Learned Output Channels for Message-Passing Methane Source Inversion

Finding methane leaks from downwind sensors is a high-dimensional inverse problem. The unknown is an emission rate at every cell of a grid of candidate source locations, often thousands of cells, and almost all of them are zero. Near the source, the link between the average plume and what a fast sensor actually records is far from Gaussian. Turbulent air breaks a plume into intermittent puffs, so a one-second reading is often near background and occasionally spikes. No one can write down the probability of that reading. It exists only as the output of a stochastic dispersion simulator. Current methods split over this problem. Grid inversions keep the dimensionality but assume Gaussian errors. Sampling methods treat the dispersion physics honestly but only for about seven parameters. Simulation-based inference (training neural networks on simulator runs to approximate posteriors) handles black-box simulators but degrades beyond a few dozen unknowns.

We propose to connect these through the structure of a message-passing algorithm. Vector Approximate Message Passing (VAMP), an iterative estimator that alternates a linear solve with simple per-coordinate corrections, needs the measurement model in only one form. It asks for the posterior mean and variance of a single receptor's mean concentration, given its noisy reading and a Gaussian guess passed from the linear step. That is a two- or three-dimensional inference problem, which is small enough for neural posterior estimation to learn reliably from simulations. Plugging this learned "output channel" into VAMP gives a grid-scale inversion that respects intractable turbulence statistics. As a side benefit, the theory that tracks VAMP's error through a one-line recursion, called state evolution, becomes a cheap predictor of how well a given sensor layout can recover sources. We will test directly whether that prediction survives on physical transport operators, which the theory does not cover.

Core Claim

We write $x$ for the grid of emission rates. The matrix $H$ maps emissions to ensemble-mean concentrations $z = Hx$ at each receptor and time. It is computed from a plume or particle dispersion model. Each reading $y_i$ is drawn from an intractable distribution $p(y_i \mid z_i, \eta)$ produced by a stochastic puff simulator, where $\eta$ collects nuisance turbulence settings such as stability class and averaging window. This is a generalized linear model: a linear mixing step followed by a nonlinear, noisy output. The VAMP extension to such models needs the output denoiser

$$g(y, r, \tau) = \mathbb{E}\left[z \mid y,\; z \sim \mathcal{N}(r, \tau)\right],$$

together with the matching posterior variance. Here $r$ and $\tau$ are the Gaussian mean and variance that the linear step passes to receptor $i$. The variance supplies the "Onsager" correction, the term that keeps the algorithm's errors behaving like fresh Gaussian noise.

Our claim is that $g$ and its variance can be learned once, by amortized neural posterior estimation. Training draws $(r, \tau)$ from a design range, $z \sim \mathcal{N}(r,\tau)$, $\eta$ from its prior, and $y$ from the puff simulator. The learned channel then turns a $10^3$–$10^4$-dimensional likelihood-free inversion into many small learned problems, coordinated by VAMP's linear solves with $H$.

We make two falsifiable predictions. First, this learned-channel VAMP will beat Gaussian-likelihood inversions on high-frequency, near-source data. The advantage should shrink as the sensor averaging window grows, because averaging many puffs makes the reading nearly Gaussian by the central limit theorem. Second, state evolution computed from the singular values of $H$ and the learned channel will predict per-iteration error. If it fails, the failure should trace to the physical structure of $H$ rather than to the learned channel.

Relation to Prior Work

Three atmospheric studies motivate the problem and show the two ways it is currently split. Mesoscale inversion: first results from the CERES campaign with synthetic data inverts regional CO₂ with an explicit particle-model Jacobian and Gaussian errors. State-wide California 2020 carbon dioxide budget estimated with OCO-2 and OCO-3 satellite data replaces fixed covariances with a Gaussian process but keeps a Gaussian likelihood around linear transport sensitivities. Both scale to grids, and neither handles a non-Gaussian measurement process. Probabilistic Inversion Modeling of Gas Emissions: A Gradient-Based MCMC Estimation of Gaussian Plume Parameters goes the other way. It treats dispersion parameters as unknowns and shows that assigning the wrong stability class badly biases the recovered rate and location. That result motivates our nuisance variable $\eta$. However, that work relies on an explicit Gaussian likelihood and a seven-parameter single source sampled with manifold Langevin MCMC. We keep its lesson and drop both restrictions.

