Machine Learning-Based Surrogate Modeling for Urban Water Networks: Review and Future Research Directions

Surrogate models replace computationally expensive simulations of physically-based models to obtain accurate results at a fraction of the time. These surrogate models, also known as metamodels, have been employed for analysis, control, and optimization of water distribution and urban drainage systems. With the advent of machine learning (ML), water engineers have increasingly resorted to these data-driven techniques to develop metamodels of urban water networks (UWNs). In this article, we review 31 recent articles on ML-based metamodeling of UWNs to outline the state-of-the-art of the field, identify outstanding gaps, and propose future research directions. For each article, we critically examined the purpose of the metamodel, the metamodel characteristics, and the applied case study. The review shows that current metamodels suffer several drawbacks, including (a) the curse of dimensionality, hindering implementation for large case studies; (b) black-box deterministic nature, limiting explainability and applicability; and (c) rigid architecture, preventing generalization across multiple case studies. We argue that researchers should tackle these issues by resorting to recent advancements in ML concerning inductive biases, robustness, and transferability. Recently developed neural network architectures, which extend deep learning methods to graph data structures, are preferred candidates for advancing surrogate modeling in UWNs. Furthermore, we foresee increasing efforts for complex applications where metamodels may play a fundamental role, such as uncertainty analysis and multi-objective optimization. Lastly, the development and comparison of ML-based metamodels can benefit from the availability of new benchmark datasets for urban drainage systems and realistic complex networks.

Automated summary

Surrogate models, or metamodels, are increasingly being used in water engineering to replace computationally expensive simulations. However, current metamodels suffer from issues such as curse of dimensionality, lack of explainability, and rigid architecture. To tackle these issues, recent advancements in machine learning, such as inductive biases and robustness, should be applied. Additionally, neural network architectures that extend deep learning methods to graph data structures show promise in advancing surrogate modeling. Furthermore, future research should focus on complex applications and the development of benchmark datasets for realistic complex networks.