FloodMAR Somerset — scheme assessment (v2)

Working draft, started 2026-08-09. Standalone: this supersedes papers/somerset_scheme/somerset_scheme.md rather than amending it.

How to read the parameter tables

Every parameter carries a confidence mark and, where it matters, a critical-path star.

A star is not “important”. It is “important and shaky”. A measured parameter the answer depends on heavily is not starred, because there is nothing to fix.

1. Introduction

The scheme diverts high winter flows from the River Tone into surface storage, treats the water, and injects it into aquifer storage for recovery in summer. The question is whether that is worth doing, and in what configuration.

The scheme is a seasonal battery, and that framing does more work than it first appears. Its value is not that it saves water — most water “lost” from an unconfined store re-emerges as river baseflow within the same catchment. Its value is that it moves water from February, when the catchment has too much, to July, when it has too little. Any loss mechanism that returns water to the river in winter therefore costs the scheme almost everything, even though it costs the catchment nearly nothing.

This is the most useful idea for reading the rest of the report, and Section 5 shows it deciding a headline number.

What changed from v1

v1 was assembled while several parameters were still being revised, and its appendix documents values that several of its own runs did not use. The drift is set out in reports/status_parameter_audit.md. v2 starts from the scenario files as the source of truth and states every parameter’s provenance.

Substantively, three things are new:

1. Published aquifer properties were read (BGS WD/97/34) and the aquifer parameters re-derived per aquifer rather than globally. See reports/status_aquifer_properties.md and status_aquifer_params.md.
2. A parameter error was found and fixed — the sandstone store was running on a Chalk recession constant. Section 5.
3. The binding constraint was identified, and it is not the one the design discussion has focused on. Section 6.

2. The local area

The modelled store sits near Taunton in the Vale of Taunton, in the Tone catchment. That matters for the reference data: BGS WD/97/34 treats the Tone catchment as one of two study areas in its South-West England chapter, so the published values are regionally apt — with one large caveat repeated throughout this report.

Geology within 15 km of the store

From the BGS 625k hydrogeology polygons, grouped by unit and productivity class:

Two things follow immediately.

The store sits on mudstone, not on aquifer. The Class 2C polygon at 0.0 km is Mercia Mudstone; the nearest productive Sherwood sandstone is 2.7 km away. This is independent corroboration, from mapping, of what the borehole logs found directly.

TRIASSIC ROCKS (UNDIFFERENTIATED) is not one thing. It covers both the Class 1A sandstone and the Class 2C mudstone, and within 15 km the mudstone is nearly four times the larger. Any calculation grouping by rock unit alone overstates the aquifer roughly fivefold. CLASS, not ROCK_UNIT, is the field that matters.

Direct evidence

The BGS index holds 2177 boreholes within 15 km of the store. That figure should not be read as an evidence base:

Median depth, over the 1879 holes that record one, is 6 m. Of 2177 boreholes, exactly one is both near the productive sandstone and deep enough to have a realistic chance of proving a section. The index is overwhelmingly shallow foundation and water-well records.

26 scanned logs were shortlisted from it; 5 transcribed, 21 still unread — including every deep record. Of the 5:

• none reach Sherwood/Otter-type sandstone
• thickest proven low-permeability cover is 76.2 m, base not reached in any hole
• no salinity is recorded in any of them

The negative result is real but weak: every one of those holes stopped inside the Mercia Mudstone, so they establish that the cover is thick and continuous, not that the sandstone beneath is absent. See reports/borehole_log_review.md.

3. Water

Inflow, diversion and the river-support constraint.

STUB — carries over from v1 pending re-run. Needs restating against the current gauge set and the exceedance-based low-flow threshold.

4. Reservoir

Surface storage: siting, stage curves, efficiency.

v1’s candidate list mixed two resolutions. Sites inside the 1 m LiDAR box were screened on fine terrain and labelled S; sites outside it were screened on 50 m coarse terrain and labelled C, and survey_c1.py had to carry a measured coarse-vs-fine bias correction to compare them at all.

That S/C split was an artefact of data coverage, not a property of the sites. The box was extended to E 315000–360000, N 115000–145000 — the largest hole-free 1 m extent available, 54 of 54 quarter-tiles present, containing the full 1 km footprint of every candidate in both lists. Measured LiDAR coverage over the box is 100.0%. Every site is now screened identically, there is no bias correction, and there is no C tier: one list, one area, one resolution.