The message-passing papers supply the method and set its limits. Vector Approximate Message Passing and Rigorous Dynamics and Consistent Estimation in Arbitrarily Conditioned Linear Systems give the algorithm, its adaptive parameter learning, and state evolution for right-rotationally invariant matrices. These are matrices whose right singular vectors behave as if randomly rotated. Asymptotic Errors for High-Dimensional Convex Penalized Linear Regression beyond Gaussian Matrices confirms error predictions for that same matrix class. These results constrain our claims: dispersion Jacobians are anisotropic and local, so the theory's guarantees do not apply to them. We treat state evolution as an empirical predictor to be tested, not as a guarantee. Multi-Layer Generalized Linear Estimation shows that message passing extends to nonlinear output channels, but it assumes the channel density is known in closed form. Replacing that closed form with a simulation-trained one is our step.

The closest structural relatives are Score-Based Turbo Message Passing for Plug-and-Play Compressive Imaging and its companion Score-Based Turbo Message Passing for Plug-and-Play Compressive Image Recovery. Both insert a learned denoiser into message passing, but on the prior side and with a known linear-Gaussian likelihood. We insert the learned component on the likelihood side, where the physics is intractable, and keep an explicit sparse prior.

The simulation-based inference literature supplies the learning machinery and warns about scale. Benchmarking Simulation-Based Inference and Sequential Neural Likelihood: Fast Likelihood-free Inference with Autoregressive Flows show that neural estimators work well in low dimensions. Simulation-based Inference for High-dimensional Data using Surjective Sequential Neural Likelihood Estimation compresses high-dimensional data, but that does not solve a high-dimensional parameter space. Amortized Simulation-Based Inference in Generalized Bayes via Neural Posterior Estimation addresses misspecification by tempering, and we reuse its idea as a robustness variant. We implement the channel estimator with BayesFlow 2.0: Multi-Backend Amortized Bayesian Inference in Python.

Research Program

We will first build a stochastic puff simulator. It releases Gaussian puffs whose dispersion depends on stability class and turbulence intensity, and it records path-averaged or point readings at a chosen window length. The ensemble mean of this simulator defines $H$ over a candidate-source grid, one column per grid cell, for any wind history. The prior on $x$ is spike-and-slab per cell: each cell is either off, or on with a log-normal rate. This prior is separable, as VAMP's denoising step requires, and it matches the physics of sparse leaks.

We then train the amortized channel estimator $q_\phi(z \mid y, r, \tau)$ with $\eta$ marginalized under its prior. We validate it against exact posteriors on a Gaussian special case, using the classifier two-sample test (C2ST) standard from Benchmarking Simulation-Based Inference. We also check that its variance agrees with the derivative of its mean with respect to $r$, scaled by $\tau$, which must hold for a correct Gaussian-prior posterior. This consistency check matters because the Onsager correction depends on the variance.

Readings within one weather block share $\eta$, so the channel is not truly per-receptor. We handle this in two ways, both following the spirit of Adaptive VAMP's parameter learning. In the marginalized variant, $\eta$ is integrated out. In the adaptive variant, a block-level $\eta$ is re-estimated between iterations from the current channel messages.

For theory, we run state evolution by Monte Carlo over the scalar channel, using the empirical singular values of $H$. We compare this with the observed per-iteration error. We add a control in which $H$ is multiplied by a random orthogonal matrix, which leaves its spectrum unchanged but makes it rotationally invariant. This control separates failure caused by the operator's structure from failure caused by the learned channel. If state evolution holds, we sweep sensor count and averaging window to produce predicted recovery-versus-layout curves, and we check each curve against full inversions.

Baselines

The naive reference is the prior mean with no data, alongside a Gaussian maximum a posteriori inversion using the standard quadratic cost from the CERES study. That Gaussian inversion uses the same $H$ and an observation error variance fitted to the simulator's marginal variance. It is the "do nothing new" point. Any method that does not beat it on near-source data has no reason to exist.

The strongest existing competitor for grid inversions is Adaptive VAMP from Rigorous Dynamics and Consistent Estimation in Arbitrarily Conditioned Linear Systems. It uses the same spike-and-slab prior but a Gaussian output channel whose variance is learned by expectation-maximization. It is the strongest because it shares everything with our method except the learned channel, and it already handles ill-conditioning. That paper reports −41.7 dB normalized mean squared error (NMSE) against −10.5 dB for SPGL1, a standard sparse-regression solver, on ill-conditioned compressive image recovery at condition number 100. That figure comes from imaging, not atmospheric data, so it frames expectations rather than serving as a target.

