AI Models for Water Networks

# AI Models for Water Networks ## Project details **Applicant:** Alexander Garzón Díaz **Faculty:** CEG **DCC members:** Jose Carlos Urra, Elviss Dvinskis **Support period:** Feb 2024 - Sep 2024 **Repository/Archive:** - [A reference commit indicating the repository's state before the support process.](https://github.com/alextremo0205/SWMM_GNN_Metamodel/commit/b22fd30e47499acab028023c3d3212624a42b4c7) - [Private development repository](https://github.com/alextremo0205/SWMM_GNN_Metamodel) - [Public repository (linked with paper)](https://github.com/alextremo0205/SWMM_GNN_Repository_Paper_version) - [Dataset in 4TU containing the data collected during the development](https://doi.org/10.4121/fec1e3de-9586-4a61-b3a1-02382592e52c) - [Code for the SWMM GNN metamodel in 4TU](https://doi.org/10.4121/989a0d3d-3b4d-47c7-8677-31c5975f9dec) **Research Background**: The project focuses on storm water systems infrastructure. More specifically, on transfarable and data efficient metamodelling of the storm water system nodal depths by using auto-regressive graph neural networks. Typically, utility companies depend on computer simulators to design, manage, and operate these storm water systems, but they are time consuming. Researchers have turned to more cost-effective models, such as meta-models, as alternatives to computationally expensive models. Given the recent rise in artificial intelligence applications, machine learning has emerged as an important approach for meta-modeling urban water networks. **Project Description:** This project applies inductive biases and transfer learning to develop stormwater system metamodels that require minimal data while maintaining high performance when applied in different contexts. **Project goals** When the support was requested, the researcher was struggling with correctly versioning the development pipeline. Additional support was needed for implementing tests, establishing proper workflows, and improving readability and documentation. During the support process, we defined and scoped our support in a tangible manner. ## Project Results In this project, we provided knowledge transfer on best practices for software development, licensing, publishing, and archiving, as well as guidance on suitable data management solutions. This included test-driven refactoring, reproducibility checks of the experimental setup, and environment and dependency management. We also provided guidance on restructuring components of the project and implementing data versioning control. Additionally, we offered advice on version control, improving documentation, and opening well-documented issues and pull requests. The code and data were published and archived in 4TU. ### Lessons learned >Research software should be focused on science and publications of code as a complementary artifact of research output by default. >Only when strictly necessary research software projects should adopt a more conventional software engineering lifecycle.