Spelling suggestions: "subject:"subsurface characterization"" "subject:"ubsurface characterization""
1 |
[en] DEEP PHYSICS-DRIVEN STOCHASTIC SEISMIC INVERSION / [pt] INVERSÃO SÍSMICA ESTOCÁSTICA COM APRENDIZADO PROFUNDO ORIENTADO À FÍSICAPAULA YAMADA BURKLE 28 August 2023 (has links)
[pt] A inversão sísmica é uma etapa essencial na modelagem e caracterização de reservatórios que permite a estimativa de propriedades da subsuperfície a partir dos dados de reflexão sísmica. os métodos convencionais usualmente possuem um alto custo computacional ou apresentam problemas relativos à não-linearidade e à forte ambiguidade do modelo de inversão sísmica. Recentemente, com a generalizaçãodo aprendizado de máquina na geofísica, novos métodos de inversão sísmica surgiram baseados nas técnicas de aprendizado profundo. Entretanto, a aplicação prática desses métodos é limitada devido a ausência de uma abordagem probabilística capaz de lidar com as incertezas inerentes ao problema da inversão sísmica e/ou a necessidade de dados de treinamento completos e representativos. Para superar essas limitações, um novo método é proposto para inverter dados de reflexão sísmica diretamente para modelos da subsuperfície de alta resolução. O método proposto explora a capacidade das redes neurais convolucionais em extrair representações significativas e complexas de dados espacialmente estruturados, combinada à simulação estocástica geoestatística. Em abordagem auto-supervisionada, modelos físicos são incorporados no sistema de inversão com o objetivo de potencializar o uso das medições indiretas e imprecisas, mas amplamente distribuídas do método sísmico. As realizações geradas com simulação geoestatística fornecem informações adicionais com maior resolução espacial do que a originalmente encontrada nos dados sísmicos. Quando utilizadas como entrada do sistema de inversão, elas permitem a geração de múltiplos modelos alternativos da subsuperfície. Em resumo, o método proposto é capaz de: (1) quantificar as incertezas das previsões, (2) modelar a relação complexa e não-linear entre os dados observados e o modelo da subsuperfície, (3) estender a largura de banda sísmica nas extremidades baixa e alta do espectro de parâmetros de frequência, e (4) diminuir a necessidade de dados de treinamento anotados. A metodologia proposta é inicialmente descrita no domínio acústico para inverter modelos de impedância acústica a partir de dados sísmicos pós-empilhados. Em seguida, a metodologia é generalizada para o domínio elástico para inverter a partir de dados sísmicos pré-empilhados modelos de velocidade da onda P, da velocidade da onda S e de densidade. Em seguida, a metodologia proposta é estendida para a inversão sísmica petrofísica em um fluxo de trabalho simultâneo. O método foi validado em um caso sintético e aplicado com sucesso a um caso tridimensional de um reservatório brasileiro real. Os modelos invertidos são comparados àqueles obtidos a partir de uma inversão sísmica geoestatística iterativa. A metodologia proposta permite obter modelos similares, mas tem a vantagem de gerar soluções alternativas em maior número, permitindo explorar de forma mais efetiva o espaço de parâmetros do modelo quando comparada à inversão sísmica geoestatística iterativa. / [en] Seismic inversion allows the prediction of subsurface properties from seismic reflection data and is a key step in reservoir modeling and characterization. Traditional seismic inversion methods usually come with a high computational cost or suffer from issues concerning the non-linearity and the strong non-uniqueness of the seismic inversion model. With the generalization of machine learning in geophysics, deep learning methods have been proposed as efficient seismic inversion methods. However, most of them lack a probabilistic approach to deal with the uncertainties inherent in the seismic inversion problems and/or rely on complete and representative training data, which is often scarcely available. To overcome these limitations, we introduce a novel seismic inversion method that explores the ability of deep convolutional neural networks to extract meaningful and complex representations from spatially structured data, combined with geostatistical stochastic simulation to efficiently invert seismicn reflection data directly for high-resolution subsurface models. Our method incorporates physics constraints, sparse direct measurements, and leverages the use of imprecise but widely distributed indirect measurements as represented by the seismic data. The geostatistical realizations provide additional