The Origin of Thunderstorm Charge Separation: Structural asymmetry and directional electron capture in ice–graupel collisions

The preprint version is available on SSRN: The Origin of Thunderstorm Charge Separation: Structural Asymmetry and Directional Electron Capture in Ice-Graupel Collisions (August 17, 2026). http://dx.doi.org/10.2139/ssrn.7309279  

 

 

The Origin of Thunderstorm Charge Separation

Structural asymmetry and directional electron capture in ice–graupel collisions

 

 

Juliet Zhong

Independent Researcher, London, United Kingdom.
ORCID: 0009-0006-5099-3671 | August 2026

 

Abstract

The microscopic origin of charge separation during collisions between ice crystals and graupel in thunderstorm clouds is a longstanding open problem. Recent measurements establishing that hexagonal ice is a strong flexoelectric material suggest one candidate mechanism, but any mechanism rooted in an intrinsic, season-independent material property must still explain why global lightning frequency is governed by convective intensity, peaking sharply in local summer while ice-bearing mixed-phase clouds exist in all seasons. Here we present a structural-asymmetry mechanism for the directionality of charge transfer between chemically identical ice surfaces. Vapour-grown ice crystals are ordered, low-defect solids whose surfaces offer few unoccupied electron-accepting states; rime-accreted graupel is a disordered polycrystalline aggregate whose grain boundaries, vacancies and frozen-in strain generate a dense manifold of deep electron traps. Collisional impact liberates surface electrons from the ordered crystal, and these are irreversibly captured by graupel trap states, producing a net directional transfer whose rate is set by convection-driven collision dynamics. The mechanism accounts for the temperature-dependent polarity reversal, the dependence on rime accretion rate, the seasonality of lightning, and discharge initiation at sub-breakdown mean fields via geometric field enhancement at crystal asperities. Five falsifiable predictions are stated.

Keywords: thunderstorm electrification; ice–graupel collision; charge separation; electron trapping; defect states; contact electrification; polarity reversal; lightning seasonality

 

 

The directionality problem

The macroscopic architecture of thunderstorm electrification is settled. In the mixed-phase region of deep convective clouds, rebounding collisions between small vapour-grown ice crystals and larger rime-accreted graupel transfer electric charge; below a reversal temperature near −15 °C the ice crystals charge positively and the graupel negatively; gravity and the updraft separate the two populations; and the resulting field culminates in dielectric breakdown [1,2,3]. What is not settled is the microscopic step at the centre of this chain: the two colliding surfaces are chemically identical H₂O, yet electrons cross the contact in a preferred direction. No symmetry of the chemistry permits this. The asymmetry must be sought elsewhere.

Wen et al. [4] demonstrated that ice Ih is flexoelectric, a property that may participate in lightning generation. The flexoelectric coefficient of ice, however, is a material constant. It is the same in January as in July. Global optical observations show that the lightning flash rate rises from roughly 35 s⁻¹ in Northern Hemisphere winter to roughly 55 s⁻¹ in Northern Hemisphere summer, and over extratropical land the winter rate collapses to nearly zero while ice-bearing clouds remain ubiquitous [5,6]. Lightning frequency tracks surface heating and convective available potential energy, not the presence of ice. This observational fact imposes a constraint on all candidate mechanisms: the rate-limiting factor of thunderstorm electrification resides in convection-driven collision dynamics, not in an intrinsic electromechanical property of the ice itself. A recent proposal by Kang et al. [7] attributes charge directionality to the diffusional transport of H+ ions and their trapping at L-orientational defects, with the asymmetry arising from the distinct formation pathways of ice crystals and graupel. The present paper advances a competing hypothesis in which the mobile carrier is the electron rather than the proton, and the trapping substrate is the electronic defect-state manifold rather than the orientational-defect population. A mechanism that satisfies the seasonal constraint by construction is presented below.

One material, two solids

Ice crystals and graupel share a chemistry and differ in a history, and the history is written into the lattice. A vapour-grown crystal is assembled molecule by molecule at deposition rates slow enough for each admolecule to reach its equilibrium site in the hexagonal Ih structure. The product possesses long-range order on the oxygen sublattice, well-developed facets, and low densities of vacancies, dislocations and grain boundaries [8,9]. Electronically it is a wide-gap insulator (Eᵧ ≈ 8.8 eV; [10]) whose ordered surface presents few localized states below the conduction-band minimum: there is nowhere for an extra electron to sit.

