- Home
- Conferences
- Conference Proceedings
- Conferences
81st EAGE Conference and Exhibition 2019 Workshop Programme
- Conference date: June 3-6, 2019
- Location: London, UK
- Published: 03 June 2019
61 - 80 of 93 results
-
-
Towards Subsurface ML Metrics
More LessSummaryUse of metrics are key in application of machine learning in any domain. Good metrics allow us to assess performance of algorithms, gain insight into the behaviour of models and understand the impact of model and parameter choices as well as data and feature selections. Shared metrics allow research and engineering communities share knowledge and communicate effectively at a high level, helping progress and reproducibility.
In applying ML in the subsurface, the first port of call is to use standard ML performance metrics such as accuracy, f1_score and r2 score. These metrics are well know but generic. In some cases they provide effective performance indicators, more so in classification tasks. However they generally don't provide much insight into why model is achieving a particular level of performance, or measure performance in terms of expected or acceptable subsurface behaviour.
In this workshop session, we aim to further the discussion on why development of a common set of meaningful subsurface metrics is important for the our community. We highlight some of the gotchas and shortcomings with typical metrics used in machine learning classification and regression tasks and we propose some potentially routes forward.
-
-
-
Ore content estimation based on spatial geological data through 3D convolutional neural networks
Authors B.W. W S R Carvalho, D. Civitarese, D. Szwarcman, P. Cavalin, B. Zadrozny, M. Moreno and S. MarsdenSummaryCurrent tools for identifying new exploration targets for gold are built for geologists to manually interpret data acquired from different sources. Scaling this approach to larger projects is not a trivial task. One possibility to tackle this problem is to use data-driven predictive modeling to discover relationships in the data which can then be applied throughout a mine to more readily identify exploration targets.
Here, we propose a methodology based on machine learning that takes as input data points in space describing measured geological information in a mine, correlates this with the level of gold mineralization in known places through a 3D convolutional neural network, and uses the obtained model to estimate the level of gold mineralization in every region of the mine that has available geological information.
We compare the obtained model with a baseline model and show that it outperforms the baseline in all the metrics used, providing a much more accurate estimate of presence of economic gold for geologists in their investigations.
-
-
-
Surface-To-Borehole Tfem Data Acquisition System Development and Field Test
More LessSummaryA surface-to-borehole time-frequency domain EM (TFEM)data acquisition system has been developed and a successful field test was conducted in the Liaohe Oilfield of China. The field test has successfully acquired downhole 3-component magnetic field and vertical electrical field surface-to-borehole EM data the first time using a 3 km long surface dipole current source at two source locations. The first long dipole current source was placed in the redial direction of the borehole and 3 km away from the wellhead, and the second dipole current source was placed in the redial direction of the borehole and 3.5 km away from the wellhead. The dipole current source injected square wave current into the subsurface to generate the underground TFEM signal. A newly developed 4-level borehole TFEM receiver array was used to record surface-to-borehole TFEM data from the depth of 1,000 m to 1,800 m in the borehole. Each level of borehole TFEM receiver contains a 3-component time domain induction coil package installed in the center of receiver tool and two electrodes are located at each end of the receiver tool with a spacing of 10 m. The surface-to-borehole TFEM data show characteristics of subsurface formation electromagnetic properties and changes at different depth.
-
-
-
Surface-to-Borehole CSEM for Waterflood Monitoring: Modeling & Data Analysis
Authors D. Colombo and G. McNeiceSummaryA full-scale 3D surface-to-borehole CSEM survey was carried out to map the position of the current waterfront around a test well in a producing oil field. Pre-survey modeling studies and experimental results show agreement in terms of measured signal levels, repeatability errors and expected sensitivity to targets. The vertical electric field shows the largest sensitivity to the spatial resistivity distributions in the reservoir. The obtained results provide the baseline for future time-lapse surveys targeting the monitoring of the water-oil saturation changes.
-
-
-
Surface-to-Borehole CSEM for Waterflood Monitoring: 3D Inversion Strategies
Authors G. McNeice and D. ColomboSummaryA 3D surface-to-borehole CSEM survey was acquired in a research well located in a producing oil field to monitor the movement of the waterfront. Variations in the data acquired in the initial baseline survey showed excellent spatial correlation with resistivity variations predicted through reservoir simulation. A 3D finite difference model of the anisotropic overburden and wellbore casing is used to invert for the resistivity of the reservoir surrounding the well. Comparison to borehole observations suggest the CSEM survey robustly recovers the reservoir resistivity distribution within 1.5 km of the well.
