Toward a Comprehensive Model of

Snow Crystal Growth Dynamics:

arXiv:1211.5555v1 [cond-mat.mtrl-sci] 23 Nov 2012

1. Overarching Features and Physical Origins

Kenneth G. Libbrecht
Department of Physics, California Institute of Technology
Pasadena, California 91125

Abstract. We describe a comprehensive model for the formation and morphological development
of atmospheric ice crystals growing from water vapor, also known as snow crystals. Our model derives
in part from empirical measurements of the intrinsic ice growth rates as a function of temperature
and supersaturation, along with additional observations and analyses of diffusion-driven growth
instabilities. We find that temperature-dependent conformational changes associated with surface
melting strongly affect layer nucleation dynamics, which in turn determines many snow-crystal
characteristics. A key feature in our model is the substantial role played by structure-dependent
attachment kinetics, producing a growth instability that is largely responsible for the formation of
thin plates and hollow columnar forms. Putting these elements together, we are able to explain the
overall growth behavior of atmospheric ice crystals over a broad range of conditions. Although our
model is complex and still incomplete, we believe it provides a useful framework for directing further
investigations into the physics underlying snow crystal growth. Additional targeted experimental
investigations should better characterize the model, or suggest modifications, and we plan to pursue
these investigations in future publications in this series. Our model also suggests new avenues for the
continued exploration of ice surface structure and dynamics using molecular dynamics simulations.

1 Introduction
Laboratory observations of snow crystals dating back to the 1930s have revealed a complex de-
pendence of growth morphologies on temperature and supersaturation [1, 2]. Under common at-
mospheric conditions, for example, ice crystals typically grow into thin plate-like forms near -2 C,
slender columns and needles near -5 C, thin-walled hollow columns near -7 C, very thin plates again
near -15 C, and columns again below -30 C. In addition, morphological complexity generally increases
with increasing supersaturation at all temperatures. These observations are often summarized in
the well-known Nakaya morphology diagram, and one example is shown in Figure 1 [1]. Extensions
to lower temperatures, as well as more detailed morphological studies, can be found in the literature
[3, 4, 5].
Although the overall features and morphological transitions seen in the Nakaya diagram have
been well established empirically, a basic physical explanation of why snow crystals exhibit this
growth behavior has been surprisingly elusive. In particular, the fact that snow crystals alternate
between plate-like and columnar forms as a function of temperature has been an outstanding problem
for nearly 75 years. The purpose of this paper is to present a new model for ice growth from water

1
Figure 1: The Nakaya snow crystal morphology diagram, showing different types of snow crystals
that grow in air at atmospheric pressure, as a function of temperature and water vapour supersat-
uration relative to ice. The water saturation line gives the supersaturation of supercooled water, as
might be found within a dense cloud. Note that the morphology switches from plates (T ≈ −2 C)
to columns (T ≈ −5 C) to plates (T ≈ −15 C) to predominantly columns (T < −30 C) as temper-
ature is decreased. Temperature mainly determines whether snow crystals will grow into plates or
columns, while higher supersaturations generally produce more complex structures.

vapor that includes physical mechanisms that can explain the observed morphological behavior.
Although our model is not simple, and details are still missing, we believe it provides a reasonable
overarching picture of the various physical processes that define the growth dynamics of atmospheric
snow crystals.
Progress toward explaining how complex structures arise spontaneously during solidification has
generally been hard won. The problem is exacerbated by the fact that numerous physical processes
are often involved over many length scales, and these work in concert to produce the observed
structures. Identifying and characterizing the most important of these processes, quantifying them
with suitable numerical algorithms, and merging the pieces into a comprehensive model of structure
formation has been an ongoing process for many decades.
It has been known since the mid-1960s, for example, that dendritic structures arise from the
Mullins-Sekerka instability during diffusion-limited growth [6]. Producing a quantitative model
of this process required substantial theoretical effort, however, culminating in the development of
solvability theory during the 1980s [7, 8]. With this we learned that the overall branching scale is
set by seemingly minor anisotropies in the surface dynamics. The surface energy anisotropy plays

2
the key role in the case of solidification from the liquid phase, while for solidification from gaseous
precursors the anisotropy in the surface attachment kinetics is typically the more dominant factor
[1, 9].
Numerical models of diffusion-limited growth were developed about the same time as solvability
theory, including front-tracking and phase-field techniques [10, 11]. These methods have enjoyed
considerable success in reproducing dendritic structures arising during solidification from the melt,
such as in metallurgical systems or ice growth from liquid water. These systems work well compu-
tationally in part because the material anisotropies are typically quite small, with surface energy
differences of perhaps a few percent between faceted and non-faceted surfaces. For such systems the
growth structures are generally smooth and free of sharp corners or edges, and numerical models
tend to be robust and stable.
In growth from the vapor phase, anisotropies in the surface attachment kinetics are often very
large, resulting in dendritic structures that are not smooth with continuous derivatives, but are
instead strongly faceted with sharp edges. In this case numerical instabilities can be problematic,
so considerably more care is required to produce stable growth models [12]. As a result, the usual
numerical techniques used for studying diffusion-limited growth have not yet been able to produce
satisfactory snow crystal structures from reasonable physical inputs.
In 2008-9, Gravner and Griffeath developed cellular automata (CA) techniques that avoided
the numerical instabilities that affected other methods [13, 14]. These CA models have generated
full three-dimensional dendritic structures that reproduced many characteristics of natural snow
crystals, including growth forms that are both branched and faceted, with sharp edges. Cellular
automata models incorporating more physically derived rules have since been demonstrated [15],
and with suitable inputs it now appears possible, at least in principle, to realize a numerical model
that can accurately reproduce snow crystal growth rates and morphologies at all temperatures and
supersaturations.
Analysis of diffusion-limited growth using these theoretical and computational tools has yielded
numerous insights into the dynamics of structure formation. For example, solvability theory nicely
explains why the tip velocity of a growing dendritic structure depends linearly on supersaturation
for solidification from vapor, while a quadratic dependence on undercooling is typical for growth
from the liquid phase [1, 9]. In addition, scaling relationships in diffusion-limited growth models
provide an explanation for the increase in structural complexity that accompanies decreasing vapor
diffusion rates [15, 16].
For snow crystal growth, these theoretical considerations generally explain why morphological
complexity increases with supersaturation, crystal size, and background gas pressure. Thus the
observed variation along the supersaturation axis on the morphology diagram in Figure 1 is fairly
well understood, at least at a qualitative level. Producing accurate model crystals using sensible
input physics over a range of conditions has not yet been accomplished, and some unusual dendritic
snow crystal structures may be quite challenging to reproduce [17]. Nevertheless, diffusion-limited
growth – the underlying physical mechanism responsible for the formation of much snow crystal
structure – is reasonably well understood, and satisfactory computational algorithms describing this
process are available.
The ability to make high-fidelity numerical models of dendritic growth is only the first step,
however, since models require inputs. For the snow crystal case, one of the most important physical
inputs comes from the many-body interactions that determine how water molecules are incorporated
into the crystalline lattice, referred to as the surface attachment kinetics. For a rough surface, this in-
corporation is essentially instantaneous for all molecules that strike the surface. But the attachment

