Geophysical Prospecting - 1 Machine learning applications in geophysical exploration and monitoring, 2023

Full‐waveform inversion, a popular technique that promises high‐resolution models, has helped in improving the salt definition in inverted velocity models. The success of the inversion relies heavily on having prior knowledge of the salt, and using advanced acquisition technology with long offsets and low frequencies. Salt bodies are often constructed by recursively picking the top and bottom of the salt from seismic images corresponding to tomography models, combined with flooding techniques. The process is time consuming and highly prone to error, especially in picking the bottom of the salt. Many studies suggest performing full‐waveform inversion with long offsets and low frequencies after constructing the salt bodies to correct the misinterpreted boundaries. Here, we focus on detecting the bottom of the salt automatically by utilizing deep learning tools. We specifically generate many random one‐dimensional models, containing or free of salt bodies, and calculate the corresponding shot gathers. We then apply full‐waveform inversion starting with salt flooded versions of those models, and the results of the full‐waveform inversion become inputs to the neural network, whereas the corresponding true one‐dimensional models are the output. The network is trained in a regression manner to detect the bottom of the salt and estimate the subsalt velocity. We analyse three scenarios in creating the training datasets and test their performance on the two‐dimensional BP 2004 salt model. We show that when the network succeeds in estimating the subsalt velocity, the requirement of low frequencies and long offsets are somewhat mitigated. In general, this work allows us to merge the top‐to‐bottom approach with full‐waveform inversion, save the bottom of the salt picking time and empower full‐waveform inversion to converge in the absence of low frequencies and long offsets in the data.

Seismic records from reflection seismic exploration in complex surface areas often contain scattered direct waves, a type of scattered waves generated by direct waves propagating on a rugged surface. For imaging reflections, scattered direct waves are regarded as noise, which lowers the signal‐to‐noise ratio of the reflected waves and deteriorates the quality of the seismic profile. Importantly, it is very challenging to separate scattered direct waves with strong energy and complex wave field characteristics owing to the rugged surface. In recent years, the rapid development of machine‐learning technology has broadened and advanced the application of deep learning in denoising seismic data. In this context, we propose an approach to using the convolutional autoencoder, a deep learning network, to intelligently separate scattered direct waves from seismic records on a rugged surface. The spectral element method is applied to simulate the elastic wave seismic records and scattered direct waves on a rugged surface. The synthetic shot gathers on the rugged ground are employed as the input for training the convolutional autoencoder network, while the simulated scattered direct waves on the uniform half‐space with the same rugged ground are employed as the label data for training. After the network training is completed, the trained convolutional autoencoder network can be applied to predict scattered direct waves from seismic records. The numerical experiments demonstrate that the proposed approach has good potential for suppressing complex scattered direct waves generated by direct waves propagating on a rugged surface.

Regularization and interpolation of 3D offset classes prior to imaging are an important and challenging step in the marine seismic data processing flow. Here we describe how to perform this task using a deep neural network, and we explain how to overcome the challenge of creating a suitable training data set. The training data set is generated by de‐migrating stacked pre‐stack depth migration images. For each offset class volume, we de‐migrate the pre‐stack depth migrated stacked image into two configurations: (i) the original survey configuration consisting of the recorded source/receiver positions and (ii) an ‘Ideal’ survey configuration with constant offset and azimuth for each 3D offset class. The training creates a 3D convolutional encoder–decoder model that will regularize and interpolate seismic data. The convolutional encoder–decoder is trained on 3D sliding windows in each 3D offset cube to map from (i) to (ii), i.e. to map the original survey configuration with irregular and sparse sampling into the fully sampled regular offset cubes suitable for offset‐based migration, such as Kirchhoff migration. Such migration algorithms rely on regular and sufficiently dense sampling to achieve constructive interference to image the structures and destructive interference to suppress migration noise. We test the new method on one synthetic and one field data example and show that it performs better than a standard regularization/interpolation method based on anti‐leakage Fourier transform, especially for the smallest offset classes. On the synthetic data, we also demonstrate that the convolutional encoder–decoder method preserves the amplitude versus offset as well as the standard method.