-
-
-
Borehole-to-Surface Time-Lapse CSEM Measurements Across a Producing Oil-Field: Repeatability, Vertical Electric Fields, 3D Inversion Including Steel-Casings
More LessSummaryControlled Source Electromagnetics (CSEM) measurements were acquired across an onshore oilfield in Northern Germany between 2014 and 2018 comprising up to 5 transmitters, 29 surface and 3 borehole receivers.
Repeatability of data is an essential prerequisite for reservoir monitoring. Our results suggest that repeatability of CSEM measurements depends on source-receiver distances, source-polarisation, and relocation errors, in particular at sites close to the source. Best repeatability was observed for receiver stations at 2–4 km distance from the source and frequencies <20 Hz. At these stations, phases and amplitudes usually agreed within ±1° and ±5% between repeat measurements.
The vertical electric field (Ez) was measured with a newly developed receiver chain, suspended in a 200 m deep observation borehole. Although amplitudes of Ez are about one to two orders of magnitude smaller than amplitudes of horizontal electric fields, Ez data are stable and show excellent repeatability within <±2° and <±5 % during the 3 years.
For 3D inversion of the field data set, we developed a new methodology which accounts for first order effects of steel-cased wells in the oil field. We demonstrate that both energised and passive well casings can strongly influence the outcome of 3D inversion.
-
-
-
Challenges in the Analysis of Surface-Borehole EM Measurements. Reviewing the Field Distortion due to Metallic Casing
Authors N. Cuevas, D. Colombo, G. McNeice and M. PezzoliSummaryIn this paper a discussion is presented of the main aspects describing the challenges expected and encountered in recording and analyzing STB/BTS data in the presence of the steel casing. To this end, numerical simulations of the casing effect are analysed in relation to the many unknown of the system, i.e. casing, overburden and reservoir properties. The discussion highlights the evident need to either remove the distorting component of the field due to the current flow in the casing, or to directly model the response of the entire system, i.e. including the localized highly conductive anomaly of the metallic casing. In the former case, an estimate of the casing effect could be obtained from the recorded data, but the accuracy may not be enough to remove the casing response while keeping a x100 weaker response of the reservoir. In the later case, aside from the numerical burden of the modelling exercise, the properties of the casing may not be known accurately enough to discriminate the subsurface response from the total field dominated by the casing effect.
-
-
-
Joint Inversion: One Mans Perspective
More LessSummaryTremendous advances have been made in the last two decades in joint inversion of multiple data sets. Coupling of all forms of geophysical data and flow data have been demonstrated. Advances in methods to link the data sets and associated parameters, both by structural and rock-physics coupling approaches have been key to the impressive results currently being demonstrated.
Looking forward it seems that applications of joint inversion to improve our reservoir flow models holds the largest benefit compared to other applications such as structural imaging. In the reservoir application, combining of both structural and rock-physics coupling in joint inversion of seismic, EM, gravity and flow data will undoubtedly be needed to maximize results.
Advances in theory and compute power will lead to stochastic MCMC based sampling techniques as the preferred method for joint inversion. MCMC techniques ability to provide a global solution, the easy of incorporating multiple and desperate a priori information, and accurate uncertainty estimation will all drive the field toward stochastic approaches.
-
-
-
Joint Assimilation of Electromagnetic and Seismic Data - a Stochastic Approach
By F. VossepoelSummaryThe complementary nature of seismic and electromagnetic (EM) data asks for joint inversion of these data sets for reservoir characterisation and monitoring. EM data contain valuable information on the reservoir lithologies and have the ability to discriminate between hydrocarbon- and brine-filled rock. As the EM signal is diffusive, the resolution of the data is generally low, and is best combined with seismic data and appropriate prior models that help constrain the solution space.
To account for uncertainties in the data in a statistically robust manner, we propose to make use of data assimilation techniques. This approach is especially attractive in monitoring applications where dynamic models provide a physically consistent prior estimate of the reservoir characteristics and its state evolution. After providing an overview of the possibilities for joint assimilation of EM and seismic data, a number of data-assimilation examples will illustrate the advantages and disadvantages of the various approaches.