3
to faceted surfaces is much slower, and in snow crystal formation the attachment kinetics are often
as important as diffusion in governing the morphological development. Therefore, understanding
the growth of faceted ice surfaces is essential for any snow crystal model.
Some relevant physical processes that occur during ice crystal growth from vapor include the
adsorption and evaporation of surface admolecules, their diffusion along the surface, the nucleation
of new molecular layers, and the Gibbs-Thomson effect. Near the melting point one must also
consider the high vapor pressure of ice, along with temperature-dependent surface structural effects
associated with surface melting [18]. Characterizing the many-body physics of these various effects
remains a significant challenge.
The surface attachment kinetics for ice are sufficiently complex that crystal growth rates cannot
yet be determined from ab initio molecular dynamics modeling, and it appears that this direct ap-
proach will be unfeasible for the foreseeable future. Nevertheless, one need not track every molecule
to create a reasonable physical picture of the crystallization process, or to realize a useful parame-
terization of crystal growth rates. Another route is to combine basic theoretical considerations with
empirical data to create a more simplified crystal growth model, hopefully one that explains the
morphology observations with as much quantitative accuracy as possible.
Statistical models of crystal growth dynamics have been developed over many decades also, and
some notable early advances in this area include the formulation of classical nucleation theory, as well
as the role of screw dislocations in crystal growth (for example, see [19]). The former is especially
important in snow crystal growth, while the effects of dislocations are often negligible [26]. Crystal
growth theory has generally enjoyed much success describing many faceted materials, especially
cases where the surface structure is very stable and the admolecule dynamics are especially simple.
Unfortunately, the ice surface has a high equilibrium vapor pressure, which means that the surface
is in a highly dynamical state. Under typical conditions, for example, ice molecules evaporate from
and redeposit onto the surface at rates of order one monolayer per microsecond, while crystal growth
rates can easily be 4-6 orders of magnitude slower. Add the conformational changes associated with
surface melting, and it is easy to see why understanding the molecular physics of ice crystal growth
presents some unique theoretical challenges.
Because the ice surface dynamics are only poorly understood, snow crystal growth models to
date have been quite speculative in nature. Some early models in the 1960s focused on surface
diffusion rates and step growth velocities, and how these depended on temperature for the principal
facet surfaces (for a review, see [20]). This approach now appears to be incorrect, as subsequent
experiments have shown that the nucleation of new molecular layers on the principal facet surfaces
is the rate-limiting step in their growth. Recent evidence for this will be discussed below.
One of the best previous attempts to create a comprehensive and physically motivated model of
snow crystal growth was that presented by Kuroda and collaborators in the 1980s [21, 22, 23]. This
model describes a qualitative picture of the influence of surface melting on crystal growth rates,
and how temperature-dependent surface melting could explain the observed spectrum of growth
morphologies. Subsequent experimental data have not supported this model, but the underlying
assumption that surface melting plays an important role appears to be sound.
More recent experimental and modeling efforts have convincingly demonstrated that layer nu-
cleation plays an important role in the formation of faceted ice surfaces (see [24, 26] and references
therein), as had been assumed in the Kuroda model and in some earlier models. Moreover, it has
certainly been demonstrated that bulk diffusion through the surrounding gas is a key factor influenc-
ing growth rates [25]. To date, however, no model has convincingly explained, even in a qualitative
fashion, the temperature dependence seen in the Nakaya morphology diagram.

4
 Our approach to this problem has been to use quantitative growth measurements to guide our
model building efforts, and to use numerical CA modeling to compare data with calculated growth
rates and morphologies over a range of conditions. We first observed growth rates at reduced
background pressures to nearly eliminate bulk diffusion effects, thus giving measurements of the
intrinsic growth rates of the ice facet surfaces. With these data in hand, we then examined growth
rates and morphologies at normal atmospheric pressures, thereby allowing a quantitative analysis of
diffusion-driven growth instabilities. Our result is a comprehensive model that agrees with the most
reliable growth data and explains the general features in the snow crystal morphology diagram.
Although our model is not without speculative elements, we feel it is a significant improvement
over previous attempts. The model provides a useful overarching picture of the underlying physics,
as well as a reasonable framework for continued discussion of this problem. For example, we have
found that a key quantity in our model is the step energy β associated with the edge of a molecular
terrace on the facet surfaces. As β is a molecular-scale, equilibrium property of the crystal surface,
its calculation should be much easier than full dynamical calculations of crystal growth rates. Thus
it may be possible to use current ab initio modeling techniques to examine some aspects of snow
crystal growth down to the molecular scale.
We have also found that our model is very useful for suggesting targeted experimental investiga-
tions that can both test the model predictions and better determine the model parameters. In such
a capacity, we believe it likely that continued investigation of this model will be fruitful, and may
yield new insights into ice growth dynamics in different circumstances. We are currently working on
several such experiments, and the results from these will be presented in future publications in this
series.

