Zarafshan Valley Groundwater Model
→ ↓ Space
Next slide
Следующий слайд
← ↑
Previous slide
Предыдущий слайд
Home / End
First / last slide
Первый / последний слайд
C
Contents rail — chapters and sections
Оглавление — главы и разделы
O
Overview grid — click a slide to jump
Обзор слайдов — щёлкните, чтобы перейти
N
Presenter notes
Заметки докладчика
P
Print / export to PDF
Печать / экспорт в PDF
L
Switch language EN / RU
Переключить язык EN / RU
?
This help
Эта справка
Ministry of Mining Industry and Geology of the Republic of Uzbekistan
Regional groundwater model of the Zarafshan valley · MODFLOW 6
Modelling and analyticshydrosolutions GmbH

Status update · September 2026

Zarafshan Valley Groundwater Model

A regional MODFLOW 6 model of the alluvial aquifer between Ravathoji and Khazara. Where the work stands, what it already shows, and what comes next.

Overview

What this presentation covers

  • 1 Why a groundwater model for the valley
  • 2 The aquifer and the data behind it
  • 3 The water balance: what enters and leaves
  • 4 The model and what it shows so far
  • 5 What the model can and cannot do yet
  • 6 Next steps and handover

Keys: → next slide, C contents, O overview, N notes, L language, P print to PDF.

Summary

The main messages

A working regional groundwater model of the Zarafshan valley exists. It runs, it is documented, and it was handed over to MMIG in May 2026 with the full data pipeline.

The water balance is built from measured data: river gauges, canal records, satellite evapotranspiration, population-based pumping and the MMIG abstraction register.

The steady-state model reproduces the regional head pattern. The remaining misfit is structural: forty percent of the head targets lie on the northern piedmont outside the mapped alluvium, and the registry pumping cannot be sustained where the model places it.

The next round settles the pumping magnitude and routing, rebuilds the calibration target set, and re-thinks the piedmont boundary before any further parameter estimation. One data request to MMIG, the annual abstraction series, would resolve the biggest open question.

1 Why a groundwater model

Why the valley needs a groundwater model

The Zarafshan between Ravathoji and Khazara is the most intensively irrigated stretch of Uzbekistan. Canal losses, field percolation, pumping and the river itself all feed one shallow alluvial aquifer.

Groundwater and the river exchange water in both directions along the whole valley. Without a model, the size and direction of that exchange can only be guessed.

MMIG monitors groundwater levels and issues abstraction permits. A regional model turns the monitoring record into a tool for asking what-if questions about pumping, recharge and river flow.

1 Why a groundwater model

One model domain from Ravathoji to Khazara

One model domain from Ravathoji to Khazara

Finding. The active model covers the valley floor from the Ravathoji hydroworks downstream to Khazara. Bedrock and the northern mountain catchments are outside the model and enter only as boundary inflows.

What it means. Results are regional. The model answers questions at the scale of districts and river reaches, not individual well fields.

1 Why a groundwater model

The model in numbers

230 × 70 kmModel domain, rotated to the valley axis
72 × 223 cells of 1 kmRegional grid, single unconfined layer
180 monthly periodsTransient simulation 2010–2024
300 river reachesStreamflow routing (SFR) along the Zarafshan
381 wellsSteady-state head targets after quality control
17 849 monthly targetsTransient head observations 2010–2024

MODFLOW 6 with FloPy, Newton-Raphson solver, open-source throughout. Every array is rebuilt from the raw data by a documented notebook pipeline.

2 The aquifer and the data

Five geology zones from MMIG maps

Five geology zones from MMIG maps

Finding. The Quaternary geology map was vectorised and mapped onto the grid: modern alluvium (Qh), upper and middle Quaternary terraces (QIII, QII), Eopleistocene deposits (QE) and bedrock. Bedrock cells are inactive.

What it means. Aquifer properties are assigned by zone. The zones follow the field mapping of the Zarafshan hydrogeological station, so they can be discussed and corrected with MMIG geologists.

2 The aquifer and the data

Why one layer is enough

Why one layer is enough

Finding. Wells screened in Quaternary deposits and wells screened deeper in Neogene deposits plot on the same head range at every depth: there is no systematic head separation between them. The MMIG cross-sections likewise show continuous alluvium without a regional aquitard.

What it means. A single hydraulic layer is structurally correct for the valley and simpler to calibrate. The aquifer bottom is the largest remaining structural unknown.

2 The aquifer and the data

Hydraulic conductivity by zone

Hydraulic conductivity by zone

Finding. Transmissivity from the MMIG well registry, filtered to reliable pumping tests, gives the prior conductivity of each zone. Modern alluvium is the most permeable; older terraces and Eopleistocene deposits are tighter.

What it means. These are starting values. Calibration adjusts them, and the sensitivity analysis later shows which zones the observations can actually constrain.

