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 review · September 2026

Groundwater in the Zarafshan valley

A regional groundwater model between Ravathoji and Khazara: why it was built, what it is built from, what it shows, where it falls short, and what should happen next.

Overview

Five parts

  • 1 Introduction and problem statement
  • 2 Data and methods of the regional modelling approach
  • 3 Results: long-term average and month-by-month
  • 4 Issues, discussion and required improvements
  • 5 Conclusions

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

Summary

Key messages

The valley's groundwater is under growing pressure: a third more people since 2010, rainfall that swings by half between years, and pumping that the register reports at eight times the population-based estimate.

A working regional model exists. It is built from every dataset available, runs in seconds, reproduces the regional water table, and simulates fifteen years month by month.

It is not yet a tuned tool. Forty percent of the observation wells lie in the foothill belt outside the mapped river deposits, and the pumping in the register cannot be placed without a drop in levels that the wells do not show.

The way forward is a settled pumping record, a cleaned set of observation wells, and a linked soil–canal–aquifer approach so that recharge is computed rather than assumed, as already done in the Chu basin.

1 Introduction and problem statement

A valley that lives on one aquifer

Between Ravathoji and Khazara the Zarafshan runs through the most intensively irrigated land of Uzbekistan. Under the fields lies a single shallow aquifer in the river deposits that the canals, the fields, the river and the towns all draw on.

Samarkand, Kattakurgan and Navoi take their drinking water from it. Farmers pump it where canals do not reach. The river gains from it in winter and loses to it in summer.

Decisions on permits, monitoring and new well fields are taken district by district, but the aquifer does not know district boundaries. A regional model is the only way to see the whole system at once.

1 Introduction and problem statement

The study area: 230 km of valley and everything the model must account for

The study area: 230 km of valley and everything the model must account for

Finding. The model covers the valley floor and terraces, 10 208 km², from the Ravathoji hydroworks to the Khazara gauge. The Zarafshan splits into the Oqdaryo and Qaradaryo branches, feeds the canal systems and the Kattakurgan reservoir, and receives three ungauged tributaries from the north, whose basins lie outside the model.

What it means. Everything the model says is regional: districts and river reaches, not individual wells. The mountain catchments enter only as inflows at the model edge.

Source: OpenStreetMap tiles and canals; SWB outline and grid; tributary basins (hydra-shed); HydroLAKES; inflow points; WorldPop city cells.

1 Introduction and problem statement

Four pressures on the same aquifer

Population growth

A third more people in the model area since 2010. Domestic pumping grows with them, and the towns' well fields sit in the most permeable part of the aquifer.

Climate variability and change

Rainfall over the valley varied between 178 and 471 mm a year in 2010–2024. Warmer summers raise crop demand; the snow-fed river peak is expected to shift earlier.

Unmeasured abstraction

The register lists eight times more drinking and industrial pumping than population norms explain, without well coordinates. Agricultural and unregistered pumping are not measured at all.

River diversion

Two thirds of the river is diverted at Ravathoji before it enters the valley. What is left in the channel sets how much the aquifer can exchange with it.

1 Introduction and problem statement

More people, more pumping

The model area is home to almost five million people. Their number grew by a third between 2010 and 2024, and domestic demand in the model grows in step.

Even the population-based estimate, which is the lower bound, rises from 0.09 to 0.12 km³ a year over the simulation period.

More people, more pumping

Source: data/population/yearly_pumping_summary.csv (WorldPop 2020 scaled by national growth).

1 Introduction and problem statement

Dry years come in runs

Rain and snow over the valley are the only recharge that does not depend on the canals. Three of the last five years were well below the 2010–2024 mean.

Recharge from rainfall is small in absolute terms, about a tenth of what falls, but the same dry years also cut the river inflow that feeds the canals.

Dry years come in runs

Source: CHIRPS v3 monthly rasters averaged over the model outline (measured in this review); the 10 % share of rain that recharges is the water-balance assumption.