2 Intrinsic Growth of the Principal Facets

An important step toward understanding snow crystal growth dynamics is to quantify the growth
rates of the principal facet surfaces as a function of temperature and supersaturation. Following the
notation in [1], we parameterize the surface growth velocities using v = αvkin σsurf , where v is the
perpendicular growth velocity, vkin (T ) is a temperature-dependent “kinetic” velocity derived from
statistical mechanics, and σsurf is the water vapor supersaturation relative to ice at the growing
surface. The attachment coefficient α, which depends on T , σsurf , and other factors, encapsulates
the attachment kinetics governing crystal growth at the crystal/vapor interface. From the definition
of vkin , we must have α ≤ 1.
For the simplest case – the growth of an infinite, clean, dislocation-free faceted ice surface in
near equilibrium with pure water vapor at a fixed temperature – the attachment coefficient is well
defined and we must have a unique α(σsurf , T ) for each facet surface. We refer to the α(σsurf , T )
for the two principal facets in this ideal case as the “intrinsic” attachment coefficients.
We determined α(σsurf , T ) through a lengthy series of measurements of crystals growing on a
substrate at low background pressure. Experimental details, the resulting data and analysis, and
references to prior work measuring α(σsurf , T ), can be found in [26]. Over the temperature range
−2 C > T > −40 C, our growth data are well described by a dislocation-free layer-nucleation
crystal growth model, which we parameterize using α(σsurf , T ) = A exp(−σ0 /σsurf ). The measured
parameters A(T ) and σ0 (T ) for the basal and prism facets are shown in Figure 2.
The data in Figure 2 are the most recent, most extensive, and we believe the most accurate
measurements to date of the intrinsic α(σsurf , T ) for the principal facets of ice. Several sources

5
Figure 2: Measurements of the intrinsic growth rates of the principal ice crystal facets. The growth
velocity normal to the surface is described by v = αvkin σsurf , where σsurf is the supersaturation at
the surface and the attachment coefficient is parameterized with α(T, σsurf ) = A exp(−σ0 /σsurf ).
The solid points show the measured A(T ) and σ0 (T ) for the basal facets, while the open points show
measurements of the prism facets, from [26].

of potential systematic errors were examined with greater care than had been done in previous
experiments, including effects from substrate interactions and bulk diffusion, and these errors were
largely eliminated. Possible errors from chemical contamination of the ice surface were examined as
well [27]. Within the stated error estimates, therefore, we assume that the data in Figure 2 represent
empirical fact that must be incorporated within our model of snow crystal growth dynamics.

2.1 Physical Origins - I

Our understanding of the detailed molecular structure and dynamics of the ice surface is not sufficient
to provide a full physical explanation of the measurements displayed in Figure 2, and ab initio
molecular dynamics simulations cannot accurately calculate ice growth rates directly [28, 29, 30].
However, the problem is greatly simplified with the realization that classical nucleation theory
dictates that the parameter σ0 (T ) derives from the step energy β(T ) associated with the edge

6
Figure 3: The step energy β(T ) associated with the edge of a molecular terrace for the two principal
facets of ice. This quantity was derived from the measured σ0 (T ) in Figure 2 using classical nucleation
theory [26]. For comparison we also show β relative to β0 = γa, where γ is the ice surface energy
and a is the molecular step height, as described in the text.

of a molecular terrace on the facet surface [19]. Using this theory, we are able to convert the
measurements to give β(T ) for both facets, as shown in Figure 3 [26].
By taking advantage of nucleation theory, we are thus able to reduce the problem of crystal
growth to one of determining of the step energy. This represents a major simplification of the
problem, since crystal growth is a complex dynamical process, while the step energy is a basic
equilibrium property of the ice surface. Although β(T ) has not yet been calculated from existing ice
surface models, this goal may be reached in the near future. As part of our model, we suggest the
following qualitative picture of how temperature-dependent surface melting can explain much of the
detail seen in Figure 3.
Beginning at the lowest temperatures in Figure 3, we see that β(T ) increases monotonically
with decreasing temperature for both facets below T = −15 C. For extremely low temperatures, we
would expect that the ice surface structure would approach that of a simple molecular layer model
(essentially a basic Kossel crystal [31]), so the step energy would approach β ≈ β0 = γa, where γ
is the ice surface energy and a is the step height. This follows because in a simple layer model the
increased surface area associated with a step is aL, where L is the step length, so the increase in
effective surface energy is γaL, giving β ≈ γa.
This reasoning suggests that β → β0 at very low temperatures, and we consider this to be a
prediction of our model. Additional experiment and theory could be used to confirm this hypothesis.
As temperature is increased, we would expect that thermal fluctuations would influence the structure
of the terrace step, causing a surface reconstruction that effectively smooths out the step, reducing
β. Thus we would expect β to decrease monotonically with increasing temperature, with β/β0 < 1,