Graupel is the same molecules assembled violently. Each supercooled droplet that strikes the rimer freezes in microseconds to milliseconds, far faster than the timescale of molecular rearrangement, locking molecules into non-equilibrium positions. The product is a porous polycrystalline aggregate riddled with grain boundaries, vacancies, interstitials, stacking faults, frozen-in strain fields and trapped air [8,11]. Every one of these defects breaks the local periodicity and injects localized states into the gap. Grain boundaries in particular carry extended networks of under-coordinated molecules and dangling hydrogen bonds, producing dense manifolds of trap states whose binding energies exceed kT at cloud temperatures by an order of magnitude or more. In semiconductor language, the ice crystal approximates a clean intrinsic surface and the graupel a heavily defected one; if the defect-state densities of polycrystalline insulators and oxide surfaces are any guide, the density of unoccupied electron-accepting states may differ between the two by orders of magnitude.

The contrast between the two surfaces can be framed in the language of chemical potential. In a well-ordered crystal at thermal equilibrium, the electrochemical potential of electrons is uniform across the surface and determined by the periodic lattice potential; the density of states near the Fermi level is that of a clean wide-gap insulator, and the surface presents a smooth, featureless energy landscape to any incoming electron. In the graupel, the situation is qualitatively different. Each grain boundary introduces a sheet of under-coordinated molecules whose local bonding environment departs from the bulk, splitting off states from the valence and conduction band edges and distributing them across the gap. Vacancies and interstitials contribute further localised levels. The cumulative effect is a rough, heterogeneous energy landscape peppered with potential wells of varying depth—a surface that is, in solid-state language, chemically identical to the crystal but electronically unrecognisable. The asymmetry that drives directional transfer is therefore not a difference in composition but a difference in the statistics of available electronic states, manufactured entirely by the freezing history and erasable by annealing.

Defect capture sets the arrow

The mechanism has three steps, each independently grounded in established solid-state physics. First, mechanical liberation: the collisional stress pulse transiently distorts the lattice at the contact, raising surface electrons above their equilibrium binding energy and ejecting a population of them into the interfacial region, the same physics as mechanoemission and fracto-emission from crystalline solids [12]. Second, asymmetric capture: a liberated electron at the interface sees two surfaces. On the graupel side it encounters a high areal density of deep unoccupied traps; on the crystal side it encounters a filled, ordered surface with almost none. The capture cross-section is overwhelmingly larger on the disordered side, in direct analogy with trap-mediated capture in defected semiconductors [13,14], a parallel that motivates the present hypothesis. Bang and Kang [15] directly measured the interaction of low-energy electrons with crystalline ice films and found efficient transmission through the ordered lattice, with electron trapping occurring preferentially at the ice surface. These observations establish that low-energy electrons can interact with ice through surface-localized trapping, although whether rimed polycrystalline ice possesses a substantially higher density of deep traps remains to be tested. Third, irreversibility: a trap of depth Eₜ ≳ 0.5 eV at 253 K has a thermal re-emission probability suppressed by a factor of order exp(−Eₜ/kT) ≲ 10⁻¹⁰, so capture is one-way on the timescale of the rebound and separation. Electrons liberated from the graupel side in the same impact face a surface with no room for them; they return. The collision is mechanically symmetric and electronically rectifying.

The energetics of this rectification can be stated quantitatively. The thermal emission rate from a trap of depth E_t follows e_n = ν exp(−E_t/kT), where ν ~ 10¹²–10¹³ s⁻¹ is the attempt frequency for phonon-assisted escape. At T = 253 K, a shallow state of 0.1 eV empties in roughly 10⁻¹¹ s and cannot retain charge. A grain-boundary state of 0.5 eV, by contrast, has a residence time of order 10⁻³ s, and one of 0.7 eV a residence time of seconds to tens of seconds — both long against the roughly 10⁻⁴ s contact duration and the subsequent separation. The irreversibility of capture is therefore not an assumption but a consequence of trap depth, and the mechanism's charge-storage capacity is bounded by the areal density of states deeper than roughly 0.5 eV. Laboratory values of charge transferred per crystal–graupel rebound are of order 1–50 fC, corresponding to 10⁴–10⁵ elementary charges across a contact area of order 10–10² μm². Reproducing these values requires an areal density of filled traps of 10¹⁰–10¹² cm⁻², which falls within the range of grain-boundary state densities measured in polycrystalline insulators and defected oxide surfaces. The measured charge per collision is therefore reproduced without free parameters at the upper end of ordinary defect densities, with the observed scatter attributed to collision-to-collision variation in contact area and local defect coverage.

A single collision transfers a statistical excess, not a deterministic packet. But the excess always has the same sign, and an active updraft hosts millions of collisions per second. The bias integrates into macroscopic charge separation: crystals stripped of electrons ride the updraft to the anvil as the upper positive charge; graupel loaded with trapped electrons descends as the mid-level negative charge. The dipole of the thunderstorm is the time-integral of a defect-density asymmetry.

The mechanism belongs to a class of phenomena whose reality has recently been established directly. Sobarzo et al. [16] showed that nominally identical insulator samples, which initially exchange charge randomly, develop a reproducible triboelectric ordering after repeated contact, correlating with nanoscale morphological changes induced by contact history. These findings demonstrate that contact history and nanoscale morphology can generate systematic charge asymmetry between identical materials; the ice-crystal--graupel pair is a natural candidate. The Workman–Reynolds freezing potential [17,18] operates on the same substrate: the rapid interfacial freezing that generates transient potentials during riming is the identical process that manufactures the defects and traps invoked here, so the two effects are complementary faces of one non-equilibrium solidification history.

Polarity reversal and the quasi-liquid layer

The mechanism does not merely permit the observed sign; it predicts the observed reversal. As the temperature rises toward 0 °C, the ice surface premelts into a quasi-liquid layer whose thickness grows from sub-nanometre to several nanometres [19]. The QLL acts on the two surfaces oppositely. On the ordered crystal it destroys the very order that starved the surface of traps: a liquid-like disordered skin is itself a manifold of localized states, so the crystal’s effective trap density rises steeply with temperature. On the graupel it does the opposite: the molecular mobility of the premelted skin anneals shallow defects and relaxes frozen-in strain, healing the trap manifold. The asymmetry that drives the transfer therefore shrinks with warming, passes through zero, and changes sign—which is precisely the polarity reversal measured near −15 °C and its dependence on liquid water content and crystal growth rate [2,20,21,22]. The same logic yields the rime-accretion-rate dependence directly: vigorous riming freezes droplets faster, locks in more disorder, deepens the trap manifold and strengthens negative graupel charging, exactly as observed in the laboratory [23,24]. Growth-rate theories such as the Relative Diffusional Growth Rate framework are recovered here as phenomenological shadows of the underlying variable, which is the defect-state density contrast of the two surfaces.

The quantitative relationship between QLL thickness and reversal temperature reinforces this account. Surface-sensitive measurements show that the QLL grows from sub-nanometre thickness below −30 °C to several nanometres near −10 °C, with a steep onset in the range −20 to −15 °C that coincides with the observed polarity reversal window. Below this onset the crystal surface remains essentially bare and well-ordered, sustaining the full trap-density asymmetry and strong negative graupel charging. As the QLL develops, the crystal surface acquires an increasingly liquid-like disordered skin whose electronic structure converges toward that of the graupel surface, progressively erasing the asymmetry that drives directional transfer. The dependence on liquid water content enters through the same variable: higher liquid water content accelerates riming, shortens the freezing timescale per droplet, and locks more disorder into the graupel structure, deepening its trap manifold and shifting the balance point to warmer temperatures. Lower liquid water content permits slower freezing and partial annealing, reducing the defect contrast and shifting the reversal to colder temperatures. Both dependencies have been measured independently in laboratory charging experiments and are recovered here as corollaries of a single underlying variable—the defect-state density contrast between the two colliding surfaces—rather than as separate empirical parameters requiring separate explanations. The additional sensitivity to relative humidity reported by Berdeklis and List [25], which modulates the vapour growth rate and hence the surface ordering of the crystal, is consistent with this picture but warrants separate quantitative treatment.

Seasonality selects the mechanism class

The seasonal constraint of the opening section now does its work. In the defect-capture mechanism the material asymmetry is a necessary condition but never the rate limiter: charge separation per unit time is proportional to the collision rate, which is set by updraft strength, ice mass flux and particle concentrations—all of which scale with surface heating and convective instability. Summer convection multiplies collisions; winter quiescence starves them; lightning follows convection because charging is collision-rated. The observed seasonality is not an additional fact to be accommodated; it is the mechanism’s signature. Any competing proposal that locates the primary cause in a temperature-independent material constant—flexoelectric, piezoelectric-like, or thermoelectric—must import the collision dynamics as an auxiliary assumption to survive the seasonal data, at which point the material constant is no longer doing the explanatory work.

