Creating geologically informed training models for deep learning inversion: Applications to potential-field data

Geological prior information is one of the ways of bringing geological realism into reservoir facies models. Models based on geologically realistic priors are more accurate and provide more robust predictions under uncertainty. Commonly geological prior information is captured from modern depositional environments or outcrops of ancient deposits. The aim of this work was to demonstrate how use of prior information improves the models of fluvial facies and reduces the uncertainty associated with the estimation of channel geometry. Geological priors are built using intelligent techniques like Artificial Neural Networks and Support Vector Regression. The data driven methods reveal the hidden relationships among the variables that form the priors and also allow handling the associated uncertainty. Furthermore, we used the intelligent prior models to predict realistic parameter combinations which may not have been observed in the available data but are still plausible and may exist in nature. The intelligent prior models were combined with multiple point statistics (MPS) simulation for a test synthetic reservoir case study. Multiple point statistics (MPS) was chosen to model channel facies because of its capability of modelling realistic geobodies and adaptability to well and seismic data. The study shows improvement in controlling simulated channel geometry and highlighting the impact on volume estimation and oil production.

Prospectivity mapping based on deep learning typically requires substantial amounts of geological feature information from known mineral deposits. Due to the limited spatial distribution of ore deposits, the training of predictive models is often hampered by insufficient positive samples. Meanwhile, data-driven mineral prospectivity mapping often overlooks domain knowledge and expert experience, leading to poor interpretability of predictive results. To address this problem, we employed the Gaussian mixture model (GMM) for spatial feature classification to expand the number of positive samples. The approach integrated the embedding of geological map knowledge graphs with geological exploration data to enhance the knowledge constraints of the prospecting model, which enabled the integration of knowledge with data. Considering the complex spatial structure of geological elements, a bi-branch utilizing the 1-dimensional convolutional neural network (CNN1D) and graph convolutional network (GCN) was used to extract geological spatial features for model training and prediction. To validate the effectiveness of the method, a gold mineralization prediction study was conducted in the Wulonggou area (Qinghai province, western China). The results indicate that, when the number of GMM spatial feature classifications was 17, the positive-to-negative sample ratio was optimal, and the embedding of the knowledge graph controlled the prediction area distribution effectively, which demonstrated strong consistency between the prospecting area and the known mineral deposits. Compared with the predictions by CNN1D, the fused prediction model of CNN1D and GCN yielded higher accuracy. Our model identified 11 classes of mineralization potential areas and provides geological interpretations for different prediction categories.

Objectives/Scope Variational autoencoders as generative models have been widely adopted through large language models in the natural language processing domain in recent years. However, the relative novelty of this architecture has resulted in limited investigation within the production engineering field. This study proposes a novel approach towards pump working condition monitoring for sucker rod pumping systems based on derived downhole dynamometer cards utilizing Conditional Variational Autoencoders (CVAE). Methods, Procedures, Process A case study comprising 17 wells in Upper Austrian oil fields of RAG Austria AG has been conducted. Using a finite element method, downhole dynamometer cards reflecting the operating condition of the pump are derived from polished rod load cell measurements at the surface. Consequently, a CVAE network is trained in two distinct configurations: utilizing labeled data encompassing pre-defined pump working conditions and employing a dataset solely comprising a healthy pump state exclusively for anomaly detection. Following the application of dimensionality reduction techniques of the generated hidden vectors, the transformed representations are employed to evaluate model performance. Results, Observations, Conclusions This study introduces a novel SRP-CVAE framework for continuous pump condition monitoring, intertwining finite element method based mechanical modeling of the sucker-rod string for downhole card derivation with a deep probabilistic generative model for pump working condition classification. It has been found that within a benchmark setting, pump condition classification accuracies can be increased by adopting a hybrid supervised CVAE architecture for automatic feature extraction incorporating both a generative and discriminative objective. The respective superior performance within this setting can be partly attributed to its inherent multi-objective optimization framework, wherein the model simultaneously optimizes reconstruction fidelity and classification accuracy through shared latent representations. In the second classification setting with the objective of solely distinguishing between a healthy or faulty pump working condition, the model is trained on datasets exclusively reflecting the respective states. The results show that the proposed SRP-CVAE model demonstrates significant robustness towards accurately detecting abnormal pump working states and generalizes well for conditions characterized by data scarcity. This may be ascribed to the inherent regularization through the Kullback-Leibler divergence regularization term effectively penalizing overly complex latent representations that deviate substantially from the imposed prior. Finally, the proposed architecture allows for synthetic data generation or augmentation by sampling from the latent space, which has proven useful for sampling rare pump working conditions. Novel/Additive Information This paper will serve as a resource to any engineer or technician in the production engineering domain who is building models for anomaly detection using sensoric data. The proposed architecture can be readily adapted to diverse anomaly detection applications or used to enhance classification accuracy in related domains, thereby advancing health, safety and environment standards as well as facilitating risk mitigation in field operations.

