Common image gather conditioning using cycle generative adversarial networks

Seismically derived amplitude-versus-angle attributes along with well constraints are the base inputs into inverting seismic into subsurface properties. Conditioning the common image gathers is a common workflow in quantitative inversion and leads to a more accurate inversion product due to the removal of post-migration artefacts. Here, we apply a neural network to condition the post-migration gathers. The network is a cycle generative adversarial network, CycleGAN, which was designed for image-to-image translation. This can be considered the same problem as translating an artefact rich seismic gather to an artefact free seismic gather. To assess the feasibility of applying the network to pre-stack conditioning, synthetic data sets were generated to train different networks for different tasks. The networks were trained to remove white noise, residual de-multiples, gather flattening and a combination of the above for conditioning. The results show that a trained network was able remove white noise providing a more robust amplitude-versus-offset calculation. Another network trained using synthetic gathers with and without multiples assisted in multiple removal. However, instability around primary preservation has been observed so the network works better as a residual de-multiple method. For gather conditioning, a network was trained with the unpaired artefact-rich and artefact-free training data where the artefacts included complex moveout, noise and multiples. When applied to the test data sets, the networks cleaned the artefact-rich test data and translated complex moveout into flat gathers whilst preserving the amplitude response. Finally, two networks are applied to real data where a gather based on the well logs is used to quantify the match between the conditioned gathers and the raw gathers. The first network used synthetic data to train the network and, when applied to real data, provided a better tie with the well. The second network was trained with synthetic gathers whose properties were constrained by real seismic gathers from near the well. As anticipated, the network trained on the representative training data outperforms the network trained using the unconstrained data. However, the ability of the first network to condition the gather indicates that a sweep of networks can be trained without the need for real data and applied in a manner analogous to the way parameters are adjusted in traditional geophysical methods. The results show that the different neural networks can offer an alternative or augmentation to the existing geophysical workflow for conditioning pre-stack seismic gathers.     

Neural networks for quick earthquake damage estimation

This paper proposes the use of neural networks to predict damage due to earthquakes from the indices of recorded ground motion. Since the relationship between ground motion indices and resulting damage is difficult to express in mathematical form, neural networks are conveniently applied for this problem. Simulated earthquake ground motions are used to have a well-distributed data set and the ductility factor from non-linear analysis of two single-degree-of-freedom structural models is used to represent the damage. A sensitivity analysis procedure is described to identify qualitatively the input parameters that have a greater influence on the damage. The result of the trained neural network is then verified by using several recorded earthquake ground motions. It is found that some instability in the prediction can occur. Instability occurs when input values exceed the range of the training data. The neural network model using PGA and SI as input give the best performance in the recall tests using actual earthquake ground motion, demonstrating the usefulness of neural network models for the quick estimation of damage through earthquake intensity monitoring.     

基于神经网络的随机地震反演方法

针对随机地震反演中存在的两个主要问题,随机实现含有噪声和难以从大量随机实现中挖掘有效信息,提出了一种基于神经网络的随机地震反演方法.通过对多组随机实现及其正演地震数据的计算,构建了基于序贯高斯模拟的训练集.这也为应用神经网络求解地球物理反问题,提供了一种有效建立训练集的方法.较之传统的神经网络反演,这种训练集不仅保证了学习样本具有多样性,同时还引入了空间相关性.数值模拟结果表明,该方法只需要通过单层前馈神经网络,就可以比较有效的解决一个500个阻抗参数的反演问题.     

Deep neural network application for 4D seismic inversion to changes in pressure and saturation: Optimizing the use of synthetic training datasets

In this work, we tackle the challenge of quantitative estimation of reservoir dynamic property variations during a period of production, directly from four-dimensional seismic data in the amplitude domain. We employ a deep neural network to invert four-dimensional seismic amplitude maps to the simultaneous changes in pressure, water and gas saturations. The method is applied to a real field data case, where, as is common in such applications, the data measured at the wells are insufficient for properly training deep neural networks, thus, the network is trained on synthetic data. Training on synthetic data offers much freedom in designing a training dataset, therefore, it is important to understand the impact of the data distribution on the inversion results. To define the best way to construct a synthetic training dataset, we perform a study on four different approaches to populating the training set making remarks on data sizes, network generality and the impact of physics-based constraints. Using the results of a reservoir simulation model to populate our training datasets, we demonstrate the benefits of restricting training samples to fluid flow consistent combinations in the dynamic reservoir property domain. With this the network learns the physical correlations present in the training set, incorporating this information into the inference process, which allows it to make inferences on properties to which the seismic data are most uncertain. Additionally, we demonstrate the importance of applying regularization techniques such as adding noise to the synthetic data for training and show a possibility of estimating uncertainties in the inversion results by training multiple networks.     

