Predicting filtration coefficient and formation damage coefficient for particle flow in porous media using machine learning

• The filtration coefficient is first estimated using machine learning methods. • A power-law relationship is established among the filtration coefficient and velocity, concentration, and pore size/grain size. • The formation damage coefficient is found to have a power-law relationship with the critical porosity fraction. • The XGBoost showed the best performance in predicting the filtration coefficient and the critical porosity fraction. The clogging of porous media with solid particle suspension flow is modeled using two empirical parameters of filtration coefficient ( λ ) and formation damage coefficient ( β ). These parameters are typically determined through coreflood tests. This study employs machine learning techniques to predict λ and β using experimental data from open literature. The prediction of β is based on critical porosity fraction ( γ ) data and a power law equation relating β and γ . Collected data were randomly partitioned into training (80%) and testing (20%) subsets. Four regression algorithms were employed, treating λ or γ as the target variable, with injection velocity ( u m ), particle concentration ( C in ), and ratio of mean pore size ( D pore ) to mean particle size ( D p ) as features. The extreme gradient boosting (XGBoost) algorithm showed the best performance. The feature C in had the highest influence on λ and γ , revealing a significant finding previously overlooked. Postmortem analyses revealed qualitative consistencies in λ results, supporting the existence of critical velocities. Furthermore, λ results showed a power law relation between λ and all three features used. An equation was formulated to estimate λ as a function of these three features. A direct prediction of β using these features was established by applying the XGBoost model to predict γ and then employing an existing power law relationship between β and γ . This study demonstrated that machine learning offers an alternative approach for predicting λ and β , which is particularly useful for initial evaluations of clogging potentials and identification of experimental conditions to focus on.