Physics-informed machine learning method for modelling transport of a conservative pollutant in surface water systems

Abstract

A reduced fresh water supply, e.g. as a consequence of climate change, can cause an increased salinization of (coastal) surface water systems and accurate, robust and computationally efficient models are needed to study these phenomena. While machine learning (ML) techniques are rapidly gaining popularity, their application is not straightforward and their ‘black-box’ structure is typically not interpretable, making the models unreliable for predictions under boundary conditions that differ from the training dataset. This paper presents a new ML model architecture, based on the LSTM-cell. The principle of mass conservation is enforced in the model’s architecture to introduce scientific knowledge with the goal of increasing the technique’s robustness. The model is compared with established reference neural networks for predicting salinization of navigable waterways in Belgium. Training data are generated with a detailed, physics-based solute transport model. The proposed model shows higher accuracy under new conditions when compared to the reference models.