Evaluation of Time-step Frequency on Prediction Accuracy Applied to Deep Learning Neural Network Surrogate Models for CO2 Storage

Abstract:

Surrogate modeling strategies are regularly used in fluid flow and transport applications. Their computational efficiency affords the ability to generate numerous realizations while providing accurate approximations to complex reservoir simulator input/output relationships. The application of deep learning (DL) approaches is considered a promising strategy for developing surrogates for complex multiphase flow problems. Supervised DL architectures have become increasingly adept at forecasting complex spatio-temporal response variables. The challenge here is that large volumes of modeling simulation realizations needed to serve as training sets to build effective DL models. The creation of extensive datasets via reservoir simulation can be both time-consuming and computationally demanding for complex multiphase flow problems. Additionally, this issue is expressed during DL model training and hyperparameter tuning as well. In this study, we developed surrogate models for CO₂ geologic storage using multilayer perceptron and long short-term memory 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. For each modeling domain, 27 unique scenarios with 72 timesteps were generated, each with varying porosity, permeability, and CO₂ injection rates as predictor variables. The response variables (i.e. reservoir pressure, CO₂ saturation, and water production rates) are forecasted using the DL approaches. We explore an adaptive approach that reduces the time-step frequency and analyzes its effect on model accuracy. The goal is to provide insight to future modeling efforts in their pursuit of suitable data based on problem complexity. This research is part of NETL’s SMART-CS initiative Task 5 aimed at developing a virtual learning environment that enables different stakeholders’ ability to explore and test CO₂ storage reservoir behavior.