2 The aquifer and the data

The river in the model

The river in the model

Finding. The Zarafshan main stem, with the Oqdaryo and Qaradaryo branches, is traced from mapped river geometry rather than from the terrain, because flow-routing on a flat irrigated plain fails. Channel widths of 40 to 500 m come from satellite cross-sections.

What it means. The river is a routed stream in MODFLOW, not a fixed boundary. Its stage responds to inflow and to groundwater exchange, which is what the study is about.

2 The aquifer and the data

Where the groundwater is observed

Where the groundwater is observed

Finding. The MMIG well registry, the ZRB monitoring wells and the monthly UGV time series were merged and quality-controlled. Coverage is dense around Samarkand and thin in the western valley.

What it means. Observations are the currency of calibration. Where they are sparse, the model is weakly constrained and results should be read as regional tendencies.

2 The aquifer and the data

What the model is built from

SourceDataCoverage
MMIG (Momigeo)Well registry, monthly groundwater levels, abstraction register, geology maps and cross-sections2000–2024
UzHydrometDaily discharge at Ravathoji2010–2024
Ministry of Water ResourcesCanal diversions and 52 gauges, discharge to the Koradaryo2014–2022
CHIRPS v3, CHELSA v2.1Monthly precipitation; temperature and precipitation climatologies1981–2025
Ragettli et al. (2025)Satellite evapotranspiration of irrigated land2017–2022
Landsat irrigation mapsIrrigated area at 30 m2016–2020
WorldPop, GHS-SMODPopulation density and settlement class2020
SRTM, HydroLAKES, OSMTerrain, reservoirs, river geometrystatic

Sensitive MMIG and ministry data stay outside the public repository and are shared with partners separately.

2 The aquifer and the data

Known data gaps, stated openly

Production-well coordinates

The MMIG abstraction register gives district totals and well counts, but coordinates only for Bukhara. Pumping around Samarkand and Navoi is therefore spread by population, not placed at well fields.

Aquifer bottom

No bedrock-depth map exists. The model uses a uniform thickness; well filter depths are only lower bounds.

River inflow record

Reliable discharge to the Koradaryo below the Ravathoji diversions exists for 2014–2017 only. Other years are filled with monthly ratios to Ravathoji.

Agricultural pumping

Sector splits are observed for Samarkand districts and transferred to Navoi. Return fractions of pumped water are bookkeeping assumptions, not measurements.

3 The water balance

Irrigated area from satellite

Irrigated area from satellite

Finding. Landsat irrigation maps for 2016–2020 give the irrigated fraction of every model cell. Irrigation dominates the valley floor and thins toward the terraces.

What it means. This fraction splits recharge and evapotranspiration between irrigated and rain-fed land. It is the single most decisive forcing in the balance.

3 The water balance

Two independent estimates of irrigation water use agree

Two independent estimates of irrigation water use agree

Finding. Top-down, the Ministry of Water Resources records how much water the canals divert. Bottom-up, satellite evapotranspiration shows how much the fields consume. The two close to within a few percent at district scale.

What it means. Both sources can be trusted in the balance. The difference between diversion and consumption is the water available to recharge the aquifer.

3 The water balance

Recharge and evapotranspiration on the grid

Recharge and evapotranspiration on the grid

Finding. Precipitation recharge from CHIRPS and irrigation return flow form the recharge array. Crop evapotranspiration, shallow-water-table evaporation and reservoir evaporation form the evapotranspiration array. Both vary month by month over 2010–2024.

What it means. The share of irrigation losses that actually reaches the water table is the key uncertain quantity. The sensitivity analysis confirms this.

3 The water balance

Where groundwater is pumped

Where groundwater is pumped

Finding. Domestic and industrial pumping is distributed with 100 m population data and Uzbek per-capita norms, then compared with the MMIG abstraction register district by district.

What it means. The population template underestimates urban well fields by an order of magnitude. The current model therefore uses the full permitted abstraction from the register, smoothed within each district.

3 The water balance

Inflow from the northern tributaries

Inflow from the northern tributaries

Finding. Oqsoy, Tusunsoy and Qorasuv are ungauged. Their runoff was transferred from 85 gauged analogue basins in the Amu Darya region and checked against a published estimate for Tusunsoy.

What it means. These tributaries enter the model as groundwater inflow along the northern edge, not as surface streams, because most of their water infiltrates on the fans.

3 The water balance

What actually enters the river below Ravathoji

What actually enters the river below Ravathoji

Finding. The river boundary uses the measured discharge to the Koradaryo, the flow that remains after all canal diversions at Ravathoji. Using the gross Ravathoji flow would count canal water twice, once in the river and once as irrigation recharge.

What it means. The river in the model starts with roughly a third of the Ravathoji flow. Canal water reaches the aquifer through the recharge arrays instead.