Seismic velocity plays an important role in imaging and identifying underground geology. Conventional seismic velocity inversion methods, like full waveform inversion, directly update the velocity model based on the misfit between the observed and synthetic data. However, seismic velocity inversion is a highly nonlinear process, and the inversion effect greatly relies on the initial inversion model. In this paper, we propose a novel network‐domain full waveform inversion method. Different from the existing network‐domain full waveform inversion methods, which use random or fixed numbers as network input, we reparameterize the low‐dimensional acoustic velocity model in a high‐dimensional inversion network parameter domain with seismic observed data as the network input. In this way, the physical information within the observed data can be directly encoded into the inversion parameters, leading to a better inversion effect than the current network‐domain full waveform inversion method. Moreover, comparison experiments on the Society of Exploration Geophysicists and the European Association of Geoscientists and Engineers Overthrust model and the Marmousi model show the advantages of the proposed method over conventional full waveform inversion from the aspects of inversion accuracy, robustness to noisy data, and more complex geological structures. These advantages may benefit from the fact that reparameterization within the inversion network domain can empower the inversion process with the regularization ability of denoising and mitigating the cycle‐skipping issue. In the end, the potential of the proposed method in terms of network initialization is further discussed.

We propose a workflow to correct migration amplitudes by estimating the inverse Hessian operator weights using a neural network–based framework. We train the network such that it learns the transformation between the migration output and true amplitude reflectivity constrained by different conditioners. We analyse the network output with a velocity model and with source illumination as a conditioner. Compared to the velocity model, source illumination as a conditioner performs better because source illumination encodes the geometrical spreading information and accounts for non‐stationarity. We further use the output of the deep neural network as a starting model for accelerating the convergence of iterative least‐squares reverse time migration. Using a deep learning framework, the proposed method combines the model domain and data domain least‐squares migration approaches to recover images with interpretable amplitudes, attenuated migration artefacts, better signal‐to‐noise ratio and improved resolution. We compare the output of the proposed algorithm with conventional least‐squares output and show that the proposed workflow is more robust, especially in the areas with weak illumination.

We propose a method to generate seismic images with corresponding fault labels for augmenting training data in automatic fault detection. Our method is based on two generative adversarial networks: one for creating a fault system and the other for generating two‐dimensional seismic images with faults as a condition. Our method can capture the characteristics of field seismic data during inference to generate samples that have properties of both field seismic data and synthetic training data. We then use the newly generated seismic images with corresponding fault labels to train a convolutional neural network for fault picking. We test the proposed approach on a three‐dimensional field dataset from the Gulf of Mexico. We use different areas in the field dataset as input to generate new training data for corresponding fault‐picking models. The results show that the generated training data from our method help in improving the fault‐picking models in the targeted areas.

Petrophysical interpretation and optimal correlation extraction of different measurements require accurate well log depth matching. We have developed a supervised multimodal machine learning alternative for the task of simultaneously matching raw logging while drilling and electrical wireline logging logs. Seven one‐dimensional convolutional neural networks are trained using different log measurements: gamma‐ray, resistivity, P‐ and S‐wave sonic, density, neutron and photoelectric factors, and their depth shift estimates are aggregated using different multimodal late fusion strategies. We test the late fusion average, late fusion weighted average, late fusion with linear and nonlinear learners and model‐level fusion. Depth matching results using the different fusion strategies applied to two unseen wells are compared using visual inspection and the mean Pearson correlation. All models perform well, increasing the correlation after depth matching. Late fusion weighted average achieves the highest scores for all log types. The late fusion weighted average results are compared to a cross‐correlation user‐assisted workflow and manual depth matching for validation. In general, the convolutional neural network fused method exhibits a lower performance than the traditional methods. For one of the wells, the cross‐correlation shows higher correlation values than the other methods but for the second well the manual depth match performs best. However, the differences in Pearson correlation values are small ranging from 0.01 to 0.1. The manual depth match performs very well for the sonic logs, which tend to require slightly larger depth shifts than the other measurements, thus a common depth shift might not always be suitable. Although our convolutional neural network fused approach is limited to estimating bulk shifts and uses constant fusion weights, its performance is similar to that of more time‐consuming methods. Our approach might be substantially improved by including dynamic shifts (stretch/squeeze) and depth‐dependent fusion weights via long‐short‐term memory recurrent neural networks.

