Physics Augmented Machine Learning Models for Determining Gas Solubility in Formation Brines for CCS and Gas Processing Applications

Abstract Gas solubility in brine is crucial input for engineering design of various chemical/petroleum processes such as oil recovery, CO2 sequestration in saline aquifers and water bearing formations, separation and utilization, corrosion in wellbores/facility/pipelines with acidic gases, and gas processing. However, with limited experimental data, intrinsic trends of gas solubility under varying conditions (pressure/ temperature/brine composition) are difficult to obtain. This study uses machine-learning-based approach to explore those intrinsic trends while identifying the contributing elements correlated with fluid properties. Gas solubility in brine depends on various thermodynamic properties of the components, pressure, temperature, and salinity. To explore effect of these factors, we employ various Machine learning (ML) tools: decision tree (DT), random forest (RF) and artificial neural network (ANN) techniques due to their stability and convergence characteristics coherent with the data utilized and develop a framework to determine solubility of a gas in brine solutions. While most existing literature are limited to very few salts (NaCl/KCl/CaCl2), this work captures combinations of various common salts (chlorides, carbonates/bicarbonates, and sulphates) as they are seen in real formation brines and water utilities. The prediction from ML models were validated against the available experimental data that were not used in training. The main results are as follows: Validation processes indicated that ML models predict the experimental trends accurately, within the relative error of 1% for gas-water systems and 3% for complex gas-brine systems. Various input features based on the thermodynamic and physical properties of gases and ions (cations and anions) were considered and main contributing features were identified. Most importantly, the framework is general, fast, convenient and can easily be extended for different gas species including greenhouse or hydrocarbon gases, as well as for variety of salts. Additionally, it can fill the gaps in experimental data for the gas-brine systems, and extrapolate to elevated pressure and temperature conditions. While ML-based approach to estimate gas solubility in brines have been developed in the literature, they are very restrictive in terms of their broadness/applicability and valid for specific gases such as CO2 as well as few salts (NaCl/KCl/CaCl2). Here, the ML-based framework covers a wide range of salts and gases, and additionally, the current framework can easily be extended to other systems.