Lifting 800 m3/d of Brine Makes Room for 0.2 Mt/yr More CO2 in a Closed Aquifer

On a synthetic closed saline aquifer, each 800 m3/d of brine that a pressure-management well lifts lets the injector put away about 0.2 Mt/yr more CO2 for five years, with the high-case pressure forecast still under the limit. Per cubic metre of brine, that is 679 kg of CO2, which at the unit's conditions is close to one reservoir volume of CO2 for each volume of water removed: what a closed tank predicts. The number matters because the storage projects that have run into trouble have run into it through pressure, and the wells that relieve pressure are the ones whose value is hardest to state.

Gorgon's limit has been pressure

Gorgon is the Chevron-operated LNG plant on Barrow Island, off the Pilbara coast of Western Australia. Its feed gas carries a lot of CO2, which the plant separates and injects into the Dupuy Formation, a saline sandstone more than 2 km below the island. Chevron's 2010 plan expected injection to cut the project's emissions by about 3.4 to 4 Mt a year of CO2 equivalent, through nine injectors from three drill centres, with four water production wells to manage reservoir pressure [1]. Injection began in August 2019, and more than 11 Mt had gone in by May 2025 [2]. In the 2024-25 financial year the system stored 1.33 Mt, its lowest year yet and about a third of its 4 Mt design rate [3].

The constraint is pressure. Reservoir pressure in the Dupuy has to stay within a range, so injection has been held back [4]. The water production wells that were meant to relieve it have produced sand, and Chevron's current optimisation work modifies those four wells to extract more water, to lower reservoir pressure and let injection rise [2][5]. The 2010 plan did not aim to balance offtake against injection; it expected a large connected aquifer to absorb much of the pressure increase [1]. A confined unit absorbs nothing. Its capacity is set by pressure rather than by pore volume [6], and in it the question of what a water well is worth comes first.

An operator needs two numbers from a storage forecast. How many megatonnes a year can the site take for the next N years with the high-case pressure under the limit? And what does one more pressure-management well add to that?

Where a producer enters the pressure model

An injection site is a tank with a mass balance and a pressure history. The pressure history sets the limit. In our MRV module, bottom-hole pressure at a well is the initial pressure, plus a closed-tank term, plus a Cooper-Jacob transient for the well's own rate history [7]. The tank term is the net volume injected divided by total compressibility times pore volume.

A brine producer enters the model in two places. The brine it lifts leaves the tank, so the tank term falls. Its rate history also adds one more transient, at its own distance and with the brine viscosity:

$$P(t) = P_0 + \frac{V_{CO_2}(t) - V_{brine}(t)}{c_t V_p} + \frac{\mu}{4\pi k h}\left[A(t) + q(t)\ln\frac{2.25\,k}{\phi\,\mu\,c_t\,r_w^2}\right] + \sum_p \frac{\mu_p}{4\pi k h}\left[A_p(t) + q_p(t)\ln\frac{2.25\,k}{\phi\,\mu_p\,c_t\,r_p^2}\right]$$

FormulaLaTeX

Copy LaTeX

Read left to right: the starting pressure, the tank filling, the injector's own transient, and one transient per producer, where A(t) carries the superposed log-time terms of each rate change. For fixed rock properties every term is linear in the rates. That is what makes strategy search cheap: each ensemble member's response to a unit rate at each well is computed once, and any new schedule of injection and offtake is then one matrix product across the whole ensemble. The permeability inside the logarithm is not linear, so it is evaluated member by member.

A synthetic unit with a producer in its data

Our test case is the MRV module's synthetic unit [8]. It has two injectors on a closed 50 km2 saline aquifer at 1,800 to 1,850 m. We added one brine producer 800 m from the gauged injector, running at 800 m3/d from day 300. The same forward model that does the match generates the pressures, with 0.03 MPa of gauge noise and a true permeability of 40 mD.

The match is an ensemble smoother with multiple data assimilation [9]. The check script holds 48 assertions, and three carry the weight. The match with the producer in the data recovers 39.97 mD, with a P10 to P90 of 39.68 to 40.27. Dropping the producer's rows from the same export gives a worse fit, so the offtake carries information. And adding a planned producer lowers every ensemble member's forecast pressure by the closed-tank plus Cooper-Jacob amount, within 2%.

The generator and the analysis share the same physics. So this shows the code works. It does not validate the model against a real aquifer.

After 1,000 days the unit has taken 1.64 Mt of CO2, known to within 1.2% from metering uncertainty, and the producer has lifted 560,000 m3 of brine. At the last reading the matched well sits 4.44 MPa above initial pressure from the tank and 1.25 MPa from its own transient, and the producer takes 0.84 MPa off. The Cooper-Jacob validity parameter for the producer is 0.007, under the 0.01 threshold, so its term can be taken at face value; a producer further away, or one that started later, would be flagged.

