Workflow for predicting undersaturated oil viscosity using machine learning

Undersaturated oil viscosity is a dominant fluid parameter to be measured in oil reservoirs due to its direct involvement in flow calculations. Since PVT experimental work is expensive and time costly, prediction methods are essential. In this work, viscosity data from in-house and literature measurements (500+ reports, 20,000+ data points) has been utilized for the first time to develop machine learning models predicting undersaturated oil viscosity using easy-to-get measurements. Several popular statistical metrics are used to judge the accuracy of each algorithm. Our goal is to introduce a complete workflow that demonstrates the integrity of the steps followed and guides in further research in predicting similar PVT properties. The workflow showcases the advantages of combining engineers expertise to the art of data driven models development, specifically on accuracy and ease of implementation, as well as their limitations.