information with higher spatial resolution than the original seismic data. When used as input to our inversion system, they allow the generation of multiple possible outcomes for the uncertain model. Our approach is fully unsupervised, as it does not depend on ground truth input-output pairs. In summary, the proposed method is able to: (1) provide uncertainty assessment of the predictions, (2) model the complex non-linear relationship between observed data and model, (3) extend the seismic bandwidth at both low and high ends of the frequency parameters spectrum, and (4) lessen the need for large, annotated training data. The proposed methodology is first described in the acoustic domain to invert acoustic impedance models from full-stack seismic data. Next, it is generalized for the elastic domain to invert P-wave velocity, S-wave velocity and density models from pre-stack seismic data. Finally, we show that the proposed methodology can be further extended to perform petrophysical seismic inversion in a simultaneous workflow. The method was tested on a synthetic case and successfully applied to a real three-dimensional case from a Brazilian reservoir. The inverted models are compared to those obtained from a full iterative geostatistical seismic inversion. The proposed methodology allows retrieving similar models but has the advantage of generating alternative solutions in greater numbers, providing a larger exploration of the model parameter space in less time than the geostatistical seismic inversion.
|
2 |
Imagerie sismique 4D quantitative en milieux complexes par l'inversion 2D de forme d'onde complète / Quantitative 4D seismic imaging in complex media using 2D full-waveform inversionAsnaashari, Amir 14 October 2013 (has links)
Le suivi temporel est un processus d’acquisition et d’analyse d’acquisitions multiples répétées au même endroit sur la même cible à différentes périodes de temps. Cela s’applique bien à l’exploration sismique quand les propriétés de la cible varient au cours du temps comme pour les réservoirs pétroliers. Cette technique de sismique, dite 4D en raison de l’intégration du temps dans la construction des images, permet une détection et une estimation des variations du sous-sol survenues lors de l’évolution en temps du milieu. En particulier, dans l’industrie, le suivi et la surveillance peuvent améliorer notre compréhension d’un réservoir de pétrole/gaz ou d’un site de stockage de CO2. Analyser la sismique 4D peut aider à mieux gérer les programmes de production des réservoirs. Ainsi, des acquisitions répétées permettent de suivre l’évolutiondes fronts de fluide injectés: on peut optimiser les programmes d’injection de fluides pour une récupération améliorée des hydrocarbures (enhanced oil recovery). Plusieurs méthodes ont été développées pour l’imagerie variable dans le temps en utilisant les informations des ondes sismiques. Dans ma thèse, je montre que l’inversion de forme d’onde complété (FWI) peut être utilisée pour cette imagerie. Cette m´méthode offre des images sismiques quantitatives haute résolution. Elle est une technique prometteuse pour reconstruire les petites variations de propriétés physiques macro-échelle du sous-sol. Sur une cible identifiée pour ces imageries 4D, plusieurs informations a priori sont souvent disponibles et peuvent être utilisées pour augmenter la résolution de l’image. J’ai introduit ces informations grâce à la définition d’un modèle a priori dans une approche classique FWI en l’accompagnant de la construction d’un modèle d’incertitudes a priori. On peut réaliser deux reconstructions indépendantes et faire la différence les reconstruits: on parle de différence parallèle. On peut aussi effectuer une différence séquentielle o`u l’inversion de l’ensemble de données de la second acquisition, dite moniteur, se fait `a partir du modèle de base et non plus à partir du modèle utilisé initialement. Enfin, l’approche double-différence conduit à l’inversion des différences entre les deux jeux de données que l’on rajoute aux données synthétiques du modèle de base reconstruit. J’étudie quelle stratégie est à adopter pour obtenir des changements vitesse plus précis et plus robustes. En plus, je propose une imagerie 4D ciblée en construisant un modèle d’incertitude a priori grâce `a une information (si elle existe) sur la localisation potentielle des variations attendues. Il est démontré que l’inversion 4D ciblée empêche l’apparition d’artéfacts en dehors des zones cibles: on évite la contamination des zones extérieures qui pourrait compromettre la reconstruction des changements 