What 30 years of a hot-granite doublet look like: lessons from the Utah FORGE geothermal doublet

At Utah FORGE, a well pair about 100 m apart in hot granite, circulated at 10 kg/s, would cool from 207 °C to 126 °C over 30 years in KAI's simulation, with a P10 to P90 of 114 to 143 °C. That makes it a fine test bed and a poor power plant. The useful part is why: what the spacing, the rate, the rock stress and the model's own misfits teach about any enhanced geothermal system.

This guide walks through those results and the lessons. For the general method, spacing, breakthrough and stress limits for any doublet, read geothermal doublet design. Every KAI number below is simulated, from the open notebooks behind the Utah FORGE case.

The site: hot granite under basin fill

FORGE, the Frontier Observatory for Research in Geothermal Energy, is the U.S. Department of Energy's field laboratory for enhanced geothermal systems, near Milford, Utah. Its purpose is to learn how to create and sustain fracture networks in basement rock, and its data are published on the Geothermal Data Repository [1]. Between October 2020 and April 2024 the team drilled two first-of-their-kind wells and created a geothermal reservoir between them [1]: 16A(78)-32, which injects, and 16B(78)-32, which produces.

The FORGE temperature model holds 5.2 km³ of granite hotter than 200 °C. In the older well 58-32, 200 °C is reached at 2,257 m, and the model matches the measured log to within 1.7 °C (root mean square) down to 2,295 m.

The two wells run roughly parallel, with 16B about 330 ft (about 100 m) above 16A [2]. In August and September 2024 an extended circulation test ran water from one to the other for almost a month; the pressures, rates and temperatures are public under CC BY 4.0 [3].

EGS reservoir simulation: how the model was built

Flow in the stimulated granite follows fractures, so the model starts from them. KAI built the natural fractures from the image log of 58-32 and the FORGE fracture models, and the stimulated network from the published discrete fracture network of the 16A and 16B wells. Each fracture adds its permeability to the cells it crosses. Fractures open and close with pressure following the Barton–Bandis relation.

Because nobody knows the fractures exactly, the model is an ensemble of 20 members, each with its own fracture density, apertures and stress direction. Each member is calibrated on the August 2024 test. Then the injector 16A injects at a set rate, the producer 16B produces at its August 2024 bottom-hole pressure, and the model runs 30 years.

Enhanced geothermal system 30-year forecast

At 10 kg/s the produced temperature starts at 206.8 °C (204.9 to 208.7 °C) and falls to 126.5 °C (113.5 to 142.9 °C) after 30 years. Heat output falls from 6.15 MWth in year 1 to 2.78 MWth (2.23 to 3.48 MWth) after 30 years. On a refined grid the cold water reaches 16B after 1.26 ± 0.13 years.

Rate trades temperature for heat. At 5 kg/s the 30-year temperature is 149.8 °C (133.4 to 164.5 °C) but the heat after 30 years only 1.88 MWth. At 20 kg/s the temperature falls to 104.7 °C (97.7 to 120.3 °C) for 3.73 MWth. The cold front simply arrives sooner, as the classic doublet solution of Gringarten and Sauty predicts [4]: halving the rate doubles the breakthrough time, but the pair is far too close for any rate to hold its temperature.

Lesson 1: a test bed, not a power plant

So what would a 30-year doublet in this granite need? KAI trained a surrogate on the full runs and asked it for the produced temperature after 30 years at other spacings. The target line is 192.4 °C, still within 10 % of the way from the starting temperature to the reinjection temperature.

At 10 kg/s, the answer is about 320 m (P50), and 460 m to be right nine times in ten. The surrogate's error is 9.7 °C on held-out runs, so read the spacings as ±20 %. And the chart rests on one large assumption: that stimulation at a larger spacing makes a connected network of the same kind as FORGE's today. That is the hard part of EGS, not the arithmetic.

