The Nonlinear Physics and Statistical Field Theory of Twilight Clouds: Hovering Dynamics and Dynamic Inertia Emergence

Appendix Emergent Noise Variance: Physical Constraint

Introduction

The coupling of cloud microphysics with climate dynamics remains one of the most challenging problems in atmospheric science. Failure to accurately parameterize clouds is responsible for the largest source of uncertainty in global climate projections [1]. Addressing this requires a more precise description of the nonlinear processes governing the onset and development of clouds, especially considering the vast spatial and temporal scale separation between global circulation and fine-scale turbulent motions.

Among all cloud types, the warm and smallest ones—often termed “twilight clouds”—are particularly difficult to capture in climate models despite their abundance [2, 3, 4, 5]. These clouds mark the transition from non-activating haze to fully activated cloud droplets. Their constituent particles reside in a precarious, marginal state highly sensitive to initial perturbations, with microphysical structures centered near the critical activation radius [6]. Complementary work has shown that these small clouds often form in buoyant humidity pockets, resulting in sub-Lifting Condensation Level (subLCL) clouds that oscillate between haze and cloud states [7, 8, 9].

Understanding the dynamics of droplet transitions in this marginal environment is the central goal of this study. To tackle this, we develop a theory that starts at the scale of an isolated moist parcel—where droplet growth is governed by Köhler thermodynamics, molecular diffusion, and mixing-induced supersaturation removal—and then lifts the resulting slow dynamics to a macroscopic cloud field through non-equilibrium statistical field theory. The unifying structure across these scales is the emergence of a slow manifold: a dynamical backbone on which the fast supersaturation environment is slaved to the slower evolution of the droplet-growth variable.

Classical Köhler theory [10] robustly explains the activation of individual cloud condensation nuclei [11], but it is not by itself a theory of marginal clouds. In real clouds, populations of growing droplets deplete and reshape the surrounding supersaturation field [11, 12, 13], while turbulent mixing and weak updrafts continuously perturb the thermodynamic balance. Detailed parcel models can incorporate these effects by coupling thermodynamics, droplet growth, and vertical motion [6, 7, 8, 9], but their complexity makes it difficult to isolate analytical mechanisms and scaling laws.

Here we focus on weak-updraft regimes, where supersaturation production is limited and droplets remain close to critical activation conditions. To expose the organizing dynamics of this regime, we adopt a reduced dynamical-systems approach [14, 15, 16], grounded in first-principles microphysics and closed by parameterizing unresolved turbulent exchange [14]. The starting point is the core Köhler diffusional-growth equation [15, 16], which is especially relevant near cloud base, where activation is initiated and where small, marginal or twilight clouds are prone to be encountered. This reduction is justified by a separation of timescales: supersaturation adjusts on microphysical timescales that are typically much shorter than the mesoscale evolution controlling the ascent. The vertical velocity may therefore be treated as quasi-steady at leading order, while retaining the essential competition between adiabatic cooling, condensational depletion of supersaturation, and unresolved mixing. Measurements in newly forming shallow cumuli show substantial supersaturation fluctuations superposed on a quasi-steady background, emphasizing the competition between turbulent mixing and condensational relaxation [17].

In our reduced parcel model approach, environmental mixing enters as a net supersaturation removal rate $\lambda$, with turbulent variability represented by stochastic fluctuations of intensity $\sigma$ (Section VI). The deterministic dynamics already reveal the central geometry: our analysis exposes that supersaturation rapidly approaches and tracks its quasi-steady nullcline, with a lag quantified below (§ 0.6), and mixing-induced removal drives a saddle-node bifurcation on this manifold (Section IV). The saddle-node creates an edge state that replaces the classical Köhler activation radius by a dynamically shifted threshold $R^{*} > R_{c}$. This defines a forbidden zone: a region below the classical Köhler equilibrium curve where deterministic intuition would predict immediate return to haze, but where the slow-fast dynamics create a metastable transition layer.

The stochastic analysis then explains why this layer can persist. Using a small-noise reduction [18, 19], we derive in Section VII an effective stochastic differential equation constrained to the slow manifold. A Lamperti transformation [20] casts this reduced dynamics into a gradient form [21, 22], exposing the potential landscape near the activation edge (Section VII B). In the supercritical mixing regime $\lambda> \lambda_{\mathrm{c}}$, the effective potential flattens near the saddle-node (Section VII C). Droplets need not immediately activate or evaporate; instead, they hover near the edge state for extended periods, producing intermittent transitions between haze and activation (Section VII D). This mechanism gives a dynamical explanation for the fleeting, optically faint character of twilight clouds and recovers the observed scale of 1–2$\mu$m droplet-size distributions in weak-updraft regimes [6].

