Rustam Akhmadiev, Milad Bader, Guillaume Barnier, Biondo L. Biondi, Robert G. Clapp, Stuart Farris, Paige Given, Fantine Huot, Joseph Jennings, Min Jun Park, Ariel Lellouch, Bin Luo, Rahul Sarkar, Adam Shugar, Joe Stitt, Jonathan Voyles, Rachael Wang, Siyuan Yuan
SEP-187 Reports (Password Protected) (May, 2022)
Generating synthetic models for machine learning - Robert G. Clapp
ABSTRACT Using synthetic data is the only realistic way to make enough labeled data for neural networks. The better these synthetic datasets replicate real data, the more effective they become. Simplex synthetic noise can produce realistic noise that can be used for things such as simulating deposition, fault planes, and compression events. Simplifying the hyper-parameter space describing specific types of features enable faster model building. The improved synthetic model generator described in this paper is both more computationally efficient and produces more realistic models than previous model generators.
Reproducibility with containers - Robert G. Clapp
ABSTRACT Reproducible research plays a key role in advancing science. Software containers can be used to perfectly reproduce a local computing environment ensuring the successful building of software. In conjunction with git repositories, containing how the results of a paper generated, an easily shareable snapshot can be generated.
Adversarial Regularizers for Seismic Inversion Applications - Joe Stitt, Stuart Farris, Robert G. Clapp, and Biondo L. Biondi
ABSTRACT Posing our regularization operator to be a deep neural network discriminator allows for a strong balance of data and knowledge-driven approaches for geophysical inverse processes. Through this balance, it has been proven that the adversarial deep neural network is able to learn a range of geologic and general model features and provide a distinct sensitivity for the amount of penalty provided to the objective function. Our regularization functional suppresses undesirable features within our unregularized model result by focusing on the model’s posed “style”. This problem is first investigated with the ill-conditioned Dix inversion by forcing a blocky style to the inputted starting interval velocity model given a range of vertical smooth to vertical blocky interval velocity training models produced from a synthetic model generator. Implementing the proven neural network into the optimization problem is the next step in a line of provided directions that will be explored in the near-future.
Identifying Geologic Facies Through Seismic Dataset-to-Dataset Transfer Learning using Convolutional Neural Networks - Joe Stitt, Adam Shugar, and Rachael Wang
ABSTRACT We combine the ideas of transfer learning and semantic segmentation of seismic facies in a dataset-to-dataset transfer learning research problem with the goal of reducing the number of inputted labeled seismic slices while also improving computational cost and accuracy of predicting geological facies. A U-Net architecture is used to train a baseline model with an offshore New Zealand dataset to transfer learned features to a second model using a Netherlands North Sea dataset. The main hyperparameter tuning goal is focused on maximizing the number of frozen layers while minimizing the number of inputted examples. With a 95% validation accuracy from the baseline model and multiple iterations of implementing the correct regularization parameters for limited overfitting, we find a semi-successful method to an 82% degree of test accuracy with 100 training and validation inputs and 78% accuracy with only ten training and validation examples. These results show that general locations of seismic facies are predicted accurately in their respective spatial areas with a minimal number of inputted examples. Further recommendations are noted to hopefully improve this result in the future to act as a fully automatic facies predictor rather than a baseline interpretation tool.
Vertical Seismic Profile (VSP) Processing using Distributed Acoustic Sensing (DAS) Data from Onshore Japan - Jonathan Voyles
ABSTRACT The recent growth in interest of distributed acoustic sensing (DAS) data is renewing research efforts in vertical seismic profiling (VSP) mainly because of its cost-effectiveness over long periods of acquisition for monitoring purposes. We now have access to DAS VSP field data from a pilot project in Ichihara, Japan. We use the near-zero-offset VSP data with repeated vibroseis sources to determine an acoustic properties profile of the subsurface. We verify our VSP processing by generating and processing realistic 2D synthetic elastic DAS data.
Moment tensor inversion of perforation shots using distributed acoustic sensing - Milad Bader, Robert G. Clapp, and Biondo Biondi
ABSTRACT Perforation shots in unconventional wells excite seismic waves that travel through the reservoir layer and get recorded by distributed acoustic sensing (DAS) fiber. These waves can be used to characterize either the elastic properties of the reservoir or the perforation itself. In both cases, the seismic source representing the perforation shot must be known. To the authors’ knowledge, there is no available representation of such a source. We propose a three-mechanisms source model with three structured moment tensors whose magnitudes are to be determined. Following a three-step procedure, we invert the resolvable moment tensor components of 100 perforation shots using DAS data. The result can be used to interpret the quality of the perforation and to conduct subsequent elastic waveform inversions.
