Software

The group turns its research on stormwater modelling into open-source software that anyone can run. Two projects, both developed in the group by Ph.D. candidate Zhonghao Zhang, are in active use: Agentic SWMM, an agentic runtime that drives EPA SWMM end to end with an audit trail, and SWMMCanada, a web service that builds a ready-to-run SWMM model from a boundary drawn anywhere in Canada. Together they close a loop from open data to model to calibrated results with uncertainty.

Agentic runtime for EPA SWMM · stable v0.9.3 · MIT licence

Agentic SWMM

Agentic SWMM is an open-source, verification-first framework for reproducible stormwater modelling with EPA SWMM. Its runtime, aiswmm, lets a modeller describe a goal in natural language while model execution stays deterministic: GIS preprocessing, model generation, SWMM runs, QA checks, plots, calibration, and climate-scenario batches each leave an inspectable artifact with provenance. Audited runs feed a modelling memory that can propose refinements to the workflow's Skills, which remain under human review. The Skills and MCP servers also work with Codex, Claude Code, OpenClaw, and Hermes.

One English sentence can fetch the real downtown Victoria storm network from SWMMCanada, run SWMM, audit the result, screen it against a design rulebook, plot the hydrograph, and export a Word report. The Greenwich Peninsula case can be replayed in the browser without installing anything.

Status

Stable v0.9.3 (August 2026); pip install aiswmm, one-line installers, or a pinned Docker image

Paper

Zhang & Valeo (2026), AI for Engineering 1(1), 5. doi:10.3390/aieng1010005

Preprint

Agentic Modelling Pipeline with OpenClaw, EarthArXiv (March 2026). doi:10.31223/X5F47G

Runs with

Its own aiswmm runtime, or Codex, Claude Code, OpenClaw, and Hermes through Agent Skills and MCP

Try the live demo Project website Source on GitHub Read the paper

Ready-to-run SWMM models from Canadian open data · v0.5.0 · MIT licence

SWMMCanada

SWMMCanada turns a boundary drawn or uploaded anywhere in Canada into a complete, ready-to-run EPA SWMM model. It pulls the Canadian open data for that area (ECCC rainfall and design-storm intensities, NRCan terrain including 1 m LiDAR where available, land cover, soil, and the city's published storm and sanitary pipes) and assembles a model.inp together with a shareable datastore package. Where a city publishes its storm network (35 cities, from Victoria and Vancouver to Ottawa and Moncton), the model uses the real pipes; elsewhere it synthesises a network from the street map with pipes sized by the rational method and real ECCC IDF intensities.

Checked against a real gauge: an uncalibrated model of the 22 km² Graham Creek basin in Ottawa, built automatically over the official drainage polygon, reproduced all twelve summer-2024 rainfall events on the correct days against the Water Survey of Canada record. Models are first-pass research baselines from open data; SWMMCanada is the upstream model builder for Agentic SWMM, which runs, calibrates, and audits them.

Status

v0.5.0 (July 2026); hosted beta at swmm.h2ox.me for small areas, or self-host with Docker

Coverage

Real municipal storm networks for 35 cities (34 Canadian municipalities plus Reykjavík); synthesised networks anywhere else in Canada

Exports

EPA SWMM 5.2, plus DHI MIKE+, InfoWorks ICM, and HEC-RAS import packages from the same datastore

Preprint

Zhang (2026), EarthArXiv. doi:10.31223/X5NR31; software archived on Zenodo, doi:10.5281/zenodo.21058544

Open the app Project website Source on GitHub Read the preprint

How the pieces fit together

SWMMCanada is the first half of a loop: it turns an area into a complete, runnable model. Agentic SWMM is the second half: it runs the model, calibrates it against observations, post-processes the results, and quantifies uncertainty, with every step recorded. Inside Canada, Agentic SWMM fetches SWMMCanada models directly; outside Canada it can synthesise a baseline network with SWMManywhere.

The same pattern is being extended to other engines. Sibling projects, at an earlier stage, include Agentic MIKE+ for headless DHI MIKE+ automation, Agentic HEC-RAS for HEC-RAS, and the Agentic Hydrology Platform, which adds an LSTM streamflow engine and orchestrates the engines under one natural-language interface.

Using and contributing

Both projects are MIT-licensed. Contributions are welcome in additional case studies, calibration and validation workflows, new data sources and MCP tools, QA testing, and tutorials. Questions about the software can go to zhonghaoz@uvic.ca or valeo@uvic.ca. If you use the tools in your work, please cite the paper or preprint listed above.