Interpretable Machine Learning Identifies Stage-Dependent Controls on Calcium Sulfate Scaling in Reverse Osmosis Membranes

Calcium sulfate scaling remains a major barrier to stable reverse osmosis (RO) operation. This study introduces an interpretable, stage-oriented framework that analyzes three complementary scaling endpoints rather than treating scaling as a single response. A literature-derived dataset of 284 observations from 40 studies was used to model induction time, surface scale coverage, and flux decline rate from eight water-chemistry and operating variables. Random forest (RF), gradient boosting regression (GBR), and extreme gradient boosting (XGBoost) were compared. XGBoost performed best across the reported metrics, although prediction quality differed among endpoints: surface scale coverage and induction time were more predictable than flux decline rate. SHapley Additive exPlanations (SHAP) and partial dependence analyses identified endpoint-specific model associations and found that ion concentrations were most strongly associated with deposition, pH and antiscalant dosage with scaling onset, and operating conditions with nonlinear hydrodynamic responses. Because the endpoints originated from different studies and the models have not been validated at full scale, the results do not establish temporal transitions or causal control mechanisms. Nevertheless, the framework provides an exploratory basis for stage-specific risk assessment and for prioritizing variables in future RO monitoring and validation studies.