Mud Invasion Zone Detection Using Retrieval-Augmented Generation (RAG): A Generative AI Approach

Abstract Mud invasion, the process where drilling mud infiltrates surrounding rock formations, poses a significant challenge in well log interpretation within the petroleum industry. This phenomenon distorts measurements of formation properties such as porosity, permeability, and fluid content, leading to inaccuracies in formation evaluation and hydrocarbon volume estimation. Traditionally, detecting mud invasion zones requires manual analysis of well log data, particularly resistivity, bulk density, neutron, and sonic logs which are time-consuming and subject to interpretation biases. This study presents an automated approach to identifying mud invasion zones in well logs using Retrieval-Augmented Generation (RAG) with advanced indexing techniques, specifically Vector Store Index and Tree Index RAG. Our approach integrates a comprehensive RAG pipeline that includes naive RAG, Vector Store Index, and Tree Index RAG, leveraging domain-specific manuals and well data stored in LAS format, transformed into text files through a custom processing pipeline. These technical resources, including annotated "bad data zones" indicating mud invasion, enable the model to contextualize well log anomalies with greater precision. Additionally, the RAG system not only detects mud invasion zones but also retrieves relevant guidance from manuals on handling these cases, enhancing the model’s practical application in real-world scenarios. To evaluate performance, we use Cosine Similarity based on TF-IDF (Term Frequency-Inverse Document Frequency), ROUGE Score, Human Feedback (HF), and processing time metrics. TF-IDF allows Cosine Similarity to measure the relevance of retrieved information against manuals, while ROUGE Score assesses the overlap between generated outputs and pre-annotated bad data zones, providing an accuracy measure. Results show that the Vector Store Index RAG configuration outperforms others in both processing time and accuracy metrics. It achieves higher relevance and accuracy and offers significantly reduced processing times compared to naive and Tree Index RAG configurations. This efficiency makes Vector Store Index particularly suitable for practical deployment in well log analysis. Automating mud invasion detection can streamline workflows, reduce human error, and ensure consistency in well log analysis. This study underscores the growing role of Generative AI in the energy sector, where vast amounts of unstructured data must be processed efficiently. Our research demonstrates how combining NLP and domain knowledge can improve well log interpretation, leading to more accurate and scalable reservoir assessments. This work represents a step toward integrating AI-driven solutions in subsurface characterization, advancing automated interpretation in the energy sector.