2020 SMART Annual Review Meeting – Virtual Poster Sessions

UT-BEG has actively participated in Task 5 of the SMART Project, in the areas of designing the testing (toy) problems, implementing machine learning models, and implementing workflows. Multiple ML models were screened under Task 5 for model comparison and baseline development. So far, we have developed the long short-term memory (LSTM) model, sparse grid model, and Gaussian process regression (GPR) model on 2D and 3D toy problems.

Fast proxy models of CO2-water flow are useful for the engineering and design of geologic carbon sequestration reservoirs.  In some important limiting cases, fast proxy models can be used in place of full-physics models, and for training of machine-learning methods.  We are developing  transient macroscopic invasion percolation (TMIP) and diffusion-limited aggregation (DLA) approaches.  Preliminary results demonstrate the promise of the methods through their ability to model two-phase displacements very rapidly.

Joint multiphysics experiments will allow us to accurately measure CO2 saturation and its effects on geophysical data. Simultaneously measured NMR data will provide saturation changes during CO2 injection while ultrasonic velocity and complex conductivity measurements determine the associated mechanical and electrical changes. Our findings will help guide machine learning algorithms when used on observation data (Seismic and EM) – both from a physical understanding and from quantitative data.

We will discuss a novel approach for rapid field-scale modeling and visualization of subsurface flow and transport and its application to unconventional oil and gas reservoirs. Our proposed approach is based on a high frequency asymptotic solution of the diffusivity equation in heterogeneous reservoirs and serves as a bridge between simplified analytical tools and complex numerical simulation.

Real-time reservoir management demands fast assimilation of observation data to reduce the uncertainty in reservoir properties and future prediction. The most computational expensive component of data assimilation workflow is forward reservoir simulation.

In this work, we tested a variational autoencoder (VAE) with gradient-based optimization approach for GCS inverse analysis with regularization. A VAE will be trained to map to a low-dimensional set of latent variables with a simple structure to the high-dimensional parameter space (i.e., original space) that has a complex structure.

We evaluate the feasibility of using physics-informed machine learning (PIML) for subsurface energy-related pressure management. Pressure management is important when fluids are being injected and or extracted from the subsurface to avoid induced seismicity and unwanted migration of fluids. We develop a PIML framework that trains a neural network to manage pressures by determining extraction rates for arbitrary reservoir

ConvLSTM has been successfully used for complete time-series profile prediction in carbon storage applications. Our study indicates that univariate models predicting single variable outperform multivariate models predicting multiple outputs simultaneously, while the modified affected range scheme shows good performance in sink/source type scenarios, such as predicting water production rate of a well.

The SMART Model Comparison App is a framework for the standardization and comparison of machine learning models and their predictive performance. The standardization of inputs, outputs and model classes allows for an easy apples-to-apples comparison of multiple machine learning models all in one place. The aim for this app is to allow different users to compare model performance by several criterion, including prediction speed, overall (global) predictive ability and grid level (local) predictive ability.

An efficient Physics-Constrained Deep Learning Model is developed for predicting three-dimensional multiphase fluid flow and transport in porous media. The model architecture fully leverages the spatial topology predictive power from deep convolutional neural networks, and it is coupled with an efficient physics-based smoother for the image processing that need continuity at pixel level. 

Graph Neural Networks (GNNs) are an emerging deep learning architecture useful for making predictions across nodes and edges of a graph. For the task of approximating full physics simulator outputs across grids of cells, the GNN architecture is a natural fit and has advantages over other modeling approaches. The Battelle team tested the GNN model on a toy CO₂ injection problem to evaluate its performance both spatially and temporally. The results were promising, but also revealed some opportunities for improving performance.

We demonstrate a physics-informed deep learning method that uses deep neural networks but also incorporates flow equations to predict a carbon storage site response to CO2 injection. A 3D synthetic dataset is used to demonstrate the effectiveness of this modeling approach. The model approximates the temporal and spatial evaluations of CO2 pressure and saturation and predicts water extraction rate over time (outputs), given the initial porosity, permeability and injection rate (inputs).