-
-
-
Deep Learning Applications in EM Imaging
Authors J. Byun, S. Oh, K. Noh, D. Yoon and S. J. SeolSummaryDeep learning is now one of the most powerful techniques for solving various scientific and engineering problems. These deep learning techniques have recently begun to be applied in the field of subsurface imaging. As a part of the effort, we have applied the deep learning techniques to the imaging of subsurface from electromagnetic (EM) data. This presentation introduces three cases of the application: salt delineation and monitoring of injected CO2 using towed streamer EM data sets and kimberlite exploration using airborne EM data set. The results with significant qualities open up the possibility of the deep learning as an alternative of the conventional inversion techniques.
-
-
-
Machine Learning for Geoscience Applications
By A. AbubakarSummaryWith roots in Artificial Intelligence (AI), machine learning has evolved over several decades with contributions from various scientific disciplines. In the nineties, remarkable techniques such as probabilistic graphical models, kernel, boosting, and random forest methods emerged and since the mid-2000s, with availability of large datasets and improvements in computational power led to advances in neural network based methods with various deep learning architectures. The latter has resulted in some remarkable innovations in the recent years, and led to wide and visible successes for a spectrum of scientific and commercial applications. With these modern methods, it is now feasible to solve problems with significant underlying complexity; and that too with remarkable accuracy and flexibility. Oil and gas industry acquires large and complex datasets for exploration and field development purposes. However, these datasets are not being optimally used to extract useful information. We believe that with the recent advances in machine learning and computational power, advanced machine learning methods can be used to not only extract useful information from these complex datasets but also reduce the man power costs to process and makes sense of these datasets. Recognizing this potential, over the past several years, we have been actively researching and developing numerous modern machine-learning applications in various domains, including the geosciences. Through examples, our focus will be on the potential of machine learning to address complex geoscientific problems such as well log processing, interpretation, correlation and seismic interpretation.
-
-
-
EM for Land-Based Reservoir Monitoring: Achievements, Challenges and Road Ahead
Authors D. Colombo, G. McNeice, R. Adams and M. DeffenbaughSummaryWe analyze the development, current challenges and the future opportunities of land-based electromagnetic (EM) techniques for reservoir monitoring. Specific reference is made to technologies connecting the surface and the borehole as the most promising approaches to 3D reservoir characterization and monitoring. A few prominent and recent experiments are providing a common base for the definition of an emerging technology for oil reservoirs as well as for CO2 storage monitoring. The path to the establishment of an effective oil field technology shall pass through engineering and standardization for which the current technical challenges need to be identified. The road ahead includes especially the solution of steel casing and pipe responses, new sensors for permanent reservoir monitoring, fiber optics, high voltage signal transmitters, multi-physics integration approaches, integration with reservoir simulators, use of novel machine learning techniques and the adoption of a multi-physics culture enabling the integration of multiple observations to better describe the reservoir dynamics.
-
-
-
Benefits of Joint Seismic and EM Inversion for Hydrocarbon Development Projects
By P. VeekenSummaryReservoir monitoring is illustrated on a Borehole CSEM case study. Ways are examined how to increase the detection power. The following parameters proved important: grid step of computation scheme, incorporation of induction effects, 3D survey design, constrained inversion, guided inversion, timelapse approach. Joint inversion of independent multi-physics datasets is useful: seismics, MT, CSEM, gravmag. It reduces the solution space and enhances the discrimination power for thin beds. Each dataset requires their own quality control. A sequential or simultaneous approach can be chosen. Prestack inversion and AVO effects give access to rock physical parameters. Analysis of the generated data is a challenge. A multi-disciplinary integrated approach is stimulated by the methodology. Joint inversion makes economic sense due to better reservoir management decisions. Improvements and innovations in the workflow should be pursued: better computers, efficient algorithms, extraction of relevant attributes, mixed attributes, principal component analysis, machine learning, neural network analysis.
-
-
-
Ten Years of Becoming Less Uncertain About the Uncertainty of Our Uncertainty Estimates…
By C. HoeltingSummaryWe present a partial review of industry's efforts over the past decade or so to produce physically-based estimates of velocity and depth uncertainty, focusing on the trade-off between velocity and anisotropy. We will discuss the necessity and the difficulty of applying constraints (beyond those inherent in surface seismic data) during this estimation process. We have chosen material and examples with the intent to help initiate a lively discussion.
-
-
-
Resolution Constraints in Bayesian McMC Travel-Time Tomography
More LessSummaryThe resolution matrix from a linearized inversion scheme is integrated into a probabilistic Markov chain Monte Carlo algorithm to provide multi-variate compensations to the random pertubartions. This comes at no additional computational cost other then the prior computation of a generalized inverse, and it improves acceptance ratio, mean step length and mixing properties. The efficiency is tested with a synthetic example from refraction seismic travel-time tomography.
