Physics-guided deep learning for seismic inversion with hybrid training and uncertainty analysis

Geophysics ◽  
2021 ◽  
pp. 1-64
Author(s):  
Jian Sun ◽  
Kristopher A. Innanen ◽  
Chao Huang
Keyword(s):  
Deep Learning ◽  
Elastic Property ◽  
Seismic Data ◽  
Degrees Of Freedom ◽  
Seismic Inversion ◽  
Training Data ◽  
Data Driven ◽  
Model Data ◽  
Data Set ◽  
Propagation Network

The determination of subsurface elastic property models is crucial in quantitative seismic data processing and interpretation. This problem is commonly solved by deterministic physical methods, such as tomography or full-waveform inversion. However, these methods are entirely local and require accurate initial models. Deep learning represents a plausible class of methods for seismic inversion, which may avoid some of the issues of purely descent-based approaches. However, any generic deep learning network capable of relating each elastic property cell value to each sample in a seismic data set would require a very large number of degrees of freedom. Two approaches might be taken to train such a network: first, by invoking a massive and exhaustive training data set and, second, by working to reduce the degrees of freedom by enforcing physical constraints on the model-data relationship. The second approach is referred to as “physics-guiding.” Based on recent progress in wave theory-designed (i.e., physics-based) networks, we have developed a hybrid network design, involving deterministic, physics-based modeling and data-driven deep learning components. From an optimization standpoint, a data-driven model misfit (i.e., standard deep learning) and now a physics-guided data residual (i.e., a wave propagation network) are simultaneously minimized during the training of the network. An experiment is carried out to analyze the trade-off between two types of losses. Synthetic velocity building is used to examine the potential of hybrid training. Comparisons demonstrate that, given the same training data set, the hybrid-trained network outperforms the traditional fully data-driven network. In addition, we performed a comprehensive error analysis to quantitatively compare the fully data-driven and hybrid physics-guided approaches. The network is applied to the SEG salt model data, and the uncertainty is analyzed, to further examine the benefits of hybrid training.

Deep learning-driven velocity model building workflow

2019 ◽  
Vol 38 (11) ◽  
pp. 872a1-872a9 ◽  
Author(s):  
Mauricio Araya-Polo ◽  
Stuart Farris ◽  
Manuel Florez
Keyword(s):  
Deep Learning ◽  
Seismic Data ◽  
Model Building ◽  
Ground Truth ◽  
Velocity Model ◽  
Training Data ◽  
Quality Data ◽  
Generative Adversarial Network ◽  
Data Set ◽  
Velocity Models

Exploration seismic data are heavily manipulated before human interpreters are able to extract meaningful information regarding subsurface structures. This manipulation adds modeling and human biases and is limited by methodological shortcomings. Alternatively, using seismic data directly is becoming possible thanks to deep learning (DL) techniques. A DL-based workflow is introduced that uses analog velocity models and realistic raw seismic waveforms as input and produces subsurface velocity models as output. When insufficient data are used for training, DL algorithms tend to overfit or fail. Gathering large amounts of labeled and standardized seismic data sets is not straightforward. This shortage of quality data is addressed by building a generative adversarial network (GAN) to augment the original training data set, which is then used by DL-driven seismic tomography as input. The DL tomographic operator predicts velocity models with high statistical and structural accuracy after being trained with GAN-generated velocity models. Beyond the field of exploration geophysics, the use of machine learning in earth science is challenged by the lack of labeled data or properly interpreted ground truth, since we seldom know what truly exists beneath the earth's surface. The unsupervised approach (using GANs to generate labeled data)illustrates a way to mitigate this problem and opens geology, geophysics, and planetary sciences to more DL applications.

