Author ORCID Identifier

https://orcid.org/0009-0000-7620-6820

Semester

Spring

Date of Graduation

2026

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Petroleum and Natural Gas Engineering

Committee Chair

Kashy Aminian

Committee Member

Sam Ameri

Committee Member

Ebrahim Fathi

Abstract

Flow regime identification in co-current upward gas-liquid flow through annular conduits remains a significant challenge in petroleum engineering, with major safety and operational implications. It is also important across industries involving the transport of multiphase fluids. Misidentifying flow regimes can introduce major operational risk, yet regime boundaries in annular gas-liquid flow are often visually complex and context dependent.

The objective of this study was to evaluate the utility of convolutional neural network (CNN) classifiers for flow regime identification. The CNN was trained using annular flow image dataset published by Texas A&M University. The dataset consists of approximately 947 RGB images organized into four flow regimes (bubbly, Taylor, slug, and churn) with substantial variability in image size, aspect ratio, lighting, and visual ambiguity. To quantify how preprocessing and architectural design influence performance of CNN, experiments were performed using dynamically parameterized CNN architectures trained with cross entropy loss and the Adam optimizer, followed by post hoc probability calibration using temperature scaling optimized with Limited memory Broyden Fletcher Goldfarb Shanno (LBFGS) method. Three progressively refined implementation pathways were used to address fixed size inputs, dynamic padding, and more efficient input handling, yielding a large architecture sweep (871 trained models) for comparative analysis.

With respect to these experiments and their hyperparameters, results improve across the implementation sequence, with moderately deeper architectures and more frequent pooling performing most consistently. Final evaluation considers predictive performance and calibration (expected and maximum calibration error), noting split sensitivity and label ambiguity as limiting factors. The thesis concludes with guidance to improve reliability, including stronger split controls, better region of interest isolation, and expanded, curated imagery to strengthen generalization for deployment.

Download

Share

COinS