flood-lab.github.io/HydroTuring · English, Español, 中文

A benchmark that asks one question of any AI hydrologic model: does it conserve what physics says it must conserve?

Not whether it fits a hydrograph. Whether its water budget closes, its energy budget closes, and its routing conserves momentum. A model passes HydroTuring only when every criterion of every probe passes.

Bring a probe or a model, join the paper. The suite is only as good as the physics people bring to it, and only as interesting as what has been put through it. Anyone may propose either — you do not need to be invited, affiliated, or known to us.

One merged probe earns co-authorship on the HydroTuring paper. So do five accepted model proposals. You do not have to be able to package a model to propose it; say so on the form and the work is assigned to someone who can, and it still counts as yours.

Start with a probe or a model, or come talk it over on Discord first. Terms in CONTRIBUTING.md.

$ ht run --model reference_bucket
reference_bucket v1.0.0  ->  ✅ PASS (OK)  [1/1 probes passed]
  ✅ PASS  mass/catchment-closure
        ✅  closure            cumulative residual 0.0000% of sum_pr (limit 5.0%)
        ✅  state_bounds       all storages stay physical
        ✅  et_plausible       cumulative ET is 0.664 of potential ET (limit 1)
        ✅  non_degenerate     partition and variability are non-trivial
        ✅  forcing_fidelity   reported forcing matches the input

The idea

A benchmark that only checks closure is trivially gamed. A set of deliberately broken models lives in this repository to prove it, and no probe is merged until it passes an exactly conservative model and catches the broken ones it names.

reference_cheater is the one worth dwelling on. Its closure residual is exactly zero on every seed, forever. Randomising the forcing cannot touch it, because the cheat is in its internal wiring rather than its memory. It is caught only because the model contract requires absolute storage states rather than tendencies, so the storage it invents has to stay physical, and it does not.

The probes

Twenty-two: sixteen under mass, five under energy and one under momentum. Each was merged only after the acceptance gate saw it pass its declared exact reference and fail a purpose-built broken one on the named criterion. Four physical models, a bucket that conserves water exactly, two hand-written FLEX models and the NWS's SAC-SMA with Snow-17, must pass every probe that can ask them anything; four of the five energy probes need outputs they do not report and are not scored for them. A probe that fails a physical model is examined before the model is; that is the first thing done with any probe pull request. Eleven of the twenty-two can be scored on a model that reports runoff and nothing else. ht list prints them; ROADMAP.md has the ten more we want, all unclaimed.