A variance-based (Sobol) analysis on the surrogate ranks rate first and spacing second across the chart; the thickness of the stimulated network comes a distant third. For the FORGE pair itself, at its own spacing and rate, that thickness is the largest uncertainty. The indices come from notebook 5 and were not recomputed after the temperature correction of notebook 7, which moves the start and the target together. The ranking matches the general rule that breakthrough time grows with the square of the spacing and falls with the rate.

Lesson 2: the stress limit decides the rate

Injecting above the minimum horizontal stress risks opening new fractures and inducing earthquakes, so KAI asked the highest rate the pair could take for 30 years while staying below it. The answer depends on which stress data you trust, and we lead with the cautious one:

• With the stress measured by minifracs in 16B and carried down to 16A (32.2 MPa there): about 4.9 kg/s (P50; P10 to P90 3.6 to 6.1 kg/s). Only 9 of 20 members can inject even 5 kg/s for 30 years.
• With the regional GDR 1222 stress gradient (40.3 MPa at the top of 16A): 9.8 kg/s (P10 to P90 9.4 to 10.7 kg/s).
• With the minifracs adjusted for cooling: 11.8 or 16.2 kg/s (an earlier phase, 8 members, before the temperature correction).

The 2024 test itself ran above even the GDR 1222 limit, by 2.5 MPa. McClure reports a sustained production rate of 28 kg/s during the 30-day test (26 kg/s in the paper's summary) [2]. That is a month of circulation with the injection pressure above the stress holding the stimulated fractures open, which is different from 30 years of injection kept below it. A minifrac in the reservoir section of 16A would settle which limit applies. The lesson for any EGS project: measure stress where you inject.

Lesson 3: check the temperature model at the wells

The FORGE temperature model was built before 16A and 16B were drilled, so the wells are a blind check. The 16A temperature log of August 2021 shows the model running +5.9 °C too hot in the 16A open hole and +8.2 °C at the depths where 16B produces. A pressure–temperature gauge in 16B agrees, about +6 °C. Correcting for it lowered the starting temperature at 10 kg/s from 213.7 °C to 206.8 °C and cut the misfit against the August 2024 test from +11.2 °C to +4.4 °C (2.7 to 5.9 °C). Most of the gap was the temperature model, which can be measured, not the fracture physics.

Lesson 4: report the misfits you cannot remove

Two misfits remain, and the case shows them rather than tuning them away.

Tracers. Nanoparticle tracers came back to 16B within about 10 hours. The model returns 5 % of a tracer in 0.19 hours (0.13 to 0.44 hours), 20 to 50 times too fast. Its flow paths hold too little water: the apertures that fit the pressures are too thin for the transport. This barely changes the heat results.

Coarse grids. On 70 m cells, the usual coarse model ends +10.8 °C (+3.1 to +14.2 °C) warmer at year 30 than the fine model. With flow upscaled from the fine model, the error drops to +2.4 °C (0.0 to +5.1 °C). A correction meant to fix coarse-model breakthrough times failed the test set before running it, leaving times 10 to 20 % early, so coarse-model breakthrough times are not quoted.

How the numbers were checked

The KAI engine is checked against exact answers, among them the Gringarten and Sauty doublet breakthrough [4], and against the open-source simulator open-DARTS [5] on a 3D doublet, where breakthrough came at 4.906 against 4.924 years (0.4 %). The FORGE notebooks assert every published number, fetch the raw repository files at run time and check them by SHA-256. Data: Utah FORGE, U.S. Department of Energy, via the Geothermal Data Repository (OpenEI), CC BY 4.0. The simulated results are KAI's, not the FORGE team's.

What KAI offers, and what it does not

KAI-GeoEnergy is a live preview. The Reservoir Lab runs a simplified solver in your browser for a first look and, once you sign in, queues the full run on the KAI engine. For your own site we build the model from your wells, logs and tests, calibrate it on what you measured and give you the 30-year forecast with its P10 to P90, the stress limit and the misfits we could not remove.

KAI does not drill, hire out equipment or run field work such as minifracs, circulation tests or tracer tests. Fault-slip, subsidence and caprock checks are still in build. The FORGE results are a demonstration on public data, simulated, and not a prediction of what the FORGE project will do.