When an air-quality station records a pollution spike, the standard way to find the source is to trace the arriving air backward in time along the wind and then ask which map cells those backward paths crossed on dirty days. The popular statistics for this, the Potential Source Contribution Function (PSCF) and Concentration-Weighted Trajectory (CWT), credit every cell along the path equally. As a result, a weak source 50 km upwind and a strong source 300 km upwind on the same wind corridor produce the same map. This is not just noise. It is a parameter degeneracy: two different source configurations give exactly the same measurement, so the inverse problem has a direction along which the data carry no information. Physicists in other fields break degeneracies like this by adding a second observable that is blind in a different direction. This proposal argues that atmospheric chemistry already supplies one. As a plume travels, oxidation steadily changes its composition: the oxidized fraction of organic aerosol rises, and fragile markers such as levoglucosan and fatty acids decay. That makes chemical age a clock that reads travel time, and on a fixed corridor, travel time is distance from the source.
The core idea is to stop treating plume aging as a confound. The source-apportionment literature currently regards it as a nuisance, the "aging paradox." We instead use it as the probe whose blind direction runs perpendicular to the transport footprint's. The difficulty is that the clock's rate depends on how much oxidant is in the air, which is itself uncertain, so distance trades off against oxidant level. That second degeneracy is curved and non-Gaussian. We handle it with simulation-based inference, meaning a neural density estimator trained on simulated pairs of sources and observations, and we test explicitly whether that machinery is needed or whether a simpler linear inversion suffices.
Core Claim
Discretize the domain into cells $j$ with emission rates $e_j$. For each arrival time $k$, a backward-trajectory or dispersion model gives a residence time $A_{kj}$ in each cell and a travel time $\tau_{kj}$ from that cell to the receptor. The receptor concentration is $c_k = \sum_j A_{kj} e_j$. When winds follow a persistent corridor, the rows of $A$ are nearly parallel, and the matrix has a near-null space that redistributes emission along the corridor. PSCF and CWT never acknowledge this null space. A second observable, the aged fraction of a reactive marker, weights the same cells by their travel time:
$$y_k = \frac{\sum_j A_{kj}\, e_j \left(1 - e^{-\kappa\, \tau_{kj}}\right)}{c_k},$$
where $\kappa$ is the effective first-order aging rate, the product of a rate constant and the mean oxidant concentration. Because $\tau_{kj}$ grows monotonically along a corridor, $y_k$ has sensitivity in exactly the direction where $c_k$ has none. If $\kappa$ were known, $y_k$ would be linear in $e$ after multiplying through by $c_k$, and the problem would be an ordinary augmented linear inversion. When $\kappa$ is unknown, the products $e_j \kappa$ make the posterior bilinear and curved, and a near source with fast oxidation looks like a far source with slow oxidation.
The claim has two parts. First, combining the footprint with a chemical clock shrinks posterior width along the corridor by a factor that can be predicted from the angle between the two observables' insensitive directions. This is the same complementarity logic that joint lensing and gas-pressure analyses use in cosmology (DES Y3 × ACT DR6 tSZ, which breaks the degeneracy between clustering amplitude and baryonic feedback by pairing two probes). Second, when the aging rate is uncertain, the remaining distance–oxidant degeneracy is better represented by a neural posterior than by a Gaussian one.
Prior Work and What It Leaves Open
Several papers motivate the proposal by documenting the gap. The Beijing 2013 haze study (Zhang et al.) ran positive matrix factorization (PMF), a factor analysis that splits measured chemistry into source profiles, on aerosol mass spectra, and then ran HYSPLIT trajectories and PSCF on the resulting factors. It found the secondary organic aerosol (SOA) fraction rising from 35% on clean days to 63% on polluted days. The Jinan molecular-marker study (Li et al.) did the same with levoglucosan and fatty acids and flagged that photochemical processing alters marker ratios independently of emission strength. Both run factorization and trajectory attribution as two uncoupled steps. Neither uses the oxidation signal to localize sources, and both inherit PSCF's smearing along the path. We reuse their measured quantities, the ratio of low-volatility to semi-volatile oxygenated organic aerosol and marker degradation ratios, as the clock.