3 The water balance

Closing the valley water balance

Closing the valley water balance

Finding. River inflow, precipitation, tributaries and canal water on one side; evapotranspiration, pumping, reservoir evaporation and outflow at Khazara on the other. The residual constrains net groundwater recharge, the quantity the model cannot observe directly.

What it means. Every flux entering MODFLOW is traceable to a data source and to this balance. That is the foundation the model rests on.

4 The model

How the model is assembled

How the model is assembled

Finding. Each MODFLOW 6 package carries one process: grid and properties, recharge, evapotranspiration, wells, the routed river, reservoirs as head-dependent boundaries, fixed heads at the valley ends and, in the transient model, storage.

What it means. The package set is deliberate and documented. Alternatives such as an unsaturated-zone package or drains were tested and rejected for stated reasons.

4 The model

The simulated groundwater surface

The simulated groundwater surface

Finding. The steady-state model, driven by 2010–2024 mean forcing, reproduces the regional gradient from about 900 m near Ravathoji to about 330 m at Khazara and converges in under a second. Against the 381 quality-controlled targets the uncalibrated model has a mean error of −7 m and a root-mean-square error of 30 m.

What it means. Fast runs make systematic testing possible: thousands of parameter combinations and dozens of conceptual scenarios have already been run.

4 The model

Which assumptions matter most

Which assumptions matter most

Finding. A global sensitivity analysis with 5 000 model runs, of which 3 091 converged, ranked seventeen parameters. The conductivity of the upper terrace, the shallow-groundwater evaporation rate, precipitation recharge, aquifer thickness and the conductivity of the middle terrace control the misfit. River-bed conductance, boundary heads, bedrock conductivity and Manning roughness barely matter for heads.

What it means. About half of the candidate parameters can be set aside before calibration. No parameter combination brings the mean absolute error below roughly 15 m, which points at structure rather than parameters. Head observations alone cannot constrain the river parameters; that needs discharge data.

4 The model

Calibration: the honest state

Calibration: the honest state

Finding. Ensemble calibration with PEST++ tightens the posterior fit within the ensemble: mean absolute error 17 m against 19 m for the prior. Re-running the best parameter set in the full model gives 27 m, worse than the uncalibrated baseline. The parameters that fit best in the ensemble do not transfer.

What it means. The remaining error is structural: uniform thickness, zone-uniform conductivity, observation quality and river-reach geometry. Fixing these matters more than further parameter tuning.

4 The model

What was tested in the model this year

A rapid steady-state scenario track tests one conceptual change at a time against the same 381 targets. The current diagnostic candidate stacks seven changes on the April model:

  • River inflow net of inter-basin transfer and irrigation intakes, so canal water is no longer counted twice
  • Full permitted pumping from the MMIG register, sector-aware and smoothed by population, instead of the much smaller population-only estimate
  • Return flow of pumped water added explicitly as local recharge, so gross abstraction stays visible
  • Gross-loss recharge partition: 90 % of canal losses and 50 % of field losses reach the aquifer
  • Shallow collector drains at 2 m depth that route drained water back to the river
  • Eastern fixed-head boundary replaced by an auditable underflow of 10 % of the Ravathoji inflow
  • Diagnostic aquifer thickness of 200 m

Only the first two bookkeeping corrections and the explicit underflow left the head fit unchanged; every later step worsened it. The candidate is therefore not promoted into the durable model.

4 The model

Current candidate: what full pumping does to the fit

With the full registry pumping and local return flows the model still converges, but the fit deteriorates from the April model. The deterioration is local and diagnostic.

  • 381 targets: mean error −16 m, RMSE 37 m, against −7 m and 30 m for the April model
  • Within 10 km of Samarkand the candidate lowers heads by 31 m on average; the observed 2010–2024 means show no such cone
  • Simulated river outflow at Khazara 54 m³/s against an observed 23 m³/s; the excess is drain return fed by the gross-loss recharge, not the pumping loop
  • Global conductivity and river-bed multipliers do not recover the fit
Current candidate: what full pumping does to the fit

Measured 14 September 2026 from scratch/ss_rapid_v1_kdown_sfr_connectivity/scenario_metrics.csv and the Iteration 3 diagnosis in GAPS.md §5; the Khazara target is read from the water-balance ledger.

4 The model

Fifteen years, month by month

Fifteen years, month by month

Finding. The transient model runs all 180 monthly periods without dry cells and with a mass-balance error below 0.02 percent. Against 15 491 monthly observations after a two-year spin-up the mean absolute error is 9.5 m. Multi-year levels are reproduced at some wells and offset by several metres at others; the seasonal swing is too small at most.

What it means. The model can already be used to ask how levels respond to a wet or dry year at regional scale. The muted seasonal amplitude points at the storage parameters, which the steady-state model cannot see.

