Mechanism-Embedded Machine Learning for Predicting Barite Induction Period Using Effective Interfacial Energy

Abstract Barite is a common inorganic scale in oilfield and is difficult to remove. Accurate prediction of the induction period is important for the safe and efficient operation of gathering and transportation systems. However, machine-learning-based prediction studies of the induction period of barite in systems containing scale inhibitors remain limited. In particular, few existing prediction methods explicitly incorporate information on nucleation mechanisms. Therefore, in this study, an effective interfacial energy descriptor was constructed based on Smoluchowski theory and introduced into the optimized support vector regression model to establish a mechanism-embedded model. The results showed that incorporating effective interfacial energy into the support vector regression (SVR) model improved its predictive performance. Specifically, the test-set R2 increased from 0.8364 to 0.857, while the RMSE decreased from 0.2853 to 0.2321 and the MAE decreased from 0.2082 to 0.1592. Meanwhile, the training-set R2 increased from 0.9286 to 0.949, while the RMSE decreased from 0.1768 to 0.1493 and the MAE decreased from 0.1054 to 0.0672. Further statistical analysis supported the rationality of using effective interfacial energy as an embedded descriptor and showed that it could provide complementary physical information for the model. These results indicate that the mechanism-embedded machine learning approach can achieve a favorable balance between prediction accuracy and result interpretability within the scope of the present dataset, providing a methodological reference for barite risk assessment, scale inhibitor screening, and dosage strategy optimization under complex water chemistry conditions.