Probe	Law	What it asks	The broken model it catches
mass/catchment-closure	mass	Does the water budget close over ten generated years?	reference_leaky, reference_cheater, reference_degenerate
mass/resolution-invariance	mass	The same month at the minute, the hour and the day: do the integrated volumes agree?	reference_fixed_step, reference_degenerate
mass/warming-response	mass	The same rain with the air 3 °C warmer and 3 °C cooler: does runoff move the way physics says, in both directions?	reference_degenerate, reference_streamflow_only
mass/causality	mass	One storm added mid-record: nothing may change before it, and runoff must answer after it.	reference_anticipating
mass/dry-down	mass	Two years without rain: runoff can only fall, and no more may drain than the catchment held.	reference_climatology
mass/steady-state	mass	Three years of the same day: does everything settle, and does the budget balance once it has?	reference_restless
mass/extreme-rain	mass	The largest storm scaled up to ten times: runoff may not fall, nor exceed the rain that was added.	reference_saturating
mass/extreme-event-closure	mass	Overlap rainfall events toward fitted 100-year depths in one median-wet year of twenty; check each event's water budget so long-record averaging cannot hide a loss.	reference_in_sample
mass/runoff-bounds	mass	Over ten years, is the runoff possible at all: at least rain minus demand minus storage, at most rain plus storage? The mass question a runoff-only model has to answer.	reference_degenerate, reference_overflowing
mass/area-invariance	mass	The same weather on the same catchment told as ten times larger: every depth must be identical.	reference_area_leak
mass/response-nonnegativity	mass	One 120 mm storm added: from that day on, runoff may never be lower than without it.	reference_overshooting
mass/antecedent-monotonicity	mass	The same storm after a dry month and a wet one: the wetter catchment runs off more, and no more than the extra water.	reference_cheater
mass/phase-counterfactual	mass	The same water falling as rain instead of snow: timing moves, the integrated volumes may not.	reference_sublimating
mass/time-origin-invariance	mass	The same weather under a 28-year calendar shift that preserves seasons and leap days: evaporation, runoff and water stores must agree.	reference_calendar, reference_degenerate, reference_leaky
mass/precipitation-counterfactual	mass	The same seed 20% wetter, 10% wetter and 20% drier: the water added or removed must be partitioned among evaporation, runoff and storage, and runoff must rise from drier to wetter.	reference_cheater, reference_leaky, reference_degenerate
mass/human-abstraction	mass	A prescribed net irrigation withdrawal must leave the budget: the same weather run with and without it, and the difference must account for exactly the abstracted volume.	reference_abstraction_blind, reference_leaky
energy/pet-consistency	energy	Evaporation reaches demand when the model's own soil is wettest, stays below it, and falls when the soil is driest.	reference_thirsty
energy/latent-heat-et-consistency	energy	The evaporation a model reports as water and the evaporation implied by the latent heat it reports: are they the same evaporation?	reference_two_head, reference_constant_lambda, reference_sublimation_blind, reference_energy_leak
energy/evaporative-partition	energy	One summer without rain under net radiation that did not change: the latent heat a drying surface gives up has to warm the air.	reference_two_head, reference_ground_dodge
energy/surface-energy-closure	energy	Does each day and night close its hourly surface energy budget, without opposite errors cancelling?	reference_diurnal_bias
energy/radiation-consistency	energy	The surface temperature a model reports and the upward longwave it reports: do they describe one surface, hour by hour, at the emissivity it was given?	reference_air_emitter, reference_no_reflection
momentum/routing-conservation	momentum	The channel store is never negative and never holds more than its hydrograph can.	reference_stuck_router

Models

models/ holds three kinds. A submitted model is there to be evaluated, and every run of one is appended to models/result.csv; the physical models' runs are archived there too. A physical model is there to test the probes: the exact bucket, two hand-written conceptual models from chrimerss/HydrologicModels and the NWS's SAC-SMA with Snow-17 must pass every probe that can ask them anything, so a probe that fails one is wrong until shown otherwise. All four report water and no energy, so all four are N/A, with reason INCOMPLETE, on the four probes that need an energy output: the three that need the latent, sensible and ground heat fluxes, energy/latent-heat-et-consistency, energy/evaporative-partition and energy/surface-energy-closure, and energy/radiation-consistency, which needs a surface temperature and its upward longwave. Not scored rather than passing: they have not violated conservation of energy, they have declined to be falsifiable about it, exactly as reference_streamflow_only does on the budget probes. A broken model is broken in one specific way, so that no criterion goes untested.

A standing counts the probes a model passed out of the probes that could score it. A probe that needs a variable the model does not report, or that the model cannot consume, is N/A for it and in neither number, so the totals differ between models.

Model	Kind	What it does	Standing
google_flood_forecast	submitted	The mean-embedding forecast LSTM behind Google Flood Hub, at the published weights. Predicts discharge and nothing else.	FAIL (VIOLATION), 5 of 11 probes passed. It passes the runoff bounds, memory, area, causality and steady state, and fails step, extreme rain, phase, warming, dry-down, and a 0.18 mm/day dip after an added storm
dhbv2	submitted	δHBV 2.0, the MHPI group's differentiable HBV: neural networks write the parameters of a bucket model that reports its stores and its evaporation.	FAIL (VIOLATION), 11 of 18 probes passed. Its learned regional-groundwater term, declared as gwex, closes the budget to 1e-8; what remains is a learned field capacity twice the catchment's, a response to doubled rain above the rain added, a runoff depth that changes with the area it is told, a third more runoff when snow falls as rain, and on mass/human-abstraction it never reads the prescribed withdrawal, so it accounts for none of the 380 mm and trips that probe's state_bounds on the same learned field capacity
wflow_sbm	submitted	Deltares' Wflow.jl SBM: a soil column with unsaturated and saturated stores, interception, snow and kinematic-wave routing, adapted in Julia and run on one representative cell at the resolution of Wflow's Moselle model.	FAIL (VIOLATION), 15 of 18 probes passed. Its water budget closes to 2e-4 of the rain and it passes the area, routing, step, calendar, causality, extreme-rain, phase, precipitation-counterfactual, warming and human-abstraction probes, the last by taking the prescribed withdrawal through Wflow's own water demand and allocation. That 2e-4 is water its river kinematic wave creates when open-water evaporation exceeds the rain on the river and the wave floors its discharge rather than drying the channel; on mass/extreme-event-closure it is 0.0016 mm on one 0.006 mm drizzle day on one seed, past that probe's 0.001 mm floor, the only one of the probe's events to fail. Its other two failures both come from its soil column: a storm that does not fill the column makes almost no runoff, and under the shipped mapping a wet month's extra water has already been evaporated down to the rooting depth when the storm arrives, so the wetter catchment runs off barely more than the dry one, about a thousandth of the storm where the probe asks for two hundredths; and once the root zone dries in a rainless spell, drainage from the unsaturated store raises a water table that sits below the roots, and lateral flow rises with no rain. Both move with how the stated soil capacity is mapped onto soil thickness and roots.
summa	submitted	SUMMA 4.0.0, a land model that solves the water and energy balances of canopy, snow, soil and aquifer with one implicit solver, run as one lumped HRU from its shipped test-case setup, with the radiation and humidity it needs mocked from each forcing row.	FAIL (VIOLATION), 8 of 21 probes passed. Its water budget closes to rounding and its surface energy budget to 0.2% of its own net radiation; what fails is a latent heat of vaporisation held at its 0 C value, a net radiation of its own that is not the probe's rn (every hourly phase, and a drought that warms the surface), a canopy that, as the shipped setup configures SUMMA, intercepts all rain and keeps what freezes near 0 C, holding up to 27 mm of ice against a 2 mm capacity on the probes that score it (also the only failure on the precipitation counterfactual), leaf area that keeps its seasons under constant weather, runoff that follows the melt rather than the rain on one seed, a 5.2% runoff response to turning snow into rain on another, Green-Ampt runoff that appears only at the hourly step, and no human water use, so the prescribed withdrawal on the human-abstraction probe is never taken
cwatm	submitted	CWatM 1.11, IIASA's Community Water Model and an ISIMIP global hydrological model, run on one grid cell with every store it carries reported.	FAIL (VIOLATION), 14 of 18 probes passed. Its budget closes to 0.03%, and its own water-demand module pumps a prescribed withdrawal out of groundwater to within 0.004%. That 0.03% is water its capillary rise creates, a few thousandths of a millimetre a day, and on mass/extreme-event-closure it fails 3 to 8 one-day drizzle events of under 0.07 mm per seed, where 0.001 to 0.005 mm more leaves or is stored than fell, against that probe's allowance of 5% of the rain or 0.001 mm, whichever is larger. It also fails on a groundwater reservoir with no dt that drains 24 times too fast at an hourly step (52% of the rain); on a 0.29 mm/day dip after an added storm, where preferential flow turns surface runoff into interflow that leaves through a slower runoff-concentration lag; and on evaporation on wet soil at 0.70 of demand, near the cap its crop coefficients set. The last two are packaging choices as much as results: this package follows the CWatM-Earth-30min template its parameters come from, and with preferentialFlow = False, the setting of the pinned model repository's own 30′ templates, both pass
lisflood	submitted	LISFLOOD 5.0.0, the EC Joint Research Centre's distributed model behind EFAS and GloFAS, stepped through its own Python framework on one representative 5 km cell and reporting every store its own water balance module counts.	FAIL (ERROR), 15 of 18 probes passed. Its budget closes to 1e-13 mm per step; it reports no heat fluxes and no surface temperature, so the four energy probes that need an energy output cannot ask it anything and are N/A. It fails the step probe because its potential infiltration is a pore-space storage multiplied by the step length, so rain falling within hours runs off at the hourly step (13.0% of the rain between PT1H and PT1D). The two ten-year probes take about 90 s per run under amd64 emulation against a 60 s budget, so their rows are ERROR from the host's speed, and those two ERROR rows alone make the verdict FAIL (ERROR). Run outside the limit, mass/precipitation-counterfactual passes every criterion. On mass/human-abstraction, LISFLOOD's own water-use rule takes the groundwater share in full and the rest only from channel water above an environmental-flow reserve, recording what the channel cannot give as shortage. On this one-cell water region it withdraws 17 to 33% of the prescription, from the sourced reserve to none, and leaves 67 to 83% where the probe allows 5%
reference_bucket	exact	conserves water exactly by construction	must pass every probe that can ask it anything; N/A on the four energy probes that need outputs it does not report
flex_lumped	physical	lumped FLEX/HBV: interception, beta-partitioned unsaturated store, fast and slow reservoirs, triangular lag	must pass every probe that can ask it anything; PASS, 18 of 18
flex_topo	physical	FLEX-Topo: plateau, hillslope and wetland units on real Wark fractions sharing one groundwater store	must pass every probe that can ask it anything; PASS, 18 of 18
sacsma_snow17	physical	the NWS's SAC-SMA with Snow-17 and a gamma unit hydrograph, ported from the legacy Fortran and checked against it	must pass every probe that can ask it anything; PASS, 18 of 18
reference_coupled	exact	the bucket with snow sublimation and a surface energy budget: every kilogram converted at the latent heat of the phase it actually underwent	must pass every criterion of the three energy-flux probes; supports daily and hourly steps
reference_abstraction_blind	broken	the same bucket, blind to the prescribed withdrawal, so the two variants come out identical	caught by human_abstraction
reference_two_head	broken	a water head and an energy head that never meet: both budgets close to 1e-15 and the latent heat implies an evaporation it never reported	caught by flux_identity
reference_constant_lambda	broken	coherent, but converts every kilogram at the same latent heat of vaporisation	caught by flux_identity
reference_sublimation_blind	broken	coherent for liquid water, but converts snow sublimation at the latent heat of vaporisation instead of sublimation	caught by flux_identity
reference_ground_dodge	broken	coherent, both budgets close, and the sensible flux never reads the soil: when the soil dries the ground flux absorbs the whole shift	caught by partition_shift
reference_diurnal_bias	broken	shifts sensible heat so the energy residual is +20 W/m2 by day and -20 W/m2 by night; the full-record residual cancels	caught by energy_closure_by_phase
reference_radiative	exact	the coupled reference with a skin: temperature diagnosed from the sensible heat flux through a fixed conductance, upward longwave that skin's emission plus the reflected sky at the emissivity it was given	must pass every criterion of energy/radiation-consistency
reference_air_emitter	broken	reports the skin's temperature but emits at the air's, reflected sky unchanged	caught by radiative_identity
reference_no_reflection	broken	reports emission alone as the total upward longwave, the reflected sky left out: 3 to 23 W/m2 that a 5% tolerance would not reliably see	caught by radiative_identity
reference_energy_leak	broken	discards 15% of net radiation; the energy counterpart of reference_leaky	caught by energy_closure
reference_leaky	broken	hides a silent 15% sink	caught by closure
reference_cheater	broken	solves for storage as whatever balances the budget	caught by state_bounds
reference_degenerate	broken	evaporates all precipitation, produces no runoff	caught by non_degenerate, response_sign
reference_fixed_step	broken	treats every row as a day whatever the step is	caught by resolution_invariance
reference_anticipating	broken	smooths runoff over a centred window, so three days of the future are in every value	caught by causality
reference_climatology	broken	emits the seasonal mean whatever falls, and keeps flowing without rain	caught by dry_down
reference_saturating	broken	caps its daily runoff, so an extreme storm adds rain and no runoff	caught by monotone_response
reference_restless	broken	a recession on an internal thirty-day clock, so it never settles	caught by steady_state
reference_overflowing	broken	reports its runoff plus 80% of the rain again, from nowhere	caught by runoff_bounds
reference_area_leak	broken	loses a share of runoff that grows with the area it is told	caught by invariance (area)
reference_overshooting	broken	a derivative term sharpens its hydrograph, so an added storm lowers later flow	caught by response_nonnegativity
reference_sublimating	broken	loses 40% of every snowfall to an unreported sublimation	caught by phase_invariance
reference_thirsty	broken	evaporates a fixed share of its soil store, never reading demand; conserves water exactly	caught by demand_consistency
reference_stuck_router	broken	a routing kernel summing to 0.9, so a tenth of every day's runoff never leaves the channel	caught by routing_conservation
reference_streamflow_only	honest limit	reports discharge only, from a store that never reads the temperature	N/A (INCOMPLETE) on budget probes; caught by response_sign
reference_in_sample	broken	removes surface runoff above a fixed 55 mm daily precipitation cutoff	caught by event_water_closure
reference_calendar	broken	a recession that drifts with the calendar year	caught by invariance (time origin)

How a case is generated

Forcing is generated fresh at run time from a recorded seed. Nothing is committed as data, so there is nothing to memorise, and a reviewer reviews a short deterministic script instead of a binary blob. Each probe runs several seeds and all of them must pass, so no model gets through on a lucky draw. Any run reproduces exactly with ht run --seed <n>.

This is an unseen-sample guarantee, not a claim that the public generator's distribution is secret. For evaluation against generators or data unavailable during training, keep a second probe tree outside the repository and add it with --probe-root /secure/hidden-probes. The model container receives only opaque case metadata and the inputs needed for inference; the host retains the probe identity, generator seed, annotations and scoring code. See docs/adapting-a-model.md.

Verdicts

A probe that can ask the model something passes or fails it, and the reason is recorded separately from the verdict, because these mean different things:

• VIOLATION the model reported its budget and the budget did not close.
• ERROR the adapter or benchmark machinery failed. This is operational, not a scientific verdict, and is the one outcome that makes ht run exit 2.

A probe that cannot ask the model anything is N/A: not scored, and neither a pass nor a fail. That happens two ways:

• INCOMPLETE the model never reported enough to be checked. Every streamflow-only model lands here on the budget probes. It has not violated conservation; it has declined to be falsifiable.
• INCOMPATIBLE the model and probe disagree on timestep, on a forcing or static input one needs and the other does not supply or declare, or on paired-perturbation support, so running them would not be meaningful.

A model passes when at least one probe could be put to it and every probe that could be put to it passes. Otherwise it fails, with the worst reason among the scored probes that did not pass; an unscored probe never supplies that reason. A model that no probe could be put to at all is N/A too, and ht run exits 1 for it as it does for a FAIL.

A scored verdict is one bit. Everything under it stays quantitative, so a paper can show that one model leaks 6% and another 40% long before anyone crosses the line.

Quick start

pip install -e '.[dev]'

ht init-probe --list-templates            # probe shapes to start from
ht init-probe --template invariance      # start a new probe
ht list                                  # probes and models
ht validate                              # schema-check everything
ht gate                                  # the probe acceptance gate
ht run --model reference_bucket          # evaluate one model
ht run --model my-model --json out.json  # machine-readable report

Without installing, ./ht (or ht.cmd in a Windows shell) runs the CLI straight from src/, as does python -m hydroturing with src on PYTHONPATH. Set HT_ASCII=1 for reports with the words and no marks, which is also what you get automatically wherever the output stream cannot carry them.

Submitting a model

Your model may be written in any language. It ships as a container plus a thin adapter that reads /io/request.json and writes /io/output/result.csv. The adapter is usually thirty lines. See docs/adapting-a-model.md, and AGENTS.md if you are having a coding agent build the sandbox for you.

We want more models, and we want the ones you think matter. Every published rainfall-runoff model, every LSTM, every foundation model with a hydrologic claim is in scope.