A multi-data training method for a deep neural network to improve the separation effect of simultaneous-source data

Within the field of seismic data acquisition with active sources, the technique of acquiring simultaneous data, also known as blended data, offers operational advantages. The preferred processing of blended data starts with a step of deblending, that is separation of the data acquired by the different sources, to produce data that mimic data from a conventional seismic acquisition and can be effectively processed by standard methods. Recently, deep learning methods based on the deep neural network have been applied to the deblending task with promising results, in particular using an iterative approach. We propose an enhancement to deblending with an iterative deep neural network, whereby we modify the training stage of the deep neural network in order to achieve better performance through the iterations. We refer to the method that only uses the blended data as the input data as the general training method. Our new multi-data training method allows the deep neural network to be trained by the data set with the input patches composed of blended data, noisy data with low amplitude crosstalk noise, and unblended data, which can improve the ability of the deep neural network to remove crosstalk noise and protect weak signal. Based on such an extended training data set, the multi-data training method embedded in the iterative separation framework can result in different outputs at different iterations and converge to the best result in a shorter iteration number. Transfer learning can further improve the generalization and separation efficacy of our proposed method to deblend the simultaneous-source data. Our proposed method is tested on two synthetic data and two field data to prove the effectiveness and superiority in the deblending of the simultaneous-source data compared with the general training method, generic noise attenuation network and low-rank matrix factorization methods.     

3D inversion of DC data using artificial neural networks

In this paper, we investigate the applicability of artificial neural networks in inverting three-dimensional DC resistivity imaging data. The model used to produce synthetic data for training the artificial neural network (ANN) system was a homogeneous medium of resistivity 100 Ωm with an embedded anomalous body of resistivity 1000 Ωm. The different sizes for anomalous body were selected and their location was changed to different positions within the homogeneous model mesh elements. The 3D data set was generated using a finite element forward modeling code through standard 3D modeling software. We investigated different learning paradigms in the training process of the neural network. Resilient propagation was more efficient than any other paradigm. We studied the effect of the data type used on neural network inversion and found that the use of location and the apparent resistivity of data points as the input and corresponding true resistivity as the output of networks produces satisfactory results. We also investigated the effect of the training data pool volume on the inversion properties. We created several synthetic data sets to study the interpolation and extrapolation properties of the ANN. The range of 100–1000 Ωm was divided into six resistivity values as the background resistivity and different resistivity values were also used for the anomalous body. Results from numerous neural network tests indicate that the neural network possesses sufficient interpolation and extrapolation abilities with the selected volume of training data. The trained network was also applied on a real field dataset, collected by a pole-pole array using a square grid (8 ×8) with a 2-m electrode spacing. The inversion results demonstrate that the trained network was able to invert three-dimensional electrical resistivity imaging data. The interpreted results of neural network also agree with the known information about the investigation area.     

Inversion of nuclear well-logging data using neural networks

This work looks at the application of neural networks in geophysical well‐logging problems and specifically their utilization for inversion of nuclear downhole data. Simulated neutron and γ‐ray fluxes at a given detector location within a neutron logging tool were inverted to obtain formation properties such as porosity, salinity and oil/water saturation. To achieve this, the forward particle‐radiation transport problem was first solved for different energy groups (47 neutron groups and 20 γ‐ray groups) using the multigroup code EVENT. A neural network for each of the neutron and γ‐ray energy groups was trained to re‐produce the detector fluxes using the forward modelling results from 504 scenarios. The networks were subsequently tested on unseen data sets and the unseen input parameters (formation properties) were then predicted using a global search procedure. The results obtained are very encouraging with formation properties being predicted to within 10% average relative error. The examples presented show that neural networks can be applied successfully to nuclear well‐logging problems. This enables the implementation of a fast inversion procedure, yielding quick and reliable values for unknown subsurface properties such as porosity, salinity and oil saturation.     

