Prediction of asphaltene onset pressure and precipitation amount using tree-based stacking–blending ensemble meta-learners: a flow assurance perspective

This is the authors' abstract. We don't add key points for this paper.

Flow assurance is critical to reliable oil production, and asphaltene precipitation is among its costliest threats, damaging the near-wellbore region, cutting output, and forcing expensive remediation. Equation-of-state models describe it rigorously but need detailed fluid characterization and heavy computation. This study aims to predict asphaltene onset pressure (AOP) and precipitation amount, as two separate regression tasks, from routinely available reservoir data. We compiled 483 AOP and 408 precipitation measurements from the literature, trained XGBoost, Random Forest, LightGBM, and Extra Trees, and combined them through stacking and blending with a Huber meta-learner. Tuning used 5-fold cross-validation on training data only, and all metrics come from a held-out test set. An ensemble performed best in each task: Stacking-Huber for AOP (R2 = 0.991, MAPE = 1.65%) and Blending-Huber for precipitation amount (R2 = 0.988, MAE = 0.0298), though margins over the best base learners were small. C1 and C5 mattered most for AOP, and pressure and bubble point pressure for precipitation amount, in line with thermodynamic expectations. These findings support rapid screening of asphaltene risk when detailed equation-of-state characterization is unavailable and can inform inhibitor design. When coupled with the pressure and temperature profile along the wellbore, the models can be used to estimate the depth at which precipitation begins and the amount of asphaltene precipitated, within the compositional range of the training data. The main challenge is generalization: records from one source may fall in both training and test sets, so scores reflect the compiled data, not unseen oils. Future work should use source-wise validation on independent datasets, cover gas-injection conditions, and embed thermodynamic constraints through physics-informed models.