Forecast 2026–2050 · dream time 2075–3000
Water Modeling to 3000
Through 2050 this is a forecast, and confidence falls at every step. After that it is dream time. One check runs through all of it, and it starts with conservation.
- 2026nowToday
Where it stands
- SWMM6 is an alpha. A stable release is targeted for 2027; that is a target, not a commitment.
- The engine is thread-safe and handle-based, with pip-installable Python bindings and an MCP server for agents.
- 2D finite-volume overland flow has shipped in the 6.0 alphas. 1D finite-volume routing is in the development code, not yet in a tagged release.
- SWMM5 is still the version consultants and regulators actually run.
CheckContinuity error in every .rpt is the first gate. Peaks, timing and depths still have to match observations.
Evidence for this baseline
- 2030+4 yrForecast
Boring, in the good sense
- 6.x is stable, and parity with SWMM5 is shown on a public corpus with a published register of expected differences.
- Classic dynamic wave and finite volume ship side by side, and old .inp files still load.
- 1D/2D urban flooding on a GPU is routine. Full shallow-water codes like ANUGA couple in through HydroCouple when convective acceleration matters.
- Agents run, diff and diagnose models as a normal workflow.
- Still open: whether it becomes the official EPA version, and how FEMA treats it.
CheckCompatibility tests against SWMM5, analytical benchmarks, and a register of every documented difference.
Where today’s code already points
Starting point in the code now
- Released: the explicit local-inertial 2D finite-volume solver (6.0.0-alpha.1), with OpenMP, CUDA, HIP and SYCL backends as plugins (alpha.2).
- Released: Python bindings (pip install openswmm) and the MCP server, since 6.0.0-alpha.1.
- Development code only: FLOW_ROUTING FV sits beside classic dynamic wave in the develop line, marked unreleased. It runs CPU-serial; no GPU plugin serves the 1D solver yet.
Extrapolation on this page
- A stable 6.x release, a public parity corpus with a difference register, routine GPU use in practice, ANUGA coupling, and agent workflows as normal practice.
What it depends on · what could delay it
Depends on
- EPA or a funded consortium carrying SWMM6 to a stable release.
- A public test corpus that both engines run, with results anyone can reproduce.
- Enough core developers to maintain 5.x and 6.x in parallel.
Could delay it
- The 2027 stable target slipping.
- Corpus disagreements that nobody is funded to resolve.
- Regulators waiting for an EPA endorsement before accepting results.
These are judgments, not dated milestones. No probability is attached on purpose.
- 2035+9 yrReasoned guess
A platform
- Uncertainty bands are a default output, and calibration is gradient-based or Bayesian.
- Sensor-fed ensembles run faster than real time, so digital twins become ordinary.
- The .inp becomes an export format. The model lives in a versioned store with provenance.
- Water quality, LID and groundwater, already in the engine, become swappable modules with interchangeable formulations, each validated on its own.
- SWMM5 is frozen and kept as the regulatory reference while SWMM6 earns acceptance.
Historical note from Robert Dickinson, firsthand: during the 1990s, SWMM was supported outside USEPA. The SWMM5 line above is a forecast; this is the precedent it echoes.
CheckEach learned surrogate carries error bounds, a tested operating range and a fallback to the engine.
Where today’s code already points
Starting point in the code now
- In the code now: a header-only Plugin SDK for input, output and report plugins, full SWMM 5.x LID controls, and a two-zone groundwater model (released in 6.0.0-alpha.1).
- Development code only: finite-volume water-quality transport works at scheme level but is unreleased, and is not yet connected to pollutants, buildup and washoff, inflows or treatment.
- The project roadmap lists groundwater on the 2D mesh, a multispecies quality engine, heat transport and LID as storage nodes. None of these carry dates.
Extrapolation on this page
- Uncertainty bands by default, sensor-fed ensembles, a versioned model store, learned surrogates, and swappable formulations validated one by one.
What it depends on · what could delay it
Depends on
- Real-time sensor feeds that utilities actually maintain.
- An agreed format for versioned model storage.
- Published error bounds for every surrogate in use.
Could delay it
- Data ownership and cybersecurity rules for live utility feeds.