基于深度卷积神经网络的地震数据随机噪声压制

本文以提高地震数据的成像质量为目标,提出一种智能的卷积神经网络降噪框架,从带有噪声的地震数据中自适应地学习地震信号。为了加速网络训练和避免训练时出现梯度消失现象,我们在网络中加入残差学习和批标准化的方法,并采用了ReLU激活函数和Adam优化算法优化网络。此外,Marmousi和F3数据集被用来对网络进行训练和测试,经过充分训练的网络不仅能在学习中保留地震数据特征,而且能去除随机噪声。首先充分地训练网络,从中提取出随机噪声,并保留学习到的地震数据特征,之后通过重建地震数据估算测试集中的波形特征。合成记录和实际数据的处理结果显示了深度卷积神经网络在随机噪声压制任务中的潜力,并通过实验验证表明了深度卷积神经网络框架有很好的去噪效果。      

Common-offset domain reflection–diffraction separation with deep learning

Separation of diffracted from reflected events in seismic data is still challenging due to the relatively low amplitude of the diffracted wavefield compared to the reflected wavefield as well as the overlap in the kinematics of reflection and diffraction events. A workflow based on deep learning can be a simple and fast alternative, but using training data made by physics-based modelling is expensive and lacks diversity in terms of noise, amplitude, frequency content and wavelet. This results in poor generalization beyond the training data without retraining and transfer learning. In this paper, we demonstrate successful applications of reflection–diffraction separation using a conventional U-net architecture. The novelty of our approach is that we do not use synthetic data created from physics-based modelling, but instead use only synthetic data built from basic geometric shapes. Our domain of application is the pre-migration common-offset domain where reflected events resemble local geology and the diffracted wavefield consists of downward convex hyperbolic diffraction patterns. Both patterns were randomly perturbed in many ways while maintaining their intrinsic features. This approach is inspired by the common practice of data augmentation in deep learning for machine vision applications. Since many of the standard data augmentation techniques lack a geophysical motivation, we have instead perturbed our synthetic training data in ways to make more sense from a signal processing perspective or given our ‘domain knowledge’ of the problem at hand. We did not have to retrain the network to show good results on the field data set. The large variety and diversity in examples enabled to trained neural networks to show encouraging results on synthetic and field data sets that were not used in training.     

Adaptive Training of Neural Networks for Automatic Seismic Phase Identification

—?A neural network module has been implemented in the Prototype International Data Centre (PIDC) for automated identification of the initial phase type of seismic detections. Initial training of the neural networks for stations of the International Monitoring System (IMS) requires considerable effort. While there are many seismic phases in the analyst-reviewed database that can be assumed as the ground-truth resource of the initial phase type of Teleseism (T), Regional P (P), and Regional S (S), no ground-truth database of noise (N) is available. To reduce analyst effort required in building a ground-truth database, an “Adaptive Training Approach” is proposed in this paper. This approach automatically selects training patterns to take advantage of the learning ability of neural networks and information on the accumulated observation database. Using this approach, neural networks were trained on the data provided by station STKA, Australia. The performance of automated phase identification has been improved significantly by the retrained neural networks. This approach is also validated by comparison with the performance using the ground-truth noise database.     

Deep Artificial Neural Networks as a Tool for the Analysis of Seismic Data

The number of seismological studies based on artificial neural networks has been increasing. However, neural networks with one hidden layer have almost reached the limit of their capabilities. In the last few years, there has been a new boom in neuroinformatics associated with the development of third-generation networks, deep neural networks. These networks operate with data at a higher level. Unlabeled data can be used to pretrain the network, i.e., there is no need for an expert to determine in advance the phenomenon to which these data correspond. Final training requires a small amount of labeled data. Deep networks have a higher level of abstraction and produce fewer errors. The same network can be used to solve several tasks at the same time, or it is easy to retrain it from one task to another. The paper discusses the possibility of applying deep networks in seismology. We have described what deep networks are, their advantages, how they are trained, how to adapt them to the features of seismic data, and what prospects are opening up in connection with their use.     

