Semester of Graduation

spring 2022

Degree

Master of Science in Petroleum Engineering (MSPE)

Department

Petroleum Engineering

Document Type

Thesis

Abstract

Compositional modeling is essential when simulating processes involving significant changes in reservoir fluid composition. It is computationally expensive because we typically need to predict the states and properties of multicomponent fluid mixtures at several different points in space and time. To speed up this process, several researchers have used machine learning algorithms to train deep learning (DL) models on data from the rigorous phase-equilibrium (flash) calculations. However, one shortcoming of the DL models is that there is no explicit consideration for the governing physics. So, there is no guarantee that the model predictions will honor the thermodynamical constraints of phase equilibrium (Ihunde & Olorode, 2022).

This work is the first attempt to incorporate thermodynamics constraints into the training of DL models to ensure that they yield two-phase flash predictions that honor the physical laws that govern phase equilibrium. A space-filling mixture design is used to generate one million different compositions at different pressures (Ihunde & Olorode, 2022). Stability analysis and flash calculations are performed on these compositions to obtain the corresponding phase compositions and vapor fraction (Ihunde & Olorode, 2022). Physics-informed neural network (PINN) and standard deep neural network (DNN) models were trained to predict two-phase flash results using the data from the actual phase-equilibrium calculations (Ihunde & Olorode, 2022).

Considering the stochasticity of the deep learning optimization process, we used the seven-fold cross-validation to obtain reliable estimates of average model accuracy and variance (Ihunde & Olorode, 2022). Comparing the PINN and standard DNN models reveals that PINNs can incorporate physical constraints into DNNs without significantly lowering the model accuracy (Ihunde & Olorode, 2022). The evaluation of the model results with the test data shows that both PINN and standard DNN models yield coefficients of determination of ~97% (Ihunde & Olorode, 2022). However, the root-mean-square error of the physics-constraint errors in the PINN model is over 55% lower than that of the standard DNN model (Ihunde & Olorode, 2022). This indicates that PINNs significantly outperform DNNs in honoring the governing physics. Finally, we demonstrate the significance of honoring the governing physics by comparing the resulting phase envelopes obtained from overall compositions computed from the PINN, DNN, and linear regression model predictions (Ihunde & Olorode, 2022).

Committee Chair

Olorode, Olufemi

DOI

10.31390/gradschool_theses.5492

Download

Included in

Engineering Commons

Share

COinS