Distributed Acoustic Sensing (DAS) converts fiber-optic cables into massive arrays of strain-rate sensors, which enables continuous monitoring of strain-rate changes at high spatial resolution (< 1 m) along the length of the wellbore. At the low-frequency limit (<mHz), DAS is extremely sensitive to quasi-static strain(rate), which makes it an excellent tool for continuous monitoring of in-situ deformation and stress during reservoir operations. In this work, we explore the application of AI deep learning approaches to low-frequency DAS-derived strain datasets from sparse boreholes to map the evolution and spatio-temporal distribution of strain within a reservoir.

The goal of this research is to obtain high-resolution geophysical attributes by posing the problem as a missing value imputation question. Here we assume that low resolution field-scale geophysical attributes represent possibly noisy, sparse observations of a higher resolution (e.g., ~10m scale) model and the remainder of that model is missing. We use deep learning, specifically Generative Adversarial Imputation Networks, to fill these missing values. We present a preliminary experiment with the Kimberlina simulation, which enables us to understand the utility, limitations, and areas of improvement for this approach.

Maximizing stimulated reservoir volume is one of the primary hydraulic fracturing concerns for economic production from horizontal shale gas or oil well. We developed a workflow to evaluate the optimal placement of multiple clusters for each stage to enhance production per lateral. The proposed engineered completion design aims to address the challenges from Marcellus subsurface geologic complexities including extensive natural fractures, variations in geomechanical properties, and reservoir characterizations.

Fast predictions are critical for the SMART platform; therefore, quantifying the tradeoff between machine-learning algorithm complexity and training/prediction time is an essential first task. We began by comparing a relatively simple tree-based ensemble algorithm (Random Forest) against a higher-complexity neural network model (MLP) using the 2D and 3D Toy Problems. Each algorithm is used to predict pressure, CO2 saturation, and water extraction rate.

The SMART Initiative will leverage a regional geologic carbon storage model developed under the CarbonSAFE Rocky Mountain Phase I program to aid in the development of fast-predictive models. The model represents the Glen Canyon Formation and associated sealing formations on the Colorado Plateau’s western edge in central Utah. The primary storage reservoir is the Navajo Sandstone, the upper member of the Glen Canyon Formation.

Work to date has involved the development of surrogate models for CO2 geologic storage using multilayer perceptron (MLP) and long short-term memory (LSTM) neural networks that are capable of accurate prediction of spatio-temporal outputs of CO2 saturation and pressure in both 2D and 3D spaces. Synthetic training datasets were developed using CMG-GEM simulations (provided by U.T. Austin).

Tracking the state of stress evolution within developed reservoirs is important for optimizing use and understanding seismic hazard. We develop a novel method to obtain a full description of the evolving stress field that utilizes machine learning to quantify the subsurface stress tensor changes using satellite observations. We train a convolutional neural network using training data produced from analytical realizations of surface displacements mapped to stress, constrained by possible stress sources in the reservoir.

In order to understand the long-term storage potential in the subsurface, the ability to quantify relative permeability and accurately simulate CO₂ propagation within the storage reservoir is necessary. Using an unsteady state flow methodology coupled with CT, empirical data on saturation and relative permeability in multiple rock types was gathered into a database called CO2BRA.

Image-to-image translation using a conditional deep convolutional generative adversarial network is applied to the task of replicating the output of a 3D multiphase flow simulator.

The MSEEL wells contain many terabytes of subsurface data that is accessible online and available for big data processing. Data availability is one of the biggest challenges faced by the SMART community for modeling and analysis. Collecting, integrating, and intuitively managing data is a time-consuming process, but one which is fundamental to further analyses.

Solving the seismic full-waveform inversion (FWI) problem can be challenging due to its ill-posedness and high computational cost. We develop a new hybrid computational approach to solve FWI that combines physics-based models with data-driven methodologies.

We exploit geological principles in enhancing the resolution of field-data derived field-scale attributes to log-scale by employing autoencoders.

The Kimberlina 1.2 Reservoir Model was created as part of the National Risk Assessment Program (NRAP) funded by the Carbon Storage part of DOE’s Fossil Energy program. We have taken the flow model results which simulate CO₂ injection into the Vedder Sandstone reservoir at approximately 3km depth, and converted hydrologic properties at 0 and 20 years after start of injection to 3D volumes of density, electrical resistivity, and seismic velocity.

This poster is going to share a material balance principle based approach to support associated CO2 storage decision making in a rapid fusion. The method mainly focuses on operational support by providing injection allocation, extraction response, and reservoir communication based upon injection and production data study.

