Air pollution and weather influence each other. Weather moves and dilutes pollutants, and thick haze in turn dims sunlight, cools the ground, and traps the lowest layer of air, which makes the haze worse. Today the only way to measure how strong this return effect is involves running an expensive coupled weather–chemistry simulation twice, once with the feedback on and once with it off, and taking the difference. That answer depends on the simulation's own assumptions. This note proposes a way to read the feedback directly from data: take a few years of hourly readings from scattered ground stations, rebuild the hidden state of the atmosphere from each station's recent history, and ask a directional question. Does the history of the weather variables contain information about pollution that it could only hold if pollution pushes back on the weather?
The tool is convergent cross-mapping, a causality test for nonlinear dynamical systems. It rebuilds an attractor from time-delayed copies of one variable. The attractor is the geometric shape the system's trajectory traces out over time, and the delayed copies are called a delay embedding. The test then checks whether nearest neighbours on that shape predict a second variable, and whether the predictions improve as more data is added. Two obstacles have kept the test away from air quality. Monitor records are gappy and irregularly sampled. The atmosphere is also strongly driven by the daily solar cycle, and a common driver of this kind is known to fool attractor-based tests. We address the gaps with a continuous-time latent model integrated by a high-order Runge–Kutta scheme under an explicit step-size constraint. We address the shared driver by conditioning on solar phase. We validate the whole pipeline on paired coupled-model runs in which the true feedback strength is set by hand.
Core Claim
Our claim is that the aerosol–radiation feedback leaves a detectable, dose-dependent, daytime-concentrated signature in the cross-map skill between sparse surface meteorology and particulate monitors. The test depends on a reconstruction that does not fold the attractor, so false neighbours must not be created. Coarse or unconstrained latent integration creates exactly such false neighbours. The result would be the first observation-based estimate of a two-way chemistry–meteorology coupling that does not rerun the physical model. It could then be checked against model-based estimates such as the roughly 8% average and over 20% peak-daytime contribution of the feedback to surface pollution reported in Aerosol–radiation feedback deteriorates the wintertime haze in the North China Plain.
The logic follows the standard reading of cross-mapping. If variable $X$ drives $Y$, then $Y$'s reconstructed attractor carries information about $X$. Let $\rho_{Y \to X}(L)$ be the correlation between true $X$ and its estimate from neighbours on $Y$'s attractor, built from a library of length $L$. Causation $X \to Y$ shows up as $\rho_{Y\to X}(L)$ rising and leveling off as $L$ grows. Under one-way, "offline" coupling, meteorology drives chemistry and only $\rho_{\text{PM}\to\text{met}}$ should converge. Under online coupling with feedback, $\rho_{\text{met}\to\text{PM}}$ should converge as well, and its gain over a forcing-matched surrogate should scale with feedback strength.
Positioning Against Prior Work
The method comes from Latent Convergent Cross Mapping. That work replaces fixed delay vectors with the hidden state of a continuous-time recurrent model (GRU-ODE-Bayes) so that cross-mapping works on sporadic, noisy series. It was tested only on double pendulums and simulated neurons, and its systems lacked strong shared periodic forcing. We reuse the latent idea but change two things. We use a different integrator with a stability constraint, and we target a forced, two-way geophysical system where the right answer can be dialed in.
The constraint comes from Numerical and statistical analysis of NeuralODE with Runge-Kutta time integration. It proves that a Runge–Kutta-discretized neural flow stays one-to-one only when the step size is below a threshold inversely proportional to the network's Lipschitz constant, which measures how sharply the network's output can change. For cross-mapping this is not a technicality. A flow that is not one-to-one folds distinct atmospheric states onto nearby latent points, and nearest-neighbour cross-mapping then reports coupling that is not there. We turn that theorem into a monitored training constraint. Adaptive Runge-Kutta Dynamics for Spatiotemporal Prediction supplies supporting evidence that higher-order stepping inside networks pays off in accuracy per parameter (KTH MSE 39.27 with 3.8M parameters). It operates on full grids and never asks a causal question.
Two lines of work motivate the gap. Sensing with shallow recurrent decoder networks shows that delay histories from a few sensors reconstruct whole fields, ozone among them. It treats ozone as a passive tracer and says nothing about feedback direction. Air Quality Arena shows that univariate foundation forecasters struggle most on reactive species and lack meteorological covariates. Both reconstruct or forecast; neither measures coupling.
On the physical side, the Wu et al. study and Circulation-regulated impacts of aerosol pollution on urban heat island in Beijing establish the feedback with the coupled simulation system WRF-Chem. They supply our ground truth and our main confound. The Beijing study shows that large-scale circulation regulates the feedback's effect, so a common synoptic driver can mimic coupling. It also reports that nighttime absorption by black carbon can reverse the daytime cooling. That finding stops us from treating night as a clean null.
Finally, Context parroting constrains our design. On 135 stationary chaotic systems, copying nearest-neighbour segments in delay space matches large foundation models, and the authors note it breaks under non-stationary forcing. This tells us that nearest-neighbour geometry is powerful, which is the premise of cross-mapping. It also tells us that forcing must be handled explicitly, which is our conditioning step.
Research Program
The work has three parts.
The first is a twin experiment with known truth. We run WRF-Chem over several North China Plain winter haze episodes in the Wu et al. setting. The aerosol optical feedback into radiation is scaled by a factor $\alpha \in \{0, 0.33, 0.67, 1\}$, with matched initial and boundary conditions. Model output is sampled only at real monitor and weather-station locations. We then add measurement noise and the empirical missingness pattern of real stations. This yields sparse, gappy records whose feedback strength is known. A second family of runs perturbs the shared synoptic driver while holding $\alpha = 0$, to measure false positives from common forcing.
The second part is the reconstruction and test. Each station variable, or each small group of co-located variables, is encoded by a latent continuous-time model. Its hidden state evolves under a learned vector field integrated with classical fourth-order Runge–Kutta between observation times, and it updates on arrival of each observation. During training we estimate the vector field's Lipschitz constant with spectral-norm bounds and cap the step size accordingly, following the Ehrhardt et al. threshold. Solar elevation and day-of-year enter as known inputs. Cross-mapping is computed on residual dynamics conditioned on solar-phase bins, which removes the shared deterministic driver. Significance is assessed against surrogates that preserve each series' diurnal and seasonal structure but scramble the phase relationship between the two variables.
The third part applies the method to real observations. Chinese and Indian station records from AQA-Data are paired with co-located surface meteorology and with NCEP FNL operational analyses for winds. The comparison of interest is the met-to-particulate cross-map gain across haze and clean periods, day and night, and dust-storm episodes. Dust storms act as natural shocks that perturb chemistry, as shown in Widespread surface ozone reduction triggered by dust storm disturbance on ozone production and destruction chemistry.
Baselines
The naive reference point is lagged cross-correlation between surface temperature or boundary-layer proxies and PM2.5, computed on the same sampled twin-experiment records with the same gaps filled by linear interpolation. It shows how much "signal" appears from co-variation alone. Beating it means separating $\alpha=0$ from $\alpha=1$ runs where it cannot.
A second baseline is linear Granger-style regression of meteorology on lagged PM2.5, controlling for lagged meteorology and solar phase. This is the method an atmospheric scientist would try first. It gets the same inputs and covariates and the same hyperparameter search budget over lag orders.
Classical cross-mapping with fixed delay vectors, with delay and dimension chosen by standard false-nearest-neighbour and mutual-information heuristics, isolates whether the latent model is needed at all. It runs on identical gap-filled data.
The strongest existing method is Latent-CCM as published, using GRU-ODE-Bayes with its own integrator and no step constraint or forcing conditioning. Its paper reports an AUC of 1.0 for detecting causal direction on the coupled double pendulum, and roughly a tenfold larger effect size than imputation baselines on NEST neuron simulations. No reported result was found on Lacuna for Latent-CCM, or any cross-mapping method, on atmospheric chemistry data. Latent-CCM gets the same training budget, latent width, and data as our model.