7
and this behavior is seen in Figure 3 for T < −15 C.
Because the step energy β is a molecular-scale equilibrium property of the ice surface, similar in
this regard to the surface energy γ, it appears promising that β may be calculable using molecular
dynamics simulations. The low-temperature region is probably the most amenable to such inves-
tigations, since for T < −15 C it appears that the surface structure is relatively simple and not
complicated by substantial surface melting. If the measured approximate power-law behavior in
β(T ) could be reproduced for the low-temperature region, this would give one confidence that our
characterization of the step energy is a reasonable one, as well as providing physical insights into
the underlying molecular origins of β(T ).
Continuing up in temperature in Figure 3, we see unusual features in both βbasal and βprism at
T ≈ −12 C. A sharp peak in βbasal (T ) is seen at this temperature, as is a broader peak (little more
than a shoulder) in βprism (T ). Both peaks are robust features in our data, and we suggest that both
are associated with the onset of significant surface melting in the ice surface. By “onset” in this
context, we mean the temperature at which surface melting first begins to significantly affect the
crystal growth dynamics. This interpretation is consistent with some other measurements of surface
melting in ice [32], although in general our theoretical understanding of surface melting in ice and
other materials is rather poor [18].
Exactly why the onset of surface melting produces peaks in β(T ) near −12 C is not known. We
can only speculate that the observed behavior may result from complex many-body dynamics at
this temperature, when the top molecular layers are just beginning to become disordered. If this is
true, then explaining these −12 C features in detail may be quite challenging. Their effect on snow
crystal growth dynamics, however, appears to be fairly small compared to the overall behavior in
β(T ) for the two facets.
Focusing now on the prism facet, we see in Figure 3 that βprism (T ) declines precipitously with in-
creasing temperatures above T ≈ −10 C. At these temperatures, surface melting produces a substan-
tial quasi-liquid layer (QLL) in ice, so the ice/vapor interface becomes replaced by an ice/QLL/vapor
interface. The data then suggest that surface reconstruction further smooths the step edge on the
prism facet, thereby reducing βprism with increasing temperatures. Here again the situation seems
ripe for additional investigation with molecular dynamics simulations. Investigations using ab initio
calculations have yielded considerable insights into ice surface melting [28, 29, 30], and and it seems
promising that examining step energies with these methods may well explain some of the features
seen in Figure 3.
For the basal facet, we see that βbasal (T ) initially drops for temperatures above T ≈ −10 C, but
then the trend reverses and βbasal (T ) exhibits a brief upward jump with increasing temperature for
T ≈ −6.5 C, finally leveling off for T > −4 C. To explain this behavior, we borrow a key feature
from the Kuroda model and suggest that extensive surface melting at higher temperatures yields
a sharp ice/QLL interface on the basal facet. We suggest that this interface is accompanied by a
βbasal that approaches some constant value as the temperature approaches the melting point and
the QLL thickness diverges. In this behavior we see an important difference between the basal and
prism facets of ice – near the melting point, the nucleation barrier essentially vanishes on the prism
facet (its effects become negligible in the crystal growth dynamics) while remaining substantial on
the basal facet.

8
2.1.1 Connection to Ice Growth from Liquid Water
This high-temperature picture is consistent with observations of ice growing from liquid water,
where one typically sees faceting on the basal facet but no faceting on the prism facet [33]. This
makes sense given our understanding that the QLL thickness diverges near the melting point, so one
expects similar nucleation barriers at the ice/water interface and the ice/QLL interface when the
QLL thickness is large.
Another prediction from our model, therefore, is that there should be a correspondence between
βbasal near the melting point as determined from two different measurement strategies – using
ice growth from water vapor, and using ice growth from liquid water. To our knowledge, this
correspondence has not been investigated to date. Experiments on ice growth from liquid water
have not yet determined βbasal accurately [33], mainly because of interference from heat diffusion
effects. But it appears that an accurate measurement of βbasal for the liquid case could be achieved
with a targeted experimental investigation. Examining the correspondence between growth from
water vapor near 0 C and growth from liquid water at low supercoolings may well yield additional
insights into the attachment kinetics in both cases.

2.1.2 Surface Diffusion and A(T )

Early attempts to model the temperature dependence in snow crystal growth focused on the 2D
diffusion of water admolecules on the ice surface, as mentioned above and discussed at length in
[20]. The measurements in Figure 2 tell us, however, that surface diffusion plays a relatively minor
role compared to nucleation dynamics. From crystal growth theory in a multinucleation model
(see [19]), we find that the growth parameter σ0 (T ) is determined mainly by the step energy β(T ),
while the prefactor A(T ) is determined mainly from the surface diffusion of admolecules and their
1/6
attachment to the terrace edges. (Note that we have ignored a factor of σsurf , as this factor is
typically inconsequential compared to the nucleation factor exp(−σ0 /σsurf ).) Since A(T ) derives
from an extrapolation of the growth data to high σsurf [26], and is further subject to the constraint
that α ≤ 1, we see that fast surface diffusion yields the limit A(T ) = 1, as is observed on the prism
facet for T ≤ −10 C and on the basal facet at all temperatures measured.
From Figure 2 we see that Aprism (T ) is less than unity only for T > −10 C, when surface melting
is well developed. We suggest that Aprism is reduced in this region because of a growth impedance
at the ice/QLL interface, although the microphysics underlying this is not known. Additional
investigation of this behavior, and comparison with other crystal systems, would likely be fruitful.
At temperatures near T = −2 C, the nucleation barrier on the prism facet is quite low, so the
main impediment to the intrinsic growth of the prism facet comes from the fact that Aprism < 1.
Because of this, the growth of the prism facet is slower than a rough surface, which results in faceting.
In growth from liquid water we see no faceting of the prism facet, however. This suggests another
prediction of our model, that Aprism (T → 0) ≈ 1, owing to the correspondence between growth
from liquid water and growth when the QLL is thick. Extending the data in Figure 2 to higher
temperatures, although challenging experimentally, could test this prediction.

