Third EAGE Conference on Seismic Inversion

We have developed a novel approach for petrophysical inversion in reservoir modeling, utilizing the Incremental Gaussian Mixture Network (IGMN) with uncertainty propagation. Traditional methods for understanding subsurface reservoir properties often struggle with data and modeling uncertainties, particularly in dealing with the diverse range of rock and fluid properties that can exhibit multimodality. This challenge has led to an increasing adoption of methods relying on Gaussian Mixture Models (GMM), albeit requiring meticulous parameterization. IGMN distinguishes itself by offering adaptability in learning and modeling multimodal data distributions, thus obviating the need for intricate parameterization. Moreover, integrating uncertainty propagation enhances estimation reliability and provides insights into prediction confidence levels. Synthetic data resembling real-world reservoir scenarios are employed to showcase the methodology’s effectiveness in terms of accuracy, robustness, and computational efficiency. These results underscore the potential of the proposed approach in tackling the challenges associated with reservoir characterization tasks.

Geothermal energy provides a sustainable energy source by utilizing Earth’s natural heat, and to unlock its potential, understanding subsurface properties like porosity and permeability is crucial. However, traditional methods for estimating these properties face challenges due to sparse exploration data. To overcome this, a hybrid approach combines machine learning with synthetic data generation, enhancing predictions.

The process begins with rock physics modeling and statistical analysis, followed by simulating pseudo-wells and generating synthetic seismic data. A deep neural network is then trained on this synthetic catalog, enabling predictions of porosity and permeability. Seismic inversion outputs in the form of elastic attributes are essential when utilizing deep neural networks that rely on post-stack seismic data, to enhance its efficiency and effectiveness.

Applied to the SCAN project in the Netherlands, the workflow demonstrates promising results. Convolutional neural networks accurately predict target properties crucial for geothermal projects, despite limited well data. By leveraging inversion outputs and machine learning, this approach revolutionizes geothermal exploration, offering a viable solution for sustainable energy production.

The study delves into seismic inversion, a pivotal technique in reservoir characterization, using the One-Dimensional Stochastic Inversion (ODiSI) method on a dataset from the West of Shetland basin. Initially focusing on binary lithofacies classification (sand and shale), the study subsequently incorporates a third facies to improve prediction accuracy.

ODiSI simultaneously estimates facies, reservoir properties, and impedances, leveraging pseudo-wells and calibrated rock physics models. It utilizes optimized chi-projection volumes for lithology and fluid discrimination, enhancing seismic anomaly differentiation.

Results from the binary classification revealed mis-predictions due to the dominance of fluid chi-projection volumes. To address this, a third facies, shaley-sand, was introduced, enabling better modelling of sand facies above and below the fluid contact.

The updated approach yielded continuous sand prediction, rectifying previous mis-predictions. It showcased varying probabilities of sand within layers and highlighted facies changes between shale and shaley-sand. Comparison with the initial inversion illustrated improved predictions below the fluid contact, enhancing gas presence modelling.

ODiSI proved promising for lithofacies prediction in challenging geological settings. Its stochastic nature and flexibility in exploring multiple scenarios offer valuable insights into reservoir facies distribution and behaviour.

The present study introduces a geophysical workflow that focuses on interpreting 4D seismic signal with the aim to support monitoring hydrocarbon production over time, potentially improving reservoir knowledge, especially in areas far from well locations. This can ultimately reduce risks associated with reservoir development and field management decisions. The Simultaneous Elastic Inversion (SEI) workflow aims to transform seismic reflectivity to elastic contrasts and to generate absolute properties that describe target rocks in terms of P-Impedance, S-Impedance, and Vp/Vs.

Once the inverted elastic properties are properly processed, the 4D interpretation workflow involves calculating the difference between the baseline and each subsequent survey. The resulting anomalies are then modelled as geobodies whose cells are populated with data from various elastic attributes in attempt to assign a unique “4D seismic signature” to each anomaly. Finally, observations from the 4D interpretation are compared and integrated with the field production history. This combined analysis improves understanding of the reservoir dynamic behavior and supports decision-making for future field development.

