Well Completions Optimized by Agentic Conversational Analytics

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 228143, “Integrating Data-Driven Insights With Domain Expertise Using Agentic Conversational Analytics for Well-Completions Optimization,” by Carla J.S. Santiago, SPE, Niven Shumaker, and Ashley Weir, SLB. The paper has not been peer-reviewed. _ The design of completions and fracturing strategies presents challenges that hinder the effectiveness of unconventional resources development, including a wide range of parameters, the nonlinear nature of certain types of formations, and the balance between production and economic viability. Multivariate data analytics is key in evaluating those resources. The complete paper aims to present a workflow that leverages a multiagent conversational system to integrate data, analytics, and domain expertise for improved completion strategies. Insights From the Fayetteville Shale The Fayetteville Shale is a mature dry-gas play in Arkansas that has long served as a testing ground for completion strategies in unconventional developments. The experience in Fayetteville illustrates that optimization in shale completions is not purely a function of more-aggressive completions strategies. Previous researchers found that fracture aperture, spacing, and connectivity vary with depth and local stress regimes, significantly influencing how fractures accept proppant. Their findings support the need for adaptive completions strategies. Other analyses in the literature suggested that the interaction between geology and completions is often nonlinear and poorly captured by static models alone. These studies have a central theme in shale development: optimization of completions is inherently multidimensional and context-dependent. Improvements in one domain may yield different results across wells, even within the same field. As demonstrated by Fayetteville’s evolution, successful completions strategies depend on integrating geological understanding, data analytics, and iterative field feedback. Methodology Overview and Technology Stack. The application has two stages. In the first stage, a well map is provided, and the completions engineer selects an area of investigation. Based on the selected area, a tree-based ensemble regressor is trained with data from the selected wells. The machine-learning (ML) model is then used to make predictions of gas volume (IP90 per foot of completed lateral length) over an n-dimensional grid for the selected features. A mesh grid search algorithm is used to obtain optimized completion parameters within the feature space. Any completion parameter can be used as features. In this work, the selected features are lateral length, fracturing fluid volume per foot of completed lateral length, and proppant amount per foot of completed lateral length. The optimized feature values that result in the maximum gas production are retrieved. Additionally, a cost model is applied on the grid space to search for the optimized feature values that result in the minimum gas-production cost. The optimizer agent used in this paper is limited to the ranges of the sets included in the training set selection and, consequently, will not suggest solutions beyond the range of training data. In a second stage, a chat window is provided where the completions engineer can perform various tasks such as understanding the data and the trained model, generating visualizations, and obtaining insights about the completions and the Fayetteville Shale in general. It is at this stage that the user can interact with the agentic framework.