This work presents comparison of multiple machine learning (ML) methods to predict CO₂ flow and production rates based on 2D and 3D synthetic problems. ML architectures have been constructed using convolutional neural networks and long-short term memory (LSTM). A number of hyperparameters and size of CNN-LSTM architectures (e.g., encoder-decoder, U-Net) are explored to evaluate the prediction accuracy. In particular, we will highlight the parameter optimization and the impact of hyperparameters on prediction accuracy.

This poster will share the proposed methodology of how to use multiple level of the data driven approach to help identify the fracture network to support the rapid decision making. The static data, hydraulic fracking data, and dynamic monitoring/testing data will be applied to the fracture network visualization and imaging regarding to the data availability over the field development stages.

This project aims to determine if DAS monitoring data can be used to detect anomalous changes in stress-related parameters (strain, stress, displacement, pressure) that could lead to fluid leakage out of reservoir via injection-induced/enhanced fractures/faults. The approach will involve 1) Test proof-of-concept using synthetic (modeling) data and a real geologic reservoir-caprock system (Chester 16 reef, Michigan); 2) Create synthetic training data set(s) that include(s) pressure, strain, stress, displacement for long-term injection period that causes fracturing/faulting and “focused” fluid migration out of reservoir; 3) test different ML algorithms.

This work presents a preliminary result to demonstrate a framework for accomplish machine learning-driven CO₂ modeling by combining a variational autoencoder (VAE) with ensemble-based data assimilation (EnDA) for data assimilation. For fast testing and optimal implementation of VAE, CO2 saturation and pressure distribution are simulated using an open-source percolation model (pyPERC) and single phase flow model (MODFLOW).

Forecasting of production performance from reservoir with complex geometry, geologic parameters distribution, well completion and operation is a challenging task. In the SACROC model, 5-spot injection scheme is applied with water alternating gas injection. In the coarse grid model, 23 production wells and 22 injection wells are placed. The water and gas injections are altered annually.

Induced seismicity, earthquakes caused by anthropogenic activity, have increased significantly in the last decade resulting from practices related to oil and gas production. Large earthquakes have been shown to promote the triggering of events due to static stresses caused by physical movement along the fault, and also remotely from the passage of seismic waves (dynamic triggering). In order to understand the mechanisms leading to earthquake failure, we investigate Prague, Oklahoma, a region where natural, induced, and dynamically triggered events occur.

Through the Plains CO₂ Reduction Partnership (PCOR) Partnership, led by the Energy & Environmental Research Center (EERC) in partnership with Denbury Resources Inc. (Denbury), a variety of geological, geophysical and geochemical data from the Bell Creek oil field of southeastern Montana was collected to evaluate incidental carbon dioxide (CO₂) storage associated with a commercial enhanced oil recovery (EOR) project.

Potential CO₂ storage reservoirs were represented by a GoM model and a Permian Basin SACROC model. Fifty geologic realizations for both sites were ranked using the diffusive time of flight (DTOF) method. Candidate model realizations at p10, p50, and p90 levels were selected for dynamic reservoir modeling.

Imaging CO₂ saturation in carbon storage monitoring requires multiphysics approaches that quantitatively integrate different geophysical data sets. Each geophysical method is primarily sensitive to one physical property, and each physical property in turn depends upon the CO₂ saturation differently. Statistical petrophysical data that characterize these physical properties in different zones in a reservoir constitute an important data set.

This poster will present an approach for combining and training physics-based and machine-learning based geomechanical models. We will briefly discuss the planned approach and the progress and challenges faced so far.

Simulation of carbon storage reservoirs have shown sensitivity to the simulation constraints, especially for parameters such as the injection well bottom-hole pressures (BHP). Since control of injection well BHPs are extremely important for safe and compliant class VI injection, Penn State and the EERC conducted sensitivity studies with current SMART-CS simulations. These studies assured that current simulations are generating expected (physical) reservoir responses as well as examined the utility of local grid refinement in cases where non-physical reservoir responses are observed.

The role played by natural fractures in unconventional resource plays varies among basins and even within a particular basin. One common theme among unconventionals, which differentiates them from conventional fractured reservoirs, is that the natural fracture population prior to drilling does not constitute a permeable flow network. This may result from a sparse, poorly connected natural fracture network, or alternatively a more substantial fracture network that has been sealed by mineralizing fluids.