The pressure limit is 0.9 of the minimum horizontal stress at the caprock, moved to the gauge, which gives 27.48 MPa. The stress gradient is assumed, because only a mini-frac or an extended leak-off test measures it. The injector's recent rate is 1,302 t/d, about 0.475 Mt/yr. At that rate the P90 pressure crosses the limit at 2.5 years, and every ensemble member is over it by year five. The site cannot keep doing what it is doing.

Pricing a producer

The history match is the expensive step, at about 10 seconds, and we run it once and store the ensemble beside the run record. Each strategy then loads the ensemble rather than rematching, and searches by bisection, in about a second, for the constant injection rate at which the chosen pressure percentile just touches the limit. Over five years, with the P90 pressure under 27.48 MPa:

Strategy	Planned brine offtake, m3/d	Sustainable rate at the P90 pressure, Mt/yr	Same rate, t/d
As is, the existing producer continues	0	0.283	776
One more producer at 1.5 km	800	0.481	1,318
Twice that offtake, at 1.5 km	1,600	0.679	1,860
Three times, at 1.5 km	2,400	0.878	2,404

The first row is the hard one: the injector is running at the equivalent of 0.475 Mt/yr, and the unit can sustain 0.283 for five years. One more producer at 800 m3/d more than closes that gap. On the fitted line, holding today's 1,302 t/d for five years takes about 775 m3/d of new offtake.

The steps are equal. Each 800 m3/d adds 0.198 Mt/yr, and across all 13 offtakes in the grid the rate is a straight line in offtake, with a slope that works out to 679 kg of CO2 per cubic metre of brine lifted. CO2 at the unit's 70 C has a density of 705 kg/m3 at the pressure the forecast starts from and 764 kg/m3 at the limit [10], so the producer frees room for 0.89 to 0.96 of a reservoir volume of CO2 per volume of brine. A closed tank swaps them one for one. The shortfall comes from the transients. A tonne more injection raises the gauge pressure through the tank and through the injector's own transient, evaluated at the wellbore; a cubic metre more offtake relieves it through the tank and through a transient evaluated 1.5 km away, which is smaller.

Planned brine offtake at 1.5 km: 800 m3/dHorizon: 5 years

pressure read at:P90, the high caseP75P50right panel:over the horizonpressure band

P90 pressure over 5 years: 0.283 Mt/yr with no new offtake, 0.481 with one 800 m3/d producer. Each m3 of brine lifted makes room for 679 kg of CO2: 0.89 to 0.96 of a reservoir volume at 70 C. Holding the injector's 1,302 t/d (0.475 Mt/yr) for 5 years needs about 775 m3/d of new offtake. Past 7.2 years one producer adds more than the unit sustains without it.

The exhibit opens on that five-year, P90 case. Move the horizon and the less obvious result appears. The unit's own headroom is a fixed pressure margin spread over the horizon, so its sustainable rate falls from 0.985 Mt/yr over one year to 0.137 over ten. A producer's contribution does not fall with it, because the brine it lifts accumulates for exactly as long as the CO2 does: one 800 m3/d producer adds 0.186 Mt/yr over one year, 0.198 over five and 0.194 over ten. Past 7.2 years, one producer adds more sustainable rate than the unit can take without it. Over a ten-year permit, the pressure-management well is worth 1.4 times the aquifer's own spare capacity.

The one-year figure is lower for a physical reason. At 1.5 km the producer's transient has not fully reached the injector within a year, so in the first year the grid gives 636 kg of CO2 per cubic metre of brine rather than 679.

The percentile control shows something less comfortable. Reading the forecast at P90 instead of P50 costs 0.001 Mt/yr here, against 0.198 for a producer, because the P10 to P90 band on the as-is forecast is 0.039 MPa wide at year five. That is true of this synthetic case only. It is parameter uncertainty under white gauge noise, and a real site's band is wider than the model knows. On a real aquifer the band also carries what the model leaves out, and once it is wide enough the choice of percentile is worth more than a well. That is why the regulator's reviewer should ask which percentile the forecast was read at. A history match that narrows its band as data arrive is not necessarily getting closer to the truth [11].

Our most trusted check is analytic. With one enormous permeability and no spread, the transients vanish and only the tank remains. The sustainable rate then has a closed form, and the bisection matches it to within 1%.

A Gorgon-shaped case, labelled as one

We built an illustrative case on public Gorgon figures: the Dupuy at more than 2 km below the island, the Basal Barrow Group shale above it, nine injectors from three drill centres, four water production wells, and first injection in August 2019 [1][2]. Everything not in a public document is an assumption, and the interface labels it as one. That covers depth to base, net-to-gross, porosity, permeability, area, initial pressure, temperature, the fracture gradient and the achieved water rate. The monitoring series are synthetic, generated from those parameters.

In this example the three drill centres inject 1.75 Mt/yr together and the four producers run at half their design water capacity. The five-year sustainable rate at the P90 pressure is 1.70 Mt/yr. Two more producers at 3,000 m3/d each take it to 2.91. These are not Gorgon's numbers. They show the shape of the answer, and an operator with the real monitoring export could get the real one in about a second per strategy.