Robust deep learning seismic inversion with a priori initial model constraint

2021 ◽  
Author(s):  
Jian Zhang ◽  
Jingye Li ◽  
Xiaohong Chen ◽  
Yuanqiang Li ◽  
Guangtan Huang ◽  
...  
Keyword(s):  
Deep Learning ◽  
Seismic Data ◽  
A Priori ◽  
Seismic Inversion ◽  
Training Dataset ◽  
Initial Model ◽  
Data Set ◽  
The Real ◽  
Model Constraint ◽  
Trained Network

Summary Seismic inversion is one of the most commonly used methods in the oil and gas industry for reservoir characterization from observed seismic data. Deep learning (DL) is emerging as a data-driven approach that can effectively solve the inverse problem. However, existing deep learning-based methods for seismic inversion utilize only seismic data as input, which often leads to poor stability of the inversion results. Besides, it has always been challenging to train a robust network since the real survey has limited labeled data pairs. To partially overcome these issues, we develop a neural network framework with a priori initial model constraint to perform seismic inversion. Our network uses two parts as one input for training. One is the seismic data, and the other is the subsurface background model. The labels for each input are the actual model. The proposed method is performed by log-to-log strategy. The training dataset is firstly generated based on forward modeling. The network is then pre-trained using the synthetic training dataset, which is further validated using synthetic data that has not been used in the training step. After obtaining the pre-trained network, we introduce the transfer learning strategy to fine-tune the pre-trained network using labeled data pairs from a real survey to acquire better inversion results in the real survey. The validity of the proposed framework is demonstrated using synthetic 2D data including both post-stack and pre-stack examples, as well as a real 3D post-stack seismic data set from the western Canadian sedimentary basin.

scholarly journals A Data-Driven Surrogate Approach for the Temporal Stability Forecasting of Vegetation Covered Dikes

Water ◽  
2021 ◽  
Vol 13 (1) ◽  
pp. 107
Author(s):  
Elahe Jamalinia ◽  
Faraz S. Tehrani ◽  
Susan C. Steele-Dunne ◽  
Philip J. Vardon
Keyword(s):  
Numerical Simulation ◽  
Water Flux ◽  
Temporal Stability ◽  
Synthetic Data ◽  
Climatic Conditions ◽  
Training Data ◽  
Data Driven ◽  
Data Set ◽  
Surface Cracking ◽  
Real Time Analysis

Climatic conditions and vegetation cover influence water flux in a dike, and potentially the dike stability. A comprehensive numerical simulation is computationally too expensive to be used for the near real-time analysis of a dike network. Therefore, this study investigates a random forest (RF) regressor to build a data-driven surrogate for a numerical model to forecast the temporal macro-stability of dikes. To that end, daily inputs and outputs of a ten-year coupled numerical simulation of an idealised dike (2009–2019) are used to create a synthetic data set, comprising features that can be observed from a dike surface, with the calculated factor of safety (FoS) as the target variable. The data set before 2018 is split into training and testing sets to build and train the RF. The predicted FoS is strongly correlated with the numerical FoS for data that belong to the test set (before 2018). However, the trained model shows lower performance for data in the evaluation set (after 2018) if further surface cracking occurs. This proof-of-concept shows that a data-driven surrogate can be used to determine dike stability for conditions similar to the training data, which could be used to identify vulnerable locations in a dike network for further examination.

Data-driven method for training data selection for deep learning

2021 ◽  
Author(s):  
C. Lacombe ◽  
I. Hammoud ◽  
J. Messud ◽  
H. Peng ◽  
T. Lesieur ◽  
...  
Keyword(s):  
Deep Learning ◽  
Training Data ◽  
Data Selection ◽  
Data Driven ◽  
Selection For ◽  
Training Data Selection

scholarly journals Augmented Data Selector to Initiate Text-Based CAPTCHA Attack

2021 ◽  
Vol 2021 ◽  
pp. 1-9
Author(s):  
Aolin Che ◽  
Yalin Liu ◽  
Hong Xiao ◽  
Hao Wang ◽  
Ke Zhang ◽  
...  
Keyword(s):  
Deep Learning ◽  
Training Data ◽  
Accuracy Rate ◽  
Data Set ◽  
Research Directions ◽  
Training Models ◽  
The Past ◽  
Future Work ◽  
Design Cost ◽  
Accuracy Rates

