Splitting the Cloud Water Response to Aerosol into Regime Switching and Within-Regime Thinning

Adding aerosol particles to a low cloud spreads its water over more, smaller droplets. That makes the cloud brighter and cools the planet. What happens next to the cloud's total water content, its liquid water path (LWP), is the largest open question in how much aerosols have cooled the climate. Smaller droplets can suppress rain, which keeps water in the cloud. They can also evaporate faster at the cloud top, which thins the cloud. Satellite studies summarize this response as one slope of log LWP against log droplet number. Those slopes have a known problem: when checked against "natural experiments" such as ship tracks and volcanic plumes, they predict thinning that is too strong.

We propose to estimate the full probability distribution of cloud water, given droplet number and weather conditions, using a deliberately small model. The distribution is written as a weighted mix of a few fixed curve shapes, each representing a cloud regime such as raining or non-raining. With this form, the average water response splits into two parts. One is clouds moving between regimes, for example rain being switched off. The other is clouds within a regime getting thinner or thicker. We treat the two satellite instruments that measure water in different ways as two disagreeing "raters" and keep their disagreement explicit instead of averaging it away. The test that matters is transport: a model fit on global climatology should predict what happened in natural experiments it never saw. This mirrors how clinical risk scores are judged on an external hospital cohort.

Core Claim

The single-slope sensitivity mixes two physically different processes that respond differently to confounding. A parsimonious conditional density model separates them, and the separation should transport better to natural experiments than either a regression slope or a flexible deep density model. Concretely, we expect three things. First, the part of the response that comes from moving clouds between regimes transports to ship-track and volcanic cases. Second, the within-regime thinning part is where most of the confounding inflation seen by Constraining the aerosol influence on cloud liquid water path lives. Third, keeping per-instrument offsets, rather than averaging instruments, is necessary for that decomposition to be stable. If a confounder-adjusted piecewise regression transports just as well and the two parts are not stable across experiments, the density representation adds nothing, and we will report that.

The Model

Let $\ell$ be log LWP, $n$ log droplet number, $m$ a small set of meteorological adjustment covariates, and $r$ the instrument supplying $\ell$. The covariates in $m$ are cloud depth, lower-tropospheric stability, and free-tropospheric humidity. The instrument $r$ is either MODIS, an optical retrieval, or AMSR-E, a microwave retrieval. We model

$$p(\ell \mid n, m, r) = \sum_{k=1}^{K} \pi_k(n, m)\, \phi_k\big(\ell - \mu_k(n, m) - b_{k,r}\big).$$

Each term is defined as follows:

  • $\phi_k$ are $K$ fixed shape atoms drawn from a small dictionary of Gaussian, Laplacian and skewed kernels, in the style of RAD.
  • $\pi_k$ are regime weights. They are produced by a shallow tree over $m$ whose leaves hold logits that are linear in $n$.
  • $\mu_k$ are within-regime locations, linear in $n$ inside each leaf.
  • $b_{k,r}$ are per-instrument offsets.

The tree depth, $K$, and the atom choice are selected by minimum description length. This is the principle, used by Conditional Density Estimation with Histogram Trees, of choosing the model that compresses the data most. It sets model size without cross-validation.

Differentiating the mean of the distribution gives the decomposition we care about:

$$\frac{\partial\, \mathbb{E}[\ell]}{\partial n} = \sum_k \frac{\partial \pi_k}{\partial n}\,\mu_k \;+\; \sum_k \pi_k\,\frac{\partial \mu_k}{\partial n}.$$

The first sum is regime switching and the second is within-regime thinning. Cloud effective radius is not conditioned on, because Stratocumulus adjustments to aerosol perturbations disentangled with a causal approach identifies it as a mediator on the path from aerosol to water. Cloud depth is conditioned on, because the same paper shows it is a confounder that produces Simpson's-paradox artifacts when omitted. Simpson's paradox here means a relationship that reverses or distorts when a hidden grouping variable is ignored.

Research Program and Positioning

The work has three parts: build the estimator, fit it to Aqua satellite data, and test transport on held-out natural experiments. Each input direction plays a different role.

The first role is motivation from aerosol-cloud work. Constraining the aerosol influence on cloud liquid water path built joint histograms and piecewise log-log fits from MODIS and AMSR-E. It found the response positive at low droplet number and negative at high droplet number, implying about a 62% offset of the brightening effect, but it also found that ship tracks and the Kilauea plume show weaker thinning. We keep its data and its non-monotonic finding. We replace static two-dimensional histograms, which empty out once meteorology is added, with a compressed conditional density. The causal study of the Namibian deck supplies the adjustment set. Its central caution motivates us: the causal water decrease was about 6% of the non-causal regression estimate. That paper is linear with Gaussian errors, which is exactly why it cannot express regime switching. Challenges in global climate models to represent cloud response to aerosols: insights from volcanic eruptions supplies a Holuhraun counterfactual constraint. Its finding that six climate models match optical depth through compensating errors is the motivation for a decomposition. A single aggregate number can be right for the wrong reasons.

The second role is the method, from density estimation. CDTree's minimum-description-length tree with histogram leaves and RAD's dictionary atoms both inform our design. CDTree has only been tested on 14 UCI tables without distribution shift. RAD has only been tested on synthetic data. Neither separates regime occupancy from within-regime location, and neither handles multiple measurement sources. We use fixed parametric atoms rather than histograms, because histogram leaves have bounded support and cannot extrapolate. Extrapolation matters because the plume cases push droplet number to the edge of the training range.