Sub-basalt imaging problems and the application of Artificial Neural Networks

In many areas of the world, the presence of shallow high velocity, highly heterogeneous layers complicate seismic imaging of deeper reflectors. Of particular economic interest are areas where potentially hydrocarbon-bearing strata are obscured by layers of basalt. Basalt layers are highly reflective and heterogeneous. Using reflection seismic, top basalt is typified by a high-amplitude, coherent reflector with poor resolution of reflectors below the basalt, and even bottom basalt. Here, we present a new approach to the imaging problem using the pattern recognition abilities of a back-propagation Artificial Neural Network (ANN). ANNs are computational systems that attempt to mimic natural biological neural networks. They have the ability to recognize patterns and develop their own generalizations about a given data set. Back-propagation neural networks are trained on data sets for which the solution is known and tested on the data that are not previously presented to the ANN in order to validate the network result. We show that Artificial Neural Networks, due to their pattern recognition capabilities, can invert the medium statistics based on the seismic character. We produce statistically defined models involving a basalt analogous layer, and calculate full wavefield finite difference synthetic seismograms. We vary basalt layer thickness and source frequency to generate a synthetic model that produces seismic that is similar to real sub-basalt seismic, i.e. high amplitude top basalt reflector and the absence of base basalt and sub-basalt events. Using synthetic shot gathers, generated in a synthetic representation of the sub-basalt case, we can invert the velocity medium standard deviation by using an ANN. By inverting the velocity medium standard deviation, we successfully identified the transition from basalt to sub-basalt on the synthetic shot gathers. We also show that ANNs are capable of identifying the basalt to sub-basalt transition in the presence of incoherent noise. This is important for any future applications of this technique to the real-world seismic data, as this data is never completely noise-free. There is always a certain level of residual (noise remaining after initial noise filtering) environmental/ambient noise present on the recorded seismics, hence, neural network training with noise-free synthetic seismic is less than optimal.     

Rock Properties and Seismic Attenuation: Neural Network Analysis

—Using laboratory data, the influence of rock parameters on seismic attenuation has been analyzed using artificial neural networks and regression models. The predictive capabilities of the neural networks and multiple linear regresssion were compared. The neural network outperforms the multiple linear regression in predicting attenuation values, given a set of input of rock parameters. The neural network can make complex decision mappings and this capability is exploited to examine the influence of various rock parameters on the overall seismic attenuation. The results indicate that the most influential rock parameter on the overall attenuation is the clay content, closely followed by porosity. Though grain size contribution is of lower importance than clay content and porosity, its value of 16 percent is sufficiently significant to be considered in the modeling and interpretation of attenuation data.     

电阻率二维神经网络反演

由于非线性特性地球物理反演一直以来都是一个比较困难的问题. 近十年来,非线性反演方法如人工神经网络、遗传算法在地球物理数据解释中得到越来越多的应用,但目前基本仍限于一维反演问题. 对于二维反问题,反演参数较多,神经网络反演运用较少. 本文利用BP神经网络优化方法,实现了电阻率二维非线性反演. 与传统线性化的迭代反演比较,神经网络反演能够克服传统方法的不足、获得更好的反演结果.     

An approximation method for solving the inverse MTS problem with the use of neural networks

The work develops the approximation approach to solving the inverse MTS problem with the use of neural networks. The inverse problem is considered in model classes of parametrized geoelectric structures, whose electric conductivity is controlled by a few hundreds of macroparameters (N ∼ 300). An approximate inverse operator of the problem is constructed for each model class as a neural network, whose coefficients are determined in the process of training on a representative sample of standard examples of forward problem solutions. The problem of determination of the model class of geolectric structures corresponding to the presented input MT data is solved with the use of the neural network classifier constructed for the available set of model classes of structures. Regularizing factors and errors of the neural network method are analyzed. The operation of the algorithm is illustrated by examples of the 2-D inversion of synthetic MT data.     

