Desenvolvimento de modelos de aprendizado de máquina para predição de estabilidade, viscosidade e tamanho médio de gotas de emulsões água em óleo

Abstract

Emulsions are essential in the petroleum industry, influencing phase separation and fluid transport processes. This study evaluated the effectiveness of machine learning (ML) models in predicting the influence of key parameters on emulsion stability, apparent viscosity, and mean droplet size (MDS). A two-step methodology was employed. Initially, experiments using thirteen (13) different oils generated a comprehensive dataset by varying water cut, temperature, and salinity concentration through a Latin Hypercube Sampling (LHS) design to analyze their effects on emulsion properties. Subsequently, the experimental data were used to train and test ML models for both classification (Decision Tree – DT, Gradient Boosting – GB, Random Forest – RF, and Multi-Layer Perceptron – MLP) and regression tasks (Polynomial Regression – PR, eXtreme Gradient Boosting – XGBoost, RF, and MLP), with hyperparameters optimized using the Particle Swarm Optimization (PSO) algorithm. SHapley Additive exPlanations (SHAP) analysis was applied to quantify the impact of input parameters on the model outcomes, revealing their relative contributions to the predictions. The ML models accurately predicted emulsion stability, with the GB model achieving the best performance, reaching an accuracy of 0.938. For mean droplet size, the XGBoost model showed the best predictive performance, achieving an R² of 0.924. For viscosity, XGBoost also stood out, reaching an R² of 0.992. The models were able to capture complex phase separation patterns and demonstrated strong correlations between predicted and experimental values for both apparent viscosity and droplet size. SHAP-based parametric analysis revealed that water content (%) was the most influential factor in predicting emulsion stability, while °API emerged as the key variable for both viscosity and mean droplet size.