1 Introduction and problem statement

Where the water table is falling

Where the water table is falling

Finding. Of 102 monitoring wells with at least eight years of data, 38 show levels falling by more than 0.1 m a year and only 5 rising. The strongest declines cluster south of Samarkand and in the eastern terraces.

What it means. The median trend is close to zero, so this is not a valley-wide depletion yet. It is a set of local declines that a regional model must be able to attribute to pumping, canal changes or dry years.

Source: data/observations/regional_gw_observations_long.csv, quality-checked values, straight-line trend per well (measured here).

1 Introduction and problem statement

How much is really pumped? Nobody knows precisely

How much is really pumped? Nobody knows precisely

Finding. For the thirteen Samarkand districts with a purpose breakdown, the MMIG register reports 27.5 m³/s of drinking and industrial abstraction; population and per-capita norms explain 3.6 m³/s. The gap sits almost entirely in the three districts with regional well fields.

What it means. Well fields export water by pipe to Navoi and Bukhara and return city water through the sewer collector. Without well coordinates, annual series and export volumes, the model cannot place this pumping, and unregistered farm pumping is not in either number.

Source: data/population/modelled_vs_observed_comparison.csv (MMIG ОТБОР register).

1 Introduction and problem statement

What the model has to answer

  • How much water reaches the aquifer from canals, fields and rain, and how much leaves through pumping, evaporation and the river?
  • Where and when does the river gain groundwater, and where does it lose it?
  • What happens to levels if pumping grows, if a dry decade comes, or if canal losses are reduced?
  • Which of these answers can be trusted today, and which need more data?

The next three parts follow these questions: the data and methods, the results, and the limits.

2 Data and methods

Everything the model is built from

Everything the model is built from

Finding. Eleven datasets from MMIG, UzHydromet, the Ministry of Water Resources and global satellite products. Groundwater levels cover 2000–2024, river flow 2010–2024, canal records 2014–2022 with reliable flow below the intakes only for 2014–2017.

What it means. Every input is traceable to one of these bars. The short bars are the weak points: the canal record and the single-year population snapshot.

Source: data/README.md and the per-folder READMEs; coverage as recorded there.

2 Data and methods

How water moves through the valley aquifer

How water moves through the valley aquifer

Finding. One unconfined aquifer between the land surface and the bedrock. It is fed by canal and field losses, a little rain, the fans of the northern tributaries and groundwater arriving from upstream; it is drained by the river, evaporation from shallow groundwater and crops, and pumping.

What it means. This concept is what the model implements. Two of its elements, the bedrock depth and the foothill margins, are also where the model is weakest, as chapter 4 shows.

Source: Conceptual drawing after the MMIG cross-sections and the model concept; no model result is shown.

2 Data and methods

Five geological zones carry the aquifer properties

Five geological zones carry the aquifer properties

Finding. The geological map of the young deposits was traced and transferred to the model grid: recent river deposits along the Zarafshan, two generations of older terrace deposits on both flanks, small patches of the oldest deposits, and bedrock, which is left out.

What it means. Wells drawing from the young deposits and from the older rock beneath them show no systematic difference in water level, so one layer is a sound choice. The zones follow MMIG's field mapping and can be refined with MMIG geologists.

Source: data/model/aquifer_properties.npz zone_grid (measured square counts).

2 Data and methods

How easily water moves and how much the aquifer stores

How easily water moves and how much the aquifer stores

Finding. How easily water moves through each zone comes from pumping tests in the MMIG well registry, one value per zone: the recent river deposits pass water fastest, the old terraces slowest. How much water each zone can release from storage started from published values.

What it means. These are starting values. Chapter 3 shows which of them the measurements can actually pin down.

Source: data/model/aquifer_properties.npz and aquifer_zone_lookup.json.

2 Data and methods

Where groundwater levels are measured

Where groundwater levels are measured

Finding. 398 wells passed quality control: 148 with monthly time series and 250 with a single measurement. Coverage is dense around Samarkand and along the northern edge, sparse west of Kattakurgan.

