Paper presented at the IADC/SPE International Drilling Conference and Exhibition, Galveston, Texas, USA, March 2024.

In today’s well construction operations, a substantial volume of data is generated and stored across multiple databases. The primary objective being to use them as a guide for future well construction optimization. However, much of this data gets lost in computer storage, and appropriate information is difficult to find at the right time. This paper shows the results of deploying a generative pre-trained transformer (GPT) large language model on an operator’s dataset to alleviate this problem.

The process starts with gathering all relevant data into a common database. In this case, the dataset included sensor data, processed data, morning reports, end of well reports, after-action reviews of non-productive times, bit forensics data and publicly available data from wells drilled by other operators. The files were pre-processed, and metadata was added appropriately to ensure appropriate indexing and training of the information. This data is then fed to the cloud platform on which the model is learnt. The model is then integrated into the data platform so that the end users can pose queries.

The dataset consisted of more than 200 wells of the operator in a region that the operator is actively drilling. Data curation was a time-consuming task that had to be performed to ensure only quality and organized information was fed to the model. Documents containing well construction related subject matter were also used in the training to provide the end user assistance with core concepts. During the test stage, a multitude of questions were posed to the platform, including questions such as: What happened the last time there was a stuck pipe in this region? What is the best ROP that could be attained in the lateral section? Significant time savings were recorded due to the ease with which information could be retrieved. A big concern was the potential for wrongs answers being provided to the questions. To alleviate this concern, all answers were accompanied by references found in the database, to give the person reviewing the answers confidence in the answers.

This paper introduces the benefits that large language models (LLMs) bring to both well planning and real-time operations. LLM offers the capability to be able to retrieve information extremely quickly and provide answers in a conversational format to user questions. This paper also provides recommendations to the industry and details some of the challenges to adopting LLMs.