The 0D parcel description, however, is intrinsically local. It captures temporal hovering, but it cannot describe how neighboring cloud elements communicate through the supersaturation field, nor how hovering states organize across space. We therefore pass from isolated parcels to a spatially extended mean-field theory. Starting from the standard fluid-dynamical description of turbulent warm-cloud microphysics, and retaining Köhler activation physics at the level of a coarse-grained droplet population, we derive a degenerate stochastic droplet–supersaturation reaction–diffusion system; see Section VIII and Supplementary Note 5. The slow field is the mean-square-radius field $\phi(x,t) = \langle R\rangle_{x,t}$, representing the local coarse-grained droplet population, while the fast field $\chi(x,t)$ is the resolved supersaturation bath sampled by that population. The system is degenerate in the sense that diffusion does not act on both fields symmetrically: the slow droplet field $\phi$ evolves locally through condensation and Köhler feedback, whereas diffusion acts only on the fast supersaturation field $\chi$, through the effective scalar diffusivity $D_{\chi}$, which includes the turbulent eddy-diffusivity contribution from the unresolved transport fluctuations. Thus, the model is not a standard two-field reaction–diffusion system with independent smoothing of both components; it is a partially parabolic slow–fast system in which neighboring droplet populations communicate indirectly through the diffusing supersaturation bath. This construction provides a leading-order closure that bypasses the need to resolve the full droplet-size distribution through bin microphysics [e.g. [23, 24, 25]], while still retaining the dominant nonlinear Köhler feedback mechanism inherited from the parcel level (Appendices A and B).

We then embed this stochastic reaction–diffusion system in the Martin–Siggia–Rose–Janssen–de Dominicis (MSR) path-integral formalism [26, 27, 28, 29, 30, 31] (Section VIII). This step is not merely a change of notation: it recasts the droplet–supersaturation system as a response theory in which the fast supersaturation bath can be eliminated at the level of the action. The resulting construction reveals a useful structural analogy with the polaron problem in solid-state physics [32]. In the Fröhlich–Gurari picture, a slow electron becomes dressed by a polarization envelope of a fast medium, thereby modifying its effective dynamics [33, 34]. In the cloud problem, the slow object is not an individual microscopic droplet, but the local coarse-grained droplet population represented by $\phi$. The corresponding fast medium is the supersaturation bath $\chi$, and the analogue of the polarization envelope is a statistically localized supersaturation halo generated by the turbulent bath response. The resulting object is therefore a collective “cloud polaron”: a droplet field dragging a supersaturation response, rather than a single droplet carrying a microscopic vapor shell.

Using a Gurari-type variational shift, we integrate out the fast supersaturation bath around its turbulently dressed slow manifold. As in Gurari’s treatment of the polaron [33], the displacement of the fast field is a variational object, fixed by a stationarity condition; it is the retarded, spatially screened response of the bath to the droplet field, and its denominator has the same structure as Gurari’s optimal phonon displacements, with the bath relaxation rate in place of the phonon frequency, the diffusive cost in place of the recoil energy, and the evolution of the droplet field in place of the motion of the electron. The turbulent fluctuations enter this bath through the supersaturation equation itself: the unresolved mixing and exchange term with the dry air is decomposed into a mean removal rate and a fluctuating component, producing a multiplicative noise of the form $\sigma\chi\eta$ at the level of the fast field $\chi$, with $\eta$ denoting Gaussian noise that is white in time and correlated in space, as specified in Eq. (66) below. Thus, turbulence is not imposed as an additive random force on the droplet equation. It acts by modulating the supersaturation bath sampled by the local droplet population.

This distinction is essential in the MSR reduction aimed at a closed $\phi$-equation. Because the noise amplitude is state-dependent, $\Sigma(\chi) = \sigma\chi$, and because the mixing-rate fluctuations have a finite correlation time, so that their white-noise limit is of Stratonovich type [35], the transformation from noise histories to supersaturation histories carries a non-trivial functional Jacobian [30, 36, 37, 38]. After regularization at the Kolmogorov scale, converting the Stratonovich equation to Itô form gives the relaxation-rate correction $-\sigma_{\Sigma}^{2}/\tau_{K}$. The equivalent midpoint formulation accounts for this correction through its prescription-dependent measure terms (Appendix C); it is included only once. Here $\sigma_{\Sigma}^{2}$ is the dimensionless mixing-fluctuation strength and $\tau_{K}$ is the Kolmogorov turnover timescale. The correction offsets part of the background supersaturation relaxation, so that turbulence dresses the Köhler slow manifold itself rather than merely adding fluctuations around it. At the parcel level, this turbulent dressing shifts the saddle-node threshold to a larger effective mixing rate and shrinks the forbidden zone. In physical terms, a stronger drying environment is required to suppress activation once the local coarse-grained droplet population is coupled to, and dressed by, its turbulent supersaturation bath.

