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

Shahab Mohaghegh

Committee Member

Samuel Ameri

Committee Member

Kashy Aminian

Abstract

Well logging is a fundamental technique in formation evaluation providing continuous, real-time measurements of geological and petrophysical properties within a well. Through the systematic analysis of well logs, engineers and geoscientists can accurately determine critical formation characteristics, including porosity, permeability, lithology, and fluid composition. Well logging is fundamental for making informed decisions and reducing uncertainties in the exploration and development of oil and gas reservoirs.

Despite its significance, the acquisition of reliable well log data in oil and gas wells is often compromised by various operational, mechanical, and formation-related challenges, as well as pressure, fluid, and equipment constraints. In high-risk scenarios where tool loss is a concern, logging operations may be avoided entirely to prevent costly recovery procedures, resulting in the absence of crucial data required for formation evaluation, reservoir modeling, and field development planning. Cased-hole logging offers an alternative to conventional open-hole logging, providing behind-barrier evaluation through tools and techniques with enhanced depth of investigation.

AI-based prediction of well logs has advanced considerably for open hole environments, with many models achieving high accuracy and field relevance. However, the field of cased hole logs prediction remains largely an underexplored domain despite their critical role in reservoir management and late-life well operations due to data scarcity, tool complexity, and interpretive challenges. This study addresses the issue of missing well log data by leveraging artificial intelligence (AI) methodologies to predict and reconstruct cased-hole integrity logs.

The study utilizes a dataset of six horizontal wells located in Bradford County, Pennsylvania targeting the Marcellus Shale. The training data consists of cased-hole measurements, including gamma ray, temperature, pressure, and four bands of acoustic power frequency curves—low, medium, high, and ultra-high—all recorded to the top of the Marcellus formation at approximately 4,500 feet within the near-vertical section of each well. A sequential model, used generally in AI-based open-hole logs prediction, was developed, but also non-sequential models were tried and compared to each other. All four developed models were trained and evaluated using cross-validation and key statistical metrics (R², RMSE, and AARE%). The RandomForest model demonstrated superior performance across all target curves, achieving a test R² of 0.9618 for log prediction, and 0.9864 for log reconstruction. However, results vary per-target curve depending on whether depth was modeled independently (non-sequential) or sequentially, capturing continuity along depth as in lithological heterogeneous dependencies, and especially when analyzing the AARE%. In several cases, the AARE-optimal model differed from the R²-optimal model, indicating that some models reduced relative error in low-magnitude intervals even when overall variance explanation was lower. The study also presents a per-target curve winner model recommendation and outlines future work to expand AI applications for cased-hole log prediction and reconstruction.

Download

Share

COinS