All of these rates hold the reservoir at constant temperature. CO2 that reaches the rock colder than the formation lowers the stress the limit is computed from and would cut every rate here; that is a separate calculation, and it is not in these numbers.

A regulator can replay every run

Each strategy is a scenario of the stored match, and its run record lets a reviewer rerun it from the stored inputs and check that the result comes back identical. A scenario that changes anything other than the strategy is rematched, and the record says so. A storage permit that rests on a pressure forecast needs this: the number, and the means to get it again.

What to ask of a storage forecast

1. At which pressure percentile is the forecast read against the limit, and what rate keeps it under the limit for the permit period?
2. What does one more pressure-management well add in Mt/yr, and is it close to volume for volume?
3. Can the forecast be rerun from its stored export and parameters to the same result, and does it say which parameters are measured and which are assumed?

Key takeaways

1. 1A brine producer enters the pressure model twice: as volume out of the shared tank, and as a Cooper-Jacob transient at its own distance.
2. 2On the synthetic unit, the match with a producer in its data recovers 39.97 mD against a true 40 mD; the generator and the match share the same physics, so this checks the code, not the model.
3. 3The injector's recent 1,302 t/d is 0.475 Mt/yr; the unit sustains 0.283 Mt/yr for five years at the P90 pressure, and about 775 m3/d of new offtake would let it keep going.
4. 4Each 800 m3/d producer adds 0.198 Mt/yr over five years: 679 kg of CO2 per cubic metre of brine, 0.89 to 0.96 of a reservoir volume at the unit's 70 C.
5. 5The producer's yield stays between 0.186 and 0.199 Mt/yr from one year to ten while the unit's own rate falls from 0.985 to 0.137, so past 7.2 years one producer adds more than the unit holds alone.
6. 6P90 against P50 is worth 0.001 Mt/yr here only because the synthetic band is 0.039 MPa wide; on a real site the percentile can be worth more than a well.

Where the model is weakest

The aquifer is closed by construction. A unit connected to a larger aquifer bleeds pressure away, which raises the sustainable rate and lowers what a producer is worth, and Gorgon's own 2010 plan assumed exactly that kind of connection [1].

The data are generated by the same physics that fits them, with white gauge noise. The 39.97 mD and the narrow band show the inversion works; they say nothing about how well a tank with Cooper-Jacob transients describes a real formation, and real gauge errors drift and correlate.

The pressure limit rests on an assumed stress gradient. If a mini-frac moved the minimum horizontal stress, every rate in the table would move with it, and the producer's yield, which depends on the tank rather than on the limit, would move much less.

The model is isothermal and has no geomechanical coupling. Cooling near the injector, and the rock's response to rising pore pressure, are both outside it.

Every producer here sits 1.5 km from the gauged injector and produces without trouble. Moving it changes its transient term, and a producer that makes sand, as Gorgon's did, delivers less than its design rate, which is the number this calculation needs as an input rather than one it can supply.

References

1. Chevron Australia. Gorgon CO2 injection project, presentation hosted by Geoscience Australia (GA16243), 2010: nine injectors from three drill centres, four water production wells, the bottom-hole pressure limit, the Basal Barrow Group shale, and the plan's expectation that a connected aquifer would absorb much of the pressure: ga.gov.au/bigobj/GA16243.pdf.
2. Chevron Australia. Gorgon carbon capture and storage fact sheet, 2025: more than 11 Mt injected from August 2019 to May 2025, and the works to expand water extraction from the four water producing wells: australia.chevron.com.
3. K. Morrison (IEEFA). World's largest CCS project hits new low on amount of carbon captured and stored. RenewEconomy, 2 December 2025: 1.33 Mt in FY2024-25 against a 4 Mt/yr design rate: reneweconomy.com.au.
4. Gorgon CCS in Australia falls further from declared target: IEEFA. Rigzone, 29 November 2024: reservoir pressure has to stay within a range, and injection has been constrained: rigzone.com.
5. H. Kombrink. Are the problems around the Gorgon CO2 storage project proof that CCS does not work? GEO ExPro, 17 December 2025: sand production in the water production wells: geoexpro.com.
6. Why a closed aquifer's capacity is set by pressure rather than pore volume: Rate is not capacity.
7. H. H. Cooper and C. E. Jacob. A generalized graphical method for evaluating formation constants and summarizing well-field history. Transactions, American Geophysical Union, 27(4):526-534, 1946: doi.org/10.1029/TR027i004p00526.
8. EarthScan injection guardian and MRV bench, synthetic data throughout, with the public source behind every Gorgon figure: available on request.
9. A. A. Emerick and A. C. Reynolds. Ensemble smoother with multiple data assimilation. Computers and Geosciences, 55:3-15, 2013, the history-matching method: doi.org/10.1016/j.cageo.2012.03.011.
10. NIST Chemistry WebBook, thermophysical properties of carbon dioxide (Span and Wagner equation of state), isotherm at 70 C, 18 to 28 MPa: webbook.nist.gov/chemistry/fluid.
11. Why a history match can narrow its band and drift from the truth: More data, a narrower band, less truth inside it.