What it means. These wells are the yardstick for the model. Where they are sparse the model is weakly checked; where several sit in the same 1 km square they cannot all be matched.

Source: data/observations/calibration_targets_steady_state.csv (MMIG registry, ZRB monitoring wells, UGV time series).

2 Data and methods

Most of the river is diverted before it enters the valley

Most of the river is diverted before it enters the valley

Finding. The Zarafshan brings 134 m³/s to Ravathoji on average. Transfers to the Jizzakh and Kashkadarya basins take 39 m³/s, the valley canals another 49 m³/s. The river below the intakes carries 46 m³/s, a third of the inflow.

What it means. The model's river starts with that lowest line. Canal water enters the aquifer as recharge through the fields, which is why the loss fractions matter so much.

Source: data/mow/mainstem_sfr_inflow.npz (UzHydromet daily discharge, Ministry of Water canal records 2014–2017, other years by monthly ratios).

2 Data and methods

Three ungauged tributaries from the north

Oqsoy, Tusunsoy and Qorasuv have no gauges. Their flow was estimated from 85 similar basins with gauges, scaled by elevation and checked against a published estimate for Tusunsoy.

They add 4 m³/s on average, mostly in March and April, and enter the model as groundwater inflow along the northern edge because their water seeps into the ground on the fans.

Three ungauged tributaries from the north

Source: data/mow/northern_tributary_inflows.csv.

2 Data and methods

Where the water is used

Where the water is used

Finding. Landsat maps for 2016–2020 give the irrigated share of every 1 km square: 3 272 km² inside the model area, concentrated on the valley floor and thinning towards the terraces.

What it means. This share splits recharge and water loss to the air between irrigated and rain-fed land. It is the most decisive single input of the water balance.

Source: data/irrigation/regional_irrigation_fraction.npz; area = sum of the irrigated shares over active squares.

2 Data and methods

What goes in and what is drawn out, every month

What goes in and what is drawn out, every month

Finding. Recharge combines rain seeping into the ground with canal and field losses, weighted by the irrigated share. Water loss to the air combines crop use, evaporation from shallow groundwater and from open reservoirs. Both maps exist for each of the 180 months.

What it means. The fraction of irrigation losses that actually reaches the water table, set at 45 percent, is the key assumption. Chapter 4 argues it should be simulated, not assumed.

Source: data/model/regional_rcha_evt_arrays.npz, long-term average maps, converted from m/d to mm/yr.

2 Data and methods

Where groundwater is pumped in the model

Where groundwater is pumped in the model

Finding. Domestic and industrial pumping is spread with 100 m population data and Uzbek per-capita norms, 3.4 m³/s in total for the April model. The September candidate instead uses the full register value of 37.8 m³/s, smoothed within districts.

What it means. Neither placement is right: the first is too small, the second puts well-field pumping where the population is rather than where the wells are.

Source: data/population/pumping_mean_annual.npz and population_regional_2024.npz (WorldPop 2020, projected).

2 Data and methods

The valley water balance, and what is left for the aquifer

The valley water balance, and what is left for the aquifer

Finding. Inflows of 8.2 km³ a year against 6.8 km³ of measured or estimated outflows leave 1.4 km³ unaccounted. With the assumed loss fractions, 0.96 km³ of that recharges the aquifer and 0.46 km³ is lost in the soil zone.

What it means. The aquifer's recharge is the one term no gauge measures. It is derived as a remainder, which is why every other term must be as good as the data allow.

Source: data/mow/regional_water_balance_constraints.json. Recharge = 45 % of canal and field losses plus 10 % of precipitation.

2 Data and methods

How the model talks to its surroundings

How the model talks to its surroundings

Finding. The river is computed along 300 channel sections, so its water level responds to inflow and to exchange with the aquifer. Fixed water levels close the two valley ends, 75 reservoir squares exchange water with the aquifer depending on their level, and pumping is applied square by square.

What it means. The northern and southern foothill edges are closed to flow and receive almost no recharge. That choice is revisited in chapter 4.