4D réels. Une étude de sensibilité, concernant l’échantillonnage en fréquence pour cette imagerie 4D, montre qu’il est nécessaire de faire agir simultanément un grand nombre de fréquences au cours d’un cycle d’inversion. Ce faisant, l’inversion fournit un modèle de base plus précis que l’approche temporelle, ainsi qu’un modèle des variations 4D plus robuste avec moins d’artéfacts. Toutefois, la FWI effectuée dans le domaine temporel semble être une approche plus intéressante pour l’imagerie 4D. Enfin, l’approche d’inversion 4D régularisée avec un modèle a priori est appliquée sur des ensembles de données réelles d’acquisitions sismiques répétées fournis par TOTAL. Cette reconstruction des variations locales s’inscrit dans un projet d’injection de vapeur pour améliorer la récupération des hydro-carbures: Il est possible de reconstituer des variations de vitesse fines causées par la vapeur injectée. / Time-lapse monitoring is the process of acquiring and analysing multiple seismic surveys, repeatedat the same place at different time periods. This seismic technique, called 4D becauseof the integration time in the construction of images, allows detection and estimation of thesubsurface parameter variations occured through a time evolution. Particularly, in industries,the monitoring can improve our understanding of a producing oil/gas reservoir and CO2 storagesite. Analyzing the time-lapse seismics can help to better manage production programsof reservoirs. In addition, repeated surveys can monitor the evolution of injected fluid frontsand can permit to optimize injection programs which are considered for enhanced oil recovery(EOR) techniques.Several methods have been developed for time-lapse imaging using seismic wave information.In my thesis, I show that full waveform inversion (FWI) can be used for time-lapseimaging, since this method delivers high-resolution quantitative seismic images. It is a promisingtechnique to recover small variations of macro-scale physical properties of the subsurface.In time-lapse applications, several sources of prior information are often available and shouldbe used to increase the image reliability and its resolution. I have introduced this informationthrough a definition of a prior model in a classical FWI approach by also considering a prioruncertainty model. In addition, I have suggested a dynamic weighting to reduce the importanceof these prior models in the final convergence. In realistic synthetic cases, I have shownhow the prior model can reduce the sensitivity of FWI to a less accurate initial model. It istherefore possible to obtain a highly accurate baseline model for 4D imaging.Once the baseline reconstruction is achieved, several strategies can be used to assess thephysical parameter changes. We can make two independent reconstructions of baseline andmonitor models and make subtraction of the two reconstructed models. This strategy is calledparallel difference. The sequential difference strategy inverts the monitor dataset starting fromthe recovered baseline model, and not from the model used initially. Finally, the doubledifferencestrategy inverts the difference data between two datasets which are added to thecalculated baseline data computed in the recovered baseline model. I investigate which strategyshould be adopted to get more robust and more accurate time-lapse velocity changes. Inaddition, I propose a target-oriented time-lapse imaging using regularized FWI including priormodel and model weighting, if the prior information exists on the location of expected variations.It is shown that the target-oriented inversion prevents the occurrence of artifacts outsidethe target areas, which could contaminate and compromise the reconstruction of the effectivetime-lapse changes.A sensitivity study, concerning several frequency decimations for time-lapse imaging, showsthat the frequency-domain FWI requires a large number of frequencies inverting simultaneously.By doing so, the inversion provides a more precise baseline model and more robust time-lapsevariation model with less artifacts. However, the FWI performed in the time domain appearsto be a more interesting approach for time-lapse imaging considering all frequency content.Finally, the regularized time-lapse FWI with prior model is applied to the real field timelapsedatasets provided by TOTAL. The reconstruction of local variations is part of a steaminjection project to improve the recovery of hydrocarbons: it is possible to reconstruct thevelocity variations caused by the injected steam.
|
Page generated in 0.0953 seconds