Monotone Model-Mismatch Correction for Transport Inverse Problems

Many scientific inverse problems ask where a pollutant, dye, heat source or tracer came from, given a few noisy concentration readings downstream. Solving them needs a forward model: a simulator that predicts readings from a candidate source. In practice that simulator is always somewhat wrong. The velocity field is approximate, diffusion is anisotropic in ways nobody measured, and the mesh is coarse. A recent fix learns a small, untrained neural network per problem instance that absorbs the gap between the nominal simulator and the data. That correction is unconstrained, so the corrected simulator can predict physically impossible things: negative concentrations, or new peaks appearing in the middle of the domain with no source there. Transport and diffusion physics forbid both. Mathematicians call this rule the maximum principle. Without sources, a concentration field's extremes sit on the boundary or in the initial data. A numerical scheme that respects this rule on the grid is said to satisfy the discrete maximum principle (DMP).

We propose putting the learned correction inside a discretization that satisfies the DMP by construction, instead of adding it to the simulator's output. Concretely, the network outputs per-edge diffusivities and per-edge transport strengths for an edge-averaged finite-volume or virtual-element scheme. With positive diffusivities, that scheme yields an M-matrix for any values the network outputs. An M-matrix is a system matrix with positive diagonal, non-positive off-diagonals and an entrywise non-negative inverse, which guarantees non-negative, non-overshooting solutions. The hypothesis is that this does more than keep outputs in range. It narrows what the correction can "explain away", so less of the true source leaks into the model-error term. We test this in the setting where simulation-only evaluation is weakest: data made by a different, higher-fidelity process than the one being inverted, plus a small real tabletop dye experiment.

Core Claim

Restricting a model-mismatch correction to the set of monotone (M-matrix) transport operators gives better source reconstructions under realistic mismatch than an unconstrained output-space correction. The comparison is at equal compute and tuning. The gain comes from the structural constraint and not merely from bounded outputs, so clipping or projecting the unconstrained method's outputs will not recover it.

Relation to Prior Work

The method-supplying work is the $\mathcal{A}$-adaptive architecture of Solving Inverse Problems with Model Mismatch using Untrained Neural Networks within Model-based Architectures. It writes measurements as $y = A_0(x) + f_\theta(x, A_0) + \epsilon$. Here $A_0$ is the nominal forward model, $x$ the unknown, $f_\theta$ an untrained network fit per instance, and $\epsilon$ noise. It alternates updates of $x$ and $\theta$ by half-quadratic splitting and reports 37.81 dB against 34.38 dB for Robust Loop Unrolling on mismatched deblurring. We keep its per-instance, untrained correction and its alternating optimization. We move the correction from an additive residual on the output into the coefficients of a discretized PDE, which that paper never considers: its operators are blurs, 1D convolutions and haze formulas.

Two numerical papers supply the monotone discretization. Edge-averaged virtual element methods builds Scharfetter–Gummel-style edge fluxes that give an M-matrix on Voronoi meshes. It reports an $L^\infty$ error nearly 100 times smaller than streamline-upwind stabilization (SUPG) at diffusion $10^{-6}$. A Stable Mimetic Finite-Difference Method for Convection-Dominated Diffusion Equations gets the same property on dual Delaunay/Voronoi meshes, but assumes the flow is the gradient of a potential and lists rotational flows as open.

Our parametrization is per edge. Transport strengths on each edge need not sum to zero around cycles, so they can in principle represent rotational corrections. Whether the M-matrix property survives this is checked explicitly in our first milestone and not assumed. Both numerical papers validate only on manufactured solutions on unit squares, which is the simulation-only evaluation gap we target.

Three papers motivate the proposal by showing structure-free learning fails out of distribution. InverseBench shows diffusion-prior solvers hallucinate prior-shaped features and destabilize PDE solvers on physics problems. Reliable Extrapolation of Deep Neural Operators shows neural operator error growing polynomially with distribution shift, reduced from 10% to 0.32% by physics-informed fine-tuning, with no guarantee against spurious extrema. Structure-Preserving PINNs enforces only a global energy-decay property, which cannot rule out local overshoots.

One caution from InverseBench constrains our scope. Its instabilities arose in wave and fluid solvers. The wave equation has no maximum principle, so we make no claim about full waveform inversion. We restrict to advection–diffusion–reaction transport, where the DMP is the right invariant.

Research Program

The forward model is steady or implicit-in-time advection–diffusion on a 2D Voronoi mesh, discretized with edge-averaged fluxes. On each edge $E$, the flux coefficient is $\alpha_E\,B(\pm\psi_E)$, where:

  • $\alpha_E > 0$ is the edge diffusivity,
  • $\psi_E$ is the integrated transport along the edge, scaled by diffusivity,
  • $B(s)=s/(e^s-1)$ is the Bernoulli function, which is positive for all real $s$.

The nominal model supplies $\alpha^0_E, \psi^0_E$. The correction network $g_\theta$, a small coordinate MLP evaluated at edge midpoints and untrained before each instance, outputs $\alpha_E = \alpha^0_E\exp(g^{(1)}_\theta)$ and $\psi_E = \psi^0_E + g^{(2)}_\theta$. The system matrix therefore has positive diagonal and non-positive off-diagonals for every $\theta$.

