.: Spill-point analysis and structural trapping capacity in saline aquifers using MRST-co2lab. Computers & Geosciences 75, 33–43 (2015) https://doi.org/10.1016/j.cageo.2014.11.002 . Accessed 2025-10-07
  27. [27]The Rust Project Developers: The Rust Programming Language. (2026). Accessed 27 August 2026. https://www.rust-lang.org/
  28. [28]Python Software Foundation: Python Language Reference (2026). https://www.python.org/
  29. [29]Equinor: Sleipner 2019 Benchmark Model. CO2DataShare. Dataset (2020). https://doi.org/10.11582/2020.00004 . https://doi.org/10.11582/2020.00004
  30. [30]Beaumont, M.A., Zhang, W., Balding, D.J.: Approximate Bayesian Computation in Population Genetics. Genetics 162(4), 2025–2035 (2002) https://doi.org/10.1093/genetics/162.4.2025 . Accessed 2026-03-24
  31. [31]Csilléry, K., Blum, M.G.B., Gaggiotti, O.E., François, O.: Approximate Bayesian Computation (ABC) in practice. Trends in Ecology & Evolution 25(7), 410–418 (2010) https://doi.org/10.1016/j.tree.2010.04.001 . Accessed 2026-03-24
  32. [32]Sisson, S.A., Fan, Y., Beaumont, M. (eds.): Handbook of Approximate Bayesian Computation. Chapman and Hall/CRC, New York (2018). https://doi.org/10.1201/9781315117195
  33. [33]Sisson, S.A., Fan, Y., Tanaka, M.M.: Sequential Monte Carlo without likelihoods. Proceedings of the National Academy of Sciences 104(6), 1760–1765 (2007) https://doi.org/10.1073/pnas.0607208104 . Accessed 2026-06-08
  34. [34]Toni, T., Welch, D., Strelkowa, N., Ipsen, A., Stumpf, M.P.H.: Approximate Bayesian computation scheme for parameter inference and model selection in dynamical systems. Journal of The Royal Society Interface 6(31), 187–202 (2009) https://doi.org/10.1098/rsif.2008.0172 . arXiv:0901.1925 [stat]. Accessed 2026-03-24
  35. [35]Salvatier, J., Wiecki, T., Fonnesbeck, C.: Probabilistic Programming in Python using PyMC. arXiv. arXiv:1507.08050 [stat.CO] (2015). https://doi.org/10.48550/arXiv.1507.08050 . http://arxiv.org/abs/1507.08050 Accessed 2026-06-08
  36. [36]Castro, M.: Approximate Bayesian Computation Made Easy: A Practical Guide to ABC-SMC for Dynamical Systems with PyMC. arXiv. arXiv:2511.21587 [q-bio] (2025). https://doi.org/10.48550/arXiv.2511.21587 . http://arxiv.org/abs/2511.21587 Accessed 2026-02-23
  37. [37]Ringrose, P., Martinez, R., Vinje, V., Mispel, J.: Estimating the multi-scale distribution of CO2 using seismic data at Sleipner. International Journal of Greenhouse Gas Control 151, 104581 (2026) https://doi.org/10.1016/j.ijggc.2026.104581 . Accessed 2026-03-24
  38. [38]Furre, A., Ringrose, P., Santi, A.C.: Observing the Invisible – CO2 Feeder Chimneys on Seismic Time-Lapse Data. In: 81st EAGE Conference and Exhibition 2019, vol. 2019, pp. 1–5. European Association of Geoscientists & Engineers, London (2019). https://doi.org/10.3997/2214-4609.201901646 . https://www.earthdoc.org/content/papers/10.3997/2214-4609.201901646 Accessed 2026-03-24
  39. [39]Martinez, R., Vinje, V., Moore, H., Hollingworth, S., Ringrose, P., Stovas, A.: Unraveling multilayer CO2 plumes using the entire wavefield: Case study from the Sleipner storage site. Interpretation 14(2), 163–183 (2026) https://doi.org/10.1190/int-2025-0016 . Accessed 2026-07-13
  40. [40]Jackson, S.J., Agada, S., Reynolds, C.A., Krevor, S.: Characterizing Drainage Multiphase Flow in Heterogeneous Sandstones. Water Resources Research 54(4), 3139–3161 (2018) https://doi.org/10.1029/2017WR022282 . Accessed 2026-09-02
  41. [41]Banerjee, S., Gelfand, A.E., Carlin, B.P.: Hierarchical Modeling and Analysis for Spatial Data, 3rd edn. Chapman and Hall/CRC, Boca Raton (2025)

Paper details

vlnc:W_ux6bfndv
Version
v1 · latest
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents