Leaderboard

Water models ranked by how close they get to the right answer, on tests where the answer is known. Every model runs the same inputs, and the suite scores them all the same way.

Updated: 18 Oct 2026
Models: 31
Tests: 31
Results: 238

Production software Research codes Worked examples Previous releases

By group

Each group is rated on its own tests. Select a group to open its board.

OpenHydro rating

Higher is better. Each dot is a model's rating and the band its 95 % interval; models whose bands overlap share a rank.

Full table

Select a column heading to sort. Group columns are the median percent difference from the right answer on that group's tests.

			95 % interval										Licence
1=	Basilisk Saint-Venant Basilisk (S. Popinet, CNRS)	1290	1190–1515	11	0.03 %	0.021 %	–	–	–	–	–	0.96 s	Open (GPL-3.0)
1=	ANUGA 4.0.1 Geoscience Australia and ANU	1219	1047–1402	15	0.11 %	0.02 %	–	–	–	–	–	27.6 s	Open (Apache-2.0)
1=	LISFLOOD-FP 8.3 · DG2 University of Bristol	1139	1066–1215	13	0.17 %	0.01 %	–	–	–	–	–	4.4 s	Open (GPL-3.0)
1=	LISFLOOD-FP 8.3 · FV1 University of Bristol	1139	1032–1264	13	0.17 %	0.02 %	–	–	–	–	–	4.6 s	Open (GPL-3.0)
1=	TELEMAC-2D openTELEMAC 9.1 openTELEMAC consortium	1128	1035–1244	15	0.11 %	0.034 %	–	–	–	–	–	25.7 s	Open (GPL-3.0)
2=	Hydra urban drainage NEER AI	1015	889–1150	13	49 %	–	<0.01 %	–	–	–	–	3.9 s	AGPL-3.0 or commercial
2=	LISFLOOD-FP 8.3 · ACC University of Bristol	987	843–1108	13	28 %	0.041 %	–	–	–	–	–	2.3 s	Open (GPL-3.0)
2=	HEC-RAS 7.0 1D · 2D SWE-ELM US Army Corps of Engineers (HEC)	961	732–1134	8	0.9 %	0.041 %	–	–	–	–	–	20.5 s	Free, closed source
2=	HEC-RAS 7.0 2D · SWE-EM US Army Corps of Engineers (HEC)	951	714–1188	7	8.4 %	0.044 %	–	–	–	–	–	24.6 s	Free, closed source
3	OpenSWMM 6 finite volume HydroCouple	876	672–1052	13	0.16 %	–	0.31 %	–	–	–	–	1.0 s	Open (MIT)
4=	OpenSWMM 6 dynamic wave HydroCouple	884	691–1037	13	55 %	–	0.027 %	–	–	–	–	4.1 s	Open (MIT)
4=	EPA SWMM 5.2.4 dynamic wave US EPA	880	700–1040	13	56 %	–	0.027 %	–	–	–	–	3.9 s	Open (public domain)
7	Landlab OverlandFlow Landlab (CSDMS)	792	640–885	9	28 %	0.58 %	–	–	–	–	–	1.3 s	Open (MIT)
–	GeoClaw Clawpack 5.9 Clawpack developers	–	fewer than 5 tests	4	0.092 %	0.19 %	–	–	–	–	–	2.1 s	Open (BSD-3)
–	HEC-RAS 7.0 pipe networks · 1D/2D US Army Corps of Engineers (HEC)	–	no test shared with the rated models	3	–	–	–	–	7.6 %	–	–	17.8 s	Free, closed source
–	OpenSWMM 6 coupled 1D/2D HydroCouple	–	no test shared with the rated models	3	–	–	–	–	0.79 %	–	–	3.5 s	Open (MIT)
–	OpenSWMM 6 coupled 1D/2D · cross mesh HydroCouple	–	no test shared with the rated models	3	–	–	–	–	10 %	–	–	8.1 s	Open (MIT)
–	EPANET 2.2 US EPA US EPA	–	no test shared with the rated models	2	–	–	–	0.6 %	–	–	–	0.56 s	Open (public domain)
–	Hydra water distribution NEER AI	–	no test shared with the rated models	2	–	–	–	0.34 %	–	–	–	0.53 s	AGPL-3.0 or commercial
–	HydroNet finite volume Penn State	–	fewer than 5 tests	2	32 %	–	–	–	–	–	–	1.9 s	Open (MIT)
–	HydroNet FVM-PINN Penn State	–	fewer than 5 tests	2	107 %	–	–	–	–	–	–	–	Open (MIT)
–	WNTR own hydraulic solver US EPA and Sandia	–	no test shared with the rated models	1	–	–	–	<0.01 %	–	–	–	0.07 s	Open (BSD-3)
–	TSNet water hammer · MOC University of Texas at Austin	–	no test shared with the rated models	0	–	–	–	0.13 %	–	–	–	–	Open (MIT)

Download this board as CSV or JSON, or every signed result as score cards.

Accuracy and run time

Rating against median run time on the finest grids (log scale). Top left is more accurate and faster. Times compare models on the same kind of machine only.

How the rating works

Every test is a match. On each test two models both ran, the one closer to the right answer wins; within 10 % of each other (or both under 0.01 %, or both inside the test's measured noise floor) is a draw. Nobody is compared on a test they did not run.
Each test counts the same. A test's matches share one unit of weight, whether 3 models ran it or 16; otherwise the most-run tests (the dam breaks) would decide the order.
One rating from all the matches. A Bradley–Terry fit, the method LMArena uses for chat models, on the Elo scale: 1000 is the average model, and 400 points more is ten times the odds of being closer on a test. Only models linked to each other through shared tests are rated together.
Physics only. The overall rating counts the Track A groups. Parity with a reference engine (Track B) can show a model is different, not wrong, and the Bellinge field test compares engines inside their own build-to-build noise, so both keep their own boards.
Runs the engine flags are not ranked. When an engine reports that its own run did not converge or lost mass (SWMM's report: over 10 % of steps unconverged or a 10 % continuity error), the result is shown with that report, not ranked.
An interval, not a decimal. The tests are redrawn with replacement 200 times and the rating refitted each time; the band covers 95 % of those fits. A model's rank is 1 + the number of models whose band lies wholly above its own.
Percent difference underneath. Each result is its percent difference from the right answer, the largest over the test's checks. The group columns and the grid below show it. Read the full method.
5 tests to be ranked. Models with fewer in the current board are listed without a rank.

Every result

Rows in rating order; each cell is the model's percent difference on that test. Select a cell to open the result.