Satellites mostly measure aerosol as a single number per location, the aerosol optical depth (AOD). AOD is the total haze in a vertical column, summed over all altitudes. Many of the things we care about depend on where in that column the particles sit. Air pollution exposure depends on the lowest few hundred metres. Cloud brightening depends on particles near cloud base. Heating of the atmosphere depends on the full vertical layering. Splitting a column total back into a vertical profile is badly underdetermined. The best current method, AODisaggregation: toward global aerosol vertical profiles, uses a Gaussian process (a probabilistic model that places a smooth random-function prior over the profile). It shapes that prior with weather variables such as humidity, temperature, pressure and updraft. Its authors note that the method fails when the amount or size of particles changes in ways the weather does not reflect. Today, every column gets a best-guess profile, whether or not that guess is trustworthy.
Non-invasive cardiology faced the same structure and handled it differently. A cheap external measurement, such as image geometry, is inverted to estimate a hidden internal quantity, such as the pressure drop across a narrowed artery. The estimate is then judged as a decision. The question is whether it can safely replace the risky catheter for this patient, or whether this patient still needs one. This note proposes to import that framing, and two concrete lessons from that literature, into aerosol profile retrieval. First, we add particle-property information to the weather-shaped prior. This mirrors how a two-step coronary model added patient-specific covariates on top of a physics surrogate. Second, we build and evaluate a referral rule that flags the columns where the inferred profile should not be trusted and where a direct measurement (lidar or aircraft) is actually needed. The deliverable is profile retrieval that knows its own failure cases. It is scored by how much error remains on the columns it keeps, at a fixed budget of columns sent for direct measurement.
Core Claim
The claim has two parts. First, the known failure mode of weather-conditioned disaggregation can be partly repaired with column-integrated particle-property covariates. Here weather variables move but particle loading or size moves independently. The covariates are the Ångström exponent (how AOD changes with wavelength, a proxy for particle size) and the fine-mode fraction. Both are available from multi-wavelength photometers and from model output. Second, and more important, the posterior uncertainty of the disaggregated profile can be turned into a calibrated referral rule. Under that rule, holding back a small fraction of columns for direct measurement substantially reduces error on the retained columns, for the decision-relevant quantities. These quantities are the share of extinction below 2 km and the presence of an elevated layer above roughly cloud-top height.
What is new is not the Gaussian process or the covariates individually. It is treating column-to-profile inversion as a triage problem, with an explicit referral budget and sensitivity/specificity on decisions. The existing work instead treats it as a regression scored only by average profile correlation.
Where the Two Directions Meet
The aerosol literature motivates the decision targets. Constraining the Twomey effect from satellite observations argues that column AOD is a poor stand-in for the particles that seed clouds. Clouds interact with aerosol at cloud base, and humid swelling there inflates AOD without adding seed particles. Estimating PM2.5 in Beijing-Tianjin-Hebei from MODIS AOD shows that column AOD cannot reliably predict ground-level fine particulate matter without vertical structure. Radiative and thermodynamic responses to aerosol extinction profiles over South Asia finds that climate models underestimate aerosol in the lowest 2 km. Simulated photochemical response to observational constraints over North China shows that photolysis responds to the actual vertical distribution rather than a uniform one. Together these papers fix our two targets: near-surface fraction, which matters for exposure and the boundary layer, and elevated-layer detection, which matters for cloud interactions. They also explain why average profile correlation is the wrong primary score.
The cardiology literature supplies the method and the evaluation design, not the physics. Coronary artery decision algorithm trained by two-step machine learning trains a Gaussian-process surrogate on synthetic flow simulations. It then adds patient biometrics in a second model, beating pure simulation on stent decisions. We reuse this structure: a weather-conditioned profile prior, followed by particle-property covariates that capture what the "fluid environment" misses. Validation and diagnostic performance of a CFD-based method for aortic coarctation reports diagnostic accuracy against a guideline rule rather than regression error alone. We adopt that habit of reporting decision accuracy against a simple rule. Deep learning prediction of pulmonary hypertension in newborns constrains our design. Its detection F1 fell from 0.92 in cross-validation to 0.78 on a held-out set, a warning that in-distribution scores flatter inversion models. We therefore require geographic and seasonal hold-outs.
We do not claim that blood flow and boundary-layer transport share equations. The file's own comparison notes that they operate in very different flow regimes. The bridge is the shape of the decision problem, and we test only that.
Research Program
The model starts from the AODisaggregation setup. A profile prior is conditioned on vertically resolved temperature, pressure, humidity and updraft. Observed column AOD then updates that prior through a log-normal observation model, the choice validated against 1,315 AERONET stations in that paper. We add a second input channel of column-level particle descriptors. These are the Ångström exponent and fine-mode fraction. They are computed from per-wavelength or per-mode optical depth in the ECHAM-HAM output; the precise form depends on which fields the archived runs contain. These descriptors modulate the profile's scale height and its near-surface weighting through a small learned map. The descriptors enter only as column totals, so the model remains usable wherever a multi-wavelength column instrument exists.
From the posterior over profiles we derive posterior distributions for the two decision quantities. The first is the fraction of extinction below 2 km. The second is a binary indicator: whether extinction above 2.5 km exceeds a threshold share of the column. That threshold is chosen once on training data to match the smoke-over-stratocumulus situations documented by CLARIFY-2017. Each column then gets a referral score: the posterior probability that the decision would flip, or the posterior interval width of the near-surface fraction.
We then study three things in turn. The first is calibration: do 80% posterior intervals contain the truth 80% of the time, separately by region and by aerosol regime? The second is the coverage–risk curve, which shows error on retained columns as the referred fraction grows. The third is a diagnosis of which columns get referred. Our hypothesis is that referred columns are concentrated where the humidity–extinction relationship is weak, which is exactly the failure mode the original authors identified.
Baselines
The naive reference is an idealized exponential profile scaled to match column AOD. The AODisaggregation paper evaluated it on more than 147,000 ECHAM-HAM columns, reporting a mean bias of −2.40 and a profile Pearson correlation of about 51% (AODisaggregation). We rerun it on our splits. Under our referral protocol it can only refer columns at random or by AOD magnitude, which gives the floor for the coverage–risk curves.
The strongest existing method is the AODisaggregation Gaussian process itself. On the same simulator it reports a mean bias of −0.167 and a profile correlation of about 71%. We reimplement it with the same predictors, the same sparse variational inference, and the same inducing-point budget and hyperparameter search as our model. Both models therefore get identical training columns, held-out regions and tuning compute. It is a strong competitor because it already produces posterior uncertainty, so it can drive a referral rule with no modification. We report it in two forms: as published, and with its own posterior used for referral.
Beating it means two things on the geographic hold-out. At a 10% referral budget, retained-column error on near-surface fraction must fall by a margin larger than the spread across five random seeds and region splits. Elevated-layer detection must also improve in sensitivity at matched specificity. Matching it on average correlation alone does not count as beating it.
Two ablations isolate the ingredients. The first adds particle covariates and makes predictions but refers by AOD magnitude only, which tests whether the covariates help accuracy. The second uses no covariates but keeps the posterior-based referral, which tests whether referral alone does the work. A third, flexible baseline is described below. We found no reported result on Lacuna for referral or coverage–risk evaluation of aerosol disaggregation, and no reported result on Lacuna for any disaggregation method scored against GASSP or CLARIFY profiles.
Why not just…?
Why not just use the AODisaggregation posterior as it is for referral? This is the most serious alternative, and it is the second ablation. If the covariate-free posterior already ranks columns well, then the particle covariates add nothing to the referral story. We would report that honestly and recast the contribution as an evaluation protocol plus a calibration study. Our expectation that it falls short comes from the original authors' own observation. Columns whose extinction varies independently of the weather variables are exactly where a weather-only prior is confidently wrong. Confidently wrong columns are invisible to an uncertainty-based referral rule. The test is decisive: we compare calibration on columns binned by the strength of the local humidity–extinction relationship.
Why not just train a neural network on 147,000 simulated columns? With that much simulator data, a flexible regressor from weather profiles, AOD and particle descriptors to extinction profiles could beat a Gaussian process on average error. That is a real threat to any accuracy claim. Our response is to include it: a multilayer network and a gradient-boosted model, each with a deep-ensemble variance for referral, under the same splits and tuning budget. The proposal's claim concerns calibrated referral under geographic shift, not the choice of function class. If an ensemble network wins there, we adopt it as the predictor and keep the covariates and the protocol. The newborn echocardiography result, where held-out F1 fell to 0.78, is why we expect flexible models to degrade most when moved across regions.
Why not just collect more direct profiles with lidar or aircraft? Spaceborne lidar has a narrow swath, and aircraft campaigns compiled in GASSP are sparse in space and time. Direct profiling is the scarce resource. A referral rule exists to tell us where to spend it, so this alternative is a premise of the proposal rather than a substitute for it.
Why bring cardiology in at all instead of standard selective prediction? Selective prediction is the general method of letting a model abstain on uncertain cases. The clinical papers contribute specific choices that generic selective prediction leaves open. They score decision quantities against a simple rule, add covariates that capture the patient-specific "fluid" (here, the aerosol type), and insist on external validation. If those choices make no measurable difference relative to plain posterior-variance abstention, the ablations will show it.
Proposed Experiments
The primary testbed is the ECHAM-HAM column set used by AODisaggregation, which provides complete 3D truth. We hold out whole continents and one season, train on the rest, and report four sets of results: profile bias and correlation for comparability with the published −0.167 and ~71%; error on near-surface fraction; elevated-layer sensitivity and specificity; and interval calibration. Each set is reported in-distribution and out-of-region, and the gap between the two is a headline number.
A secondary, real-world check uses GASSP aircraft profiles and CLARIFY-2017 smoke-above-cloud profiles as independent truth. Here particle descriptors come from collocated photometer data where available. This step tests whether the simulator-trained referral rule still ranks real columns sensibly. We do not use it for tuning.
The decisive check is the coverage–risk comparison at referral budgets of 5%, 10% and 20%. The proposal succeeds only if our model's retained-column error on near-surface fraction and elevated-layer detection is clearly below that of the AODisaggregation posterior used for referral, on held-out regions and across seeds. It must also do so disproportionately on weather-decoupled columns. If the gains appear only in-distribution, or only in average correlation, the core claim is refuted.
References
- AODisaggregation: toward global aerosol vertical profiles
- Constraining the Twomey effect from satellite observations: issues and perspectives
- ESTIMATING PM2.5 IN THE BEIJING-TIANJIN-HEBEI REGION USING MODIS AOD PRODUCTS FROM 2014 TO 2015
- Radiative and thermodynamic responses to aerosol extinction profiles during the pre-monsoon month over South Asia
- Simulated photochemical response to observational constraints on aerosol vertical distribution over North China
- Coronary artery decision algorithm trained by two-step machine learning algorithm
- Validation and Diagnostic Performance of a CFD-Based Non-invasive Method for the Diagnosis of Aortic Coarctation
- Deep Learning Based Prediction of Pulmonary Hypertension in Newborns Using Echocardiograms
- The CLoud–Aerosol–Radiation Interaction and Forcing: Year 2017 (CLARIFY-2017) measurement campaign
- In-situ constraints on the vertical distribution of global aerosol