2.2 Further Investigations of Intrinsic Growth Rates

Although the data in Figure 2 represent a significant step forward in understanding the intrinsic
growth behavior of the principal facets of ice, extending these measurements to both lower and

9
higher temperatures would be beneficial. The low-temperature regime is probably the easiest to
model with molecular dynamics simulations, so better measurements here could facilitate progress
in characterizing the ice surface in the absence of surface melting. Reproducing the observed σ0 (T )
with ab initio calculations would be a substantial achievement, even if only in the low-temperature
regime, given our current poor understanding of the molecular structure and dynamics of the ice
surface.
At the high-temperature end of the spectrum, comparing ice growth from water vapor and
from liquid water would also likely be fruitful. Better data are needed for both systems, and the
theory describing the correspondence needs to be developed. Comparing the ice/water and ice/QLL
interfaces would certainly help develop our knowledge of these interfaces, as well as improving our
overall understanding of surface melting in the thick-QLL limit.
Although the physical picture we have painted above is imperfect and somewhat speculative, we
believe it provides a useful framework for further investigations. Our intention in defining this model
is in part to create a plausible qualitative description of the underlying physics that is consistent
with the experimental data. In spite of its shortcomings, we believe that our model of the intrinsic
growth behavior does make some useful predictions, described above, as well as suggesting additional
targeted investigations, both experimental and theoretical.

3 Structure Dependent Attachment Kinetics

The measured intrinsic growth rates, as parameterized above and displayed in Figure 2, immediately
appear to contradict many details in the Nakaya morphology diagram. For example, we see that
σ0,basal < σ0,prism at -15 C, implying that αprism < αbasal for all supersaturations at this temper-
ature. This inequality suggests that columnar prisms would be the preferred growth morphology,
while it is well established that thin plates form at this temperature.
This discrepancy is explained in our model with the phenomenon of structure-dependent attach-
ment kinetics (SDAK) [35, 36, 37], which causes the attachment coefficient α(σsurf ) to depend
on the mesoscale morphological structure of the ice surface itself. For example, our SDAK model
at -15 C assumes that αprism on a thin plate edge is higher than the corresponding “intrinsic”
αprism for a large faceted surface. The increased αprism then reverses the above inequality, yielding
αprism > αbasal and resulting in the growth of thin plates.
More specifically, in our SDAK model the increase in αprism on the edge of a thin plate arises
because σ0,prism decreases as the width of the final molecular terrace on the prism surface approaches
molecular dimensions. Observations indicate that the radius of curvature of the edge of a thin plate
at -15 C is typically R ≈ 0.5 µm, so the width of the last molecular terrace on the prism surface is
roughly w ≈ (aR)1/2 ≈ 40a, where a ≈ 0.3 nm is the size of a water molecule. The fundamental
premise of our SDAK model is that surface melting and perhaps other structural effects will become
altered on such a narrow terrace, affecting the nucleation dynamics. Since the nucleation of new
molecular layers on this thin terrace largely determines the growth rate of the prism edge, it follows
that SDAK effects play an important role in governing ice growth morphologies.
Here again, our detailed understanding of the ice surface is not sufficient to define a incontro-
vertible model of structure dependent attachment kinetics from first principles. However, growth
measurements do demonstrate the existence of this phenomenon [37], and these same measurements
can characterize the SDAK effects to some degree. We proceed with the understanding that our
SDAK model is quite preliminary at present. Additional measurements and theoretical work will be

10
Figure 4: A schematic depiction of the SDAK instability described in the text. The growth of the
corner of a faceted ice crystal prism (left) is dominated by the nucleation of terraces on the basal and
prism facets (center). If the nucleation rate increases as the width of the top basal terrace decreases
(top right), in keeping with the SDAK model, then the accelerated growth narrows the basal surface,
accelerating the growth still more. The resulting positive feedback generates the growth of a hollow
columnar crystal. If the same SDAK effect is more prevalent on the prism facet (lower right), then
the instability leads to the growth of a thin plate from the top edge of the prism.

necessary to understand the importance of SDAK effects on both facets as a function of temperature
to a satisfactory level.

3.1 An Edge-Enhancing SDAK Instability

A key feature of our SDAK hypothesis is that it leads to an edge-enhancing growth instability. The
essential mechanism is that as a thin edge begins to form, σ0 decreases and thus further increases
the edge growth rate. The enhanced growth causes the edge to sharpen, which again increases the
growth rate. This positive feedback yields a growth instability that enhances the formation of sharp
edges.
To examine the SDAK instability in more detail, consider the growth of an initially isometric
prism on a substrate, depicted in Figure 4. We assume the presence of an inert background gas
surrounding the crystal, so the growth is partially diffusion limited. If the crystal is growing slowly
(center diagram in the figure), then molecular terraces nucleate slowly near each corner of the
crystal, where σsurf is locally highest, and steps propagate away from the corner. The corner itself
is rounded from the Gibbs-Thomson effect. For slow growth, this is essentially the standard model
of diffusion-limited faceted crystal growth, resulting in slightly concave faceted surfaces.
Consider now the highest molecular terrace on either facet next to the growing corner. As the
supersaturation is increased, new terraces nucleate more frequently, so the average width of the top
terrace decreases. As the SDAK effect reduces σ0 on narrower terraces, the nucleation rate increases

11
and in turn the more rapid growth further decreases the width of the top terrace.
At this point a competition occurs between growth on the basal and prism facets, as shown in
the pair of diagrams on the right side of Figure 4. If the SDAK effect preferentially reduces σ0
on the basal facet (top right diagram), then the basal growth is especially enhanced. Because the
growth is also diffusion-limited, the fast growth on the basal facet depletes the water vapor supply
from the nearby prism facet. This decreases the nucleation rate on the prism facet, which causes
the average width of the top terrace to increase, which in turn increases σ0,prism . The combined
effect is that σ0,prism increases to essentially its intrinsic value while σ0,basal grows ever smaller as
the basal edge grows sharper. The final result is a hollow column morphology with thin basal edges.
Alternatively, the same instability could favor the prism facets, as seen in the lower right diagram
in Figure 4. In this case a thin plate-like crystal would form on the isometric prism, producing a
“plate-on-pedestal” morphology [37].
Note that the SDAK instability nicely explains the abrupt transitions between plate-like and
columnar growth seen in the morphology diagram. Relatively small changes in the surface attach-
ment kinetics with temperature can be amplified via the SDAK instability to yield very substantial
changes in the final crystal morphologies.
Note also that the SDAK instability is essentially an extension of the well-known Mullins-Sekerka
instability in diffusion-limited growth [6]. The latter is well known for producing dendritic branching
during solidification, but alone it does not explain the formation of thin plate-like or hollow columnar
crystals. Thin-edge morphologies require strong anisotropies in the attachment kinetics – namely
αprism ≫ αbasal for thin plates or αprism ≪ αbasal for hollow columns. The SDAK instability
provides a natural mechanism to generate these strong anisotropies.