The seasonal asymmetry becomes even more striking when examined regionally. Tropical regions between 30°S and 30°N account for roughly 78% of global lightning, and continental surfaces collectively host about 88% of all flashes, leaving only about 12% over the open ocean—despite the ocean hosting mixed-phase ice clouds with far weaker updrafts due to lower surface temperature gradients [5]. Within the tropics, the diurnal cycle of lightning over land peaks sharply in the late afternoon, tracking the daily maximum of convective available potential energy with a lag of one to two hours, while the diurnal cycle over ocean is weak and nearly flat. If an intrinsic property of ice were the rate-limiting factor, the land-ocean contrast and the diurnal phasing would require separate explanations; in a collision-rated mechanism they are automatic consequences of the same variable—updraft intensity—that governs the seasonal cycle. The hierarchy of timescales is consistent: seasonal, land-ocean, and diurnal modulations of lightning frequency all track convective forcing rather than ice microphysics, and the defect-capture mechanism predicts this triple correlation by construction. The constraint does not falsify flexoelectricity as a contributing channel of charge generation during individual impacts; it demotes every intrinsic-property mechanism from cause to modulation.

The constraint can be stated more sharply. Over boreal North America and Eurasia poleward of 30°, the ratio of summer to winter land flash density exceeds an order of magnitude, while satellite cloud-phase products confirm that mixed-phase clouds containing both ice crystals and graupel persist through all seasons in these regions. If the rate-limiting step of electrification were an intrinsic property of ice—flexoelectric, thermoelectric, or otherwise—the flash rate would track ice-cloud occurrence, not surface temperature. The observed decoupling between ice availability and lightning frequency is the empirical signature of a collision-rated mechanism and constitutes a necessary condition that any viable theory of thunderstorm electrification must satisfy.

Breakdown at sub-critical mean fields

A complete account of lightning must also confront the initiation paradox: balloon soundings through active storms measure maximum fields of order 10⁵ V m⁻¹, an order of magnitude below the conventional breakdown threshold of clear air at cloud altitude. The microphysical picture developed here dissolves rather than deepens this paradox, because the same particles that separate the charge also concentrate the field. An ice crystal is a faceted dielectric body terminating in edges, corners and dendritic branches whose radii of curvature are of order micrometres. The electric field at such an asperity is geometrically enhanced above the ambient value by a factor that scales with the dielectric constant and the aspect ratio of the protrusion: for ice, whose static dielectric constant exceeds 100, an aspect ratio of 10 can produce substantial local field enhancement, with near-contact hydrometeor geometries contributing additional gap-closing amplification. The statistical ubiquity of such configurations in the mixed-phase zone ensures that local breakdown conditions are met at many sites simultaneously even when the volume-averaged field remains far below the uniform-gap threshold. Positive streamers therefore ignite from crystal tips at ambient fields of 1–2 × 10⁵ V m⁻¹ [26,27], seeding conductive channels that runaway electron processes [28] and stepped-leader dynamics [29] subsequently extend across the full inter-charge gap. The initiation problem and the charging problem thus share one physical answer with two aspects: convection controls how many ice particles collide, which sets the charging rate, and how many crystal asperities populate the high-field region, which sets the initiation probability. Both halves of the lightning chain are convection-rated, and the summer maximum of flash frequency follows twice over. This unification distinguishes the present mechanism from proposals that require an external trigger—such as cosmic-ray seed electrons—to bridge the gap between measured and threshold fields, a requirement that introduces a stochastic element difficult to reconcile with the high spatial and temporal regularity of lightning in vigorous convective systems.

The cosmic-ray hypothesis also struggles to account for the land-ocean contrast: the cosmic-ray flux is spatially uniform to first order, yet lightning is overwhelmingly concentrated over land where convective updrafts are strongest.

Falsifiable predictions

The mechanism is stated so as to be killed by experiment.

(1) Charge scales with defect density: for chemically identical ice targets, the charge transferred per collision increases monotonically with the defect density of the receiving surface, controllable via freezing rate; a null or inverse correlation refutes the mechanism.

(2) Symmetry yields silence: two well-annealed single crystals of matched orientation and defect density exchange no systematic charge; any residual transfer is small and random in sign.

(3) Disorder suffices without water: an ice target deliberately disordered by quenching, grinding or irradiation charges negatively against an ordered crystal in dry conditions, with no supercooled liquid present—separating the structural mechanism cleanly from riming chemistry.