4 The model

Seasonal amplitude points at specific yield

Seasonal amplitude points at specific yield

Finding. Five hundred transient runs varied the specific yield of each zone together with conductivity and recharge; 480 converged. The specific yields of the modern alluvium and the two Quaternary terraces control the seasonal amplitude. The best run cuts the young-alluvial amplitude error from 0.39 to 0.23 m, about forty percent, with lower specific yield than the literature values.

What it means. Literature specific yields are too high for this valley. The transient information constrains parameters the steady-state model cannot see, which is why the next calibration must be joint.

4 The model

Where the river gains and where it loses

Where the river gains and where it loses

Finding. Over the fifteen simulated years the river is a net gaining stream for most of the year: groundwater discharges to the Zarafshan in autumn and winter, and the river loses water to the aquifer during the summer irrigation peak when canal water raises the water table.

What it means. This seasonal reversal is the headline quantity for management. Its magnitude is still uncertain because it is constrained by heads only and because some river reaches are longer than one cell; discharge at Khazara is the missing check.

5 Capabilities and limits

What the model can and cannot do today

Can do now

  • Reproduce the regional groundwater surface and run 2010–2024 month by month
  • Account for every inflow and outflow with a traceable data source
  • Show the seasonal reversal of river–aquifer exchange
  • Test conceptual changes and pumping scenarios in seconds
  • Be run and modified by MMIG staff with open-source tools

Cannot do yet

  • Sustain the registry pumping at the Samarkand well fields without an unobserved 30 m cone
  • Represent the northern piedmont wells, which sit outside the mapped alluvium
  • Give calibrated absolute levels better than about 15 to 20 m mean error
  • Quantify river–aquifer exchange with confidence: no discharge check yet
  • Represent variable aquifer thickness or bedrock depth

5 Capabilities and limits

Where the remaining error comes from

Forty percent of the targets, 153 of 381, are static wells on the northern piedmont outside the mapped Quaternary. They sit on Neogene sediments with borrowed conductivity and almost no recharge, and they are 27 m too low on average. This is a boundary-concept problem, not a parameter problem.

Many observations share one 1 km cell: 206 targets fall into 50 cells, with a median within-cell spread of 12 m. Two stacked cells alone carry a third of the squared error. Part of the misfit is observation support that no single-layer model can reduce.

The registry pumping is credible as a magnitude, but the model cannot sustain it at the Samarkand fan without a drawdown cone the observations do not show. Either the fan receives inflow the model lacks, or the effective abstraction over 2010–2024 was lower than the 2026 register, or both.

Some river reaches are longer than one cell, so their leakage passes through a single cell. The fix is known and scheduled before any flux calibration.

6 Next steps and handover

Next steps

Modelling, in this order

  • Settle the pumping: bracket 2010–2024 abstraction between the 2010 and 2026 registers, route well-field exports and city sewage explicitly instead of local return, and test a Samarkand fan concept with mountain-front inflow
  • Rebuild the calibration target set: one target per cell, edge and above-ground wells excluded, static and time-series wells weighted separately
  • Give the northern piedmont its own concept: mountain-front recharge and a Neogene zone, or treat those wells as a separate aquifer
  • Promote the bookkeeping corrections that did not hurt the fit into the durable model; then drains and return flow as one calibrated sub-system

Validation

  • Compare simulated river outflow with observed discharge at Khazara, about 23 m³/s
  • Split river reaches at cell boundaries before any flux calibration
  • Joint steady-state and transient calibration only once the gate on the cleaned target set passes

Requests to MMIG

  • The annual groundwater abstraction series 2010–2025 by district, from the same reporting line as the 2026 register
  • Coordinates and rates of the production wells and well fields for Samarkand and Navoi provinces, and the pipeline exports to Navoi and Bukhara
  • Any bedrock-depth or aquifer-thickness information: deep boreholes, geophysics, hydrogeological sections
  • Continued monthly monitoring records beyond 2024

6 Next steps and handover

What MMIG has already received

  • A public code repository with the full notebook pipeline, from raw data to the transient model, installable with open-source tools
  • Twenty documented notebooks with bilingual guidance, one per pipeline step
  • A read-only model viewer for browsing heads, budgets and river exchange without writing code
  • The four-day handover workshop of May 2026 at the Geology University in Samarkand, with pre-translated slides and setup guides
  • An open issues register that records every known gap, decision and calibration iteration

Sensitive MMIG and ministry datasets are shared with partners out of band and never enter the public repository.

6 Next steps and handover

A working platform, not a finished product.

The model runs, its data are traceable, and its limits are written down. Settling the pumping record with MMIG and cleaning the target set are the two steps that turn it into a tool for permit and monitoring decisions in the valley.

hydrosolutions GmbH · Zurich · September 2026

1 / 1