Machine\nLearning Models to Predict Early Breakthrough\nof Recalcitrant Organic Micropollutants in Granular Activated Carbon\nAdsorbers

Granular activated carbon (GAC) adsorption\nis frequently used to\nremove recalcitrant organic micropollutants (MPs) from water. The\noverarching aim of this research was to develop machine learning (ML)\nmodels to predict GAC performance from adsorbent, adsorbate, and background\nwater matrix properties. For model calibration, MP breakthrough curves\nwere compiled and analyzed to determine the bed volumes of water that\ncan be treated until MP breakthrough reaches ten percent of the influent\nMP concentration (BV10). Over 400 data points were split into training,\nvalidation, and testing sets. Seventeen variables describing MP, background\nwater matrix, and GAC properties were explored in ML models to predict\nlog<sub>10</sub>-transformed BV10 values. Using the ML models on the\ntesting set, predicted BV10 values exhibited mean absolute errors\nof ∼0.12 log units and were highly correlated with experimentally\ndetermined values (<i>R</i><sup>2</sup> ≥ 0.88).\nThe top three drivers influencing BV10 predictions were the air-hexadecane\npartition coefficient and hydrogen bond acidity (Abraham parameters <i>L</i> and <i>A</i>) of the MPs and the dissolved organic\ncarbon concentration of the GAC influent water. The model can be used\nto rapidly estimate the GAC bed life, select effective GAC products\nfor a given treatment scenario, and explore the suitability of GAC\ntreatment for remediating emerging MPs.