Machine-Learning-Based Automation of Formation Testing Data Interpretation and Analysis

Abstract

Abstract Formation testers have become an essential component of every well data acquisition campaign for exploration and appraisal. They are used for reservoir pressure and mobility profiling, as well as downhole fluid analysis and sampling, microfracturing, and other applications. Hundreds of formation testing jobs and thousands of pretests are performed on a daily basis around the globe, whether they be a on wireline or while drilling. Each pretest is handpicked by an engineer to classify the test type and select the representative reservoir pressure and mobility. The operator collects all processed data and analyses them in an integrated fashion—at the well and multiwell levels—to obtain insights. The whole process is still very manual, and it could take up to 2 weeks in the best-case scenario. The aim of this paper is to introduce artificial-intelligence (AI) components to the data interpretation and analysis workflow for formation testing data. We aim to leverage human-in-the-loop automation to an entire workflow using machine-learning (ML) algorithms, reducing turnaround time from weeks to minutes while also generating new insights. Using a combination of traditional and ML algorithms, automation will affect: raw-data ingestion process; pretest interpretation (extraction of reservoir pressure, mobility, and other key features along with classification of pretests into different types), well-level analysis (automated detection of reservoir zones using well log data for gradient selection along with optimal pressure gradient selection); and multiwell-level analysis and generation of ready-to-go data for reservoir characterization and simulation. The AI solution includes many supervised and unsupervised ML models, as well as traditional algorithms for calculating and extracting key features. The AI workflow was tested on a variety of previously unseen data from various wells drilled in clastic and complex carbonate reservoirs with varying reservoir pressure and mobility. The results showed that the ML model was very accurate in predicting pretest types, distinguishing reservoir zones using well logs, and selecting the best fluid gradients automatically. The solution was built in an enterprise-scale AI/ML data science studio that allows us to deploy it across the entire organization. The novelty of the solution is an automation of the entire workflow using a mixture of AI and traditional algorithms from pretest to multiwell analysis and an integration of the results into static and dynamic models for further history-matching workflows. The use of this product in real-world data would result in an order-of-magnitude reduction in the time engineers spend on this task and lightning-fast insights.