Of the 1350 km² box, 346.2 km² (25.6%) is excluded by statutory designation or built-up area, leaving 1003.8 km² searchable. 2852 footprints reached the 5.0 Mm³ target.

The candidates

The coverage gap was hiding the best site

The new top-ranked site, S1 at E353505 N124745, is C1. C1 sat at E353525 N124725 — the same ground, now resolved at 1 m instead of inferred from 50 m terrain plus a bias correction.

It is not marginally better, it is better on every axis that matters:

Three of the top eight — S1, S5 and S8 — lie in the eastern ground that had no LiDAR before and were therefore invisible to the old fine screen. Three western sites from the old list (E324505 N141495, E333255 N135745, E328755 N139995) drop out on equal-footing comparison.

This is the clearest possible vindication of removing the coverage artefact: the previous “best fine-screened site” was best only among the sites we happened to have terrain data for, and the modelled store sits on the second-ranked one.

What this is not. Terrain screening only — the ground is the right shape and nothing more. No ground investigation, peat depth, seepage, land ownership, agricultural classification, archaeology, flood-risk sequential test or road diversions. S4 carries a conveyance route crossing an SSSI/SAC. Separately, no site was screened for minor watercourses, which remains an open gap from v1.

5. Aquifers

The model no longer carries one global set of aquifer numbers. Values are anchored per aquifer on published regional data and adjusted only where our own data can justify it.

Method

The reference gives good regional values and states how they vary across a region; it does not give site values. Our own data is far closer to the site but measures different things. So the two are used for different jobs, and neither is allowed to do the other’s:

Values stay inside the published range unless there is a specific local reason to leave it. No departure is made silently.

Sherwood Sandstone (sandstone_local)

Why the lower half of the published range. WD/97/34 records core porosity falling steadily northward — over 33% in the southern Otter Sandstone outcrop to “generally less than 20% in the north of the region” — as cementation increases. Our polygons sit at the northern end of that gradient, in fact beyond the northern limit of the figure’s own data. Reading the lower half is a defensible reading of the reference’s stated direction. It is not a measurement at our northing, and is marked as an extrapolation.

The caveat governing this whole section. The reference’s data for this aquifer is overwhelmingly from the Otter Valley, 30–50 km south. The reference itself says there has been “little work in the Tone catchment”. Every hydraulic number above is a regional value imported into an area its authors flag as data-poor, along the direction of their own stated gradient. They are better read as upper bounds for the Tone catchment than as central estimates.

The leakage correction, and why it matters less than it looks. v1 ran this store at 0.0167 /day, the observed median Chalk recession constant. That is legitimate for a Chalk store and wrong for a sandstone one. Derived from this aquifer’s own T and Sy it is 0.0018 /day, about nine times lower.

Effect on 06_sited_S1, Mm³/yr:

Read the last row before the others. The correction moves total useful output by 2%. What it does is reallocate — river baseflow into summer irrigation recovery. This is the seasonal-battery point made concrete: 80% of leakage was already returning to the river, so the leaky store was not wasting water, it was returning it in the wrong month.

One qualitative change is worth stating separately. On the Chalk constant the sandstone store never once held enough water to support the river across 16.5 years. v1 reported that as a null result. It was a parameter error.

Chalk (chalk_confined)

Confined stores are limited by pressure, not volume. Injecting V into footprint A raises head by dh = V / (A·S), and at S = 0.0052 that rises fast. The model now enforces a second ceiling, A·S·dh_max, and reports which ceiling binds.

The target body is the White Chalk Subgroup, 2143 km², nearest edge 34.9 km from S1 — Salisbury Plain, matching both the 42 km conveyance and the reference section. Over a 180-day season the pressure mound spreads to radius sqrt(4·D·t) with D = T/S = 269,000 m²/d, giving ~13.9 km and ~609 km². At that footprint the head-limited capacity is 95.9 Mm³, so volume binds at 20 Mm³ and head does not. Head rise storing 20 Mm³ is 6.3 m. The conclusion survives down to 128 km² — 6% of the body — so it is not sensitive to the footprint estimate.

The declaration therefore changes no result today. It is in place so the constraint is enforced if capacity, footprint or aquifer choice change later.

Not yet in the model

Upper Greensand: 219.5 km², Class 1B, 6.5 km from the store. A real local aquifer, closer than the Chalk and larger than the local Chalk outcrop, with no representation in the model and no reference anchor read (WD/97/34 Ch.5 covers Lower Greensand). The most significant known omission on the aquifer side.