Bayesian flux inversions supply the strongest existing machinery and also constrain the design. The Paris inversion (Lian et al.-style 6-year Paris CO₂ inversion) couples a 1 km WRF-Chem transport model to seven stations and uses observed boundary-layer heights to break a confusion between low mixing and high emissions. That is itself a degeneracy broken by an auxiliary observable, though it is linear-Gaussian and works on aggregated city budgets. The CHIMERE satellite system (local- to national-scale inversion) shows quantitatively that uncertainty reduction collapses for diffuse sources and under strong winds. That is exactly the regime where corridor smearing is worst, and the system offers no mechanism to recover it. We differ from both by adding a chemistry-derived second observable aimed at the along-wind null space, and by testing whether a non-Gaussian posterior is needed.
The probabilistic receptor model BAMF (Bayesian auto-correlated matrix factorization) attacks a different degeneracy, the rotational ambiguity of PMF, using temporal priors and Hamiltonian Monte Carlo, but it has no spatial transport. The 21-cm forest work (likelihood-free inference with spline flows) supplies the method. It shows normalizing flows recovering parameters within 1σ in a degenerate, non-Gaussian regime where Fisher and Gaussian MCMC analyses fail. It also warns that such flows misbehave out of the training distribution, which constrains how we validate on real data.
Research Program
We will build three components. The first is a synthetic test bed for "observing system simulation experiments," where the true sources are known. We take HYSPLIT backward-trajectory ensembles driven by real reanalysis winds for the Beijing January 2013 and Jinan winter 2016 receptor sites. We place known sources in the domain, including deliberately adversarial pairs such as a weak near source and a strong far source on the same corridor. We forward-simulate concentrations and aged fractions with first-order aging, drawing $\kappa$ from a realistic range and adding measurement noise matched to the reported instruments.
The second component is an identifiability analysis. For each trajectory set we compute the singular spectrum of $A$ alone and of $A$ stacked with the linearized clock operator. From this we report the angle between their insensitive directions and the predicted posterior shrinkage along each corridor. This produces an identifiability map: a statement of where on the map a source location can and cannot be resolved, which PSCF never provides.
The third component is an inference ladder with three rungs. The first rung is an augmented linear-Gaussian inversion with $\kappa$ fixed. The second is the same inversion with $\kappa$ marginalized by MCMC. The third is neural posterior estimation, using a spline normalizing flow trained on simulated pairs following the 21-cm design, with $\kappa$, the boundary-layer height scaling, and the emissions all inferred jointly. Moving up the ladder is justified only if the lower rung visibly fails.
Baselines
The do-nothing reference is PSCF and CWT computed on the same trajectories and the same threshold or concentration data, applied either to total mass or to each PMF factor. This is the two-step approach of the Beijing and Jinan studies. We score it by converting its map to a source location estimate, taking the cell of highest value per corridor. On this along-corridor localization question we found no reported result on Lacuna for PSCF or CWT accuracy against known truth. The synthetic test bed therefore supplies the reference number, and beating it means a statistically lower median along-corridor distance error on the adversarial configurations.
The strongest existing competitor is a regularized Bayesian linear inversion on the footprint alone, in the style of the Paris and CHIMERE systems, with Gaussian priors on emissions. It is strong because its prior does select a single answer within the null space, and it can use multiple stations. It runs on identical trajectories, noise, and prior covariance, with its prior hyperparameters tuned by the same marginal-likelihood search we give our method. The closest published numbers come from the CHIMERE system. Uncertainty reduction exceeds 50% only for point sources above roughly 2 MtC/yr and diffuse sources above roughly 10 MtC/yr, and it drops to 74% at 13.5 m/s winds versus 97% at 1.8 m/s (CHIMERE inversion). These figures are for satellite CO₂ columns, not aerosol receptors, so they set the regime rather than a directly comparable value. Beating this baseline means two things. Our posterior must be narrower along the corridor without losing calibration, with nominal 90% intervals covering the truth about 90% of the time. And it must keep doing so when the baseline's prior is deliberately mis-centered, which exposes whether that baseline was resolving the degeneracy or merely restating its prior.