基于卷积神经网络的断层预测方法

针对传统相干体属性在预测断层时存在断层假象以及易受噪声影响等缺点,本文提出一种利用卷积神经网络进行断层预测的方法。首先构建适合实际工区断层特征的卷积神经网络模型,然后利用部分分频地震数据和人工解释出的断层标签进行网络模型训练,最后把训练好的模型应用到整个三维地震数据中进行断层预测。实际地震数据预测结果表明基于卷积神经网络断层预测结果与地震数据吻合较好,并且在断层细节刻画上要优于传统地震相干体属性方法。      

Estimation of river flow by artificial neural networks and identification of input vectors susceptible to producing unreliable flow estimates

Reliable river flow estimates are crucial for appropriate water resources planning and management. River flow forecasting can be conducted by conceptual or physical models, or data-driven black box models. Development of physically-based models requires an understanding of all the physical processes which impact a natural process and the interactions among them. Since identification of the relationships among these physical processes is very difficult, data-driven approaches have recently been utilized in hydrological modeling. Artificial neural networks are one of the widely used data-driven approaches for modeling hydrological processes. In this study, estimation of future monthly river flows for Guvenc River, Ankara is conducted using various artificial neural network models. Success of artificial neural network models relies on the availability of adequate data sets. A direct mapping from inputs to outputs without consideration of the complex relationships among the dependent and independent variables of the hydrological process is identified. In this study, past precipitation, river flow data, and the associated month are used to predict future river flows for Guvenc River. Impacts of various input patterns, number of training cycles, and initial values assigned to the weights of the connections are investigated. One of the major weaknesses of artificial neural networks is that they may fail to generate good estimates for extreme events, i.e. events that do not occur at all or often enough in the training data set. It is very important to be able to identify such unlikely events. A fuzzy c-means algorithm is used in this study to cluster the training and validation input vectors into regular and extreme events so that the user will have an idea about the risk of the artificial neural network model to generate unreliable results.     

基于深度降噪自编码神经网络的中国大陆地壳厚度反演

本文采用基于数据驱动的深度降噪自编码网络构建了瑞雷面波群速度、相速度频散特性与地壳厚度的正反演函数关系,并利用最新频散模型反演了中国大陆的地壳厚度.对于神经网络架构体系的评价,除了考虑传统意义上的测试误差、训练误差之外,本文还用已知物理原理的正演结果与网络预测结果进行比较;在设计网络构架时,同时考虑地球模型和面波频散的...     

Automatic 1D inversion of multifrequency airborne electromagnetic data with artificial neural networks: discussion and a case study

Artificial neural networks were used to implement an automatic inversion of frequency‐domain airborne electromagnetic (AEM) data that do not require a priori information about the survey area. Two classes of model, i.e. homogeneous half‐space models and horizontally layered half‐space models with two layers, are used in this 1D inversion, and for each data point the selection of the class of 1D model is performed prior to the inversion, also using an artificial neural network. The proposed inversion method was tested in a survey area situated in Austria, northwest of Vienna in the Bohemian Massif. The results of the inversion were compared with the geological setting, logging results, and seismic and gravimetric measurements. This comparison shows a good correlation between the AEM models and the known geological and geophysical data.     

MT+, Integrating Magnetotellurics to Determine Earth Structure, Physical State, and Processes

As one of the few deep-earth imaging techniques, magnetotellurics provides information on both the structure and physical state of the crust and upper mantle. Magnetotellurics is sensitive to electrical conductivity, which varies within the earth by many orders of magnitude and is modified by a range of earth processes. As with all geophysical techniques, magnetotellurics has a non-unique inverse problem and has limitations in resolution and sensitivity. As such, an integrated approach, either via the joint interpretation of independent geophysical models, or through the simultaneous inversion of independent data sets is valuable, and at times essential to an accurate interpretation. Magnetotelluric data and models are increasingly integrated with geological, geophysical and geochemical information. This review considers recent studies that illustrate the ways in which such information is combined, from qualitative comparisons to statistical correlation studies to multi-property inversions. Also emphasized are the range of problems addressed by these integrated approaches, and their value in elucidating earth structure, physical state, and processes.