In the past decades, due to the low design cost and easy maintenance, text-based CAPTCHAs have been extensively used in constructing security mechanisms for user authentications. With the recent advances in machine/deep learning in recognizing CAPTCHA images, growing attack methods are presented to break text-based CAPTCHAs. These machine learning/deep learning-based attacks often rely on training models on massive volumes of training data. The poorly constructed CAPTCHA data also leads to low accuracy of attacks. To investigate this issue, we propose a simple, generic, and effective preprocessing approach to filter and enhance the original CAPTCHA data set so as to improve the accuracy of the previous attack methods. In particular, the proposed preprocessing approach consists of a data selector and a data augmentor. The data selector can automatically filter out a training data set with training significance. Meanwhile, the data augmentor uses four different image noises to generate different CAPTCHA images. The well-constructed CAPTCHA data set can better train deep learning models to further improve the accuracy rate. Extensive experiments demonstrate that the accuracy rates of five commonly used attack methods after combining our preprocessing approach are 2.62% to 8.31% higher than those without preprocessing approach. Moreover, we also discuss potential research directions for future work.

Techniques for Fast Screening of 3D Heterogeneous Shale Barrier Configurations and Their Impacts on SAGD Chamber Development

SPE Journal ◽  
2021 ◽  
pp. 1-25
Author(s):  
Chang Gao ◽  
Juliana Y. Leung
Keyword(s):  
Distance Measure ◽  
Flow Simulation ◽  
Training Data ◽  
Distance Measures ◽  
Data Driven ◽  
Data Set ◽  
Flow Simulations ◽  
Steam Chamber ◽  
Reservoir Models ◽  
Tracking Model

Summary The steam-assisted gravity drainage (SAGD) recovery process is strongly impacted by the spatial distributions of heterogeneous shale barriers. Though detailed compositional flow simulators are available for SAGD recovery performance evaluation, the simulation process is usually quite computationally demanding, rendering their use over a large number of reservoir models for assessing the impacts of heterogeneity (uncertainties) to be impractical. In recent years, data-driven proxies have been widely proposed to reduce the computational effort; nevertheless, the proxy must be trained using a large data set consisting of many flow simulation cases that are ideally spanning the model parameter spaces. The question remains: is there a more efficient way to screen a large number of heterogeneous SAGD models? Such techniques could help to construct a training data set with less redundancy; they can also be used to quickly identify a subset of heterogeneous models for detailed flow simulation. In this work, we formulated two particular distance measures, flow-based and static-based, to quantify the similarity among a set of 3D heterogeneous SAGD models. First, to formulate the flow-based distance measure, a physics-basedparticle-tracking model is used: Darcy’s law and energy balance are integrated to mimic the steam chamber expansion process; steam particles that are located at the edge of the chamber would release their energy to the surrounding cold bitumen, while detailed fluid displacements are not explicitly simulated. The steam chamber evolution is modeled, and a flow-based distance between two given reservoir models is defined as the difference in their chamber sizes over time. Second, to formulate the static-based distance, the Hausdorff distance (Hausdorff 1914) is used: it is often used in image processing to compare two images according to their corresponding spatial arrangement and shapes of various objects. A suite of 3D models is constructed using representative petrophysical properties and operating constraints extracted from several pads in Suncor Energy’s Firebag project. The computed distance measures are used to partition the models into different groups. To establish a baseline for comparison, flow simulations are performed on these models to predict the actual chamber evolution and production profiles. The grouping results according to the proposed flow- and static-based distance measures match reasonably well to those obtained from detailed flow simulations. Significant improvement in computational efficiency is achieved with the proposed techniques. They can be used to efficiently screen a large number of reservoir models and facilitate the clustering of these models into groups with distinct shale heterogeneity characteristics. It presents a significant potential to be integrated with other data-driven approaches for reducing the computational load typically associated with detailed flow simulations involving multiple heterogeneous reservoir realizations.