The third role is a constraint from the linear-robustness literature. Predicting Long-Term Allograft Survival in Liver Transplant Recipients showed a six-variable elastic-net model rising to a concordance of 0.847 on an external Canadian cohort, while DeepHit, a deep survival model, fell to 0.749. In-distribution the ordering had been reversed. We take the design rule, not the domain: keep the model linear in $n$ within leaves and small by compression, and judge it on an external test.

The fourth role is a lens from inter-observer variability. More is more: leveraging multi-rater information for whole slide images grading via virtual expert panel found that keeping all pathologists' labels, rather than a majority vote, improved F1 from 72.9% to 78.2% and cut calibration error from 0.183 to 0.050. It also found the full panel beat a "golden" subset. We apply the same idea to instruments. MODIS and AMSR-E are treated as two raters with structured offsets, and neither is designated as truth.

Baselines

Naive reference. The do-nothing reference is ordinary least squares of log LWP on log droplet number over the same Aqua sample, with no covariates. Its transport error on the natural experiments defines the scale of the problem.

Strongest regression competitor. This is the piecewise log-log regression of Gryspeerdt et al., rerun on our exact sample. We augment it with the same meteorological covariates, including cloud depth, so that it differs from our model only in the density representation. It is the strongest competitor because it already captures the non-monotonic response. It achieves this with a handful of parameters, and on transport it is the model a climate scientist would actually use. The file's reported figure for it is the roughly 62% offset of the brightening effect. No reported transport-error number for it on ship tracks or Kilauea was found on Lacuna; the paper reports only that observed thinning was "substantially weaker."

Linear causal baseline. We reimplement Wright's path model from the Namibian causal study on the same variables. Its reported result is that the causal LWP decrease is about 6% of the non-causal estimate. It is run on SEVIRI data for in-domain comparison and on Aqua data for transport.

Flexible density baseline. This is a mixture density network with the same inputs, trained with the same data splits and a tuning budget matched in wall-clock time to our minimum-description-length search. It represents the high-capacity alternative.

Ablations. These isolate the ingredients. We drop the per-instrument offsets, averaging MODIS and AMSR-E LWP. We use AMSR-E alone, which is the "golden rater" option. We replace the atoms with CDTree histogram leaves. We fix tree size instead of using compression.

What counts as beating a baseline. Beating the regression baseline means lower error in predicted mean LWP change on at least two of the three natural experiments. The three are ship tracks, Kilauea, and the Holuhraun counterfactual. The improvement must be larger than the bootstrap interval over plume segments. In addition, in-distribution negative log-likelihood must be no worse than the mixture density network's by a pre-stated margin.

Why not just…?

Why not just use a piecewise log-log regression with cloud depth added? This is the real threat, and it may well match us on the mean response. Our claim rests on the decomposition, which a regression slope cannot provide. The regression gives one number per droplet-number segment and cannot say whether the plume changed how many clouds rain or how thick non-raining clouds are. The decisive experiment is to fit both models and compare their transport errors. If they tie, we check whether the two decomposition terms are stable across the three natural experiments. If the terms are not stable, we concede that the density adds nothing beyond interpretation.

Why not just fit a deep conditional density model with all meteorological inputs? The liver-transplant external test and the Holuhraun compensating-error finding both show that flexible models can fit the training distribution while mislearning its structure. The plume cases sit at the edge of the droplet-number range, where flexible models extrapolate arbitrarily. We will test this directly with the matched mixture density network rather than assume it.

Why not just use the physically cleaner instrument, AMSR-E, and ignore MODIS? AMSR-E avoids optical assumptions, so this is reasonable. However, its coarse footprint mixes broken cloud fields, and the pathology panel result suggests that dropping "worse" raters discards informative variance. The single-instrument ablation settles it. If AMSR-E-only transports equally well with stable decomposition terms, the per-instrument offsets are unnecessary, and we will drop them.

Why not just use Wright's linear causal paths? That approach handles confounding correctly but averages the rain-suppression and entrainment regimes into one coefficient. We keep its adjustment set and test whether allowing regime structure changes the transported answer.

Proposed Experiments and Decisive Checks

Training. We fit all models on Aqua 2007–2009 collocated MODIS and AMSR-E data from marine liquid clouds. Regions and times overlapping the Kilauea plume, known shipping lanes used for track identification, and the Holuhraun period are excluded.

Transport test. For each natural experiment, we set droplet number to the observed polluted values and meteorology to the observed values. We then predict the LWP distribution and compare it with the observations. For Holuhraun, the comparison is against the counterfactual clean-versus-polluted difference from Wang et al. The metrics are error in mean LWP change, Wasserstein distance between predicted and observed polluted distributions, and the fraction of the change assigned to regime switching.

In-domain check. We use SEVIRI data over the Namibian deck and check whether our model, conditioned on time since perturbation, reproduces the Fons et al. sign change from positive to negative over several hours. We then check whether that change appears mainly in the regime-switching term, as the rain-suppression account predicts.

Falsifying check. The key check is stability: if the ratio of regime switching to within-regime thinning differs in sign between ship tracks and Holuhraun, the decomposition is not physical.

Secondary, conditional portability check. If precomputed slide features for the DEPP pathology cases are accessible, we would fit the same regime-plus-rater model to four-grade labels. This time we would include the 145 "indefinite for dysplasia" cases that Grove et al. excluded, and check calibration against their reported 0.050 calibration error. This is a check that the rater-offset mechanism is not climate-specific; it is not a central claim.