A surrogate model for the variable infiltration capacity model using physics-informed machine learning

Abstract

In this study, a physics-informed machine learning-based surrogate model (SM) for the variable infiltration capacity (VIC) model was developed to improve simulation efficiency in the Yarlung Tsangpo River basin. The approach combines the empirical orthogonal function decomposition of low-fidelity VIC models to extract spatial and temporal features, with machine learning techniques applied to refine temporal feature series. This allows for accurate reconstruction of high-fidelity spatial simulations from the results of the low-fidelity model. Using the SM built from the 1.0°-resolution VIC model as an example, the study highlights the challenges and solutions associated with low-fidelity simulations. The SM significantly improves accuracy, achieving a Kling–Gupta efficiency of 0.88, an Nash–Sutcliffe efficiency of 0.97, and a PBIAS value of-6.21% with reduced computational demands. Additionally, different machine learning methods impact the performance of the SM, with the support vector machine regression model performing best in these methods. SMs from varying low-fidelity resolutions maintain similar accuracy, but higher resolutions notably enhance computational efficiency, reducing time by 86.31% when compared to the high-fidelity VIC model. These findings demonstrate the potential of the SM to enhance VIC model simulations while reducing computational requirements.