Diagnostic assessment of a deep learning system for detecting atrial fibrillation in pulse waveforms

Heart ◽  
2018 ◽  
Vol 104 (23) ◽  
pp. 1921-1928 ◽  
Author(s):  
Ming-Zher Poh ◽  
Yukkee Cheung Poh ◽  
Pak-Hei Chan ◽  
Chun-Ka Wong ◽  
Louise Pun ◽  
...  
Keyword(s):  
Atrial Fibrillation ◽  
Deep Learning ◽  
Test Data ◽  
Predictive Value ◽  
Characteristic Curve ◽  
Performance Comparison ◽  
Learning System ◽  
Training Data ◽  
Validation Data ◽  
Data Set

ObjectiveTo evaluate the diagnostic performance of a deep learning system for automated detection of atrial fibrillation (AF) in photoplethysmographic (PPG) pulse waveforms.MethodsWe trained a deep convolutional neural network (DCNN) to detect AF in 17 s PPG waveforms using a training data set of 149 048 PPG waveforms constructed from several publicly available PPG databases. The DCNN was validated using an independent test data set of 3039 smartphone-acquired PPG waveforms from adults at high risk of AF at a general outpatient clinic against ECG tracings reviewed by two cardiologists. Six established AF detectors based on handcrafted features were evaluated on the same test data set for performance comparison.ResultsIn the validation data set (3039 PPG waveforms) consisting of three sequential PPG waveforms from 1013 participants (mean (SD) age, 68.4 (12.2) years; 46.8% men), the prevalence of AF was 2.8%. The area under the receiver operating characteristic curve (AUC) of the DCNN for AF detection was 0.997 (95% CI 0.996 to 0.999) and was significantly higher than all the other AF detectors (AUC range: 0.924–0.985). The sensitivity of the DCNN was 95.2% (95% CI 88.3% to 98.7%), specificity was 99.0% (95% CI 98.6% to 99.3%), positive predictive value (PPV) was 72.7% (95% CI 65.1% to 79.3%) and negative predictive value (NPV) was 99.9% (95% CI 99.7% to 100%) using a single 17 s PPG waveform. Using the three sequential PPG waveforms in combination (<1 min in total), the sensitivity was 100.0% (95% CI 87.7% to 100%), specificity was 99.6% (95% CI 99.0% to 99.9%), PPV was 87.5% (95% CI 72.5% to 94.9%) and NPV was 100% (95% CI 99.4% to 100%).ConclusionsIn this evaluation of PPG waveforms from adults screened for AF in a real-world primary care setting, the DCNN had high sensitivity, specificity, PPV and NPV for detecting AF, outperforming other state-of-the-art methods based on handcrafted features.

Improving seismic QP estimation using rock-physics constraints

Geophysics ◽  
2018 ◽  
Vol 83 (3) ◽  
pp. MR187-MR198 ◽  
Author(s):  
Yi Shen ◽  
Jack Dvorkin ◽  
Yunyue Li
Keyword(s):  
Seismic Data ◽  
Closed Form Solution ◽  
Rock Physics ◽  
Seismic Inversion ◽  
Analytical Relation ◽  
Gas Reservoir ◽  
Seismic Section ◽  
Data Set ◽  
Attenuation Model ◽  
Text Model

