Streamlining and Automating Drilling Fluids Advisory System with an End-To-End Machine Learning Pipeline

ABSTRACT The Drilling Fluid Advisory System is a decision support tool that optimizes and recommends the amounts of additives that needs to be added to the drilling fluid to achieve desired properties in real time. This article presents an automated end-to-end pipeline that was used to prepare data, train MIMO neural nets, optimize hyper-parameters and validate the performance of these models thereby enabling on-the-fly configuration of input products, fluid systems, and output properties. The Fluid Advisory System comprising of various components, including a data pipeline for loading historical data, a training pipeline for creating a digital twin of fluid property management, an automated predictive analytics tool for model selection, and a prescriptive model for determining additive quantities. The workflow orchestration involves codifying and automating the flow of data from start to finish of the model training and validation lifecycle in addition to optimizing concentration of additives based on current properties. Aiding human decision-making process using the Fluid Advisory System does not just allow autonomous management of drilling fluids. Ultimately, it allows us to integrate with automation hardware systems that handle the entire fluid management cycle including automated flow loops and mixing hoppers. Automating the training and inference pipeline has significantly accelerated and streamlined the process of reconfiguring and retraining models with different chemicals, treatments, and properties. All stages of the ML system being handled by a single process makes it easier to monitor, document, and compare model performance, leading to increased efficiency, rapid experimentation, and continuous model delivery. As volume and variety of data increases, training can be easily scaled up. This approach introduces domain-specific data validation and a centralized repository for standardizing, storing, and accessing features, enabling the discovery and reuse of up-to-date engineered features. By transitioning from level 0 to level 1 machine learning engineering culture and practices, the proposed framework established a foundation for rapid and reliable pipeline updates, as well as continuous integration and delivery, essential for transitioning software systems from development to production.