(4) Annealing erases the effect: sub-melting thermal annealing of a rimed target reduces the transferred charge in proportion to the healed defect population, measurable independently by calorimetry or diffraction linewidth.

(5) The reversal temperature is a QLL observable: it shifts predictably under surface-active dopants that alter premelting onset, mapping the polarity reversal onto an independently measurable surface property [19].

Prediction (2) is the sharpest: the flexoelectric mechanism, which requires only bending, permits charge transfer between two perfect crystals under asymmetric impact geometry, whereas defect capture forbids systematic transfer there. The two mechanisms are distinguishable on a benchtop; the H+ diffusion mechanism of Kang et al. [7], which does not require electronic trap states, makes a different prediction for this geometry.

The experiments require a cold chamber, controlled ice growth, and femtocoulomb electrometry—all standard since the classic charging studies [2,25]. What has changed since those studies is the frame: identical-material contact electrification is now an established field with quantitative surface-characterization tools, and the thunderstorm is its oldest unread experiment.

Methods

Trap energetics and irreversibility. The mechanism requires trap depths of order 0.5–0.7 eV to ensure residence times exceeding the collision contact duration. If deep traps associated with under-coordinated molecules at grain boundaries and vacancy clusters in ice fall in this range, as analogy with other hydrogen-bonded and polycrystalline solids suggests, the irreversibility step follows as a quantitative consequence.

Order-of-magnitude charge per collision. Grain-boundary state densities in polycrystalline insulators and defected oxide surfaces lie at 10¹²–10¹⁴ cm⁻², and rimed ice—solidified far from equilibrium—plausibly saturates this range. The measured charge per collision is therefore reproduced without free parameters at the upper end of ordinary defect densities, with the observed scatter attributed to collision-to-collision variation in contact area and local defect coverage. The same accounting bounds the mechanism: it cannot deliver more charge than σₜᵣₐₚA per contact, a ceiling testable by high-liquid-water-content riming experiments.

Field enhancement at asperities. For a prolate dielectric asperity of semi-axes a ≫ b in an ambient field E₀, the tip field scales as Eₜᵢₚ ≈ E₀(1 + (εᵣ−1)/(εᵣ n)) with depolarization factor n → (b/a)² ln(2a/b) for needle-like geometry; for ice (εᵣ ≈ 100 at low frequency) an aspect ratio of 10 yields substantial enhancement, and near-contact hydrometeor pairs add a further gap-closing factor [30]; the combined effect can approach two orders of magnitude. Ambient fields of 1–2 × 10⁵ V m⁻¹ [26,27] therefore reach local values above the ∼3 × 10⁶ V m⁻¹ threshold at altitude, sufficient for positive streamer inception from crystal tips.

Seasonality analysis. Flash-rate seasonality is taken from the OTD/LIS optical climatology [5] and the WGLC gridded climatology [6]. The diagnostic is the ratio of summer to winter land flash density in the extratropics (poleward of 30°), which exceeds an order of magnitude over boreal North America and Eurasia, compared against the season-independence of mixed-phase cloud occurrence in the same regions from standard satellite cloud-phase products. The contrast quantifies the statement that ice availability is not the limiting variable.

 

Acknowledgements

 

The author used Claude (Anthropic), GPT (OpenAI), and Gemini (Google) to assist with manuscript drafting, literature synthesis, and language editing. All conceptual frameworks, logical arguments, and final conclusions were developed and verified by the author, who assumes full responsibility for the integrity of the work.

 

Declarations

 

Funding: This research received no external funding.

Conflicts of interest: The author declares no conflicts of competing interests.

Data availability: No new observational data were generated or analysed in this study. All referenced datasets are publicly available from the sources cited.

Author contributions: Juliet Zhong: conceptualisation, formal analysis, writing.

 

References

 

[1] Reynolds, S. E., Brook, M. & Gourley, M. F. Thunderstorm charge separation. J. Meteorol. 14, 426–436 (1957).

[2] Takahashi, T. Riming electrification as a charge generation mechanism in thunderstorms. J. Atmos. Sci. 35, 1536–1548 (1978).

[3] Saunders, C. P. R. Charge separation mechanisms in clouds. Space Sci. Rev. 137, 335–353 (2008).

[4] Wen, X. et al. Flexoelectricity and surface ferroelectricity of water ice. Nat. Phys. 21, 1587–1593 (2025).

[5] Christian, H. J. et al. Global frequency and distribution of lightning as observed from space by the Optical Transient Detector. J. Geophys. Res. 108, 4005 (2003).