Beating it means three things together. We need a higher AUC for distinguishing $\alpha>0$ from $\alpha=0$ runs, a higher rank correlation between cross-map gain and $\alpha$, and a lower false-positive rate on the shared-driver runs at matched sensitivity.
Two ablations isolate the key ingredients. The first swaps RK4 for forward Euler at the same wall-clock cost, and also removes the Lipschitz-based step cap. The second removes solar-phase conditioning. In each ablation we report the false-neighbour rate in latent space, measured against neighbour relations in the true WRF-Chem state, alongside detection accuracy.
Why not just…?
Why not just run WRF-Chem with feedback on and off, as Wu et al. did? That remains the best way to attribute the feedback inside a model, and we use it for ground truth. But its answer inherits the model's boundary-layer and microphysics schemes, which are known sensitivities of coupled modeling. It also cannot tell us whether the real atmosphere shows the coupling at the strength the model claims. An observational test is a check on the model, not a replacement for it. The real-data stage is designed so that the two estimates can disagree.
Why not just regress temperature on lagged PM2.5 with meteorological controls? This is a genuine threat, and it is why linear regression is a full baseline rather than a straw man. The case against it is that the feedback acts through a nonlinear, state-dependent channel. It matters most during stable, polluted daytime conditions, which is why Wu et al. find the peak contribution in daytime. Meanwhile the Beijing heat-island study shows circulation modulating both variables at once. A linear model with a fixed lag structure confounds the two. If the twin experiment shows regression separating the $\alpha$ levels as well as our method, we will report that and drop the claim that attractor geometry is needed.
Why not just use classical fixed-delay cross-mapping? On dense hourly series it may work nearly as well, and that would be a useful finding. The latent model earns its place only if gaps and irregular sampling matter. We therefore run both methods under increasing missingness, from none up to the station-specific patterns in the real records, and report where the curves cross.
Why not just forecast with a foundation model plus covariates and read off feature importance? Better forecasts do not answer a directional question. Air Quality Arena shows these models struggle on reactive species. Context parroting shows that large models can flatten dynamics while matching error metrics, so attribution from them would be fragile. We do not pursue this route.
Decisive Checks
Four results would settle the claim.
First, in the twin experiment, the solar-phase-conditioned met-to-PM cross-map gain must rise monotonically with $\alpha$. It must also stay within the surrogate band for all $\alpha=0$ runs, including those with a strengthened shared synoptic driver. If the shared-driver runs produce detections, the method fails as designed.
Second, the integrator ablation must show that Euler stepping, or RK4 without the step cap, raises the latent false-neighbour rate. That rise must coincide with spurious detections at $\alpha=0$. Otherwise the high-order integration component is decoration and will be removed.
Third, on real data, the detected gain should concentrate in daytime haze episodes, mirroring the daytime peak reported by Wu et al. Its nighttime behavior should differ in structure rather than simply vanish, consistent with the black-carbon effect noted by Wang et al. A gain that is uniform across day and night would point to residual forcing contamination rather than feedback.
Fourth, the observational gain, calibrated through the twin experiment's $\alpha$-to-gain curve, should yield a feedback strength we can place beside the model-based 8% figure. Agreement or disagreement there is the result the project exists to produce.
References
- Aerosol–radiation feedback deteriorates the wintertime haze in the North China Plain
- Latent Convergent Cross Mapping
- Numerical and statistical analysis of NeuralODE with Runge-Kutta time integration
- Adaptive Runge-Kutta Dynamics for Spatiotemporal Prediction
- Sensing with shallow recurrent decoder networks
- Air Quality Arena: A Large-Scale Multi-Region Ground Monitoring Dataset and Benchmark for Air Quality Forecasting with Time-Series Foundation Models
- Circulation-regulated impacts of aerosol pollution on urban heat island in Beijing
- Context parroting: A simple but tough-to-beat baseline for foundation models in scientific machine learning
- Comprehensive Monitoring of Air Pollution Hotspots Using Sparse Sensor Networks
- Widespread surface ozone reduction triggered by dust storm disturbance on ozone production and destruction chemistry