The same reduction also determines how the local dressing acquires a spatial character. When the zero-diffusion bath limit is relaxed, the supersaturation response is no longer pointwise: neighboring coarse-grained droplet populations interact through the non-local operator $\mathcal{L}_{\phi}^{-1} = [\Gamma^{*}(\phi) - D_{\chi}\nabla^{2}]^{-1}$ through the non-local operator $\mathcal{L}_{\phi}^{-1} = [\Gamma^{*}(\phi) - D_{\chi}\nabla^{2}]^{-1}$, where $\Gamma^{*}(\phi)$ is the turbulently dressed supersaturation relaxation rate toward the dressed slow manifold $\chi_{0}^{*}(\phi)$, while $D_{\chi}$ is the effective supersaturation diffusivity. The Bessel Green’s function of $\mathcal{L}_{\phi}^{-1}$ defines then the screening length, $L_{\mathrm{scrn}}=\sqrt{D_{\chi}/\Gamma^{*}(\phi)}$, over which the supersaturation halos attached to neighboring coarse-grained cloud elements interact effectively. Written in local form, this screened response yields the pseudo-parabolic operator $-\nabla\cdot\left(\epsilon(\phi)\nabla\partial_t\phi\right)$, with coefficient $\epsilon(\phi)=D_{\chi}/\Gamma^{*}(\phi)=L_{\mathrm{scrn}}^{2}$, which records the spatial footprint of the dragged supersaturation halo: a change of the droplet field at one point alters the growth tendency over the whole screening disk. When the finite relaxation time of the bath is retained at first order in the ratio of slow to fast rates, the halo also lags behind the droplet field, and this lag renormalizes the coefficient of $\partial_t\phi$. This is the cloud analogue of the polaron effective mass [33, 34], and, as for the polaron, it is set by the ratio of the bath-mediated coupling rate to the bath relaxation rate. Together, the footprint and the lag form the dynamic-inertia operator, which confers on the macroscopic droplet field a memory of the supersaturation bath it drags (Figure 18).

The same stationarity condition that fixes the Gurari displacement also yields the retarded, nonlocal bath response from which the screening and the lag both follow, so that the two components of the dynamic inertia are the spatial and temporal faces of a single dressed object.

The resulting non-Markovian droplet-field closure is a nonlinear pseudo-parabolic stochastic partial differential equation (SPDE) for $\phi$ of Barenblatt–Zheltov–Kochina (BZK) type [39, 40], reminiscent of pseudo-parabolic models arising in heat conduction in materials with two distinct temperatures [41], fluid seepage through fissured rock formations [39], and radiative transfer theory [42]. Its stochastic forcing is not imposed externally: it is the colored macroscopic noise left behind by the elimination of the supersaturation bath.

This structure also resonates with recent progress on the reduction of SPDEs forced by multiplicative noise acting in the unresolved, or orthogonal, component of the dynamics. In such settings, stochastic parameterization formulas show that fluctuations in the eliminated sector are transmitted to the resolved large-scale modes as non-Markovian stochastic coefficients in the reduced equation [43, 44]. These coefficients depend explicitly on the history of the noise path, and their memory content is determined self-consistently by the noise intensity and by the way the unresolved fluctuations interact with the resolved variables through the nonlinear terms of the SPDE. The present cloud reduction provides a physically explicit realization of this mechanism: multiplicative turbulent supersaturation fluctuations in the fast bath become, after elimination, colored and memory-bearing stochastic forcing in the slow droplet-field equation.

Other spatial terms arising in the non-Markovian closure are equally structured. At long wavelengths, the halo-mediated coupling acts on the droplet field as $\mathcal{K}_{\phi}=2\nabla\cdot\left(\epsilon(\phi)\nabla\chi_{0}^{*}(\phi)\right)$, where $\chi_{0}^{*}(\phi)$ is the turbulently dressed slow manifold, built on the mean-field projection $\bar{f}(\phi)$ of the Köhler curve (Appendix A). Its sign is set by the slope of the dressed slow manifold, which increases with $\phi$ in the haze regime and decreases beyond a maximum located close to the Köhler critical radius. In the haze regime the coupling acts diffusively, smoothing unactivated aerosol fluctuations (Section IX G). Beyond the maximum, the same coupling becomes anti-diffusive, reinforcing spatial contrasts between cloudy and dry regions; the pseudo-parabolic operator keeps this backward diffusion well posed [45, 46]. In this sense, the term has a “rich-get-richer” structure: activated coarse-grained droplet populations draw supersaturation from their surroundings and sharpen their contrast with neighboring haze-like regions. This provides a thermodynamic mechanism for edge sharpening and pattern formation, complemented by a nonlinear squared-gradient term that emerges from the expansion of $\mathcal{K}_{\phi}$ and has the form that, for an interface height, generates Kardar–Parisi–Zhang (KPZ) growth [47, 48, 49]; see Section IX G.