Start with a model proposal. The form accepts AI-based, AI+physics and physics models, with separate links for the model code and optional pretrained weights, and asks whether there is a time window you want the test to run over.

You do not need to be able to package it yourself. The packaging status field decides who the issue is assigned to and nothing else: ready means it is yours to finish, help needed means the maintainer takes it. Either way the proposal counts as yours, and five accepted proposals earn co-authorship on the benchmark paper the same as one merged probe. Knowing which models are worth putting through the benchmark is a judgement about the field, and it is not one the maintainer can make alone.

Once a proposal is accepted, fork and build; the pull request comes from the fork and closes the issue.

A submitted model is scored on the largest flood event of the generated record rather than on all ten years of it: by default a month for a daily model and a week for an hourly one, with the full spinup in front, located by the probe's own reference model. That is what keeps a model that takes seconds per forecast inside the probe's time budget. window_days in model.yaml changes it; full asks for the whole record.

The container runs with no network and never sees the probe code, so a model cannot read the tolerance it is being judged against. Every evaluation is appended to models/result.csv.

A model that fails is worth proposing, and so is one most probes cannot score. INCOMPLETE is the current state of nearly every published rainfall-runoff model on the budget probes — it has not violated conservation, it has declined to be falsifiable — and recording that honestly is a large part of what this is for.

Contributing a probe

Please submit one. A benchmark with four probes tests four things; the reason this repository is open is that the physics worth testing is wider than any one group knows. If you have spent time with a conservation law that AI models get wrong, that law is a probe, and we would rather have it from you than approximate it ourselves.

Contributors of merged probes are co-authors on the benchmark paper. The threshold is one probe, merged and passing the acceptance gate, or five accepted model proposals. This is how model intercomparison projects have always worked in this field: you contribute an experiment, you are an author on the paper that reports it. CONTRIBUTING.md has the full terms — author order, the right to decline, and what happens before anything is submitted.

ROADMAP.md lists the probes we want, each with a difficulty and a note on what it discriminates. Claiming one is easier than inventing one, but inventing one is welcome too; propose it first so nobody builds it twice. The two we most want are the cross-budget consistency probes: a model can close its water budget and its energy budget while being incoherent between them, and nothing in the suite currently notices.

Propose first, then fork, then open a pull request — the same three steps for a probe or a model. The proposal issue is where we find out whether a probe discriminates, which is cheaper to learn in a paragraph than in three hundred lines. Once it is labelled accepted it is assigned to you and nobody else will build it.

Then do not start from a blank page:

ht init-probe --list-templates           # the shapes available
ht init-probe --template <kind>          # writes probe-draft.yaml
#   ... fill in the fields ...
ht init-probe --from probe-draft.yaml    # creates probes/<law>/<slug>/
ht gate --probe <law>/<slug>             # prove it discriminates

The templates cover conservation over one case, extrapolation in space and in time, counterfactual response, and invariance. Each scaffolds into a probe that already passes the gate against a placeholder case, so you can watch it separate the reference models before writing any physics, then replace the case with yours. docs/writing-a-probe.md has the details.

The gate is the only bar that matters, and it is a technical one: your probe must pass an exact physical model and catch the deliberately broken ones. See GOVERNANCE.md for how disagreements about tolerances get settled.

Community

Questions, probe ideas, a model you would like evaluated, a tolerance you think is wrong: the HydroTuring Discord is where that conversation happens, before and alongside the issues.

Status

Suite 0.1.0, pre-release. Twenty-two probes, sixteen mass, five energy, one momentum, synthetic track only. More energy and momentum probes, and the real-data track, are next. The harness runs paired cases and scores labelled regimes, so the generalisation probes on the roadmap — extrapolation in space and time, counterfactual response, invariance — are unblocked and unclaimed. Scores are only comparable within a suite version.

Layout

src/hydroturing/   harness: protocol, runners, criteria, scoring
probes/<law>/<id>/ probe.yaml, generate.py
models/<name>/     model.yaml, Dockerfile, adapter
schemas/           JSON schemas for both spec files