Reservoir characterization involves integration of different types of data to understand the subsurface rock properties. To incorporate multiple well log types into reservoir studies, estimating missing logs is an essential step. We have developed a method to estimate missing well logs by using a bidirectional convolutional long short-term memory (bidirectional ConvLSTM) cascaded with fully connected neural networks. We train the model on 177 wells from mature areas of the UK continental shelf (UKCS). We test the trained model on one blind well from UKCS, three wells from the Volve field in the Norwegian continental shelf, one well from the Penobscot field in the Scotian shelf offshore Canada, and one well from the Teapot Dome data set in Wyoming. The method takes into account the depth trend and the local shape of logs by using ConvLSTM architecture. The method is examined on sonic log prediction and can produce an accurate prediction of sonic logs from gamma-ray, density, and neutron porosity logs. The advantages of our method are that it is not applied on an interval by interval basis like rock-physics-based methods and it also outputs the uncertainties facilitated by dropout layers and Monte Carlo sampling at inference time.
In unconventional plays, wells are drilled at an unprecedented rate. This, together with technical challenges in terms of complex stratigraphy, multiple play types, variable rock properties, and various elements of pore pressure, geomechanics, fracturing, and diagenesis, calls for more sophisticated, faster, consistent, and wider ranging analytical tools. Given the scale of the work — i.e., the number of wells — performing classical workflows for petrophysics, pore pressure, and geomechanics prediction can be impractical (if not impossible) due to turnaround considerations. Also such workflows might not use any preexisting regional studies efficiently. In principle, a machine learning approach can mitigate these shortcomings. We show that a supervised deep neural network approach can be an alternative innovative tool for petrophysical, pore pressure, and geomechanics analysis enabling the use of all the previously interpreted data to devise solutions that simultaneously integrate wide-ranging wellbore and wireline logs. Beyond that, a similar approach is taken to predict a certain number of attributes solely from seismically derived properties, which allows one to compute volumetric models. The application of such an algorithm is shown on a Permian case study in which the automatic neural-network-based algorithms achieve reasonable accuracy in a fraction of the time.
Conventional reservoir-characterization techniques utilize amplitude-variation-with-offset (AVO) analysis to invert for the elastic parameters or directly for the physical properties of reservoirs. However, the quality of AVO inversion is degraded by errors in the velocity model, inaccurate amplitudes, and structural complexity. Whereas full-waveform inversion (FWI) potentially represents a much more powerful tool for reservoir characterization. FWI strongly relies on the accuracy of the initial model and suffers from parameters trade-offs. Here, we use a probabilistic Bayesian framework to supplement data fitting with rock-physics constraints based on geologic facies obtained from borehole information (well logs). The advantages of the facies-based FWI are demonstrated on a structurally complex isotropic elastic model and on a 3D layered VTI (transversely isotropic with a vertical symmetry axis) medium. In particular, the tests show that our algorithm can operate without ultra-low-frequency data required by conventional FWI and can reduce crosstalk between the medium parameters.
Seismic data are uniformly sampled images of geologic structures, and they spatially cover a big area as opposed to well log data, which are often scattered sparsely within the seismic cube. A common strategy to extend the coverage of the well log data is to interpolate them away from the boreholes on the seismic image grid. Image-guided interpolation is a form of blended neighbor interpolation in which the structural information that guides the interpolation is calculated from a seismic image and represented by structure tensors. Deriving this structural information from only the seismic image might not lead to desirable interpolated results, especially in cases of complex structure, a low-quality seismic image, and few well logs. In such cases, constraining the structure tensors with additional information is beneficial. We have developed a three-step hybrid method in the log-Euclidean domain to use interpreted horizons as well as the seismic image to construct the final structure tensor field. Also, we determined how to incorporate the curvature attributes from horizons for better performance of the algorithm in the presence of faults. Synthetic and real data examples determined the ability of this new technique to achieve better conformance of interpolated properties with the seismic image and interpreted horizons.
Machine learning has been around for decades or, depending on your view, centuries. To consider the tools and underpinnings of machine learning, one would need to go back to the work of Bayes and Laplace, the derivation of least squares, and Markov chains, all of which form the basis and the probability construct used pervasively in machine learning. There has been a flood of progress between 1950 (with Alan Turing's proposal of a learning machine) and early 2000 (with practical applications of deep learning in place and more recent advances such as AlexNet in 2012). Deep learning has demonstrated tremendous success in a variety of application domains in the past few years, and with some new modalities of applications, it continues to open new opportunities. The recent popularity and emergence of machine learning in the oil and gas industry is likely due to the abundance of unused or overlooked data and the economic need to extract additional information from the data currently used. Additionally, there is an unprecedented availability of computing power, easy-to-use coding libraries, and application programming interfaces, as well as recent and significant advances in various flavors of neural networks. In this paper, we will attempt to show how machine learning can assist geoscientists in performing routine tasks in a much shorter time frame. We assert that there is a great opportunity for geoscientists to learn from machines, use these techniques to quality check their work, and gain nuanced insights from their data. Another advantage is that these approaches lead to the optimization of machine learning workflows by providing more accurate training data sets thus driving continuous learning and enhancement of the model.
Deep neural networks offer numerous potential applications across geoscience; for example, one could argue that they are the state-of-the-art method for predicting faults in seismic data sets. In quantitative reservoir characterization workflows, it is common to incorporate the uncertainty of predictions. Thus, such subsurface models should provide calibrated probabilities and the associated uncertainties in their predictions. It has been shown that popular deep learning-based models often are miscalibrated, and due to their deterministic nature, they provide no means to interpret the uncertainty of their predictions. We compare three different approaches for obtaining probabilistic models based on convolutional neural networks in a Bayesian formalism, namely, deep ensembles, concrete dropout, and stochastic weight averaging Gaussian (SWAG). These methods are consistently applied to fault detection case studies in which deep ensembles use independently trained models to provide fault probabilities and concrete dropout represents an extension to the popular dropout technique to approximate Bayesian neural networks. Finally, we apply SWAG, a recent method that is based on the Bayesian inference equivalence of minibatch stochastic gradient descent. We provide quantitative results in terms of model calibration and uncertainty representation, as well as qualitative results on the synthetic and real seismic data sets. Our results indicate that the approximate Bayesian methods, concrete dropout and SWAG, provide well-calibrated predictions and uncertainty attributes at a lower computational cost compared to the baseline deep ensemble approach. The resulting uncertainties also offer a possibility to further improve the model performance, as well as enhancing the interpretability of the models.