Fifth EAGE Conference on Petroleum Geostatistics

Among many seismic data denoising methods, the KSVD denoising method is an effective method. However, due to its ill-conditioned problem, it is necessary to introduce regularization terms to improve the KSVD method, and the automatic optimization of regularization parameters is extremely important. Particle swarm optimization algorithms are widely used in parameter optimization, but the inertia parameters of conventional particle swarm optimization algorithms are generally fixed, which leads to the decline of the search efficiency in the later stage. In this paper, the adaptive dynamic particle swarm optimization algorithm is used to improve the setting of regularization approximation parameters, and the advantages of the AKSVD method for weak signal identification are used to propose the RAKSVD denoising method with the optimized regularization parameters of the adaptive dynamic particle swarm algorithm. Model testing and practical applications show that this method can not only achieve the expected denoising effect, but also pay more attention to the protection of weak seismic signals. After seismic denoising, the weak seismic signals are basically not distorted, which is conducive to the extraction and recognition of weak signal. At the same time, the computational efficiency of this method has also been improved.

This paper investigates the implications of ensemble size and the importance of localization in ensemble history matching. Increasing the ensemble size is known to reduce the impact of sampling errors, but determining the appropriate size, N, to ensure a strong signal-to-noise ratio remains challenging. Ensemble history-matching methods use the correlation between uncertain parameters and predicted measurements to compute linear updates to the input parameters. However, the nonlinear relationship reduces the correlations.

Practical limitations in computational resources and simulation time restrict the ensemble size, often leading to spurious updates in the posterior ensemble when computing the global analysis. To address this issue, we introduce a consistent correlation-based localization method. Instead of using physical distances, this method selects measurements based on their correlation strength with the updated parameter.

The paper presents examples using the REEK model, demonstrating how increasing the ensemble size improves the correlation functions and how localization reduces spurious updates and avoids underestimating the posterior variance.

The paper also indicates the need to include uncertainties in historical-rate controls and account for measurement error correlations when computing update steps. An ensemble size of around 200 is suggested for the current example to ensure physically significant correlations.

Carbon capture and sequestration in subsurface porous formations is a key technology to reduce CO2 emissions to reach the world net-zero emissions by 2050. Modelling studies to identify injection conditions which do not imply fault reactivation and induced seismicity are mandatory to support project implementation. Fault stability analysis requires to couple full-physics reservoir and geomechanical simulations, limiting the possibility to run multiple sensitivities due to the high computational cost. In this work, a cost-efficient workflow suitable for optimization and uncertainty quantification based on Functional Kriging (FK) surrogate models is proposed. Two different FK techniques are implemented and applied to two synthetic models, simulating CO2 sequestration in a depleted sweet gas reservoir using a compositional model with CH4 and CO2 as components. Field gas injection rate, fault modelling parameters but also geomechanical properties are considered as key variables to study fault reactivation. The proxies are trained on output from few simulations to estimate the time dependent Coulomb Failure Function on each element of the fault triangular mesh monitoring the fault behavior during the fluid injection. The results show that FK proxies accurately replicate fault stability indicators with low average errors on a set of blind simulations.

In this work we propose a novel procedure for covariance identification, based on the cross-validation of the correlation between primary and secondary responses within the context of multi-fidelity functional surrogate for uncertainty quantification. The method is validated on a synthetic reservoir model of realistic size and complexity. The outputs are obtained by simulating two different levels of fidelity, referred to as FINE level (or high-fidelity) and COARSE level (or low-fidelity). These outputs are considered as the realizations of a stochastic process and are used for the surrogate construction, by means of the Universal Trace-Cokriging approach. This method incorporates the information embedded in both the primary (FINE) and the secondary (COARSE) data. We show that the surrogate accurately predicts outputs at unknown points of the uncertainty space, but its construction requires a consistently reduced computational effort with respect to the single-fidelity Trace Kriging predictor. As a further contribution we introduce two possible transformation strategies for the analysis of simulator outputs characterized by complex functional shapes.

Nonstationary gaussian random fields are more challenging to work with than their stationary counterparts, as standard methods are not generally applicable to them. One type of nonstationary that is of interest to geomodellers is locally varying anisotropy (LVA), whereby dominant directions of correlation are location-dependent. We describe a practical way to generate realisations of a nonstationary GRF with locally varying anisotropy. Our approach is to generate realisations of stationary gaussian random fields first, and then blend them together in a way that achieves the desired effect. We briefly demonstrate the use of the method on a truncated gaussian model of facies. The approach can be extended beyond locally varying anisotropy, to more general types of nonstationary.

