Hydrological Processes

Graphical Abstract We test a hypothesis that mass conservation constraints restrict a model's ability to compensate for disinformation from input data. Our results are presented generally in terms of constraints enforced on deep learning (DL) and conceptual model architecture. Our findings demonstrate: Conservation may not be a good foundation for watershed scale hydrological theory. Disinformative data is not generally a major source of modelling error. DL models compensate for systematic biases in the input data on a per-event basis.