3.2 Physical Origins - II

The SDAK effect is difficult to fully evaluate using growth measurements alone, in part because of
the SDAK instability. For example, the strong SDAK effect on the prism surfaces at -15 C promotes
the growth of thin plates at this temperature, which in turn means that the basal facets are quite
large. If an SDAK effect were present on the basal facets at -15 C, we could not see it because it is
essentially overshadowed by the stronger SDAK effect on the prism surfaces. The bifurcation into
plate-like or columnar growth shown in Figure 4 means that only the dominant SDAK effect can be
characterized by growth measurements at a given temperature.
We propose that the strong SDAK effect on the prism facet at -15 C is related to the precipitous
drop in σ0,prism for temperatures above -10 C. When a prism facet becomes very narrow at -15
C, we suggest that the molecular binding in the top terrace is effectively reduced relative to the
binding on a larger facet surface. This reduced binding causes an increase in surface melting, which
in turn causes a decrease in σ0,prism , as one normally sees at higher temperatures. Thus the function
σ0,prism (T ) is essentially shifted in temperature when the prism terrace becomes narrow. Although
the details of this behavior are uncertain, we suggest that σ0,prism (T ) on the top terrace of a thin
edge looks something like the dotted line shown in the lower panel in Figure 5.
With this picture, we see that the SDAK effect will be less prevalent on a prism facet below -20
C, because a larger effective shift in temperature would be needed to substantially reduce σ0,prism ,
and such large shifts are not provided by the SDAK mechanism. Thus in Figure 5 we see that
σ0,prism,SDAK ≈ σ0,prism at -20 C. Put another way, the SDAK effect on the prism facet is most
prevalent when the temperature is close to the drop-off in σ0,prism associated with surface melting.
At temperatures below -20 C, reducing the width of the prism facet is not sufficient to induce surface

12
Figure 5: In this graph the solid lines show the intrinsic σ0,basal (T ) and σ0,prism (T ) on large facet
surfaces, using the data shown in Figure 2. The dotted lines show our proposed σ0,basal,SDAK (T )
and σ0,prism,SDAK (T ) when the terrace width is very small, as a result of the SDAK effect, as
described in the text.

melting and its accompanying drop in σ0,prism . The end result is that the tendency to produce thin
plate-like crystals is reduced below -20 C, as is observed.
On the other hand, we would expect that the SDAK effect at -10 C on the prism facet would be
substantial, since the intrinsic σ0,prism (T ) continues to drop at still higher temperatures. However,
once σ0,prism is reduced to values of order 0.1 percent, the nucleation barrier becomes so low that
any additional reduction would have little effect on the growth rates. At the same time, σ0,basal
drops with increasing temperature near -10 C as well. When both σ0,basal and σ0,prism are low,
then the anisotropy in αprism /αbasal will decrease, so thin plates no longer form. Putting the pieces
together, this qualitatively explains why the formation of thin plates peaks at temperatures around
-15 C, as seen in the morphology diagram.
Note that this overall physical picture suggests that there might exist a weaker SDAK effect
on the basal facet at -15 C, since σ0,basal also drops with increasing temperature. However, as
mentioned above, because the SDAK instability favors the growth of thin plates at this temperature,
the SDAK effect on the basal facet has little influence on the overall crystal growth morphology.
Because this weaker SDAK effect on the basal surface at -15 C is quite uncertain, and moreover it
is overshadowed by the SDAK effect on the prism surface at this temperature, is was not included
in Figure 5. Nevertheless it may exist, and it may be measurable by other means, so we suggest it

13
as a possibility for future investigation.
At temperatures near -5 C, our model includes an SDAK instability favoring rapid basal growth
and the formation of needles and hollow columns. However, the underlying physical mechanism is
quite different from the SDAK effect on the prism surface near -15 C. In our discussion of σ0,basal (T )
above, we suggested that the formation of a relatively sharp ice/QLL interface on the basal facet
caused a positive dσ0,basal /dT at -6.5 C, and likewise caused βbasal to be considerably higher than
βprism at higher temperatures. We further propose that the sharp ice/QLL interface is relatively
fragile when the QLL is thin, and therefore the sharp interface becomes unstable when the width of
the basal facet is reduced to near molecular dimensions.
In the absence of a sharp ice/QLL interface, we would then expect that σ0,basal would diminish
with increasing temperature, so that σ0,basal is no longer much larger than σ0,prism . In other words,
we would expect that σ0,basal (T ) on a narrow basal terrace would extrapolate from the behavior at
-7 C and lower temperatures as shown in the top panel Figure 5. The details of this extrapolation
are uncertain, but are not terribly important for the present discussion; all that matters is that
the reduction in σ0,basal on a narrow basal terrace near -5 C is sufficient to bring about the SDAK
instability that promotes columnar growth.