-
-
-
Near-Real Time 3D Seismic Velocity and Uncertainty Models from Ambient Noise, Gradiometry and Neural Network Inversion
Authors A. Curtis, R. Cao, S. Earp, X. Zhang, S. De Ridder and E. GalettiSummaryProducing seismic wave speed models of the Earth's interior with full uncertainty estimates is a grand challenge of geophysics. It is relatively easy to produce uncertainty estimates by linearising (approximating) the nonlinear physics relating models to data, but in strongly nonlinear problems such estimates can be almost worthless. Nonlinear solutions are usually calculated using Monte Carlo methods, requiring weeks of computation due to the high dimensionality of parameter spaces. In addition, using seismic interferometry to obtain reliable surface wave dispersion data from ambient noise often requires several days of recordings.
Clearly both recording and computation timescales must be reduced dramatically to allow ambient noise tomography in near-real time. Recording times must be reduced by changing methods used to obtain dispersion curves. Computation time is constrained by two mathematical results: the ‘curse of dimensionality’ precludes exhaustive Monte Carlo search in high-dimensional parameter spaces, and “No-Free-Lunch” theorems state that improvements over exhaustive search require substantial additional a priori information. Nevertheless, we show that recording times can be reduced to the order of minutes, and that common a priori physical assumptions plus a separation of up-front and real-time computation allow 3D velocity models and uncertainties to be obtained in less than an hour.
-
-
-
Evolutionary Algorithms: from optimization to uncertainties?
More LessSummaryThe main goal of this study is to assess the potential of Evolutionary Algorithms to solve highly non-linear and multi-modal tomography problems (such as traveltime tomography) and their ability to estimate reliable uncertainties. Classical tomography methods apply derivative-based optimization algorithms that require the user to determine the value of several parameters (such as regularization level and initial model) prior to the inversion as they strongly affect the final inverted model. In addition, derivative-based methods only perform a local search dependent on the chosen starting model. Global optimization methods based on Markov Chain Monte Carlo that thoroughly sample the model parameter space are theoretically insensitive to the initial model but turn out to be computationally expensive. Evolutionary algorithms are population-based global optimization methods and are thus intrinsically parallel, allowing these algorithms to fully handle available computer resources. We apply three evolutionary algorithms to solve a refraction traveltime tomography problem, namely the Differential Evolution, the Competitive Particle Swarm Optimization and the Covariance Matrix Adaptation - Evolution Strategy. We apply these methodologies on a smoothed version of the Marmousi velocity model and compare their performances in terms of optimization and estimates of uncertainty.
-
-
-
Surrogate-Based Forward Uncertainty Propagation for Large-Scale Seismic Wave Propagation
Authors P. Sochala, F. De Martin and O. Le MaîtreSummaryThe goal of uncertainty quantification in a forward problem is to estimate the uncertainties in the model output induced by uncertainties in model inputs.
-
-
-
Use of Tomography Velocity Uncertainty in GRV Calculation
SummaryRecently, new seismic products for quantifying uncertainties associated with delivered seismic products have emerged. It is therefore important for reservoir risk analysis to pass the uncertainties along with the data delivered. In an exploration context, identifying potential traps by combining structural and sedimentology information is a challenging process due to the lack of well data to validate the potential presence of reservoir in the lead. In this paper, we illustrate the integration of tomography velocity uncertainties in a resource evaluation workflow and demonstrate the impact on gross rock volume (GRV) distributions for the ranking of potential prospects.
-
-
-
Combining Ensemble Transform Kalman Filter and FWI for Assessing Uncertainties
Authors J. Thurin, R. Brossier and L. MétivierSummaryFull Waveform Inversion (FWI) is an iterative inversion method whose purpose is to retrieve high-resolution models of subsurface physical parameters. Because FWI relies on the solution of a non-linear ill-posed inverse problem, uncertainty estimation is a crucial issue in practical applications, both in seismology and exploration seismic. While uncertainty assessment is a strongly desired feature for FWI, it remains a challenging problem. In this presentation, we investigate uncertainty estimation within the framework provided by ensemble data-assimilation strategies. We combine the Ensemble Transform Kalman Filter and FWI. We review the concepts underlying our ETKF-FWI method, discuss its limitations and appeals for uncertainty estimation, and illustrate it on a 2D multiparameter inversion of an exploration scale field dataset.
-