tmain.tex - seaice-experiments - sea ice experiments using Granular.jl
 (HTM) git clone git://src.adamsgaard.dk/seaice-experiments
 (DIR) Log
 (DIR) Files
 (DIR) Refs
 (DIR) README
 (DIR) LICENSE
       ---
       tmain.tex (20080B)
       ---
            1 %\documentclass[11pt,a4paper]{article}
            2 \documentclass[11pt]{article}
            3 
            4 %\usepackage{a4wide}
            5 \usepackage[margin=1.4in]{geometry}
            6 
            7 \usepackage{graphicx}
            8 %\usepackage[german, english]{babel}
            9 %\usepackage{tabularx}
           10 %\usepackage{cancel}
           11 %\usepackage{multirow}
           12 %\usepackage{supertabular}
           13 %\usepackage{algorithmic}
           14 %\usepackage{algorithm}
           15 %\usepackage{amsthm}
           16 %\usepackage{float}
           17 %\usepackage{subfig}
           18 \usepackage{rotating}
           19 \usepackage{amsmath}
           20 
           21 \usepackage[T1]{fontenc} % Font encoding
           22 \usepackage{charter} % Serif body font
           23 \usepackage[charter]{mathdesign} % Math font
           24 \usepackage[scale=0.9]{sourcecodepro} % Monospaced fontenc
           25 \usepackage[lf]{FiraSans} % Sans-serif font
           26 \usepackage{textgreek}
           27 
           28 \usepackage{listings}
           29 
           30 \usepackage{hyperref}
           31 
           32 \usepackage{soul} % for st strikethrough command
           33 
           34 %\usepackage[round]{natbib}
           35 \usepackage[natbib=true, style=authoryear, bibstyle=authoryear-comp, 
           36 maxbibnames=10,
           37 maxcitenames=2, backend=bibtex8]{biblatex}
           38 \bibliography{/Users/ad/articles/own/BIBnew.bib}
           39 
           40 
           41 \begin{document}
           42 
           43 \title{Discrete-element simulation of sea-ice mechanics:\\
           44     Parameterization of contact mechanics}
           45 
           46 \author{\small Anders Damsgaard}
           47 \date{\small \today}
           48 
           49 \maketitle
           50 
           51 \section{Rationale}
           52 Lagrangian parameterizations of sea ice offer several advantages to Eulerian 
           53 continuum methods.  The continuum assumption of sea-ice mechanics becomes 
           54 increasingly questionable as spatial resolution and cell size approaches the 
           55 size of ice floes.  Ice-floe scale discretizations are natural for Lagrangian 
           56 models, which additionally offer the convenience of being able to handle 
           57 arbitrary sea-ice concentrations.  This is likely to improve model performance 
           58 in ice-marginal zones with strong advection.  Furthermore, phase transitions in 
           59 granular rheology around the jamming limit, such as observed when sea ice moves 
           60 through geometric confinements, includes sharp thresholds in effective 
           61 viscosity, which are typically ignored in Eulerian models.
           62 Granular jamming is a stochastic process dependent on having the right grains in 
           63 the right place at the right time, and the jamming likelihood over time can be 
           64 described by a probabilistic model
           65 \citep[e.g.][]{Tang2009}.  Difficult to parameterize in continuum formulations 
           66 \citep[e.g.][]{Rallabandi2017}, jamming comes natural to dense granular systems 
           67 simulated in a Lagrangian framework, and is a very relevant process complicating 
           68 sea-ice transport through narrow straits \citep[e.g.][]{Kwok2010}.
           69 However, computational performance is a central concern for coupled 
           70 ocean-atmosphere models, with traditional discrete-element method (DEM) 
           71 formulations being too costly for global-scale models.  Here we investigate the 
           72 behavior of a Lagrangian sea-ice model and assess how mechanical 
           73 parameterizations of differing complexity affect collective behavior at scales 
           74 much larger than the individual ice floes.