3.3 Further Investigations of Structure Dependent Attachment Kinetics

Our understanding of SDAK effects and the SDAK instability are rather poor at present, so much
work is needed to develop a clearer picture of this phenomenon. The measurements in [37] present a
rather strong case that the SDAK instability does exist, at least near -15 C, but little has been done
at other temperatures. Our model clearly predicts that the SDAK instability should be present
at temperatures near -5 C, and demonstrating this experimentally seems quite straightforward.
Investigations of growth behaviors at other temperatures should also be fruitful, and in principle
one could map out the SDAK effects over a range of conditions.
Observations of growth behaviors as a function of supersaturation, temperature, air pressure, and
initial conditions could all provide useful insights. As was demonstrated in [37], however, quantitative
comparisons between growth data and numerical models are essential for making progress in this
area. Careful numerical modeling of diffusion-limited growth, accurate growth measurements, and
well-defined experimental conditions, are all necessary to disentangle the various processes affecting
snow crystal growth rates and morphologies.

4 Discussion
The goal of the present paper is not to have the final word on this subject, but rather to take a
significant step forward in defining a comprehensive physical model of snow crystal growth dynamics.
Toward this end we believe we have identified many key features in such a model, and we have pro-
posed plausible physical origins for various aspects of the model behavior. Our model is admittedly
complex, and we have used some ad hoc assumptions to fill in aspects that are not otherwise well
constrained. On the plus side, however, our model makes many quantitative predictions that can be
tested with targeted experimental investigations. Moreover, our model also suggests theoretical in-
vestigations, particularly using molecular dynamics simulations, that may yield important insights.
The usefulness of any model is its predictive power, and time will tell how well the model stands up
to additional scrutiny.

14
 The first important feature in our model is simply that the transport of water molecules via
diffusion through the surrounding background gas plays a substantial role in guiding the growth and
morphological evolution of a snow crystal. This conclusion is certainly not new, but the recent de-
velopment of physically realistic cellular automata methods appears to be a significant breakthrough
in modeling the diffusion-limited growth of crystals that are both faceted and branched, as are snow
crystals.
A careful look at the diffusion equation shows that the growth of snow crystals under ordinary
atmospheric conditions is almost always strongly affected by diffusion effects. Only for micron-scale
crystals growing at quite low pressures, or for very slowly growing crystals, is this statement not
correct. Moreover, about the only way to model the diffusion-limited growth of snow crystals with
adequate accuracy is by employing numerical methods. Without careful diffusion modeling, it is
extremely difficult, if not impossible, to experimentally measure snow crystal growth parameters
in a reliable fashion. This basic fact has not been adequately appreciated in many previous snow
crystal growth experiments.
A second key feature in our model is the central role of the surface attachment kinetics, especially
the intrinsic attachment coefficients of the principal facets, αbasal and αprism . A successful model
must first describe how these parameters depend on temperature T and supersaturation σsurf ace , and
must provide a reasonable picture of the underlying physical processes responsible for the observed
behaviors in αbasal (T, σsurf ace ) and αprism (T, σsurf ace ). This task, by no means simple, is the first
step toward defining a comprehensive model of snow crystal growth dynamics.
In our model we rely on the empirical growth data to define the intrinsic attachment coefficients.
These data are well described by a dislocation-free layer-nucleation crystal growth model, which we
parameterize as α(σsurf ) = A exp(−σ0 /σsurf ), and the measured parameters A(T ) and σ0 (T ) for the
basal and prism facets are shown in Figure 2. We believe these data are substantially more reliable
than previous attempts, in part because careful attention was given to understanding systematic
experimental effects arising from particle diffusion, substrate interactions, and other sources.
We attempted to paint a reasonable physical picture explaining the intrinsic growth behavior
in our discussion above. While we believe this picture is satisfactory, there may well be other
interpretations of the data. This is clearly an area for additional experimentation and theoretical
investigations. The ice surface has a complex structure and it is far from being in a static state.
Understanding in detail how the structure and dynamics of the ice surface affect crystal growth rates
remains a significant challenge.
A third key feature in our model is the importance of structure-dependent attachment kinetics,
and this is perhaps the most novel and controversial aspect of what we have presented above. The
evidence supporting the SDAK phenomenon has become quite strong, and we believe it is now
difficult to deny that SDAK effects are present in snow crystal growth, and that they play a central
role in the formation of thin plate-like crystals near -15 C. The details in what we presented above
may be inaccurate at some level, and the physical origins we put forth for the SDAK effect may
require revision, but we believe that the SDAK effect must be an essential part of any comprehensive
snow crystal growth model. Further experimental and theoretical investigations into the SDAK effect
and the SDAK instability are likely to provide important new insights into this phenomenon.
Experimental and theoretical efforts focusing on structure formation in ice over the past several
decades have continually pushed the remaining frontiers ever closer to the molecular scale. The
SDAK instability, the parameterization of the attachment coefficients for the principal facets, as
well as the properties of surface melting, all arise from the detailed molecular dynamics at the
crystal surface. Since much progress has been made recently in molecular dynamics simulations of

15
the ice surface, it appears promising that additional investigations along these lines may reveal new
insights into ice growth behavior, and especially why the ice surface properties vary with temperature
as they do. How these advances apply to other crystal systems, and to our understanding of surface
molecular dynamics in general, remains to be seen.