Elastic full-waveform inversion in pre-stimulated unconventional reservoir using distributed acoustic sensing - Milad Bader, Robert G. Clapp, and Biondo Biondi
ABSTRACT Distributed acoustic sensing (DAS) is essential for characterizing and monitoring unconventional oil and gas reservoirs. However, its use in full-waveform inversion for high-resolution imaging of the reservoir formation remains limited. Following previous work on modeling and inverting the seismic sources representing perforation shots, we apply full-waveform inversion using DAS data to estimate the elastic properties of the unconventional reservoir layer before hydraulic stimulation. We utilize various techniques to reduce the source effects on the inversion. Our inverted P-wave velocity is consistent with an independent study based upon dispersion analysis of leaky waves.
Realistic synthetic data generation using Neural Style Transfer: Application to automatic fault interpretation - Min Jun Park, Joseph Jennings, Bob Clapp, and Biondo Biondi
ABSTRACT The robust and automatic detection of faults within seismic images remains a challenging issue within the field of seismic interpretation. Recently, supervised deep learning-based approaches using synthetics have shown much promise towards accomplishing a robust and automatic fault detection algorithm. However, training only on synthetic data often results in noisy and low-quality predictions due to the massive gap in data features that can exist between real and synthetic images. In order to overcome this issue, we propose a Neural Style Transfer workflow for incorporating real image features into synthetic images prior to training. With our workflow, we demonstrate that the updated synthetic images after Neural Style Transfer have similar characteristics to a real 2D seismic image from the Gulf of Mexico and are more suitable to be used as training images than the original synthetic images. To verify the effectiveness of our workflow, we train a model on the synthetic data generated from our workflow and a model from the original synthetic images and compare the results of automatic fault interpretation on a 2D image from the Gulf of Mexico. Our comparisons show that the model trained on the updated images provides less noisy and more accurate predictions than the model trained on the original images.
Automatic Distributed Acoustic Sensing (DAS) Seismic Event Detection through Convolutional Neural Networks: Microseismic Event Evolution Over Time - Paige Given, Fantine Huot, Bin Luo, Ariel Lellouch, Robert G. Clapp, Tamas Nemeth, Kurt Nihei, and Biondo L. Biondi
ABSTRACT Distributed Acoustic Sensing (DAS) data is substantial in size (approximately 14TB per day for modern collection), making the development of automatic seismic detection methods a necessity. Additionally, the analysis of these seismic events over time can serve as important information for engineers hoping to improve hydraulic fracturing strategies. We implement a python script that systematically imports data, processes it, creates sliding windows throughout time and space, and inputs these windows into a Convolutional Neural Network (CNN) that automatically determines the probability of an inputted window having a coherent seismic event. The proposed workflow allows us to detect seismic events in a given dataset with superhuman efficiency and accuracy. Comparing a subset of data from a single stimulation stage to that of a subset collected approximately 1-week after well stimulation, we found a drop in the number of windows predicted to have microseismic events in the data from 3.67% to 2.04%. By analyzing the post-stimulation data and comparing it to data during stimulation, we can analyze event distribution and fracture evolution over time.
The Lippmann-Schwinger equation for a background velocity depending only on depth - Rahul Sarkar and Biondo Biondi
ABSTRACT In this paper, we propose a technique to solve the Lippmann-Schwinger equation for a background velocity depending on a single coordinate (such as depth), extending the method developed for the linearly varying background case in Sarkar and Biondi (2021). The linear system that results is solved using Krylov subspace-based iterative methods, making it possible to solve the Lippmann-Schwinger equation for perturbations of arbitrary magnitude. Depending on the iterative solver, the computational cost of each iteration is determined by the cost of computing the forward and adjoint actions of the integral kernel appearing in the Lippmann-Schwinger equation. The particular case of using GMRES as the iterative solver is discussed, and it is shown that the solution obtained at the kth iteration is the best in the subspace spanned by k − 1 orders of scattered wavefields in a specific sense. To demonstrate this, we provide 2D numerical examples illustrating the differences between the GMRES solution after k iterations and the result of summing k terms of the Born-Neumann series. Some computational aspects of Krylov subspace-based iterative solvers are demonstrated using 2D numerical experiments, such as the impact of a good initial starting solution on the number of iterations taken to solve the Lippmann-Schwinger equation and comparisons of different iterative solvers. Moreover, computational tests suggest that solving the Lippmann-Schwinger equation is almost always cheaper than solving the Helmholtz equation in terms of the iteration complexity of the Krylov subspace solvers, and in some cases even less in terms of actual wall times.
