An EPANET metamodel based on Simplicial Convolutional Networks

Metamodels reproduce the response surface of physics-based models while significantly reducing simulation times. Such techniques are widely employed in water distribution system analysis since they enable the application of computationally expensive methods in designing, controlling, and optimizing water networks. Recent works proposed graph neural networks as candidates for metamodels. These models bear inductive biases as one can draw analogies between links and nodes in the graph with the pipes and junctions. This implies that new metamodels using this approach can be applied to an unseen water network topology without re-training. However, there is no evidence of the transferability properties of those metamodels so far. This work introduces Simplicial Convolutional Networks (SCNs), which offer the potential of developing transferable metamodels that can generalize across different systems.  We test the suitability of SCNs to estimate pipe flowrates and nodal pressures emulating steady-state EPANET simulations. We compare the accuracy of SCN metamodels against graph neural networks on several benchmark water networks available in the literature. Moreover, we show that SCNs are able to generalize better than graph neural networks by evaluating and measuring the performance of the metamodel in an unseen setting.