Source: data/model/regional_model_setup_v2.json (river and fixed-level data), regional_reservoir_mask.npy, regional_wel_data.npz.

2 Data and methods

The model in three dimensions

The model in three dimensions

Finding. Seen as a block with the vertical stretched about 30 times, the model is a thin slab that follows the land surface down the valley: 150 m thick everywhere, 1 300 m above sea level at Ravathoji and 310 m at Khazara, over a length of 230 km.

What it means. The geometry is simple by design: one layer, one thickness. The bedrock depth that would give the slab its true shape is the largest structural unknown.

Source: data/model/regional_model_setup_v2.npz (top, bottom, active area); drawn in model-grid coordinates.

2 Data and methods

From raw data to a working model

From raw data to a working model

Finding. Twenty processing steps in fixed order turn the raw data into model inputs, run the long-term average model with its tests and adjustments, then the month-by-month model, and hand everything over through an open code repository, a viewer and a workshop.

What it means. Every figure in this review was regenerated from those steps on 14 September 2026. MMIG staff can do the same.

Source: the project's documented processing chain.

2 Data and methods

The model in numbers

72 × 223 squares of 1 kmOne layer, 10 208 active squares
180 monthly periodsMonth-by-month simulation 2010–2024
300 river reachesRiver computed along its channel, 40–500 m wide
381 wellsObservation wells with a long-term mean level
17 849 monthly observationsMonthly level observations, 144 wells
< 1 s per long-term runEnables thousands of test runs

Built with MODFLOW 6, the standard open-source groundwater code, and open-source tools throughout.

3 Results

The simulated groundwater surface follows the valley

The simulated groundwater surface follows the valley

Finding. With average 2010–2024 inputs the model reproduces the regional gradient from about 900 m near Ravathoji to about 330 m at Khazara, with the flat central plain around Kattakurgan and steeper slopes on the flanks.

What it means. The regional picture is right. The question is how well individual wells are matched, and where not.

Source: data/model/regional_ss_workspace/zrb_regional_ss.hds (run of 2026-09-14).

3 Results

Where the model is too high or too low

Where the model is too high or too low

Finding. Against 381 wells the untuned model is 7 m too low on average; the typical error at a well is 19 m, and the largest errors push the overall error measure to 30 m. Red dots south of Samarkand are too high; the dark blue cluster on the northern edge is 30 to 60 m too low.

What it means. The pattern is structural: the northern cluster lies outside the mapped river deposits, in the foothill belt the model treats as almost dry.

Source: model water levels at the well squares; wells from calibration_targets_steady_state.csv without the flagged inconsistent wells (measured here).

3 Results

Which assumptions matter for the water levels

Which assumptions matter for the water levels

Finding. 5 000 combinations of model settings were run, 3 091 of them successfully. Five settings control the fit: how easily water moves through the upper and middle terraces, the evaporation rate from shallow groundwater, recharge from rain, and aquifer thickness. River-bed leakiness, the fixed edge levels, bedrock and channel roughness barely matter.

What it means. Water-level measurements cannot pin down the river settings at all; that needs river-flow data. And no combination brings the typical error below about 15 m.

Source: scratch/lhs_results.csv (test-run sample); influence ranking against the typical error in the young river deposits (measured here).

3 Results

No version beats the April model yet

No version beats the April model yet

Finding. The best of 3 091 tested settings brings the typical error down to 16 m, but with levels now too high on average. The set found by automatic tuning software, re-run in the full model, is worse than the April model at 27 m. The September candidate with the full register pumping sits at 25 m.

What it means. Tuning the settings has reached its limit. The remaining error comes from the model's structure and from the observation wells themselves, not from the setting values.

Source: the executed processing steps of 2026-09-14; scratch/lhs_results.csv; scratch/ss_rapid_v1_kdown_sfr_connectivity/scenario_metrics.csv.

