Stuck at Home: Machine‐Learning Models Predicting Solute Concentrations of One Stream Failed to Predict Solute Concentrations in Other Streams

ABSTRACT Machine‐learning models have been surprisingly successful at predicting stream solute concentrations, even for solutes without dedicated sensors. It would be extremely valuable if these models could predict solute concentrations in streams beyond the one in which they were trained. We assessed the generalisability of random forest models by training them in one or more streams and testing them in another. Models were made using grab sample and sensor data from 10 New Hampshire streams and rivers. As observed in previous studies, models trained in one stream were capable of accurately predicting solute concentrations in that stream. However, models trained on one stream produced inaccurate predictions of solute concentrations in other streams, with the exception of solutes measured by dedicated sensors (i.e., nitrate and dissolved organic carbon). Using data from multiple watersheds improved model results, but model performance was still worse than using the mean of the training dataset (Nash–Sutcliffe Efficiency < 0). Our results demonstrate that machine‐learning models thus far reliably predict solute concentrations only where trained, as differences in solute concentration patterns and sensor‐solute relationships limit their broader applicability.