Running Adaptive VAMP against our method is also the central ablation. The two differ only in the output denoiser. Both get identical $H$, priors, iteration counts, and damping, and both receive a hyperparameter search of the same size.

For small instances, we use manifold Langevin MCMC in the style of Probabilistic Inversion Modeling of Gas Emissions: A Gradient-Based MCMC Estimation of Gaussian Plume Parameters. It jointly samples a few sources and their dispersion parameters under a Gaussian likelihood. We also run long-chain pseudo-marginal MCMC that uses simulator draws. This serves as the reference posterior, against which we measure C2ST. The Chilbolton study reports no numerical accuracy figure (no reported result found on Lacuna for M-MALA rate or location error on Chilbolton), so we will rerun this baseline ourselves.

A direct neural posterior estimator trained on the full grid, at the same simulation budget, represents off-the-shelf simulation-based inference. On sbibm's Two Moons, sequential neural methods reach C2ST 0.5 with $10^5$ simulations. We expect far worse at our dimensionality.

Beating the baselines means specific things. Against Adaptive VAMP, we need lower NMSE and higher source-detection F1 on high-frequency data, together with nominal 90% interval coverage where Adaptive VAMP undercovers. We also need that advantage to shrink toward zero at long averaging windows. Against MCMC, we need C2ST close to MCMC's own run-to-run C2ST at under a tenth of its wall time.

Why not just…?

Why not just inflate the Gaussian error variance in a standard or variational inversion? This is the real threat. If turbulent intermittency only widens the spread of readings, a larger variance in the observation error matrix $R$ absorbs it, and Adaptive VAMP already learns that variance automatically. Our method can only win if the channel's shape matters: skew, zero-inflation, and dependence of the spread on $z$ itself. The averaging-window sweep settles this. At windows of several minutes we predict no gap. If the gap is also absent at one-second windows on Chilbolton-like data, the idea fails, and we will report that.

Why not just reduce dimension and run neural posterior estimation or ABC directly? For one to three sources with known candidate locations, this may well match us, and we will show where. Benchmarking Simulation-Based Inference shows rejection and sequential ABC lose to neural methods, and neural methods themselves degrade as unknowns grow. We therefore sweep grid size from a few dozen to $10^4$ cells at a fixed simulation budget, and we report the crossover point instead of asserting one.

Why not just use gradient MCMC with explicit plume physics? The Chilbolton study shows that joint sampling of dispersion parameters removes stability-class bias. That study also required Hessian-based manifold moves for seven parameters, which does not extend to thousands of cells. Its Gaussian likelihood also cannot represent intermittency. On the small instances we treat it as the accuracy reference, not as a competitor to beat.

Why not just put a learned prior into message passing, as score-based turbo methods do? These methods improve the prior step, whereas our bottleneck is the likelihood. The two are complementary. We keep the prior simple and explicit so that any gain can be attributed to the channel.

Proposed Experiments

The synthetic benchmark places $K$ sparse sources on grids of $30^2$ to $100^2$ cells under recorded wind sequences. Observation layouts range from 4 to 40 point sensors plus path-averaged lines, and averaging windows run from 1 second to 10 minutes. Stability class is drawn per block and is deliberately misreported to the Gaussian baselines at a controlled rate.

We report five groups of metrics:

  • NMSE of the emission field;
  • F1 for source detection within one cell;
  • relative error of total emission rate;
  • coverage of 50% and 90% intervals;
  • simulator calls and wall time.

The real-data test uses the Chilbolton controlled methane release, whose true rates and locations are known. There we run a grid formulation and compare rate and location error against our rerun of the manifold MCMC baseline and against Adaptive VAMP.

Three checks will decide whether the idea is worth pursuing. The first is whether the learned channel beats Gaussian Adaptive VAMP at short windows and ties it at long windows. The second is whether posterior coverage is near nominal on held-out misspecified weather, both with and without the generalized-Bayes tempering of Amortized Simulation-Based Inference in Generalized Bayes via Neural Posterior Estimation. The third is whether state evolution predicts per-iteration NMSE to within about 1 dB on plume Jacobians, or only after the random-rotation control. That last outcome tells the atmospheric community whether a one-line recursion can be trusted for sensor-network design, or whether the directional structure of transport operators calls for new theory.