Can machine learning and traditional hydrology models work together, rather than in opposition? This preprint, co-authored by Upstream Tech's Alden Keefe Sampson, explores a practical hybrid approach: using the outputs of a calibrated conceptual model as inputs to an LSTM.
The team tested this post-processing strategy using the Sacramento Soil Moisture Accounting Model with Snow-17 (SAC-SMA) as the upstream model, feeding its outputs into an LSTM trained across the CAMELS benchmark basins. The results were nuanced: the SAC-SMA model improved substantially when post-processed by the LSTM, particularly in snow-dominated catchments. However, the standalone LSTM still performed comparably or better overall — and the post-processing approach offered modest benefits primarily around long-term bias correction, since LSTM models are not inherently constrained by water conservation principles.
The paper is a thoughtful examination of where hybrid physics-ML approaches add value, and a useful reference point for understanding the relative strengths of data-driven and process-based modeling in streamflow forecasting.
