Does Fracture Closure Memory Make Drawdown Scheduling Worth Optimizing?

When oil or gas is pumped out of tight, fractured rock, the fluid pressure inside the rock falls. The weight of the overlying rock then squeezes the cracks that carry the flow. Flow through a crack scales roughly with the cube of its width, so a crack squeezed to half its width carries about one eighth of the flow. This produces a known trap. Pumping harder, meaning lowering the pressure at the bottom of the well, steepens the pressure gradient that drives flow, but it also crushes the channels that flow depends on. In a Xinjiang tight-oil field study, Feng et al. found that a model which accounted for crack closure recommended holding bottom-hole pressure at 20–25 MPa. A model that ignored closure recommended an aggressive 4–6 MPa drawdown instead. They reached this by manually sweeping constant pressures.

The obvious next project is to wrap a genetic algorithm around a stress-sensitive dual-porosity simulator and let it search time-varying pressure schedules. A genetic algorithm is a population-based search that recombines and mutates candidate solutions and needs no gradients. The obvious project risks a hollow result, because a skeptic can reasonably expect a two-parameter rule such as "keep drawdown below X" to do nearly as well. This note argues that the answer depends on one physical property that current models mostly assume away: whether crack closure has memory. If closure is reversible, meaning permeability depends only on current pressure, the problem is nearly myopic and simple rules should suffice. If closure is partly irreversible, as with plastic collapse or proppant crushing, then an early pressure decision permanently sets later conductivity. The schedule problem then becomes non-myopic and rugged, which is exactly the setting where population search should pay off. Proppant is the sand pumped in to hold hydraulic fractures open.

Core Claim

The value of optimizing a full time-varying drawdown schedule, compared with the best simple rule, is governed by how irreversible fracture closure is. We will show this in a dual-porosity model with explicit fractures, in which the matrix-to-fracture transfer also shrinks as fractures close. We predict three things. First, under fully reversible closure the gap between a genetic-algorithm schedule and a tuned two-parameter rule will be within run-to-run noise. Second, the gap will grow with the irreversible fraction of closure. Third, the optimal schedules under irreversible closure will show a qualitatively different shape: a slow early ramp that protects natural fractures before they are lost, followed by aggressive late drawdown.

If the first prediction fails, the conclusion is simply that genetic algorithms are useful here. If the second fails, the conclusion is that simple rules are enough everywhere. Both outcomes are informative for practice.

Mechanism and Model

We represent the reservoir in two layers. A dual-porosity model treats the rock as two overlapping continua: a tight matrix that stores most of the fluid, and a fracture network that carries it. Large hydraulic fractures are added as explicit planes embedded in the grid, following the embedded discrete fracture approach of Wang et al. (2020). That work coupled separate exponential stress sensitivities to the matrix, hydraulic fractures and natural fractures. Its natural fractures, with a coefficient of 0.08 MPa⁻¹, shut completely in late production. It ran only forward simulations with predefined step-down schedules, and its closure was reversible.

We add memory by letting fracture permeability depend on both the current pressure $P$ and the lowest pressure the cell has ever seen, $P_{\min}(t)$:

$$k_f = k_{f0}\exp\!\big(-\gamma\,[\,(1-\lambda)(P_0-P) + \lambda (P_0-P_{\min}(t))\,]\big).$$

Here $k_{f0}$ is initial fracture permeability, $P_0$ is initial pressure, $\gamma$ is the stress-sensitivity coefficient, and $\lambda\in[0,1]$ is the irreversible fraction. Setting $\lambda=0$ recovers the standard reversible law. A variant replaces the smooth $\lambda$ term with a threshold collapse below a critical pressure. This is motivated by the 35 MPa plastic-collapse limit that the fault-karst Block S study identified from field history matching. That study used a one-dimensional connected-unit model, not a dual continuum.

The second modification addresses a known weakness of dual-porosity models: the matrix-fracture transfer uses a fixed geometric shape factor. We scale the transfer coefficient with the current effective fracture aperture $w/w_0$:

$$\tau = \sigma_0\,\Big(\frac{w}{w_0}\Big)^{m}\frac{k_m}{\mu}(P_m-P_f).$$

Here $\sigma_0$ is the initial shape factor, $k_m$ is matrix permeability, $\mu$ is viscosity, $P_m$ and $P_f$ are matrix and fracture pressures, and $m$ is an exponent we vary from 0 (the classical static kernel) to 1. With this change, closing fractures throttles matrix drainage as well as fracture flow.

We implement both modifications as pressure-history-dependent transmissibility multipliers in MRST, the MATLAB Reservoir Simulation Toolbox used for the synthetic cases in Wang et al. (2022). This keeps each forward run to minutes on grids comparable to the 3,573-cell embedded-fracture model of Wang et al. (2020).

Research Program

We build three synthetic test reservoirs. None of the field datasets in the file are public, so we replicate their published physics rather than their data.

The first is a Xinjiang-like tight-oil multi-stage fractured horizontal well. It uses Feng et al.'s stress-sensitivity range of 0.01–0.05, with 0.03 as the base case, so that the constant-pressure optimum of 20–25 MPa can serve as a reproduction check.

