Metamodeling of a physically based pesticide runoff model with a long short-term memory approach

• Physically based pesticide fate models are computationally demanding. • LSTM metamodeling of MHYDAS-Pesticide was tested for simulating pesticide runoff. • LSTM model accounts for weather, pesticide properties, application date and rate. • The metamodel predicts hourly discharges and concentrations at the annual scale. • The metamodel greatly reduces computation time, enabling large-scale applications. Predicting the temporal variability in pesticide runoff from agricultural fields is essential for assessing the impacts of agricultural practices on surface water quality and the associated environmental risks. Richards-based models are valuable tools for this purpose, particularly in contexts characterized by high temporal variability in rainfall intensity, such as in the Mediterranean region. These methods can be applied to a wide range of operational applications at both the field and catchment scales. However, their operational capability is limited, primarily due to their high computation times. In this study, we assessed a metamodeling approach for a Richards-based model designed to simulate the hourly variability in water and pesticide runoff over a year while reducing computation times. The proposed approach is based on long short-term memory (LSTM) models and is applied to a field-scale pesticide runoff model coupling Richards and convection-dispersion equations with the uniform mixing cell concept for pesticide remobilization and an overland flow routine. The resulting metamodel accounts for meteorological conditions, compound properties, and pesticide application dates and rates. Although the LSTM architecture is not optimized, the metamodel reduces the computation time by 90% compared with that of the initial model. The metamodel shows high overall performance in simulating hourly discharges and pesticide concentrations, with Nash–Sutcliffe efficiency values of 0.99, confirming the ability of the LSTM models to handle hydrochemical time series. The results nonetheless indicate room for improvement in specific areas, such as simulating runoff occurrence or concentrations for compounds with low half-life values. Furthermore, the results encourage further consideration of adding static parameters as inputs to LSTM models to increase their generalizability. Overall, this study provides key insights for deploying methods to assess the impacts of agricultural practices on a large scale