[6] Kaplan, J. O. & Lau, K. H.-K. The WGLC global gridded lightning climatology and time series. Earth Syst. Sci. Data 13, 3219–3237 (2021).

[7] Kang, H., Williams, E. & Jayaratne, R. A microscopic mechanism for the separation of charge in thunderstorms. J. Geophys. Res. Atmos. 131, e2025JD046158 (2026).

[8] Hobbs, P. V. Ice Physics (Clarendon, Oxford, 1974).

[9] Bartels-Rausch, T. et al. Ice structures, patterns, and processes: a view across the icefields. Rev. Mod. Phys. 84, 885–944 (2012).

[10] Kobayashi, K. Optical spectra and electronic structure of ice. J. Phys. Chem. 87, 4317–4321 (1983).

[11] Petrenko, V. F. & Whitworth, R. W. Physics of Ice (Oxford Univ. Press, 1999).

[12] Dickinson, J. T., Donaldson, E. E. & Park, M. K. Fracto-emission from crystalline materials. J. Mater. Sci. 16, 2897–2908 (1981).

[13] Shockley, W. & Read, W. T. Statistics of the recombinations of holes and electrons. Phys. Rev. 87, 835–842 (1952).

[14] Watkins, G. D. Intrinsic defects in silicon. Mater. Sci. Semicond. Process. 3, 227–235 (2000).

[15] Bang, J. & Kang, H. Transmission and trapping of low-energy (1–10 eV) electrons in crystalline ice films. J. Phys. Chem. C 124, 15862–15869 (2020).

[16] Sobarzo, J. C. et al. Spontaneous ordering of identical materials into a triboelectric series. Nature 638, 664–669 (2025).

[17] Workman, E. J. & Reynolds, S. E. Electrical phenomena occurring during the freezing of dilute aqueous solutions and their possible relationship to thunderstorm electricity. Phys. Rev. 78, 254–259 (1950).

[18] Wilson, P. W. & Haymet, A. D. J. The Workman–Reynolds "Freezing Potential": a new look at the inherent physical process. J. Mol. Liq. 223, 1168–1172 (2016).

[19] Slater, B. & Michaelides, A. Surface premelting of water ice. Nat. Rev. Chem. 3, 172–188 (2019).

[20] Jayaratne, E. R., Saunders, C. P. R. & Hallett, J. Laboratory studies of the charging of soft hail during ice crystal interactions. Q. J. R. Meteorol. Soc. 109, 609–630 (1983).

[21] Saunders, C. P. R., Bax-Norman, H., Emersic, C., Avila, E. E. & Castellano, N. E. Laboratory studies of the effect of cloud conditions on graupel/crystal charge transfer in thunderstorm electrification. Q. J. R. Meteorol. Soc. 132, 2653–2673 (2006).

[22] Pereyra, R. G., Avila, E. E. & Castellano, N. E. Charge separation in thunderstorm conditions. J. Geophys. Res. 113, D17203 (2008).

[23] Saunders, C. P. R. & Peck, S. L. Laboratory studies of the influence of the rime accretion rate on charge transfer during crystal/graupel collisions. J. Geophys. Res. 103, 13949–13956 (1998).

[24] Emersic, C. & Saunders, C. P. R. Further laboratory investigations into the relative diffusional growth rate theory of thunderstorm electrification. Atmos. Res. 98, 327–340 (2010).

[25] Berdeklis, P. & List, R. The ice crystal–graupel collision charging mechanism of thunderstorm electrification. J. Atmos. Sci. 58, 2751–2770 (2001).

[26] Marshall, T. C., McCarthy, M. P. & Rust, W. D. Electric field magnitudes and lightning initiation in thunderstorms. J. Geophys. Res. 100, 7097–7103 (1995).

[27] Stolzenburg, M. & Marshall, T. C. Charge structure and dynamics in thunderstorms. Space Sci. Rev. 137, 355–372 (2008).

[28] Gurevich, A. V., Milikh, G. M. & Roussel-Dupré, R. Runaway electron mechanism of air breakdown and preconditioning during a thunderstorm. Phys. Lett. A 165, 463–468 (1992).

[29] Dwyer, J. R. & Uman, M. A. The physics of lightning. Phys. Rep. 534, 147–241 (2014).

[30] Crabb, J. A. & Latham, J. Corona from colliding drops as a possible mechanism for the triggering of lightning. Q. J. R. Meteorol. Soc. 100, 191–202 (1974).