Development of hybrid gradient boosting decision tree learning algorithms for accurate prediction of crude oil and nitrogen interfacial tension

This study explains a highly accurate predictive model for the interfacial tension (IFT) between crude oil and nitrogen, an important parameter for nitrogen-based gas injection in petroleum reservoirs based on a dataset of 148 experimental points, of which 90% (133 points) were used for training the models and the remaining 10% (15 points) for testing their performance. We used a gradient boosting decision tree (GBDT) algorithm, which was optimized by applying four different methods: Gaussian Processes Optimization (GPO), Batch Bayesian Optimization (BBO), Evolutionary Strategies (ES), and Bayesian Probability Improvement (BPI). These models were trained in experimental data and evaluated their performance using statistical and graphical analyses. The data proved suitable for model construction. A sensitivity examination revealed that pressure, crude oil API gravity and temperature, all negatively impact IFT, while pressure is the most significant feature. Among the evaluated models, the GBDTBBO demonstrated superior accuracy. It demonstrated the highest R-squared values and lowest error rates, and it correctly predicted the trends of IFT in relation to changes in pressure, temperature, and crude oil API.