3 Results

Step by step towards a more realistic concept

Step by step towards a more realistic concept

Finding. Five changes to the model concept were stacked on the April model in May 2026: canal intakes taken out of the river, canal and field losses counted in full with shallow drainage channels, a thicker aquifer, and the full register pumping with local return flows. Each step after the first worsened the fit to the wells, and the river outflow at Khazara stayed at 50 to 59 m³/s against an observed 23.

What it means. Two lessons: the excess river flow comes from the fully counted losses that the drainage channels return to the river, and the register pumping cannot be sustained where the model places it.

Source: scratch/ss_rapid_scenarios and scratch/ss_rapid_v1_kdown_sfr_connectivity scenario_metrics.csv (May 2026); observed outflow from the water-balance constraints.

3 Results

Fifteen years month by month

Fifteen years month by month

Finding. The month-by-month model runs all 180 months without any part drying out and with a water-accounting error below 0.02 percent. Against 15 491 observations, leaving out the first two years of settling-in, the typical error is 9.5 m. Some wells are matched closely, others are offset by several metres, and one shows the seasonal swing the others lack.

What it means. The model can already say how the regional water table responds to wet and dry years. It cannot yet reproduce the seasonal rhythm of most wells.

Source: data/model/transient_residuals.csv (run of 2026-09-14).

3 Results

The seasonal swing is too small

At 139 wells the simulated yearly rise and fall of the water table is on average a fifth of the observed one. Only a handful of wells swing too much.

Five hundred month-by-month runs traced this to the storage settings: less releasable storage in the river deposits and terraces brings the seasonal rise and fall closer to the observations.

The seasonal swing is too small

Source: Yearly high minus low per well, averaged over 2012–2024 (measured here from the month-by-month results); 480 of 500 test runs succeeded, best seasonal error 0.23 m against 0.39 m for the starting values.

3 Results

What controls the seasonal rise and fall

What controls the seasonal rise and fall

Finding. Five hundred month-by-month runs varied the storage and flow settings together; 480 succeeded. How much water the recent river deposits and both terrace generations can release from storage controls the seasonal rise and fall, with a smaller opposite effect from how easily water moves through the middle terraces.

What it means. The published storage values are too high for this valley. Month-by-month data pin down what the long-term average model cannot see, so the next tuning must use both.

Source: scratch/lhs_transient_results.csv; influence ranking against the seasonal rise-and-fall error in the young river deposits (measured here).

3 Results

The river gains groundwater most of the year and loses it at the irrigation peak

The river gains groundwater most of the year and loses it at the irrigation peak

Finding. Over the whole river and all fifteen years the aquifer releases about 10 m³/s to the Zarafshan in the gaining months. In June and July, when canal water raises the water table, the direction reverses and the river loses up to 30 m³/s to the aquifer.

What it means. This seasonal reversal is the headline result for management. Its magnitude is uncertain because it is constrained by levels only; the measured river flow at Khazara is the missing check.

Source: data/model/regional_tr_workspace_p2 water accounts, river exchange summed over all sections (run of 2026-09-14; measured here).

4 Issues, discussion and improvements

Where the error sits

Where the error sits

Finding. 153 of the 381 observation wells, forty percent, are single-measurement wells in the northern foothill belt outside the mapped river deposits. Their typical error is 20 m too low and their spread the widest of any group. Wells in the river deposits and the terraces are centred near zero.

What it means. The model was built for the valley aquifer. It is being judged, to a large degree, on wells in a different unit that it represents with borrowed properties and no recharge.

Source: candidate well errors joined to calibration_targets_steady_state.csv (map_geology); diagnosis in the project issue log (GAPS.md).

