Generative AI and Large Language Model Assisted Causal Discovery and Inference for Driving Process Improvements

Data-driven process management coupled with machine learning have been successful in driving commercial value to oil and gas operators by offering insights into process disruptions and their root causes. One frequently used approach is to analyze causes of process disruptions exclusively from historical data. In general, specific insights in the form of high correlation between certain process performance indicators and a well-defined measure of production inefficiency is often confounded as responsible causal factors. While this may yield some insights, the complexity of processes, measured in terms of number of entities involved and their interrelationships, requires a more nuanced approach that must include the context of the specific process. Thus, data analysis must be augmented with significant inputs from experts. Causal Inference provides a conceptual framework and tools for doing such analysis. In causal analysis, we embed this specific knowledge of subject matter experts using causal graphs consisting of process features (nodes) and their dependency (directed edges). For complex processes however, constructing causal graphs could be non-trivial due to ambiguity over which nodes to include and the plausible direction of their relationships. With the advent of foundational Large Language Models (LLM), there is an opportunity to mitigate this problem by utilizing the enormous information it encodes. Tools and technologies now exist to customize the response of LLM using retrieval of information from a corpus of specific high-quality knowledge in the form of related literature and data. It can therefore be used to assist the domain expert in building and finetuning the causal graph, and in simpler cases, can completely automate this step. In this work, we propose a two-step approach to combine the power of LLMs and Causal Analysis for analyzing inefficiencies in production processes. In the first step, we implement a Retrieval Augmented Generation (RAG) enhanced LLM prompting on a curated dataset designed to answer specific questions on relationship between process performance indicators. The outcome of this step is a directed acyclic graph encoding dependency of process performance indicators. Domain experts can validate or potentially refine the LLM-generated causal graph based on their domain knowledge for eliminating spurious hallucinations. In the second step, we use an appropriate causal inference method on the refined causal diagram and historical production data to estimate the causal effect of process variable contributing to disruptions or inefficiencies. Thus, by combining human expertise with machine learning, this framework offers a comprehensive approach for optimizing production processes.