Our goal is to accurately estimate attenuation from seismic data using model regularization in the seismic inversion workflow. One way to achieve this goal is by finding an analytical relation linking [Formula: see text] to [Formula: see text]. We derive an approximate closed-form solution relating [Formula: see text] to [Formula: see text] using rock-physics modeling. This relation is tested on well data from a clean clastic gas reservoir, of which the [Formula: see text] values are computed from the log data. Next, we create a 2D synthetic gas-reservoir section populated with [Formula: see text] and [Formula: see text] and generate respective synthetic seismograms. Now, the goal is to invert this synthetic seismic section for [Formula: see text]. If we use standard seismic inversion based solely on seismic data, the inverted attenuation model has low resolution and incorrect positioning, and it is distorted. However, adding our relation between velocity and attenuation, we obtain an attenuation model very close to the original section. This method is tested on a 2D field seismic data set from Gulf of Mexico. The resulting [Formula: see text] model matches the geologic shape of an absorption body interpreted from the seismic section. Using this [Formula: see text] model in seismic migration, we make the seismic events below the high-absorption layer clearly visible, with improved frequency content and coherency of the events.

scholarly journals Utilizing the Road Mark Training Set from Ground-Based Mapping System to Airborne Imagery in Deep Learning Framework

2019 ◽  
Vol 1 ◽  
pp. 1-1
Author(s):  
Tee-Ann Teo
Keyword(s):  
Neural Network ◽  
Deep Learning ◽  
Spatial Resolution ◽  
Image Features ◽  
Training Data ◽  
Training Set ◽  
Data Set ◽  
The Road ◽  
Mapping System ◽  
Close Range

<p><strong>Abstract.</strong> Deep Learning is a kind of Machine Learning technology which utilizing the deep neural network to learn a promising model from a large training data set. Convolutional Neural Network (CNN) has been successfully applied in image segmentation and classification with high accuracy results. The CNN applies multiple kernels (also called filters) to extract image features via image convolution. It is able to determine multiscale features through the multiple layers of convolution and pooling processes. The variety of training data plays an important role to determine a reliable CNN model. The benchmarking training data for road mark extraction is mainly focused on close-range imagery because it is easier to obtain a close-range image rather than an airborne image. For example, KITTI Vision Benchmark Suite. This study aims to transfer the road mark training data from mobile lidar system to aerial orthoimage in Fully Convolutional Networks (FCN). The transformation of the training data from ground-based system to airborne system may reduce the effort of producing a large training data set.</p><p>This study uses FCN technology and aerial orthoimage to localize road marks on the road regions. The road regions are first extracted from 2-D large-scale vector map. The input aerial orthoimage is 10&amp;thinsp;cm spatial resolution and the non-road regions are masked out before the road mark localization. The training data are road mark’s polygons, which are originally digitized from ground-based mobile lidar and prepared for the road mark extraction using mobile mapping system. This study reuses these training data and applies them for the road mark extraction using aerial orthoimage. The digitized training road marks are then transformed to road polygon based on mapping coordinates. As the detail of ground-based lidar is much better than the airborne system, the partially occulted parking lot in aerial orthoimage can also be obtained from the ground-based system. The labels (also called annotations) for FCN include road region, non-regions and road mark. The size of a training batch is 500&amp;thinsp;pixel by 500&amp;thinsp;pixel (50&amp;thinsp;m by 50&amp;thinsp;m on the ground), and the total number of training batches for training is 75 batches. After the FCN training stage, an independent aerial orthoimage (Figure 1a) is applied to predict the road marks. The results of FCN provide initial regions for road marks (Figure 1b). Usually, road marks show higher reflectance than road asphalts. Therefore, this study uses this characteristic to refine the road marks (Figure 1c) by a binary classification inside the initial road mark’s region.</p><p>To compare the automatically extracted road marks (Figure 1c) and manually digitized road marks (Figure 1d), most road marks can be extracted using the training set from ground-based system. This study also selects an area of 600&amp;thinsp;m&amp;thinsp;&amp;times;&amp;thinsp;200&amp;thinsp;m in quantitative analysis. Among the 371 reference road marks, 332 can be extracted from proposed scheme, and the completeness reached 89%. The preliminary experiment demonstrated that most road marks can be successfully extracted by the proposed scheme. Therefore, the training data from the ground-based mapping system can be utilized in airborne orthoimage in similar spatial resolution.</p>

Sign in / Sign up

Export Citation Format

Share Document