We test the spatial reduction directly in a regime where the bath’s temporal memory is neglected but its spatial response is retained. The resulting Markovian closure evolves only the mean-square-radius field, with the same screened operator acting on drift and forcing (Section IX F). Starting from a uniform edge state, it reproduces the coupled droplet–supersaturation system’s progression from hovering through connected haze regions to activated cloudlets (Figure 19). Agreement extends to the evolving distribution of local root-mean-square radius and the number and area of cloudlets (Figure 20). These comparisons show that eliminating the bath can preserve both population statistics and spatial organization, providing a numerical test of the screened field theory. A second comparison starts from uniform haze near the saddle-node threshold (§ 0.11 B). Keeping the screened deterministic drift unchanged, we restore the residual bath’s finite relaxation (making the closure non-Markovian) and find improved agreement with the parent system in the frequency and spatial extent of recurrent hovering (Figure 21 and Figure 22). The bath sets the fluctuation-memory timescales without additional fitting. This experiment isolates the contribution of colored forcing to the non-Markovian reduction.

By uniting slow-fast parcel microphysics, turbulence modeling and non-equilibrium statistical field theory, this work provides a first-principles route from Köhler activation to a macroscopic stochastic cloud-field equation. The theory explains how turbulent mixing creates hovering near a saddle-node edge state, how turbulent dressing shrinks the forbidden zone, how screened supersaturation halos couple neighboring cloud elements, and how delayed bath relaxation generates dynamic inertia and colored noise. These mechanisms identify the operators through which subgrid supersaturation fluctuations organize the morphology and persistence of twilight clouds.

The reduction principle behind these results has a scope well beyond cloud physics. It transposes to fields the polaron mechanism by which a slow particle, by polarizing a fast medium, acquires a dressed energy and an effective mass. Its premise is generic: a slow field with little or no intrinsic transport, coupled to a relaxing, diffusive environment that it modifies. Eliminating that environment within the MSR formulation dresses the local dynamics, induces screened interactions through the environmental halo, renormalizes the inertia of the slow field, and colors its fluctuations. The same premise underlies dryland vegetation–water feedbacks, in which slowly spreading biomass redistributes soil water and organizes vegetation patterns [50, 51], and populations of oscillators or cells communicating through a rapidly diffusing chemical, whose adiabatic elimination yields a screened nonlocal coupling of the same form [52]. It extends naturally to bacterial colonies whose growth and mobility depend on a diffusing nutrient [53] and to transport in heterogeneous media, where distributed exchange times generate memory [54]. The polaron picture also has a classical counterpart in plasma physics, the dressed test particle with its shielding cloud [55], which links the reduction to dense and complex plasmas. In each setting, the slow-field closure framework described in this work offers a systematic route from the coupled dynamics to an effective equation for the slow field that retains the interactions, memory, and fluctuations inherited from its environment (Section XI).

The remainder of this paper is structured as follows. Sections III and IV introduce at the parcel level, the reduced droplet-growth framework and the effect of mixing-induced supersaturation removal, notably in the creation of a saddle-node activation threshold. Section V analyzes the slow manifold and the exponential supersaturation relaxation rate to it. Sections VI and VII derive the reduced stochastic dynamics and use the Lamperti transformation to characterize hovering. Section VIII then lifts the parcel theory to the MSR field level. In Section VIII B, we construct the isolated slow-droplet action $\mathcal{S}_{\phi}$, the isolated fast-bath action $\mathcal{S}_{\chi}$, and the condensation coupling action $\mathcal{S}_{\mathrm{int}}$. We then integrate over the fast supersaturation bath using a Gurari-type variational construction [33] in Section VIII C. The first resulting closure, derived in Section VIII C.6, is obtained under two simplifying assumptions: a Markovian approximation, in which the bath relaxes instantaneously relative to the slow droplet field, and a zero-diffusion bath limit, in which the spatial redistribution of $\chi$ is neglected at leading order. These restrictions are relaxed in Section IX, where finite bath relaxation and spatial supersaturation redistribution generate the non-Markovian pseudo-parabolic droplet-field equation, including dynamic inertia, colored noise, and Köhler-mediated anti-diffusion. Section X tests the spatially correlated Markovian closure and the effect of residual-bath fluctuation memory against the coupled field system. Section XI discusses the implications for twilight clouds, subgrid cloud modeling, and response-based cloud closures.

The Fleeting and Visually Faint Nature of Twilight Clouds