The first milestone is proving, or empirically certifying on distorted meshes, that the resulting matrix stays an M-matrix. This should hold on Voronoi meshes by the argument in the EAVE paper. We will characterize where it breaks on general polygons, which that paper leaves open.

The unknown source $x \ge 0$ and $\theta$ are fit jointly by the same alternating scheme as $\mathcal{A}$-adaptive, with matched regularization on $x$. Differentiation through the sparse linear solve uses the adjoint method.

Baselines

The naive reference is inversion with the nominal operator $A_0$ alone, using the same monotone discretization and no correction. It measures how much mismatch hurts and sets the floor every correction must clear.

The strongest existing competitor is $\mathcal{A}$-adaptive itself (paper). Its additive residual network sits on top of the same nominal monotone solver, with the same source regularizer, iteration count and optimizer. Its network gets a parameter count and hyperparameter search budget equal to ours, both chosen on a held-out set of mismatch instances. Its only reported number is 37.81 dB PSNR on CelebA deblurring, a different task. For transport PDE inverse problems, there is no reported result found on Lacuna for $\mathcal{A}$-adaptive. We will run it ourselves. Beating it means lower source $L^2$ error and lower error on held-out sensors, averaged over mismatch families, with confidence intervals over at least 30 instances per family that exclude zero.

A second classical competitor is nominal inversion using a SUPG forward model instead of the monotone one. This shows whether the monotone discretization alone matters. The EAVE comparison predicts it does in convection-dominated regimes.

Two ablations isolate the key ingredient. The first uses the same edge-coefficient parametrization with the positivity map removed, so edge diffusivities can go negative and the M-matrix property can fail. This separates "correct the operator, not the output" from "keep the operator monotone". The second is $\mathcal{A}$-adaptive followed by projecting the reconstruction and predictions onto physical bounds at every iteration. This separates the structural constraint from mere bounded outputs.

Why not just…?

Why not just clip or project the unconstrained correction's outputs? This is the most serious threat, and the projected-$\mathcal{A}$-adaptive ablation tests it directly. Clipping enforces $u \ge 0$ and $u \le u_{\max}$ pointwise. It does not prevent in-range spurious interior extrema. It also does not stop the residual network from fitting part of the true source signal as "model error", which is an identifiability problem bounds do not touch. Third order maximum-principle-satisfying DG schemes also shows why naive clipping is worse than structural limiting: it destroys conservation and accuracy. If the projected baseline matches our method on source error, we will report that the benefit is bounds alone, and the core claim fails.

Why not just use a finer or better-stabilized nominal solver, such as SUPG or a refined mesh? Better discretization reduces discretization error. It does not touch physics mismatch such as an unmodeled swirl or anisotropy, which no mesh refinement recovers. The SUPG baseline and a fine-mesh nominal baseline quantify how much of the gap is discretization versus physics. The EAVE comparison against SUPG suggests SUPG will be worst in convection-dominated cases.

Why not just train a neural operator or diffusion prior on many simulated mismatches? Such models inherit their training distribution. InverseBench documents prior bias, and the DeepONet extrapolation study documents error growing with Wasserstein shift. Our correction is fit per instance, so it needs no mismatch-distribution training set. We include one amortized comparison: a DeepONet trained on our synthetic mismatch families and tested on a held-out family and on the real dye data. This checks whether this reasoning holds here.

Proposed Experiments

The main protocol avoids the "inverse crime", where the same simulator both generates and inverts the data. Observations come from a fine-mesh reference solver with physics absent from the inversion model. The inversion uses a coarse nominal model. We use four mismatch families:

  • an added rotational velocity component,
  • unmodeled anisotropic diffusion,
  • a spatially varying reaction term,
  • colored, spatially correlated sensor noise instead of i.i.d. Gaussian noise.

Sources are sparse non-negative blobs. Sensors number 50 to 400, echoing the sparse-sensor regime of the notched-Poisson study in Reliable Extrapolation of Deep Neural Operators. Diffusivity ranges from $10^{-2}$ to $10^{-6}$, matching the convection-dominated tests of the EAVE and mimetic papers. We also reuse the 2D diffusion–reaction random-field setup from that DeepONet study as a forward-prediction check.

Metrics are:

  • source $L^2$ error,
  • source support recovery,
  • prediction error at held-out sensors,
  • the fraction of nodes violating bounds,
  • the count of spurious interior local extrema,
  • the fraction of source mass absorbed by the correction term, which measures identifiability.

For real validation, which none of the cited numerical or transport-ML papers perform, we will build a small flow cell. Fluorescent dye is injected at known locations into water with a pump-driven, roughly measured flow, and a camera provides concentration images. Sparse pixels serve as sensors and the full images as held-out ground truth. There is no reported result found on Lacuna for any method on such data.

The decisive checks are two. First, our method must beat $\mathcal{A}$-adaptive-with-projection on source error in the rotational-flow and anisotropic families and on the dye data. Second, the unconstrained edge-parametrization ablation must fall between the two. That pattern would show the M-matrix constraint, not operator-space parametrization or bounds alone, carries the gain. If the projected baseline ties our method, the thesis is refuted.