References
[1] Libbrecht, K. G., “The physics of snow crystals,” Rep. Prog. Phys., 68, 855-895 (2005).
[2] Nakaya, U., “Snow Crystals: Natural and Artificial,” (Harvard University Press) (1954).
[3] Bailey, M., and Hallett, J., “Ice Crystal Linear Growth Rates from -20 degrees to -70 degrees
C: Confirmation from Wave Cloud Studies,” J. Atmos. Sci. 69, 390-402 (2012).
[4] Takahashi, T., et al., “Vapor Diffusional Growth of Free-Falling Snow Crystals between -3 and
-23 C,” J. Meteor. Soc. Japan 69, 15-30 (1991).
[5] Fukuta, N. and Takahashi, T., “The growth of atmospheric ice crystals: A summary of findings
in vertical supercooled cloud tunnel studies ,” J. Atmos. Sci. 56, 1963-1979 (1999).
[6] Mullins, W. W., and Sekerka, R. F., “Stability of planar interface during solidification of dilute
binary allow,” J. Appl. Phys. 35, 444-451 (1964).
[7] Trivedi, R., and Kurz, W., “Dendritic growth,” Inter. Mater. Rev. 39, 49-74 (1994).
[8] Brener, E. A., “Three-dimensional dendritic growth,” J. Cryst. Growth 166, 339–346 (1996).
[9] Libbrecht, K. G., Crosby, T., and Swanson, M., “Electrically enhanced free dendrite growth in
polar and non-polar systems,” J. Cryst. Growth 240, 241-249 (2002).
[10] Juric, D., and Tryggvason, G., “A Front-Tracking Method for Dendritic Solidification,” J.
Comp. Phys. 123, 127-148 (1996).

[11] Karma, A., and Rappel, W. J., “Phase-Field Method for Computationally Efficient Modeling
of Solidification with Arbitrary Interface Kinetics,” Phys. Rev. E 53 R3017-R3020 (1996).
[12] Barrett, J. W., Garcke, H., and Nuernberg, R., “Numerical Computations of Faceted Pattern
Formation in Snow Crystal Growth,” Phys. Rev. E 86, 011604 (2012).
[13] Gravner, J., and Griffeath, D., “Modeling snow crystal growth II: A mesoscopic lattice map
with plausible dynamics,” Physica D237, 385-404 (2008).

[14] Gravner, J., and Griffeath, D., “Modeling snow-crystal growth: A three-dimensional mesoscopic
approach,” Phys. Rev. E 79.011601 (2009).
[15] Libbrecht, K. G., “Physically derived rules for simulating faceted crystal growth using cellular
automata,” arXiv:0807.2616 (2008).
[16] Gonda, T., “The growth of small ice crystals in gases of high and low pressures,” J. Meteor.
Soc. Japan 54, 233-240 (1976).

16
[17] Libbrecht, K. G., “Identification of a Novel ‘Fishbone’ Structure in the Dendritic Growth of
Columnar Ice Crystals,” arXiv:0912.2522v1 (2009).
[18] Dash, J. G., Rempel, A. W., and Wettlaufer, J. S., “The physics of premelted ice and its
geophysical consequences,” Rev. Mod. Phys. 78, 695-741 (2006).
[19] Saito, Y., “Statistical Physics of Crystal Growth,” World Scientific Books (1996).
[20] Hobbs, P. V., “Ice Physics,” Oxford University Press (1974).
[21] Kuroda, T., “Growth kinetics of ice single crystal from vapour phase and variation of its growth
form,” J. Meteor. Soc. Japan 60, 520-534 (1982).
[22] Kobayashi, T., and Kuroda, T., “Snow crystals”, in Morphology of Crystals—Part B, editor I.
Sunagawa (Tokyo: Terra Scientific), 645–743 (1987).
[23] Fukuta, N., and Lu, Q. J., “Surface Microphysical Mechanism for Ice Crystal Growth Habit
Development,” J. Atmos. Res. 32, 31-43 (1994).
[24] Nelson, J., “Growth mechanisms to explain the primary and secondary habits of snow crystals,”
Phil. Mag. A 81, 2337-2373 (2001).
[25] Westbrook, C. D., and Heymsfield, A. J., “Ice crystals growing from vapor in supercooled clouds
between -2.5 C and -22 C: Testing current parameterization methods using laboratory data,”
J. Atmos. Sci. 68, 2416-2429 (2011).
[26] Libbrecht, K. G. and Rickerby, M. E., “Crystal growth in the presence of surface melting: novel
behavior of the principal facets of ice,” arXiv:1208.5982 (2012).
[27] Libbrecht, K. G., and Bell, R., “Chemical influences on ice crystal growth from vapor,”
arXiv:1101.0127 (2011).
[28] Nada, H., van der Eerden, J. P., and Furukawa, Y., “A clear observation of crystal growth of
ice from water in a molecular dynamics simulation with a six-site potential model of H2O,” J.
Cryst. Growth 266, 297-302 (2004).
[29] Carignano, M. A., Shepson, P. B., and Szleifer, I., “Molecular dynamics simulations of ice
growth from supercooled water,” Mol. Phys. 103, 2957–2967 (2005).
[30] Matsumoto, M., Saito, S., and Ohmine, I., “Molecular dynamics simulation of the ice nucleation
and growth process leading to water freezing,” Nature 416, 409-413 (2002).
[31] Mutaftschiev, B., “The Atomistic Nature of Crystal Growth,” Springer-Verlag (2001).
[32] Dosch, H., Lied, A., and Bilgram, J. H., “Glancing-angle X-ray scattering studies of the pre-
melting of ice surfaces,” Surf. Sci. 327, 145-164 (1995).
[33] Maruyama, M., “Roughening transition of prism faces of ice crystals grown from melt under
pressure,” J. Cryst. Growth 275, 598-605 (2005).
[34] Wolf, P. E., et al., “Crystal growth and crystal curvature near roughening transitions in hcp
4
He,” J. Phys. 46, 1987-2007 (1985).

17
[35] Libbrecht, K. G., “Explaining the formation of thin ice crystal plates with structure-dependent
attachment kinetics,” J. Cryst. Growth 258, 168-175 (2003).
[36] Libbrecht, K. G., “Observations of an edge-enhancing instability in snow crystal growth near
-15 C,” arXiv:1111.2786 (2011).
[37] Libbrecht, K. G., “An edge-enhancing crystal growth instability caused by structure-dependent
attachment kinetics,” arXiv:1209.4932 (2012).

18