Pricing Resistivity Monitoring in Liquidity Terms for Fault-Bounded CO₂ Storage

A carbon storage operator who is unsure whether a nearby fault will leak is pushed toward caution. The cautious choice is to inject slowly, pause often, and store far less than the reservoir could hold. In the best-known planning study of this problem, the leak-free policy fills only 27% of capacity, while an expert fixed schedule fills 82% but leaks. When the storage project was financed with short-term debt that must be rolled over, that caution has a second cost. Every pause cuts revenue while debt payments continue, and a project that is solvent on paper can still default for lack of cash. This proposal asks whether a specific sensor removes enough uncertainty about the fault to change that outcome. The sensor is time-lapse electrical resistivity: repeated measurements of how easily current flows through the ground, which change sharply when salty brine or resistive CO₂ moves. The project will measure the sensor's worth in dollars of avoided default, rather than in image quality.

The claim is testable in simulation. Resistivity data are fed into the operator's belief about the fault's hydraulic state, and a planner then chooses injection rates under both a leakage limit and a cash-flow limit. If the resistivity data identify the uncertain fault properties that pressure gauges and microseismic catalogs miss, the planner can inject more without leaking, and the probability of a liquidity default should fall. If resistivity only repeats what pressure already reveals, it is not worth buying, and we will say so.

Core Claim

Time-lapse resistivity is worth more to a storage operator than its imaging quality suggests, because it constrains the parameters that govern brine leakage. It can do so without first producing a resistivity image. Its value is best expressed as a reduction in the liquidity reserve needed to keep default probability below a target.

The argument starts from Lü & Salo-Salgado, who build a 2D fault-leakage model in which fault clay smears are generated stochastically. Their sensitivity analysis finds that different quantities depend on different unknowns. Fault stability depends mostly on the initial stress ratio. Pressure buildup depends mostly on fault permeability. Brine leakage depends mostly on reservoir porosity. A pressure gauge therefore tells the operator about fault permeability, but it leaves the leakage driver poorly pinned down. Resistivity responds directly to where conductive brine has moved, so it is a natural candidate for the missing channel.

The crystalline basement enters as the setting that makes this physically plausible. Dense crystalline rock is a poor conductor, so faults rooted in the basement and carrying brine should stand out against it. Airborne electromagnetic surveys over the Qingchengzi district already show this contrast, separating resistive granite bodies from conductive fault zones (Fu et al.). The Illinois Basin–Decatur site is the concrete analog: CO₂ was injected into sandstone lying directly on Precambrian basement, and the site's faults have been located from microseismicity (DeFault). Microseismicity shows where a fault slips. It does not show whether the fault passes fluid.

What is new is the combination of three things, none of which appears in the file. The first is resistivity as an observation inside a decision process for storage, used to update the belief directly rather than through a reconstructed image. The second is a fault leakage model with coupled flow and rock mechanics. The third is an objective that constrains cash-flow timing rather than a single scalar reward.

Research Program

Physics testbed. The testbed extends the MRST-based model of Lü & Salo-Salgado. It is a 2D slice with a storage aquifer, sealing caprock, and a normal fault containing clay smears, driven by the same five uncertain parameters: fault permeability, reservoir permeability, stress ratio, porosity, and rock compressibility. One change is made: a resistive basement layer is placed beneath the aquifer, and the fault is rooted in it.

Resistivity forward model. Brine saturation, salinity, and clay content from each flow snapshot are mapped to resistivity with a standard saturation–resistivity rock-physics relation that includes a clay surface-conduction term. Cross-hole and surface DC resistivity data are then simulated with a forward solver of the type used by Naeh et al. and PhResNet. Field noise is modeled explicitly: electrode contact error, correlated drift, and missing electrodes. This matters because every learned inversion in the file struggles when moving from simulated to real data. DeFault needed a dedicated domain-adaptation step for that reason.

Belief update without imaging. A resistivity image is not the planner's goal. The planner needs a posterior over the five fault parameters. Following the belief-tracking scheme of Corso et al., we use sequential importance resampling, which maintains a weighted set of candidate parameter settings and reweights them as data arrive. Here the likelihood is computed on the raw apparent-resistivity data. Forward runs are made affordable by a Flow Map Learning surrogate of the kind Lü & Salo-Salgado validated: a neural network trained to step the simulator forward in time, which reproduced their 10th, 50th, and 90th risk percentiles at one-tenth of the Monte Carlo cost. A second surrogate maps surrogate states to predicted resistivity data. Working on raw data avoids the non-uniqueness and blurred contacts of image-based inversion. The trade-off is that the surrogate must be accurate in data space, and we test this directly.

Financial layer. The operator borrows a fixed capital amount with short-term debt that rolls over each period, and earns revenue per tonne stored. A rollover fails when the cash-coverage ratio falls below a covenant threshold. A detected leak adds a remediation liability and forces a halt. This is the maturity-mismatch mechanism described for heavy industry (Duo Xu et al.) and the timing trap documented for Silicon Valley Bank (Sui et al.): a firm with valuable long-lived assets fails because cash arrives too late. Those papers do not model physical triggers. Here the trigger is a throttled injection schedule.

