[Paper Review] On the geometry of aggregate snowflakes

Axel Seifert, Christoph Siewert+3 more|arXiv (Cornell University)|Jan 15, 2026

Precipitation Measurement and Analysis0 citations

TL;DR

The paper develops a stochastic, physically based parameterization of aggregate snowflake geometry (max dimension, aspect ratio, cross-sectional area) as a function of mass, monomer number, and monomer habit, enabling realistic Lagrangian particle modeling of snowflake aggregation.

ABSTRACT

Snowflakes play a crucial role in weather and climate. A significant portion of precipitation that reaches the surface originates as ice, even when it ultimately falls as rain. Contrary to the popular image of symmetric, dendritic crystals, most large snowflakes are irregular aggregates formed through the collision of primary ice crystals, such as hexagonal plates, columns, and dendrites. These aggregates exhibit complex, fractal-like structures, particularly at large sizes. Despite this structural complexity, each aggregate snowflake is unique, with properties that vary significantly around the mean - variability that is typically neglected in weather and climate models. Using a physically based aggregation model, we generate millions of synthetic snowflakes to investigate their geometric properties. The resulting dataset reveals that, for a given monomer number (cluster size) and mass, the maximum dimension follows approximately a lognormal distribution. We present a parameterization of aggregate geometry that captures key statistical properties, including maximum dimension, aspect ratio, cross-sectional area, and their joint correlations. This formulation enables a stochastic representation of aggregate snowflakes in Lagrangian particle models. Incorporating this variability improves the realism of simulated fall velocities, enhances growth rates by aggregation, and broadens Doppler radar spectra in closer agreement with observations.

Motivation & Objective

• Characterize the geometric variability of aggregate snowflakes beyond mean behavior using a large synthetic dataset.
• Develop a stochastic parameterization for maximum dimension, aspect ratio, and cross-sectional area that depends on mass, monomer count, and habit.
• Provide correlation structure between geometric properties suitable for integration into Lagrangian particle models and remote sensing forward operators.
• Demonstrate how incorporating variability affects fall velocities, growth by aggregation, and Doppler radar spectra.
• Offer a practical implementation framework for McSnow and related Lagrangian Monte Carlo simulations.

Proposed method

• Use a physically based aggregation model to generate millions of synthetic aggregates across monomer numbers N from 2 to 2048 and mean monomer sizes from 50 to 500 μm.
• Define and compute geometric properties: maximum dimension Dmax, vertically projected area A, axes Lx, Ly, Lz, and normalized metrics of Dmax and area ratio q.
• Normalise Dmax with mean monomer diameter to obtain hat{D}_{max} and show it is approximately lognormal for fixed N; introduce D_norm to collapse mean mass–size behavior across N.
• Parameterize the mean and std of the lognormal D_norm and its correlation with the aspect ratio phi via empirical power laws and piecewise expressions (Eqs. 3–9, 10–13).
• Model joint distributions P(D_norm, phi) and P(D_norm, q) with a lognormal base and lognormal conditionals (Eqs. 19–26), enabling efficient Monte Carlo sampling in LPMs.
• Implement the parameterization in McSnow, a Lagrangian particle model, to propagate aggregate geometry with mass, N, and habit as inputs.

Figure 1: Scatter plots showing the mass-size relations for aggregates of needles, plates, and dendrites. Colors indicate monomer number $N$ , and the regression line represents the power-law fit with fractal exponent $p$ for $N>5$ .

Experimental results

Research questions

• RQ1 What is the statistical distribution of the maximum dimension of aggregate snowflakes for a given mass and monomer number?
• RQ2 How do key geometric properties (maximum dimension, aspect ratio, cross-sectional area) correlate and vary with monomer habit and size?
• RQ3 Can a tractable stochastic parameterization capture the observed variability and correlations of aggregate geometry for use in Lagrangian microphysical models?
• RQ4 How does incorporating geometric variability affect simulated fall speeds, growth by aggregation, and radar-relevant observables?

Key findings

• For a given monomer number N and mass, the maximum dimension D_max, when nondimensionalized, follows an approximately lognormal distribution (hat{D}_{max}); distribution tails suggest beta could also be considered, but lognormal is used for computational efficiency.
• A D_norm scaling with N via a universal fractal exponent eta collapses mean mass–size behavior; larger D_norm corresponds to more elongated aggregates, smaller D_norm to more compact ones.
• The mean horizontal aspect ratio phi_h(N) and its dependence on habit show that needles start around 0.5 and increase with N, plates start near 0.9 and decrease toward 0.73, with mixtures following interpolated relations (Eqs. 9–13).
• The cross-sectional area ratio q decreases with N and depends on monomer habit; parameterizations (Eqs. 14–18) quantify how q varies for needles, plates, dendrites, and mixtures.
• There is a consistent negative correlation between D_norm and phi (r ≈ -0.60 to -0.65 depending on N and habit), and a meaningful but habit-dependent correlation between D_norm and q (Eqs. 20–26; Fig. 9).
• Joint PDFs are approximated with a product form P(D_norm)P(phi|D_norm) and P(q|D_norm), enabling efficient stochastic generation of coherent aggregate geometries in Monte Carlo frameworks.

Figure 2: Histograms of the nondimensional maximum dimension $\hat{D}_{\text{max}}$ for aggregates of needles and aggregates of plates for different monomer numbers $N$ . Solid lines represent lognormal distributions with the same mean and standard deviation as the data. The dendrite dataset looks s