This study presents a novel approach to handle structural uncertainties in the interpretation of stratigraphic structures using a transdimensional sampler, the reversible jump Markov chain Monte Carlo (RJMCMC) algorithm. Moving beyond conventional reservoir simulation models which use a fixed number of layers, the layering size is treated as variable. Hence, the number and position of model layers are unknowns to be determined by the inverse problem. The proposed method extends the application of one-dimensional RJMCMC by introducing a dip parameter, enabling the simulation of sedimentary wedges and erosional features around a well, which is not typically accounted for in standard layer-cake models used for instance in well test interpretation. To keep the model dimension low, a two-dimensional, piecewise constant permeability field is determined for each layering configuration. This inverse method, tested on a synthetic well log, facilitates robust interpretations of geological reservoir geometries and could improve the prediction of permeability spatial distribution in sedimentary settings. Demonstrating successful recovery of target models, the study opens interesting perspectives both for near-well subsurface models and full field reservoir models, and identifies the need for future work on allowing interface intersections to simulate more complex stratigraphic structures.

Pressure transient analysis (PTA) is a standard tool for interpreting reservoir response in well pressure to production and injection operations and characterizing reservoirs in subsurface engineering. It involves studying flow regimes and the associated pressure changes to extract information about well and reservoir parameters. New opportunities to support reservoir decision-making have opened with more abundant availability of PTA data. A wide deployment of permanent downhole gauges (PDG) in wells provided the wealth of high-frequency data for PTA and has encouraged researchers to explore the integration of PTA into current well monitoring and reservoir evaluation and modelling techniques.

This paper describes an automated method designed and developed to detect and classify reservoir flow regime features using pressure transient data obtained from production and injection wells. The proposed is focused on feature recognition model and solving the optimization problem for pattern classification. The method is applied and validated on real field data. The method presented in this study demonstrated a high level of accuracy, successfully identifying flow regime features with an 89% success rate based on a single pressure transient within the chosen real well dataset.

This contribution presents a new workflow for compaction analysis and top seal evaluation by wireline logging data. Sonic and density logs from nine key wells in the Vienna Basin (Austria), providing a continuous depth record from 0 to 3500 m true vertical depth (TVD), were filtered by natural gamma ray, resistivity, delta rho, and caliper to bit size, to identify mudstone intervals with reliable sonic and density signals. The calculated sonic- and density-porosity depth-trends were then quality-checked with core petrophysical data, e.g., helium-porosity and hydrocarbon column heights (HCHs) calculated based on true displacement radii from mercury intrusion capillary porosimetry (MICP). Furthermore, both porosity logs were corrected by their regressions with corresponding helium-porosity values from equal depths to obtain reliable log-based compaction models. Based on these models and the relationship between helium-porosity vs. HCH from MICP, log-based HCH vs. TVD curves were generated. In conclusion, both sonic- and density-porosity trends represent the core petrophysical data well. The HCH model based on density-porosity changed more significantly upon correction, after which both sonic- and density-based HCH trends plot similarly and allow for upscaling of the core petrophysical data to all wells with available wireline logs.

The estimation and uncertainty quantification of channelized reservoirs in a data assimilation framework is very hard to achieve due to the geometrical and topological characteristics of the facies fields. The channelized structure is broken after the updated step of the data assimilation process, and this causes unrealistic updated reservoir models. Geological realism can be achieved in two ways, either by learning or by conditional sampling from the prior distribution. In this study, we train a convolutional autoencoder (CAE) to reconstruct the channelized structure of a reservoir with three facies types namely channel, levee, and shale, where there is a topological transition between them. The training is done with a set composed of pairs of images, of which one is perturbed, and the other has a correct geometrical and topological structure. The CAE is linked with a parameterization of the facies fields and became part of it. The enhanced parameterization is further coupled with ES-MDA for history matching. The results show that the updated facies fields keep the prior channelized structure, and the indicators of a good history matching are fulfilled. This means good results in terms of facies estimation, uncertainty quantification, data match, and prediction capabilities of the updated ensemble.

This paper is presenting a comparison of two methods to quantify uncertainties on production profiles for unconventional reservoirs: the traditional method with type curves fitting and statistical analysis and the geostatistical method with time series decomposition and geostatistical interpolation. Both are applied on a true case study with several hundred wells data base. The results are in favor of geostatistical approach which provides very promising validations. some way forwards are suggested.