Response of microbial nitrogen removal to sinuosity in river bends: mechanisms and development of physics-informed neural networks model.

Understanding the underlying mechanism of biogeochemical processes in river bends is important for decisions on sinuosity in river construction to improve the self-purification capability. Complex flow deeply modified by sinuosity in river bends creates diverse habitats, making biogeochemical processes spatially heterogeneous and predictive model development challenging. Here, we investigated for the first time the response mechanisms of microbial communities of rivers in affecting nitrogen removal to sinuosity from the perspectives of functional genes and enzymes through an indoor experiment. It was found that sinuosity affected nitrogen removal through environmental heterogeneity (p < 0.05), which was governed by the hydrodynamics of river bends. The environmental heterogeneity, which meant suitable habitats for microbial taxa that drove different nitrogen transformation processes, promoted collaboration among microbial taxa performing different nitrogen transformation functions (narG, nirK/S, and hzo genes correlated significantly, p < 0.05). The structural equation model revealed that the nitrogen removal rate was determined not only by the abundance of microbial taxa carrying nitrogen transformation genes but also by the effect of environmental factors on gene expression (synthesis of nitrogen transformation enzymes). A hybrid model consisting of physics constraints and a neural network (physics-informed neural network, PINN) was employed to predict nitrogen removal. The neural network comprises a convolutional neural network model (CNN, which could retain more spatial distribution information) and a multilayer perceptron (MLP, which could capture nonlinear relations efficiently). The root mean square error value of the PINN model on the training set was 0.66 and on the test set was 0.67, demonstrating the feasibility of using the PINN to predict the nitrogen removal in river bends. Our study revealed the underlying mechanism of microbial nitrogen removal in river bends. The developed model could be applied to the decision-making regarding the sinuosity of river construction and could also serve as a reference for system prediction involving spatial distribution data.