Realizing the Need for Agentic AI for Subsurface Data and Workflows

Subsurface data consists of various types, and it is used in various business workflows. The efforts spent on performing data checks and running the workflows are extensive. Advancements in machine learning and natural language processing have been beneficial in extracting information form unstructured data. Especially, the AI Agents now play very useful roles and capable of performing repetitive tasks. This abstract presents some of the Agentic AI supported concepts, where data is extracted, quality checked and used in Subsurface workflows. One example of Agentic AI workflow, where the information is extracted and data is made searchable from unstructured Cased ed Hole logs reports. These well logs support decisions related to Well Production. Following Steps are performed to achieve the use case objective. Data Extraction: Automatically extracting text and data from scanned documents and other documents (pdf, word, excel)Digital Data Availability: Comparing data from documents with digital data formats like LAS and DLIS.Natural Language Processing: Using NLP to understand query intent from available tools.Large Language Models: Using available LLM models and developing & training new models using extracted information.Retrieval Augmented Generation: Enhancing model accuracy by retrieving reliable data sources like SharePoint or custom file locations.Web User Interface: Allowing users to interact with AI agents for information retrieval. Having an Agentic AI chatbot developed using Large Language model, extracts information from the knowledge source and then provides information formatted using LLM to the business users. The chatbot helps support data management engineers in quality checking the data. This model is trained based on techniques like Summarization, prompt and classification. As a part of the presented work, various file types for the cased hole logs were collected with focus on categories like Production logging tool, Reservoir Saturation tool and Injection Logging tool. After Running the Natural language Processing techniques, different sections of the unstructured documents were extracted. A large language model to be developed based on the extracted information. To keep the LLM updated with the new data, Retrieval Augmented Generation is used. One of the examples of Chatbot for unstructured data prompts is mentioned here: The chatbot answers the following questions: Summarize the Well **NAME* Overview of Well **NAME* Which Wells have RST data Which Wells have PLT data Which Wells have ILT data Which Section contains cased hole data Show me the tool names Show me the plots What are the available Cased Hole tools available in Well **NAME* Show me the data available in Reservoir Name Show me the Well Sketch of Well **NAME* Show the Well Schematic of Well **NAME* Depending on the number of wells and required Petrophysical processing for any Subsurface analysis, it takes months to completed the workflow. Business users must go through all the unstructured data and reports. This is a very time-consuming process. Having a chatbot agent to find relevant information and a quality checked database support business workflows. In future, the large language model can be fine-tuned to incorporate more data categories. Utilization of Generative AI techniques supports the business and keeps the information up to date.