Game Changer in Chemical Reagents Design for Upstream Applications: From Long-Term Laboratory Studies to Digital Factory Based On AI

Abstract The modern oil production industry necessitates an ever-expanding use of various chemical agents to optimize extraction processes, enhance recovery efficiency, and address operational challenges such as corrosion, scaling, and emulsion stability.Development of hard-to-recover oil reservoirs and extraction of residual oil from mature oilfields are the main reasons for the widespread application of chemical enhanced oil recovery (EOR) and improved oil recovery (IOR) techniques. Each oilfield and reservoir have unique conditions and fluid properties. To successfully implement chemical EOR (Enhanced Oil Recovery) and IOR (Improved Oil Recovery) technologies, it is essential to select or develop specific chemicals or chemical formulations tailored to the particular conditions of the reservoir. However, development of such reagents takes a lot of time and resources, it is limited by available initial reagents and precursors in the market. This work presents a new approach for developing chemicals using chemoinformatics models and artificial intelligence. The proposed approach consisted of three stages: generation of synthetically available molecules using the "Virtual Factory" algorithm, prediction of properties using a graph neural network with selection of the best candidates, and laboratory synthesis and validation of the predicted properties. The Virtual Factory algorithm used 159 of the most widely produced precursors on the local market and 10 templates of most common in practice reaction. The result of the algorithm was 5731 digital structures of surfactants of five various chemotypes. Graph neural network was designed to predict to predict interfacial tension (IFT) which is the most crucial property for surfactant flooding technology. A training set consists of 55,477 data points of molecule structural properties along with 218 IFT measurements for different reservoir conditions and oil properties for the training set. From 5731 generated digital surfactants structures, 4 surfactants with the lowest IFT among various chemotypes for oilfield X conditions were selected based on graph neural network predictions. The surfactants were synthesized in the laboratory and IFT was measured for oilfield X conditions to validate the neural network prediction.