A Novel Workflow of Near Real-Time Fluid Proper-Ties Prediction by Integrating Field Data, EOS Model, and Supervised Machine Learning

Abstract The study presents a near real-time workflow for accurately predicting reservoir fluid properties by integrating machine learn-ing (ML) with limited field data. The proposed approach could create significant business value through bypassing expensive and time-consuming workflow involving fluid sampling, transportation and storage, laboratory analysis, and Equation of State (EoS) modeling. This is particularly valuable for unconventional reservoirs, where instantaneous fluid property characterization is essential because of the limited peak production period. In such cases, acquiring high-quality PVT samples and laboratory data is often chal-lenging, and traditional black oil correlations (BOC) frequently fail to provide reliable predictions for volatile oils, gas condensates and near-critical fluids. Five available field data (production gas-oil ratio (GOR), stock tank oil gravity, reservoir temperature, field separa-tor temperature and pressure) were identified as ML model inputs, while outputs included saturation pressure (Psat), formation vol-ume factors (FVF), solution GOR (Rso), and viscosity at reservoir conditions during depletion. The existing PVT laboratory data was insufficient for training robust models. A novel approach was developed to expand the database by first generating synthetic fluids covering wide range of fluids observed in the target field and then incorporating them into well-calibrated EoS models to create an input-output database. The new database was expanded tenfold and used for training and validating the ML models. Multiple supervised ML models were developed and evaluated through comparison against classical BOC and actual PVT lab data. The ML models trained from the workflow were robust in predicting fluid properties over a wide GOR range (<100 to >10,000 scf/stb) with accuracy exceeding BOC. Compared to PVT laboratory data, the ML models predicted the properties within acceptable error margins. Given the high cost and complexity of traditional fluid properties prediction workflow, this approach represents a cost-effective and efficient alternative. The workflow was successfully implemented in an unconventional field to predict Psat, FVF, and Recovery Factor (RF) for more than 1000 wells in a few minutes. The workflow can be extended to any field practice, especially for the reservoirs where only limited and/or low-quality fluid samples are available for PVT analysis.