4 Issues, discussion and improvements

What is missing in the data

Pumping

  • Coordinates and rates of production wells for Samarkand and Navoi; the register has them only for Bukhara
  • Annual abstraction series 2010–2025 by district; only the 2026 status exists
  • Piped exports to Navoi and Bukhara and sewer returns
  • Agricultural and unregistered pumping: no data at all

Aquifer geometry

  • Bedrock depth: no map; the model uses one uniform thickness
  • Filter depths give only lower bounds
  • No pumping-test data for the foothill belt and the oldest deposits

Surface water

  • Reliable river flow below the Ravathoji intakes only for 2014–2017
  • Canal deliveries per system, not per field
  • Khazara river flow used as a yearly average; monthly series needed for checking

Observations

  • 250 of 398 wells have a single measurement
  • Several wells share one 1 km square with levels differing by up to 74 m
  • 47 wells report levels above land surface: elevations need checking
  • No soil-moisture or tracer data for the soil between the surface and the water table

4 Issues, discussion and improvements

Four conceptual problems

Pumping placement. The register magnitude is plausible, but placed by population the model lowers water levels around Samarkand by 31 m on average, a 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.

The foothill edges. The northern and southern margins are closed to flow and receive almost no recharge, yet forty percent of the observation wells sit there. Recharge from the mountain front, a separate zone for the older rock, or a separate treatment of those wells is needed.

Recharge as a prescribed number. Canal and field losses reach the aquifer through a soil layer the model does not represent. The assumption that 45 percent seeps through, the 90/50 split of canal and field losses and the 2 m drainage channels are bookkeeping choices, and the step-by-step tests show they decide both the fit to the wells and the river outflow.

River sections longer than one square. Their leakage is forced through a single square, distorting the exchange locally. The fix is known and must come before any tuning against river flow.

4 Issues, discussion and improvements

The missing link: the soil between canal and aquifer

In this valley the aquifer is fed mainly by irrigation. How much of the diverted water seeps down, how much evaporates and how long it takes to arrive depends on soils, crops, canal lining and the state of the water table. The present model replaces all of that with fixed fractions.

The setting tests and the step-by-step concept tests both point at these fractions as what controls the fit to the wells and the river outflow. No amount of tuning the aquifer settings can compensate for a wrong recharge input.

Simulating the land surface and the canal system explicitly, and letting the aquifer model receive recharge from them, is the natural next stage. This is what was proposed and implemented for the Chu River basin plan.

4 Issues, discussion and improvements

How it was done in the Chu basin: one water cycle, three engines

How it was done in the Chu basin: one water cycle, three engines

Finding. In the Chu basin plan the natural hydrology (SWAT+), the managed canal network (TaqSim) and the valley aquifer (MODFLOW 6) are three engines that hand water to each other every day. Water seeping through the soil and canal losses become groundwater recharge; river gains and drain returns flow back to the river.

What it means. The Zarafshan valley has the same anatomy: mountain inflow, a diverted river, canals, irrigated plain and a shallow aquifer. The pattern transfers directly.

Source: Chu River Basin Plan 2027–2031 modelling showcase, hydrosolutions 2026; figure reused unchanged.

4 Issues, discussion and improvements

The Chu water system as one picture: which engine owns what

The Chu water system as one picture: which engine owns what

Finding. Snow and glacier melt, the reservoir, canals and irrigation, pumping, recharge, groundwater flow and the springs that return water to the river, drawn as one block of landscape. The brackets show the reach of each engine and the coupling zone where they hand water to each other.

What it means. Replace the Kyrgyz Ala-Too with the Zarafshan headwaters and the Chu with the Zarafshan, and the picture is the one this valley needs.

Source: Chu River Basin Plan 2027–2031 modelling showcase, hydrosolutions 2026; illustration reused unchanged.

4 Issues, discussion and improvements

How the Chu engines are wired together

How the Chu engines are wired together

Finding. Every transfer between the engines is booked once, with source, receiver and date, and the groundwater account closes to a tiny remainder. The aquifer model receives simulated recharge instead of a prescribed array, and the river receives simulated gains and losses.

What it means. The linking code, the accounting checks and the tuning workflow exist and are documented. Reusing them for the Zarafshan is an engineering task, not new research.

Source: Chu River Basin Plan 2027–2031 modelling showcase, hydrosolutions 2026; figure reused unchanged.

4 Issues, discussion and improvements

Proposed next stage for the Zarafshan

Proposed next stage for the Zarafshan

Finding. A land-surface and soil model for the irrigated plain and the mountain catchments, an allocation model of the Ravathoji diversions, canals and reservoirs, and the existing MODFLOW 6 model as the aquifer engine. Recharge, water loss to the air and river exchange become computed exchanges.