The second is a Wang-2020-like case with oval hydraulic fractures plus dense natural fractures, whose natural-fracture coefficient reaches 0.08 MPa⁻¹.

The third is a carbonate-like case that includes a threshold collapse at 35 MPa.

For each reservoir we sweep the irreversible fraction $\lambda\in\{0,0.25,0.5,1\}$ and the transfer exponent $m\in\{0,1\}$. We optimize ten-year net present value, meaning discounted revenue minus costs. The decision variable is a piecewise-constant bottom-hole-pressure schedule with 20 monthly-to-quarterly control periods, bounded by rate and pressure constraints. We also repeat the optimization with 5 and 40 periods to see whether the gains come from schedule shape or from dimensionality. Every optimizer is run over five geological realizations and five random seeds.

Baselines

The do-nothing reference is the aggressive static drawdown that a closure-blind model recommends. Following Feng et al., we find it by optimizing on a copy of each reservoir with stress sensitivity switched off, then evaluate it on the stress-sensitive model. This measures what ignoring closure costs.

The second baseline is Feng et al.'s method itself: a fine sweep over constant bottom-hole pressure, run on the true stress-sensitive model. Feng et al. report the optimal pressure (20–25 MPa) but no net-present-value figure, so there is no reported result found on Lacuna for net present value on this benchmark. We will reproduce the pressure window as a sanity check.

The strongest simple competitor, and the one this proposal must beat, is a tuned two-parameter physical rule. Bottom-hole pressure tracks average reservoir pressure minus a drawdown cap, and the cap relaxes linearly after a switch time. The cap and the switch time are both optimized by grid search using the same simulation budget given to the genetic algorithm.

The strongest existing automated method is StoSAG, a stochastic simplex gradient approximation. Wang et al. (2023) used it on a dual-permeability fracturing-design problem and reported nearly 20% higher net present value than the reference field design, converging in 30–50 iterations. Their problem had no production-phase closure, so that number is context, not a target. We run StoSAG on our schedule problem with the same forward-run budget as the genetic algorithm. We also run CMA-ES, an evolution strategy that adapts a Gaussian search distribution, because it is the natural stronger evolutionary competitor.

The key ablation repeats every optimizer with $\lambda=0$ and with $m=0$. This isolates whether any schedule advantage comes from closure memory or from aperture-coupled transfer.

We count the genetic algorithm as beating a baseline only if its net-present-value advantage exceeds twice the across-seed standard deviation at equal budget, consistently across realizations. The central prediction is that this holds against the two-parameter rule when $\lambda\ge0.5$ and fails when $\lambda=0$.

Why not just…?

Why not just use a single-porosity model with a pressure-dependent permeability multiplier? Wang et al. (2020) showed that natural fractures, hydraulic fractures and matrix close on different timescales, with natural fractures shutting completely at 0.08 MPa⁻¹. One homogenized multiplier cannot represent the protect-the-natural-fractures-first behavior we predict. We test this directly by running the optimized schedules from a single-porosity equivalent model on the dual-porosity model and measuring the net present value lost.

Why not just enforce a tuned static pressure rule, as Feng et al. did? This is a real threat, and it is why it appears as our strongest simple baseline rather than a straw man. Our claim is conditional. Under reversible closure we expect the rule to be sufficient, and we will say so if it is. The file offers only an argument, not evidence, that static rules fail as reservoir pressure declines. The $\lambda$ sweep is the experiment that settles it.

Why not just use gradient methods or StoSAG, which converge in far fewer runs? Closure memory introduces a running-minimum operation, and threshold collapse introduces cliffs. Both make the objective non-smooth in ways that break adjoint gradients and can mislead simplex gradients. That remains a hypothesis. The equal-budget comparison with StoSAG, reported for each $\lambda$, tests it. CMA-ES may outperform a plain genetic algorithm, and if so we report the finding as "population search is needed," not "genetic algorithms are needed."

Why not just train a deep surrogate and search on it quickly? The CNN-LSTM surrogate of Wang et al. (2022) cut evaluation time from hours or days to about 30 seconds, but it was trained on single-porosity media with static permeability. Our forward runs are already cheap enough to optimize directly. As a side check, we train a sequence surrogate on schedules drawn from the $\lambda=0$ case. We then measure its error, and the net-present-value loss of schedules optimized on it, once irreversible collapse becomes active.

Decisive Checks

The main result is a curve: the net-present-value gap between the genetic algorithm and the tuned two-parameter rule, plotted against $\lambda$, with seed-level error bars, for each of the three reservoirs. The idea is supported if the curve starts near zero and rises. It is refuted if the curve is flat, whether flat at zero or flat above zero. Two secondary checks matter.

The first is shape. Optimized schedules under irreversible closure should consistently start gently and end aggressively, and running them on the reversible model should cost measurably less than running reversible-optimal schedules on the irreversible model. This asymmetry is the signature that memory causes the difficulty.

The second is the transfer exponent. Turning $m$ from 0 to 1 should shift the optimal ramp timing. If it does not, aperture-coupled transfer is unimportant for control, and that finding would simplify future models.