Acoustic waveform inversion with the Lippmann-Schwinger equation constraint - Rahul Sarkar and Biondo Biondi
ABSTRACT The theory for performing acoustic waveform inversion in the frequency domain is developed in both 2D and 3D, with the Lippmann-Schwinger equation as the constraint. The Lippmann-Schwinger equation that we consider has the special structure that the background velocity depends only on depth, in which case an efficient method exists to compute the forward and adjoint actions of the integral kernel. The inversion is treated as a joint optimization problem, where both the model to be inverted and the wavefields for each source and frequency are simultaneously treated as optimization variables. Here we explore the penalty method formulation of the problem, and a two-step alternating minimization strategy to solve it is presented, where each step involves solving a linear least squares problem. Finally numerical examples in 2D illustrating the inversion method is presented.
One-way extended full-waveform inversion (eFWI) in the frequency-domain - Rustam Akhmadiev, Biondo Biondi and Robert Clapp
ABSTRACT The problem of full-waveform inversion with the frequency-domain model extension is posed with a non-linear extended modeling operator based on the one-way wave extrapolation. This approach avoids introducing extended perturbation of the model, eliminates the need for solving an inner least-squares problem associated with similar methods, while having analogous benefit of data-residual matching. We describe the operators involved in the extended modeling, illustrate images in the frequency-extended domain and show the results of the one-way full-waveform inversion on synthetic examples.
Improved Deep Deconvolution Auto-Encoder for traffic analyses with DAS - Siyuan Yuan, Robert G. Clapp, and Biondo Biondi
ABSTRACT Road-side DAS that captures the high spatial-temporal resolution subsurface strain response has shown great potential in traffic monitoring. As DAS records a complex mixture of various types of signals and noises in urban environments, a robust, accurate, and high-resolution car-detection algorithm is critical for DAS to be feasible in real-world scenarios. A previous study designs a self-learning Deconvolution Auto-Encoder (DAE) to deconvolve the car impulse response from the data and use beamforming for speed estimation. However, such an approach is inaccurate in real-world scenarios where cars’ speeds vary. We redesign the DAE model to perform spatial deconvolution. We apply our DAE model for car tracking and speed monitoring. We also investigate the potential usage of DAS to identify vehicle types and traffic patterns.
Ensemble regression for velocity model building with convolutional neural networks - Stuart Farris, Guillaume Barnier, and Robert Clapp
ABSTRACT For full waveform inversion (FWI) to avoid local minima and converge to a useful earth model solution, it must start from an initial model with accurate low wavenumber components. Unfortunately, the band-limited nature of seismic data makes extracting accurate low wavenumber information about the subsurface extremely difficult, especially in geologic regimes with complex overburden such as salt or basalt formations. To overcome this problem, we propose a deep learning framework that uses a convolutional neural network (CNN) to form an ensemble of low wavenumber model predictions which can be integrated to form a starting model which is sufficient for FWI to avoid cycle skipping even in the presence of complex geology. The ability of neural networks to approximate nonlinear functions allows us to find a direct mapping from a small set of 4-8Hz seismic shot gathers to an accurate low wavenumber prediction of the earth model that produced those shots. We find this direct mapping by training the CNN with exclusively synthetic training data produced with a realistic earth model generator and an elastic wave equation formulation. Given overlapping sets of streamer gathers from a large seismic survey, the trained network can predict an ensemble of overlapping earth model predictions. The ensemble provides a distribution of possible earth property values at each point in the subsurface which can be combined to form a single, more accurate model. We illustrate, on two synthetic benchmark datasets that contain complex salt and band-limited data, the ability of our deep learning ensemble approach to find a starting model for FWI.
Procedural textures for synthetic, three-dimensional salt models - Stuart Farris and Robert Clapp
ABSTRACT We use procedural noise to create realistic 3D geologic salt bodies for use as training labels in supervised learning algorithms. Perlin and fractal noise generators, implemented in a python module, are harnessed to efficiently make realistic and reproducible surfaces that emulate the top and bottom of salt bodies seen in passive margins such as the deep waters of the Gulf of Mexico. This methodology was designed to be easily reproduced and usable by via a SEP git repository.