Planning objective. The planner, POMCPOW (a tree-search solver for partially observed decision problems), chooses the injection schedule $a_{1:T}$ to solve

$$\max_{a_{1:T}} \; \mathbb{E}[M_T] \quad \text{s.t.} \quad P(\text{leak}) \le \epsilon, \quad P(\exists t: C_t < D_t) \le \delta .$$

Here $M_T$ is the CO₂ mass stored by the horizon $T$, $C_t$ is available cash in period $t$, $D_t$ is the debt service due in that period, and $\epsilon$ and $\delta$ are the tolerated leak and default probabilities. The constraints are enforced through Lagrangian penalties tuned by bisection.

The headline quantity is the minimum initial cash reserve $R^\star(\delta)$ that meets the default target under each monitoring package. The difference in $R^\star$ between two packages is the value of the added sensor, in money.

Baselines

The naive reference point is an operator who follows the expert fixed schedule with no adaptive monitoring. Corso et al. report 0.10 leak fraction, 0.82 trap efficiency, and a return of −97.91 for this schedule in their spillpoint benchmark. We rerun the schedule in the new testbed and add default probability to the reported quantities.

The strongest existing method is the POMCPOW planner with importance-resampling beliefs from Corso et al.. On their 2D spillpoint benchmark it reports 0.00 leak fraction, 0.27 trap efficiency, and a return of 2.24. We run it in two settings: on their benchmark with their reward, to reproduce the reported numbers, and in our testbed with its original observations and our constrained objective. All planners receive the same particle count, the same tree-search budget per step, the same surrogate, and the same penalty-tuning procedure. Beating it means two things. At leak probability at or below 0.01, trap efficiency must be clearly above its level, with a target above 0.5. Default probability, or $R^\star$ at a fixed default target, must also be lower, with 95% bootstrap intervals over at least 500 held-out true geologies that do not overlap.

The key ablations isolate each ingredient. One planner observes pressure only. One observes pressure plus microseismic locations, simulated with a 42 m location error to match the median error DeFault reports against expert catalogs. One observes pressure, microseismicity, and resistivity. A further ablation swaps the likelihood on raw data for belief updates on resistivity images produced by a PhResNet-style unsupervised inversion. The last ablation removes the cash constraint, so the planner maximizes expected net present value alone. This tests whether financial timing changes the chosen policy.

No reported result was found on Lacuna for default probability or reserve requirements in any subsurface storage setting, so those rows start from our own reruns.

Why not just…?

Why not just rely on downhole pressure, which is cheaper? This is the most serious threat. Pressure buildup is the quantity most sensitive to fault permeability (Lü & Salo-Salgado), and fault permeability may be the unknown that matters most to the planner. The pressure-only and pressure-plus-resistivity comparison settles it. We also compute posterior contraction for each parameter separately. If adding resistivity does not narrow the porosity and fault-permeability posteriors beyond pressure alone, or does not move $R^\star$, the claim fails and the conclusion is that pressure suffices.

Why not just finance the project with patient long-term capital? Matching long-lived illiquid assets with long-term funding is the stability prescription for real estate (Shimizu), and it removes rollover risk entirely. It does not remove the physical throttling. It also costs more, through longer-term funding, a larger reserve, or both. Expressing results as $R^\star$ makes the two remedies directly comparable: resistivity monitoring is worth buying only when the reserve it saves exceeds its cost. We report the break-even monitoring cost.

Why not just invert the resistivity data into images with a standard iterative solver and let engineers interpret them? Image inversion is non-unique and blurs contacts. Naeh et al. found that the standard iterative solver was slower and less accurate at depth than their network. PhResNet found that linear inversion could not separate adjacent anomalies it located to 0.1 m. The image-based ablation measures what the detour through an image costs in decision quality, as opposed to image error.

Why not just add microseismic monitoring, which already runs at Decatur? DeFault shows that microseismicity locates slipping fault planes quickly: about 4 seconds of inference, with a 42 m median error. Slip, however, mostly reflects the stress ratio, while leakage mostly reflects porosity. The ablation with microseismicity but no resistivity tests whether seismic data alone lift trap efficiency. We expect the two channels to complement each other rather than one replacing the other.

Proposed Experiments

The first check is fidelity. On 200 held-out full-physics runs, the combined flow and resistivity surrogates must predict apparent-resistivity data within the simulated noise level before any planning results are trusted. The second check reproduces the Corso numbers on their own benchmark, to confirm the planner implementation. The main experiment sweeps the default target $\delta$ over {0.01, 0.05, 0.1} and three debt tenors, reporting leak fraction, trap efficiency, default probability, and $R^\star$ for every monitoring package.

The decisive result is the $R^\star$ gap between pressure-only and pressure-plus-resistivity at matched leak probability. A robustness test then repeats the gap under three conditions: doubled field noise, 20% dead electrodes, and a clay-rich fault gouge that is conductive even when sealed. The last condition is the main physical confound, because it could make a sealed fault look like a leaking one.

Two findings would count against the proposal. The first is a gap that disappears under realistic noise. The second is an unconstrained, value-maximizing planner that already picks near-default-free schedules. Either would redirect the recommendation toward cheaper sensors or better financing.