AI Safety is a major concern in many deep learning applications such as autonomous driving. Given a trained deep learning model, an important natural problem is how to reliably verify the model's prediction. In this paper, we propose a novel framework --- deep verifier networks (DVN) to detect unreliable inputs or predictions of deep discriminative models, using separately trained deep generative models. Our proposed model is based on conditional variational auto-encoders with disentanglement constraints to separate the label information from the latent representation. We give both intuitive and theoretical justifications for the model. Our verifier network is trained independently with the prediction model, which eliminates the need of retraining the verifier network for a new model. We test the verifier network on both out-of-distribution detection and adversarial example detection problems, as well as anomaly detection problems in structured prediction tasks such as image caption generation. We achieve state-of-the-art results in all of these problems.

Mineralization in the Limerick Basin, located in southwest Ireland, uniquely associates with volcanic rocks, unlike other mineralized zones in the Irish midlands, where mineral systems align with large-scale normal faults. To better visualize the subsurface structures influencing Limerick’s mineralization, we conducted 2D gravity inversion incorporating geological and petrophysical constraints.Conventional methods of deterministic inversion involve adding a model norm term to the data misfit term in the objective function to regularize an ill-posed problem and obtain stable solutions. While previous studies on constrained deterministic inversion have modified the model norm to include prior information or constraints in geophysical inversion, the complex nature of geological priors makes encoding this information mathematically and computationally challenging. To tackle this problem, we implemented a deep generative model, specifically a conditional variational autoencoder (cVAE)-based inversion framework, to incorporate structural constraints derived from drill hole and petrophysical data.Initially, we tested this framework on a synthetic case by training the cVAE on thousands of 2D density models comprising geological features analogous to the field geology and populated with density values consistent with the drill core measurements acquired from the study area. Artificial drill holes were created to fix the depths of geological units at the drill hole contact points across all training models, ensuring that the predicted models adhered to prior constraints. Following training, we tested the network on some test data, which showed that the predicted models successfully captured the structural and petrophysical property constraints. The geometries of the geological features were also well recovered.We applied this method to gravity data from a NW-SE trending profile crossing the western part of Limerick Syncline. Thousands of density models were generated using drill hole data, incorporating measured rock densities for training. Since the profile’s central and deeper sections lacked sufficient geological data for direct validation of the results, we implemented a hypothesis-testing approach. In each hypothesis, geological features were added to the training density models based on prior geological knowledge of the study area. If simulated data from an inverted model failed to match field data, more geological features were added to the training models in the next hypothesis, and the workflow was repeated to achieve a low data misfit.The inversion provided three key insights into the study area’s geology. First, it identified potential volcanic intrusions in the southern Limerick Syncline, possibly extending from depths below the basement. Second, it estimated the syncline’s geometry in areas with limited geological constraints. Third, it revealed a sharp vertical displacement in stratigraphy, indicating a potential south-dipping fault in the northwest portion of the syncline. This fault may have influenced mineralizing fluid migration, playing a critical role in mineral deposit localization.

During summer 2008 a broad magnetic survey performed in Markazi province, Iran, to study the magnetic properties of a hidden fault zone in order to better locate it and calculate its unknown geological parameters. The total field magnetic anomaly along several profiles crossing the fault strike was observed using proton magnetometers as base station and field observers. The location of the profiles was determined based on geological information previously obtained. We submitted each profile anomaly to an inversion scheme. This inversion method uses a Marquardt-Levenberg algorithm to calculate the fault unknown parameters such as depth, dip angle and geometrical characteristics along each profile with an iterative approach. Mapping the profile anomalies, we detected the fault zone extent and chose the proper wavelet to perform the inversion. Results along each profile show good correlation with each other and with the geological information known from pervious geophysical and geological studies.

