Rank-Adaptive Random Probing for Mass-Conserving Multiscale Darcy Solvers

Simulating groundwater or reservoir flow means solving pressure and velocity equations on rock whose permeability varies by six or more orders of magnitude over fine scales. Multiscale methods make this affordable. They build a few special basis functions on each coarse block and run the solve on the coarse grid. In practice, though, building those basis functions is the expensive part. Each coarse neighborhood normally requires many local fine-scale solves, called snapshots, one per boundary degree of freedom. A separate line of mathematics shows why most of that work is wasted. Elliptic operators smooth out rough inputs, so the space of local solutions has only a handful of significant directions. We propose a basis builder for mass-conserving (mixed) multiscale methods that probes each neighborhood with random boundary fluxes. It keeps drawing probes only until a cheap randomized error check says the local solution space has been captured. As a result, each block spends compute in proportion to its own intrinsic dimension, not in proportion to its boundary size.

The core idea fits in two sentences. Random boundary data captures the low-rank local solution space of high-contrast Darcy flow with far fewer local solves than exhaustive snapshots. Letting each neighborhood stop on its own, once a held-out probe is reproduced, matches the work to the number of high-conductivity channels crossing that neighborhood. The open question is whether this carries over to the mixed velocity–pressure formulation that subsurface transport requires, and whether adaptive stopping beats a well-tuned fixed sample budget.

Core Claim

Two pieces of prior work set up the claim. Randomized Oversampling for Generalized Multiscale Finite Element Methods applies a small, fixed number of random boundary conditions on oversampled neighborhoods for primal (continuous Galerkin and interior penalty) GMsFEM. It reports keeping accuracy while using "only 13.46% of the snapshots" against a full offline space of dimension 931. Random Sampling and Efficient Algorithms for Multiscale PDEs explains why this works. Random boundary data acts as a sketch of the local Green's operator, and that operator has low numerical rank because its Kolmogorov N-width decays fast. Both papers motivate our claim. Neither addresses mixed formulations, and neither sets the sample count per neighborhood from data.

The mass-conserving side is the setting we target. Mixed Generalized Multiscale Finite Element Methods and Applications builds paired velocity and pressure spaces that conserve mass locally. Iterative oversampling technique for constraint energy minimizing generalized multiscale finite element method in the mixed formulation adds constraint energy minimization (CEM) and uses deterministic Richardson iterations over growing oversampling regions. Our claim has two parts. First, random flux probes, each a local mixed solve, yield snapshot spaces that conserve mass exactly by construction. Second, a rank-adaptive stopping rule cuts total offline local solves well below both full snapshots and the best fixed random budget, with no loss in velocity accuracy, on channelized media with contrast up to $10^6$.

The physical hypothesis behind the adaptive rule comes from the GMsFEM literature. In high-contrast media, the number of near-zero local eigenvalues tracks the number of high-conductivity channels that cross the neighborhood boundary. Intrinsic dimension is therefore spatially heterogeneous, and a uniform budget is necessarily wrong somewhere: too large in matrix-dominated blocks, too small in channel-dense ones.

Research Program

In each coarse neighborhood $\omega$, we draw Gaussian random normal-flux data on the oversampled boundary and solve the local mixed Darcy problem. Each solve gives a velocity–pressure snapshot $s_j$. After $k$ probes, let $P_k$ be the energy-orthogonal projector onto their span. We then draw $r$ fresh probes (typically $r=4$) and stop when

$$\max_{1\le i\le r} \frac{\| (I-P_k)\, s_{k+i} \|_{E}}{\| s_{k+i} \|_{E}} \le \tau,$$

where $\|\cdot\|_E$ is the local energy norm of the mixed problem and $\tau$ is a user tolerance. If the check fails, we add the test probes to the basis and repeat. The resulting compressed space feeds the usual local spectral problem that selects multiscale velocity bases. In the CEM variant, the same probing replaces the snapshot step that defines the auxiliary space. The Richardson-based decaying correction of Cheung et al. is kept, so that we isolate one change at a time.

