Fluid-Advisory System Offers Recommendations for Various Systems

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 221076, “Autonomous Drilling-Fluid-Management System: Expansion of Fluid-Advisory System Across Multiple Fluid Systems,” by Fathima H.A. Razak, SPE, Elizaveta V. Onegova, and Sergei Postovalov, Baker Hughes, et al. The paper has not been peer reviewed. _ This research aims to develop a fluid-advisory system that provides recommendations for the optimal amounts of chemical additives needed to maintain desired fluid properties in various drilling-fluid systems. Building on the foundation laid by the fluid-advisory system for water-based fluids described elsewhere in the literature, this project expands its scope to include both water- and oil-based muds. To achieve this goal, a machine-learning model is created to predict target fluid properties, with separate models developed for water- and oil-based muds. An integrated optimization framework determines the ideal quantities of chemical additives. Methodology The fluid-advisory system calculates the optimal amount of a chemical additive needed to adjust a fluid property, such as mud weight, rheology, pH, hardness, chloride concentration, oil/water ratio, concentration of low-gravity solids, excess lime, and electrical stability. The system was developed using both a physics-based and a data-driven approach. With the first approach, the amount of a chemical additive required to be added to the mud is calculated using formulae derived from a mass-balance equation. The data-driven approach is used when the relationship between the fluid property and its modifier is more complex. The relevant data is collected and processed, then used to train and validate a multi-input/multi-output neural-network model that predicts fluid properties from the mud composition. Finally, an inversion problem is formulated to calculate the amount of a chemical additive based on the values of the current and target fluid properties. The model-building process includes feature engineering, feature selection, hyperparameter optimization, model training, model scoring, and performance evaluation, all of which were streamlined and automated to accelerate the reconfiguring and retraining of models. The focus of this paper is on expanding the fluid-advisory system to handle multiple fluid systems. The interdependency between additives and properties poses a complex optimization problem because of the high dimensional and noisy nature of mud data, making it challenging to qualify most data points for model-training purposes. Using data from laboratory experiments, which are comparatively well-defined, limited, and specific to scope, adds to the issue of insufficient data. To address this inadequacy, several measurements were adopted, including synthetic data oversampling, regularization, and transfer learning. Synthetic Data Oversampling Two approaches are adopted to tackle the issue of insufficient data. The first involves specifying additives that need an increased number of samples and applying interpolation to add new points between the measured values. Although precision and smoothness may be limited, this approach offers benefits such as ease of implementation, high performance in terms of execution and memory, and minimal possibility of overshooting or oscillations.