6. Plant and energy

The binding constraint

The most consequential measurement in the report, and it is ours rather than the reference’s. Over 6025 days of 06_sited_S1:

Aquifer headroom never binds — not once in 16.5 years. The scheme is throttled by the rate at which water can be put into the ground, three days in four. It is not throttled by how much the aquifer can hold.

That reframes much of the design discussion. Debate over 10 vs 20 vs 30 Mm³ of aquifer capacity has been debate about a parameter that does no work in the model. Meanwhile recharge.max_daily_m3 — the parameter that binds most of the time — is assumed, with no source behind it. It is the clearest ★ in the report: highest sensitivity, weakest provenance.

Recovery capacity against the published resource

Recovery is 30,000 m³/d per store, so every two-store configuration — every portfolio and sited run, i.e. the preferred designs — pumps 60,000 m³/d. WD/97/34 p.286 puts the resource of the whole Tone catchment at about 17,500 m³/d, “most of which currently supports environmental flows”. The scheme total is 3.4× that.

This is not a straight contradiction. MAR recovers water it injected; it is not drawing on the natural resource in the same sense, and the two figures are not directly commensurable. Two objections survive:

1. Deliverability — whether the aquifer can physically pass 60,000 m³/d to pumps is a transmissivity question, and these transmissivities are not large. Never checked.
2. Licensing — a regulator assessing abstraction in a catchment whose published resource is 17,500 m³/d, most of it committed to environmental flows, will not treat a 60,000 m³/d application as a formality.

Energy — STUB. Not modelled. Conveyance lift, injection pressure and recovery pumping all carry energy costs no run accounts for. The 42 km chalk option is the obvious candidate for this to change a ranking.

7. System recommendation

STUB — pending the LiDAR re-run and a consistent re-run of all scenarios.

Nothing here should be written until every scenario has run on one parameter set. Six runs currently carry superseded values (reports/status_parameter_audit.md) and the site list is being rebuilt.

8. Discussion: relieving the injection constraint

An idea, not a result, and deliberately not modelled.

Section 6 establishes that injection rate binds 76% of days while aquifer headroom never binds. That invites an obvious question: can the injection rate be raised?

Direct injection is pressure-limited. Pushing water down a well into a confined unit fights the storage coefficient — compressing water and rock, head rising fast, ceiling set by what the confining layer tolerates before it leaks off or fractures.

Surface spreading is not. Infiltrating into an unsaturated shallow unit fills pore space at a specific yield of order 0.1 rather than a storage coefficient of order 0.005 — roughly twenty times more volume per metre of head — with no injection-pressure ceiling. Water then percolates to the deeper unit over months, which is the slow-release behaviour the scheme wants.

Why this is not in the model, and should not be. The physics is favourable; the practicality is unexamined, and the practicality decides it:

• Land. Spreading needs area of the right sort — permeable, flat, available, not already doing something. A rate calculation assuming a square kilometre is available is not an engineering case, it is arithmetic.
• It only works where an outlet naturally serves the ground. Water has to reach the spreading area by gravity from somewhere the scheme already puts it, or pumping cost eats the gain.
• New loss terms. Evaporation from open basins, lateral shallow flow away from the target unit, clogging of the infiltration surface — none represented in the model.
• It is not a parameter change. Raising recharge.max_daily_m3 and declaring the problem solved would be wrong. Spreading changes which aquifer is the right target — favouring shallow unconfined units like the Upper Greensand over the deep confined Chalk — so it changes the optimal portfolio, not one number.

How it should be evaluated. As its own scenario family, with an explicit second recharge mechanism alongside well injection, its own loss terms, and a candidate spreading area identified from terrain and land-cover data rather than assumed. It would then be compared on the same footing as the well-injection portfolios. Adding it as a parameter tweak would import a large amount of uncertainty for an apparent gain, which is the wrong trade.

Note also that raising injection alone does not run far: reservoir volume already binds on the other 24% of days, so the gain saturates. Any serious version couples spreading to reservoir sizing.

9. Known unknowns

Ordered by how much they could move a conclusion.

Items 1, 4 and 9 are starred: the answer is sensitive to them and their provenance is weak. Items 2, 3 and 8 are the geological questions, and only drilling or reading the remaining logs moves them.

← All notes · More from Aquifer Storage