Comparison of mechanistic and PLS-based regression models to predict radiocaesium distribution coefficients in soils

Radiocaesium sorption interaction descriptors were examined in 30 soil samples from Spain. Mechanistic and regression models were used to predict the solid-liquid distribution coefficients of radiocaesium (K(d)(Cs)) based on soil properties, and the obtained values were compared with the experimental ones, which were derived from batch experiments. The batch experiments used two contact solutions: one simulated the composition of the soil solution, and the other was the wash-off from the soil. Several mechanistic models of different complexity were tested based on the Radiocaesium Interception Potential (RIP), with satisfactory agreement between experimental and predicted values. A simplified model based on either the RIP, or the clay content and K status of the soil was proposed. Various multivariant regression models, which were constructed using the Partial Least Square Regression (PLS), were also evaluated. The RIP, clay content, and the K and NH(4)(+) contents were also identified by the regression models as the most relevant soil parameters to predict the K(d). As seen for the mechanistic models, the goodness of fit of the regression models was demonstrated by an excellent agreement between experimental and predicted values.