An analysis was performed on 4D inversion methods for time-lapse synthetic data generated with a petroelastic model of a carbonate reservoir. The aim was to understand their characteristics, capabilities, and limitations. Aspects such as result accuracy, algorithm robustness, computational requirements, and ease of implementation were investigated. The focus was on the applicability of these methods in the context of brazilian pre-salt reservoirs, where impedance variations are low, on the order of 2%. The 4D methods used were simultaneous inversions applied sequentially (Sparse Spike approach) or jointly (Model Based technique), and elastic inversion. The obtained results showed the importance of tests with synthetic data to identify limitations of commercially available methods when applied to challenging time-lapse data.

We used a 2D DAS seismic line that was recently acquired in West Africa to explore to which extent DAS data can be used for seismic inversion, given the fundamentally different measurements of DAS in comparison to geophone recordings. The fact that DAS data can be used quantitatively opens the door for using DAS for reservoir monitoring in particular for CCS projects.

The dominance of shales combined with the shape and preferred orientation of intrinsically anisotropic clay minerals within these shales means that Vertical Transverse Isotropy (VTI) can be expected to impact seismic amplitude variations with offset in all clastic environments. It is normal to predict and mitigate the kinematic effects of VTI during seismic processing, but the dynamic effects are often not explicitly included in seismic inversion. The assumption of isotropy, suggesting that the anisotropy can be mitigated through appropriate wavelet estimation, may be reasonable depending on the objectives of the inversion and the relationships between the elastic and anisotropic contrasts. These relationships will depend on the geological environment. In an example from a Deltaic environment the relationships are highly correlated and the errors in assuming anisotropy are sufficiently small to not impact the identification of gas sands. In a Turbidite example, isotropy cannot be assumed without significantly impacting the interpretation of inversion results, and anisotropy needs to be handled explicitly. 2D modelling prior to inversion is a critical workflow element to allow for an understanding of the impact of any potential anisotropy in achieving inversion objectives.

Pseudo-well modeling and trace matching Hit Cube inversion has been established as a method that can allow geological, elastic, stratigraphic and diagenetic details into models without actively using a background model. In unconventional domain, the difference in seismic properties are very subtle over large swath of unconventional land base.

In this study we will show how overburden and underlying lithologies could affect the seismic signature usually missed when we are not testing all the probabilities. Pseudo-well modeling also allows simulation of combination of parameters producing thousands of property combinations.

Two of the major unconventional plays of Western Canada are investigated.

We present a novel stochastic sampling algorithm, referred to as optimised rejection sampling, for inverting seismic data for facies probabilities. The new algorithm combines a bias-free rejection sampling element combined with an optimisation step to find high predictability facies vectors that when averaged provide estimates of facies probabilities. We show results on both modelled and real datasets that demonstrate significant improvement over conventional sampling algorithms.

In this paper, we present an efficient scattering approach for frequency domain full-waveform inversion (FWI) in 2D transversely isotropic media with a vertical axis of symmetry (VTI). The forward model is based on a transformation of the elastodynamic wave equation into an equivalent integral equation that can be solved in a reasonable efficient manner by using an FFT-accelerated Krylov subspace method. The ill-posed nonlinear multi-parameter inverse problem is formulated as a local optimization problem and we use the L-BFGS method to find the minimum of an objective function that includes a hybrid multiplicative-additive multi-parameter regularization term in addition to the multi-component data mismatch term. The L-BFGS method accounts for approximate Hessian information and this is essential when performing a multi-parameter FWI. To further reduce the effects of multi-parameter cross-talk, we parametrize the 2D VTI medium by four independent A-parameters. An advantage of using the A-parameters in this context is methods for constructing prior models for anisotropic FWI have recently been developed on the basis of A-parameter based ray tomography. We demonstrate the performance of our algorithm by a numerical example related to the Hess model, which includes a high-contrast salt body embedded in a VTI sedimentary environment.