We will analyze three things. The first is whether random flux data needs to be divergence-compatible on the boundary, which would require projecting out the mean flux. The second is how the chosen $k$ per neighborhood correlates with channel counts extracted from the permeability field. The third is whether $\tau$ translates into a predictable global velocity error. The analysis relies on the N-width arguments of Chen et al.; we expect a bound that depends on the spectral gap of the local problem rather than on contrast directly.

Baselines

The naive reference point is the fine-scale mixed finite element solve on the full grid. It sets the ground-truth pressure and velocity and the wall-clock cost any multiscale method must beat. Classical single-phase upscaling with block-effective permeability serves as a second naive reference. Its known failure mode, numerical dispersion of saturation fronts when sub-block velocity detail is lost (Aarnes & Efendiev), will be shown on the downstream transport test.

The strongest existing method for the accuracy target is full-snapshot Mixed GMsFEM (Mixed Generalized Multiscale Finite Element Methods and Applications), together with the iterative CEM mixed method of Cheung et al.. Their reported numerical error tables are not available here: no reported result found on Lacuna for either method's error values. We will therefore rerun both on our grids. Beating them means matching their velocity $L^2$ error within 10% relative while using at most a third of their local solves.

The strongest competitor on cost is fixed-budget random oversampling in the style of Calo et al.. It reports 13.46% of snapshots at preserved $L^2$ and energy accuracy for primal GMsFEM. We will port it to the mixed formulation ourselves. This baseline is the key ablation, because it differs from our method only in using a uniform $k$. For fairness, we sweep its budget over a grid and report its best point after the fact, which gives it oracle tuning that our method does not get. Beating it means a lower total number of local solves at equal global velocity error, across contrasts.

A second ablation replaces the held-out-probe stopping rule with a spectral-decay rule (stop when the singular values of the snapshot matrix drop below a threshold). This tests whether the held-out check is what matters. All methods share the same fine grid, coarse partition, oversampling width, linear solver, and hardware.

Why not just…?

Why not just use fixed-budget random oversampling, already shown to work? This is the real threat. If an oracle-tuned uniform budget matches our total cost, then adaptivity adds only the convenience of not tuning, and the claim narrows to "random probing extends to mixed formulations." The ablation above settles it. We expect adaptivity to win where channel density varies strongly across the domain, as in SPE10-like layers, and to tie on statistically uniform fields such as Gaussian Karhunen–Loève (KL) fields. We will report both cases.

Why not just train a neural operator to predict the basis? Locally Subspace-Informed Neural Operators predicts whole local subspaces with a Grassmannian loss, reaching 1–3% error and a 60x–100x speedup in basis generation. Standalone F-FNO and GNOT exceed 100% error out of distribution. That approach still needs 200–400 ground-truth subspaces to train, produced by the very procedure we are accelerating. It also targets primal formulations on Cartesian patches. For a single geologic model queried under many well configurations, there is nothing to amortize the training over. Our method complements it as a cheaper label generator, and we will report the cost of producing its training set both ways.

Why not just do global POD on solution snapshots? Li, Zhang and Zhao report 1–2% error with 4–15 basis functions. However, their snapshots are global fine-scale solves, which they themselves note are prohibitive for large 3D models. Their bases also do not conserve mass locally. We include global POD at a matched fine-solve budget, and we expect it to lose on the local conservation residual.

Proposed Experiments

The test problems are the 2D and 3D Darcy sets in PDEBench, plus SPE10-like channelized layers of the kind used by AMS-Net, with contrasts rescaled to $10^4$ and $10^6$. Metrics are pressure and velocity relative $L^2$ error against the fine solve, the maximum local mass-conservation residual, total local solves, and offline wall clock. As a downstream check, we run saturation error in two-phase incompressible transport driven by each method's velocity.

The first decisive check is that, at contrast $10^6$, the adaptive method reaches full-snapshot velocity accuracy with fewer total solves than the oracle-tuned fixed budget. The second is that its per-neighborhood $k$ correlates with channel crossings. If adaptive stopping does not beat the oracle-tuned fixed budget, we drop the adaptivity claim. If random flux probes miss thin channels, which would show up as error rising with contrast at fixed $\tau$, we will test channel-aligned probe distributions before concluding that randomization fails for mixed spaces.