What it means. The regional model built so far is not lost: it becomes the groundwater engine of the coupled system. The long-term average work identified exactly the inputs the linked system would replace.

Source: Pattern: the SWAT+ – TaqSim – MODFLOW 6 coupling of the Chu basin plan (hydrosolutions, 2026); engines for the Zarafshan are a proposal, not an implementation.

4 Issues, discussion and improvements

What was agreed with MMIG in Tashkent, May 2026

What was agreed with MMIG in Tashkent, May 2026

Finding. The working session fixed five management scenarios and their combinations, the output table each run has to deliver, and two data points on pumping: the population-based estimate is only the drinking and industrial part, and agricultural pumping needs MMIG's own data. It also flagged the systematic bias between simulated and observed levels as the first problem to solve.

What it means. The scenarios are the questions the model must answer; the bias diagnosis in this review is the work that makes the answers trustworthy.

Source: Whiteboard of the Tashkent working session with MMIG, May 2026 (photograph).

4 Issues, discussion and improvements

Five scenarios and their combinations

Each run reports river inflow, river outflow, the piezometric water level from the model and its change against 2024. In the long-term average runs inflow equals outflow; in the month-by-month runs the difference is the change in the water stored in the aquifer.

ScenarioDriverChange tested
S1River and climateRiver inflow at Ravathoji reduced from 140 to 100 m³/s
S2PopulationPopulation of about 3 million grows by 20 %; domestic and industrial pumping scale with it
S3AgricultureGroundwater pumping for irrigation reduced by 40 m³/s
S4ClimatePrecipitation halved
S5ClimateEvaporation demand of the air increased by 25 %
S1 + S2CombinedLess river inflow and more people
S1 + S2 + S3CombinedLess river inflow, more people, less irrigation pumping

Not yet run: under the project workflow, scenario runs start once the long-term average model passes its readiness check on a cleaned set of observation wells (chapter 4, improvements 1–3).

4 Issues, discussion and improvements

Required improvements, in order

1 Settle the pumping record

  • Obtain the annual abstraction series 2010–2025 and well-field coordinates from MMIG
  • Bound the pumping over the model period between the 2010 and 2026 registers
  • Route well-field exports and city sewage explicitly instead of local return

2 Clean the set of observation wells

  • One value per model square, weighted by well count and spread
  • Exclude edge and above-ground wells; separate static from time-series wells
  • Report all metrics on the cleaned and the full well set

3 Fix the foothill edges and the river

  • Recharge from the mountain front and a zone for the older rock along the northern edge, or a separate aquifer for those wells
  • Split river sections at the square boundaries
  • Check river outflow against monthly Khazara flow

4 Link the soil and canal system

  • Set up SWAT+ for the plain and the mountain catchments and an allocation model for the canals
  • Pass recharge and river exchange to the existing groundwater model daily, as in the Chu basin
  • Only then: joint tuning to levels, seasonal rise and fall, and river flow

5 Conclusions

Conclusions

The Zarafshan valley aquifer is under measurable pressure: more people, unmeasured abstraction and a run of dry years, with falling levels at a third of the monitoring wells.

A regional groundwater model of the valley exists, built from every available dataset, fully reproducible, and already handed over to MMIG with training. It reproduces the regional water table and the seasonal reversal of river–aquifer exchange.

Its accuracy is limited by structure, not by settings: the pumping record, the foothill wells, and above all a recharge that is assumed rather than computed.

The next stage is clear. Settle the pumping record with MMIG, clean the set of observation wells used for tuning, and link the aquifer model to a soil and canal model as in the Chu basin. The five scenarios agreed in Tashkent are then run and reported in the agreed table, and the model becomes a tool for permit and monitoring decisions in the valley.

5 Conclusions

A working platform, ready for its next stage.

hydrosolutions GmbH · Zurich · September 2026

1 / 1