Machine Learning for Magnetite Nanoparticles Contaminant Adsorption under Uncertainty: Revealing Data Quality Limits
Abstract
Despite the rapid adoption of machine learning (ML) in materials discovery, its application to water contaminant adsorption remains fundamentally constrained. In this study, we systematically evaluated whether current peer-reviewed literature on magnetite nanoparticle (MNPs) adsorbents contains sufficient descriptor information to support ML-based predictions of Langmuir adsorption capacity (qm). A structured data set compiled from 50 peer-reviewed studies yielded 137 observations, of which only ∼37% reported adsorption capacity (qm). Critical descriptors were severely underreported: Brunauer–Emmett–Teller (BET) surface area (23%), particle size (47%), and magnetization (14%). Under leave-one-out cross-validation (LOOCV), a complete-case subset with BET surface area and adsorbent dose (N = 18) achieved convergent, significant predictive performance across all model classes (R2 = 0.71–0.72). Without BET surface area, linear models failed regardless of sample size, though Random Forest, a nonlinear model, retained significant signal from operational variables alone (R2 = 0.61), indicating that BET is necessary for model-agnostic prediction rather than all predictive signal. A supplementary imputation-based sensitivity analysis did not recover the performance achieved with the directly measured incomplete descriptors. Hence, descriptor completeness, not algorithm selection, is the primary limitation in ML-driven MNPs adsorbent design, and standardized reporting practices are a prerequisite for realizing ML’s predictive potential in water treatment.