Grey Stephen Nearing, Alden Keefe Sampson, Frederik Kratzert, Jonathan Frame 

Machine learning is becoming an increasingly important part of streamflow forecasting, but as these models to date lack a physical basis, there is a potential that they may produce values that are not realistic. We tested a simple post-processing strategy that uses the outputs from a calibrated conceptual model (the Sacramento Soil Moisture Accounting Model with Snow-17; SAC-SMA) as inputs into a a Long Short Term Memory Network (LSTM). Overall, the SAC-SMA model was improved substantially, while post-processing offered only minor improvements relative to the standalone LSTM. SAC-SMA performance was improved in catchments with more snow. The standalone LSTM was improved in terms of long-term bias, which is likely because the LSTM is not constrained by conservation principles.