Flood protection systems are complex, interconnected engineered systems, where failure at one location means the failure of the entire system. Earthen levees, the systems’ major component, are at risk from many causes of failure including seepage, overtopping, erosion and instability due to seismic loading. Levees stretch for long distances and are formed through various geologic processes and human activities over time, however information regarding soil properties is collected only at limited point locations and varies significantly both laterally and with depth. Prediction of levee performance in locations where no soil data is available becomes a limitation for system risk assessment studies. This study attempts to test the hypotheses that spatial variability of soil properties is correlated to regional variables such as distance from nearest river segment, river meandering sinuosity index and surface geology. A geostatistical ordinary kriging approach was used for developing these correlations. The specific areas used for data collection and analysis and model development in this study were sub-sections of the larger Sacramento River Flood Control Project (SRFCP) in northern California. Soil strength parameters of identified levee stratigraphy layers were statistically analyzed using a geostatistical ordinary kriging approach and correlated to preselected regional variables. Global observations that applied across the study area included the increasing trend of undrained shear strength for cohesive soils, Su, with increasing distance from the river, and decreasing trend of Su with increasing river Sinuosity Index levels. Only local trends were observed in the relation of friction angle of cohesionless soils, ?, with Sinuosity Index, as well as in the relation of Su and ? with geological formations.

We processed in time and depth two seismic lines from an area of extremely complex geology in the Turtle Mountain area of Southern Alberta, Canada. The time processing was designed to attenuate noise and enhance signal in the data. To develop a velocity model for pre‐stack depth migration (PSDM) we integrated all available sources of velocity and geological information. We used the subweathering velocity models derived from refraction statics analysis together with constant velocity migrations for velocity information in the shallow section. The locations of velocity pull‐ups on the time migrated sections acted as guides to the extent of high‐velocity carbonates carried in the hangingwall of a major thrust fault. We integrated the mapped surface geology, geological cross‐sections, well depths and interval velocities from sonic logs into the velocity model. The depth processed sections show a more realistic geometry than the time sections for the reflectors at depth. The seismic data cannot image the shallow, steeply dipping strata in Turtle Mountain itself. Our interpretations of the depth sections are based upon published geological models, surface geology maps, well data and the seismic character where the reflections are imaged adequately.

Summary Seismic impedance inversion is a technique that converts seismic data interface information into stratigraphic information, serving as a bridge between seismic data, well logging, and geological information. Although the existing deep learning impedance inversion methods have improved inversion accuracy, they do not involve the quantitative evaluation of uncertainty in the inversion results. Based on these effects, we improved the existing CNN impedance inversion model by incorporating Bayesian theory and approximate Bayesian inference. We implemented two methods for impedance inversion: one based on a Bayesian convolution neural network (BCNN), and the other is a CNN based on Monte Carlo (MC) sampling and Dropout using approximate Bayesian inference. These methods allow for uncertainty modeling and incorporation of uncertainty into the inversion model. The uncertainty of the inversion results is quantified through the output probability distribution. The results of field data applications clarify the effectiveness and characteristics of the two methods. MC Dropout CNN method is superior in efficiency due to approximate theory, while BCNN shows advantages in accuracy, resolution, and uncertainty quantification.

A Hylogger-Itrax core-scanner comparison for multi-scale high-resolution petrophysical characterisation workflow

The effect of mineral composition on the sorption of cesium ions on geological formations

One of the methods to decrease CO2 emission is the capture of the concentratedly emitted carbon dioxide and its storage in geological formations. The depleted oil and gas fields have already proven that they are capable of trapping the fluids for millions of years, so there’s the biggest chance of long‐time safe carbon dioxide storage in these reservoirs. Considering the CO2 storage in geological formations two important questions arise. How much CO2 can be stored in a selected depleted or partially depleted oil or gas reservoirs? How much CO2 can be injected through an existing injection well? To calculate the injectivity of the existing well, the mobility of the CO2 at reservoir conditions must be known. In the RIAES a new laboratory equipment was developed to measure the mobility of the supercritical state CO2 gas in porous reservoir rock samples. With this laboratory equipment some test measurements were performed, to determine the mobility of pure carbon dioxide and polluted carbon dioxide gas. The main part of the developed laboratory devices, and as an example, the results of test measurements are shown in this paper.

