Projects

An iterative two-step scheme for fast geologic scenario falsification

Contributors: Anyue Jiang

Github: CNN_SS

Paper: Inverting subsurface flow data for geologic scenarios selection with convolutional neural networks

We have explored the robustness of Variational Autoencoder when dealing with complex and diverse geologic features

Contributors: Anyue Jiang

Github: VAE_Robustness

Paper: History Matching under Uncertain Geologic Scenarios with Variational Autoencoders

Model predictive control with dynamic deep neural network model for operation optimization on the surface power plant.

The objective is to develop a customized predictive analytics tool to enable efficient surveillance, control, and automation of geothermal operations. By integrating state-of-the-art machine learning algorithms with advanced multi-physics monitoring data acquisition systems, this subtask aims to develop into real-time model predictive control algorithms to improve the efficiency of energy production operations in geothermal reservoirs.

Contributors: Wei Ling, Yingxiang Liu

Github: Geothermal

Papers: Deep Learning for Prediction and Fault Detection in Geothermal Operations

Control Optimization Framework combined with a simulation model and proxy model for Patua Geothermal Field.

Combine the optimization framework with a simulation model and with the proxy model; Apply an Ensemble-based optimization strategy to reduce the computational expense.

Contributors: Anyue Jiang, Zhen Qin

Papers: 

The geomechanical risks of CO2 injection have drawn more attention in recent years and coupled flow and geomechanical simulation models have been increasingly used to study the geomechanical effects during CO2 injection. Optimization of the CO2 storage performance must, therefore, incorporate the geomechanical risks to ensure environmentally sound and safe operations.

In this project, We present an optimization framework for geologic CO2 storage under geomechanical risks, where coupled flow-geomechanics-plastic failure simulations are used to quantify the risks of injection-induced ground deformation and reservoir rock failure. Multi-objective optimization is formulated and solved to maximize CO2 storage while minimizing the two forms of geomechanical risks. The optimization decision variables include the location of each injection well. Multiple numerical experiments with increasing complexity are presented to demonstrate the performance of the proposed framework. The results reveal optimal decisions that are different from those obtained from the flow-only simulation that disregards the geomechanical risks associated with CO2 injection. When geomechanical risks are considered, the wells may not necessarily be concentrated in areas with the highest storage capacity because that may lead to rock failure and/or unacceptable levels of ground surface uplift. Overall, the observations from this study reveal important differences in optimization results and conclusions when geomechanical risks associated with geologic CO2 storage are considered.

Contributors: Fangning Zheng, Atefeh Jahandideh, Birendra Jha, Behnam Jafarpour

Papers: Quantification and Incorporation of Geomechanical Risks in Optimization of Geologic CO Storage Using Coupled-Physics Models,

Optimization of CO2 Storage under Geomechanical Risk with Coupled-Physics Models

My research focus lies in integrating deep learning-based reservoir surrogate models to calibrate reservoir models and optimizing decision variables while accounting for geologic uncertainty. The main goal is to use the surrogate models to assist in quick decision-making that can help to increase efficiency in field development planning.

Contributors: Ulugbek, Syamil Mohd Razak, Atefeh Jahandid

Github: link

Papers: Efficient Data Assimilation with Latent-Space Representations for Subsurface Flow Systems 

This project focuses on accurate description of the errors associated with flow simulations at multiple scales of resolution. The goal is to accurately predict and emulate flow responses from high-fidelity models by using lower-resolution, less precise prior geologic models, which are corrected via deep learning-based proxy models. By combining domain expertise in complex subsurface dynamical systems and state-of-the-art techniques in deep learning, we can solve these complex problems, with applications to uncertainty quantification, history matching, production optimization, and closed-loop reservoir management.

Contributors: Misael Morales

Github: Scale-Relationships

Many predictive models are built using simulated data where the forecasts are generated using a set of uncertain geologic models, under the assumption that the models are representative, and the computations are scalable. The issue arises when these models are too large or don’t scale well, and we require fit-for-purpose dynamical proxy models to accurately predict the complex reservoir responses. For this, we use state-of-the-art deep learning techniques to bring models and responses to a latent space and maintain spatial information as well as a description of the dynamical behavior of the system in order to have precise future predictions for real-time decision support and reservoir management.

Contributors: Syamil Mohd Razak, Misael Morales, Jodel Cornelio, Mahammad Valiyev

Realistic reservoir simulation models involve complex geomodels with millions of grid blocks, where even a single simulation can take days or weeks to complete, rendering many of the above workflows computationally prohibitive. Fast proxy models are designed to provide approximate predictions at a fraction of the computation time required by full-scale reservoir simulation. The conventional trade-off in developing proxy models is between speed and accuracy or fidelity (relative to the simulation model). Recent advances in machine learning (ML) have prompted the development of ML-based architectures as fast proxy models. However, an important issue in adapting data science and machine learning approaches to reservoir modeling is the need to incorporate domain knowledge and physical insight. Such incorporation can lead to better formulation, design, and solution of machine learning techniques in scientific domains and enhance their interpretability. In this project, the construction of physics-informed neural network proxy models to predict oil production in waterflooding experiments is explored, which provides accurate prediction as well as estimation of inter-well connectivity. The result shows that even the utilization of simple engineering insight can largely improve the performance of the neural network.

Contributors: Junjie Yu

Papers: Engineering Design of Neural Network Architectures for Estimation of Inter-Well Connectivity and Production Performance