The key ablation removes the clock while keeping everything else, which isolates the contribution of chemistry. A second ablation fixes $\kappa$ at its prior mean rather than inferring it, which isolates whether the flow's handling of the curved distance–oxidant valley matters.
Why not just…?
Why not just add more receptor sites and triangulate? This is a real threat. Two stations viewing a source from different angles produce footprints whose insensitive directions cross, which is the textbook way to break this kind of degeneracy. Our answer is empirical. Using CARE-China site geometry (Liu et al.), we compare along-corridor error for single-site plus clock, two-site without clock, and two-site plus clock. If two sites without clock match single-site plus clock everywhere, the contribution narrows to the single-site campaigns that make up much of the literature, such as Beijing and Jinan, and we will say so. We expect complementarity: the clock still helps when sites share a dominant wind corridor, which is common during regional stagnation episodes.
Why not just use the Bayesian prior to pick among degenerate solutions? A prior does not break a degeneracy. It replaces missing data with an assumption. The mis-centered-prior experiment tests this directly. If the footprint-only inversion follows its prior along the corridor while the clock-augmented one follows the truth, the prior was hiding the problem.
Why not just use a linear inversion with a literature aging rate, and skip the neural flow? It may well be enough. That is why it is the first rung of the ladder. The flow is retained only if, with $\kappa$ uncertain over a realistic range, the Gaussian posterior's coverage fails on the curved valley. This mirrors the failure the 21-cm study reports for Gaussian methods in its non-Gaussian regime. If coverage holds, we report the simpler method as the result.
Proposed Experiments
The decisive synthetic check uses the adversarial pairs. Weak-near and strong-far configurations on one corridor give near-identical concentration series, so PSCF and the footprint-only inversion should be unable to tell them apart, while the clock-augmented posterior should separate them. We will measure how this separation degrades as the uncertainty in $\kappa$ widens and as boundary-layer errors are injected. Success is a clear separation at realistic aging-rate uncertainty, measured as the posterior probability assigned to the correct configuration. Failure would be separation that holds only when $\kappa$ is known to within about 20%. That failure would itself be useful: it would tell field campaigns that co-measuring an oxidant proxy is necessary.
On real data, we will apply the method to the Beijing 2013 aerosol mass spectrometer factors, using oxidized-to-semivolatile organic ratios as the clock, and to the Jinan levoglucosan and fatty-acid ratios. Because no ground truth exists there, the checks are consistency tests against outside information. Recovered source distances should be the same when computed from two independent clocks, organic oxidation state and marker decay. Inferred near-receptor emissions should fall during the COVID-19 control period in a province where control measures were documented (Guangdong COVID-19 study), wherever matching receptor chemistry is available.
Flow posteriors on real data will be accepted only if the observed summaries fall inside the simulated training distribution, checked with a held-out density score, given the out-of-distribution failures documented for such flows.
References
- Constraints on cosmology and baryonic feedback with joint analysis of Dark Energy Survey Year 3 lensing data and ACT DR6 thermal Sunyaev-Zel'dovich effect observations
- Characterization of submicron aerosols during a month of serious pollution in Beijing, 2013
- Characteristics and Sources of Organic Aerosol Markers in PM2.5
- Can we use atmospheric CO 2 measurements to verify emission trends reported by cities? Lessons from a 6-year atmospheric inversion over Paris
- A local- to national-scale inverse modeling system to assess the potential of spaceborne CO 2 measurements for the monitoring of anthropogenic emissions
- A novel probabilistic source apportionment approach: Bayesian auto-correlated matrix factorization
- Deep learning-driven likelihood-free parameter inference for 21-cm forest observations
- Characteristics of PM 2.5 mass concentrations and chemical species in urban and background areas of China: emerging results from the CARE-China network
- The Impact of COVID-19 Control Measures on Air Quality in Guangdong Province