Multiphysics inversion exploits different types of geophysical data that often complement each other and aims to improve overall imaging resolution and reduce uncertainties in geophysical interpretation. Despite the advantages, traditional multiphysics inversion is challenging because it requires a large amount of computational time and intensive human interactions for preprocessing data and finding trade‐off parameters. These issues make it nearly impossible for traditional multiphysics inversion to be applied as a real‐time monitoring tool for geological carbon storage. In this paper, we present a deep learning (DL) multiphysics network for imaging CO2 saturation in real time. The multiphysics network consists of three encoders for analysing seismic, electromagnetic and gravity data and shares one decoder for combining imaging capabilities of the different geophysical data for better predicting CO2 saturation. The network is trained on pairs of CO2 label models and multiphysics data so that it can directly image CO2 saturation. We use the bootstrap aggregating method to enhance the imaging accuracy and estimate uncertainties associated with CO2 saturation images. Using realistic CO2 label models and multiphysics data derived from the Kimberlina CO2 storage model, we evaluate the performance of the deep learning multiphysics network and compare its imaging results to those from the deep learning single‐physics networks. Our modelling experiments show that the deep learning multiphysics network for seismic, electromagnetic, and gravity data not only improves the imaging accuracy but also reduces uncertainties associated with CO2 saturation images. Our results also suggest that the deep learning multiphysics network for the non‐seismic data (i.e., electromagnetic and gravity) can be used as an effective low‐cost monitoring tool in between regular seismic monitoring.

Deep‐learning inversion has recently drawn attention in geological carbon storage research due to its potential of imaging and monitoring carbon storage in real time, significantly improving efficiency and safety of carbon storage operations. We present a deep‐learning full waveform inversion method that after the neural network has been trained can image CO2 saturation and its uncertainty in real time. Our deep‐learning inversion method is based on the U‐Net architecture with the neural network trained on pairs of synthetic seismic data and CO2 saturation models. Accordingly, our training establishes a mapping relationship between seismic data and CO2 saturation models and once fully trained directly estimates CO2 saturation as a function of subsurface location. We further quantify uncertainties of CO2 saturation estimates using the Monte Carlo dropout method and a bootstrap aggregating method. For this proof‐of‐concept study, the CO2 training models and data are derived from the Kimberlina 1.2 model, a hypothetical 3D geological carbon storage model that is constructed based on various geological and hydrological data from the Southern San Joaquin Basin, California. We perform deep‐learning inversion experiments using noise‐free and noisy training and test data sets and compare the results. Our modelling experiments show that (1) the deep‐learning inversion can estimate 2D distributions of CO2 fairly well even in the presence of Gaussian random noise and (2) both CO2 saturation imaging and uncertainty quantification can be done in real time. Our results suggest that the deep‐learning inversion method can serve as a robust real‐time monitoring tool for geological carbon storage and/or other time‐varying reservoir/aquifer properties that result from injection, extraction, and/or other subsurface transport phenomena.

Blended acquisition, which allows multiple sources almost simultaneously fired, has become an effective way for accelerating seismic data acquisition. In order to use conventional processing methods for imaging, deblending is necessary for this special acquisition. Convolutional neural network‐based deblending methods provide a novel end‐to‐end framework for source separation. We proposed a field‐data‐based augmentation method that uses shuffled deblending noise as the features to be learned and take the inaccurate labels as the output of the network. Synthetic data experiments show that a network trained on data set with the proposed data augmentation method has higher accuracy for deblending even if the labelled data are noisy. Besides, 2D discrete wavelet transform, which has the advantage of multiscale decomposition and dimensionality reduction, is introduced to accelerate the computation of the network. The data augmentation method for data set generation and the computational speedup method for network training/predicting are also applied to field data. The results from synthetic and field data all confirm the performance of our methods.