<p>3D modelling of complex and irregular geological bodies is an expanding discipline that combines two-dimensional cartographic and structural data managed with GIS technology. This study presents a complete workflow developed to process geological information to build a 3D model of major stratigraphic, structural and tectonic boundaries. The investigated area is located in the western part of the Aar Massif (external Central Alps, Switzerland) characterized by pronounced topographic (600–<4000 m) relief, making it prone for surface based 3D depth constructions. The workflow comprises four major steps:</p><p>(1)  <strong>Generation of 2D polylines in a map view</strong>: a two-dimensional dataset of sequences of polylines has been generated in ArcGIS (10.3.1) defining the starting dataset for the major stratigraphic and tectonic boundaries of the bedrock units. This dataset has been compiled and integrated by using: (i) GeoCover vector datasets 1:25 000 of the Swiss Geological Survey; (ii) The Geological Special Map 1:100 000 of the Aar Massif and the Tavetsch and Gotthard Nappes of the Swiss Geological Survey; (iii) data from literature; and (iv) additional field work conducted for this study in key-locations.</p><p>(2) <strong>Projection of 2D information onto 3D digital elevation model</strong>: with the 3D structural modelling software Move (Petex/Midland Valley; 2019.1) the boundaries have then been projected on a digital elevation model (swissALTI3D) with 2 m resolution.</p><p>(3) <strong>Construction of tectonic cross sections</strong>: the use of geometric arguments as well as structural measurements allows for projection of these boundaries into a dense regularly spaced network of 2D cross-sections.</p><p>(4) <strong>Interpolation of 3D surfaces</strong>: the surface and cross-sections boundaries can be interpolated by applying 3D projection and meshing techniques resulting in a final 3D structural model.</p><p>Generally, steps (2–4) require iterative adaptations particularly in the case of surface areas being covered by glaciers or unconsolidated Quaternary sediments. In the model, special emphasis is given to visualize the current structural disposition of the western Aar Massif as well as the relative geometric and overprinting relationships of the deformation sequence that shaped the investigated area throughout the Alpine deformation. Finally, since in the investigated area underground data are scarce, an assessment of the relative uncertainties related to input data and is intended to be performed following the approach proposed by Baumberger (2014) and Ferńandez (2005). The workflow presented here offers the chance to gain validation approaches for domains only weakly constrained or with no surface data available, by generating a 3D model that integrates all accessible geological information and background knowledge.</p><p> </p><p>REFERENCES</p><p>Baumberger, R. (2014): Quantification of Lineaments: Link between internal 3D structure and surface evolution 328 of the Hasli valley (Aar massif, central alps, Switzerland), University of Bern, PhD Thesis, unpublished.</p><p>Ferńandez, O. (2005): Obtaining a best fitting plane through 3D georeferenced data, Journal of Structural Geology 27, pp. 855–858</p>

The stability of mooring systems in floating offshore wind turbines (FOWTs) is essential, as any deterioration affects platform performance and overall efficiency. Machinelearning monitoring of FOWTs requires comprehensive dynamic response data derived from various mooring systems. Data collection should encompass a variety of health states, operational scenarios, and metocean conditions. Data is scarce in practice, especially concerning damage-associated information. Additionally, the existing imbalanced dataset lacks clear labels, making it difficult to differentiate between healthy (majority) and damage-associated (minority) instances. We propose a novel deep generative model (DGM) to efficiently deal with the challenge of unsupervised labeling of imbalanced data in the initial stage. It will subsequently be used to augment the labeled data from the previous stage and generate damage-associated data for a new (target) mooring system based solely on the healthy records of that system. The conditional hierarchical variational autoencoder (CHVAE) is developed using a diffusion probabilistic architecture during a pretrain-finetune training procedure. During the pre-training stage, trained on imbalanced, unlabeled data, we employed an unsupervised feature engineering approach in the latent space created to identify the healthy (majority) and damaged (minority) data. In the fine-tuning stage, the nonlinear relationship between healthy and minority-damaged responses of a mooring system is obtained, considered as the source domain across diverse sea states. Subsequently, by employing the healthy data from the target mooring system, the CHVAE can generate real-scale damaged responses of the system across diverse operational and environmental conditions. An analysis is conducted to evaluate the similarity between the simulated records produced by OpenFast on the OC4-DeepCWind FOWT benchmark and those generated by the CHVAE framework, utilizing visual, statistical, and behavioral methods. Furthermore, the proposed DGM’s performance is first assessed using MNIST benchmark image dataset to evaluate its effectiveness for unsupervised labeling and data augmentation. The generated records for unobserved random sea states closely resemble real-world dynamic behaviors during downstream binary classification, illustrating the effectiveness and versatility of the proposed DGM, CHVAE.

Fund transfer fraud detection: Analyzing irregular transactions and customer relationships with self-attention and graph neural networks