- Procurement cycles that still buy desktop licences.
- A surrogate failing badly in a high-profile flood.
These are judgments, not dated milestones. No probability is attached on purpose.
- 2040+14 yrSpeculation
A kernel
- Nobody runs SWMM. People ask questions, and agents build the model from GIS, LiDAR and as-builts, then keep it calibrated.
- Century-long, city-scale continuous simulation across climate ensembles is the default.
- Learned closures sit inside the physics core where they have earned a certificate.
- SWMM6, SWMM5+, Rust rewrites and vendor forks all count as SWMM if they conform to the written specification the suite tests.
CheckConservation, accuracy and robustness are tested separately. Passing proves conformance to a spec, nothing more.
What it depends on · what could delay it
Depends on
- A written specification of what SWMM computes, not only a reference code.
- Trust in agent-built models, earned through audited track records.
Could delay it
- Liability rules that still require a named engineer to build the model by hand.
- GIS and as-built data too poor for agents to work from.
These are judgments, not dated milestones. No probability is attached on purpose.
- 2050+24 yrSpeculation
The twin is infrastructure
- A live twin is city infrastructure, owned, budgeted and maintained like the pipes it describes.
- Permits are checked against the certified twin instead of a stamped design report.
- Design storms come from climate ensembles refreshed every year, not from a table printed decades earlier.
- Modelers spend their time on questions and judgment rather than on building input files.
CheckTwins are scored every year against independent observations, including whether the uncertainty bands cover them.
What it depends on · what could delay it
Depends on
- Cities budgeting twins as infrastructure rather than one-off studies.
- Permitting rules that accept a certified twin.
- Long observation records to score against.
Could delay it
- Funding cycles that treat models as projects that end.
- No independent body to certify twins.
These are judgments, not dated milestones. No probability is attached on purpose.
- 2075+49 yrDream time
The network computes itself
- Every pipe, inlet and curb carries its own sensor and its own small solver.
- The simulation never starts or stops. It runs a few minutes ahead of the water.
- Calibration stops being a job, because model and system never drift apart.
- Design storms give way to designing against every plausible century at once.
CheckMass balance closes at every sensor node, and every forecast is scored against what the water then did.
- 2100+74 yrDream time
The programmable catchment
- Pavement changes its permeability on command, and storage moves to where the rain will fall.
- Rain itself gets nudged, block by block.
- The model stops predicting and becomes the controller.
- Every drop has provenance from cloud to sea. Where rivers hold legal standing, the model is how they testify.
CheckNo control action runs until its mass balance and its downstream effects are audited.
- 2200+174 yrDream time
One water, any gravity
- There is no 1D, 2D or 3D. Resolution follows the physics from pore to ocean basin in one continuous simulation.
- Watersheds get modeled before they exist, on Mars and in orbit.
- In spin-gravity habitats, downhill curves and Coriolis bends every gutter.
CheckOne continuity check at every scale, pore to basin.
- 2500+474 yrDream time
Hydrology runs backwards
- Describe the river you want: its sound, its fish, how it floods every tenth spring.
- The inverse model returns the terrain, soil and climate that produce it.
- Any storm in history can be replayed on any city of any era.
CheckInverse designs that break conservation are rejected before anyone builds them.
- 3000+974 yrDream time
No floods, only choreography
- Model and water are the same object.
- Nobody remembers when there was a difference.
CheckSomeone still checks the continuity error.
Engines come and go. Invariants travel.
Conservation of mass and momentum is where every check starts, not where it ends. A model can close its mass balance and still miss the peak, the timing or the flood extent. What reaches the year 3000 is the habit of checking.What sets the pace of the forecast
- EPAWhether SWMM6 becomes the official EPA version, and on what schedule.
- PeopleHow few core developers carry the engine, the docs and the tests.
- FundingWho pays for maintenance, verification and release work after the excitement fades.
- The standardWhether a public test corpus and a register of expected differences become the shared definition of SWMM. This is the piece the 2030 test suite can define.
X thread
9 posts
Each post carries its tier, so a copied post keeps its forecast or dream-time label.
Print or save as PDF
Your browser's own Print command prints the full timeline, with the thread only when its section is open.