           75 
           76 
           77 \section{Contact models}
           78 The presented experiments compare jamming behavior between two differing 
           79 ice-floe contact models.  Common to both models, the resistive force 
           80 $\boldsymbol{f}_\text{n}$ to axial compressive strain between to cylindrical ice 
           81 floes $i$ and $j$ is provided by (Hookean) linear-elasticity based on the 
           82 overlap distance $\boldsymbol{\delta}_\text{n}$:
           83 \begin{equation}
           84     \boldsymbol{f}_\text{n}^{ij} =
           85     A^{ij} E^{ij} \boldsymbol{\delta}_\text{n}^{ij}
           86     \quad \text{when} \quad
           87     0 > |\boldsymbol{\delta}_\text{n}^{ij}| \equiv |\boldsymbol{x}^i - 
           88     \boldsymbol{x}^j| - (r^i + r^j)
           89     \label{eq:f_n}
           90 \end{equation}
           91 where the contact cross-sectional area $A^{ij} = R^{ij} \min(T^i, T^j)$ is 
           92 determined by the harmonic mean $R^{ij} = 2 r^i r^j/(r^i + r^j)$ of the ice-floe 
           93 radii $r^i$ and $r^j$, as well as the smallest of the involved ice-floe 
           94 thickness $T^i$ and $T^j$.  The linear elastic force resulting from axial strain 
           95 of a distance $|\boldsymbol{\delta}_\text{n}^{ij}|$ of the contact is scaled by 
           96 the harmonic mean of Young's modulus $E^{ij}$.  The stiffness scaling based on 
           97 Young's modulus is scale invariant \citep[e.g.][]{Obermayr2013}, and is based on 
           98 the principle that the elastic stiffness of the ice itself is constant, 
           99 regardless of ice-floe size.
          100 
          101 Nonlinear elasticity models based on Hertzian contact mechanics may 
          102 alternatively be applied to determine the stresses resulting from contact 
          103 compression \citep[e.g.][]{Herman2013}.  However, with nonlinear models the 
          104 numerical stability of the explicit temporal integration scheme will depend on 
          105 the stress and packing state of the granular assemblage, and will under 
          106 localized compressive extremes require very small discretization in time.
          107 
          108 As demonstrated later, models based compressive strength alone are not 
          109 sufficient to cause granular jamming.  We explore two different additions to the 
          110 contact model presented in Eq.~\ref{eq:f_n}.  The first approach is typical to 
          111 DEM models and is based on resolving including shear resistance through 
          112 tangential (contact parallel) elasticity and associated Coulomb frictional 
          113 limits.  An alternative approach, fundamentally complementary to compressive 
          114 elasticity and shear friction, is tensile strength of ice-floe contacts which 
          115 leads to a cohesive bulk granular rheology.
          116 
          117 \subsection{Contact-parallel elasticity with Coulomb friction}
          118 The contact-parallel (tangential) elasticity is resolved by determining the 
          119 contact travel distance $\boldsymbol{\delta}_\text{t}$ on the contact plane for 
          120 the duration of the contact $t_\text{c}$:
          121 \begin{equation}
          122     \boldsymbol{\delta}_\text{t} = \int_0^{t_\text{c}}
          123     \left[
          124         (\boldsymbol{v}^i - \boldsymbol{v}^j) \cdot \hat{\boldsymbol{t}}^{ij} -
          125         R^{ij} \left(\omega^i + \omega^j\right)
          126         \right]
          127     \label{eq:d_t}
          128 \end{equation}
          129 where $\boldsymbol{v}$ and $\omega$ denotes linear and rotational velocity, 
          130 respectively.
          131 The deformation distance $\boldsymbol{\delta}_\text{t}$ is corrected for contact 
          132 rotation over the lifetime of the interaction, and is used to determine the 
          133 contact-tangential elastic force:
          134 \begin{equation}
          135     \boldsymbol{f}_\text{t} = \frac{E^{ij} A^{ij}}{R^{ij}}
          136     \frac{2 (1 - \nu^2)}{(2 - \nu)(1 + \nu)} \boldsymbol{\delta}_\text{t}
          137 \end{equation}
          138 with $\nu$ being the dimensionless Poisson's ratio characteristic for the 
          139 material.  However, the frictional force above is restricted to the 
          140 Coulomb-frictional limit, relating the magnitude of the normal force to the 
          141 magnitude of the tangential force:
          142 \begin{equation}
          143     |\boldsymbol{f}_\text{t}| \leq \mu |\boldsymbol{f}_\text{n}|
          144     \label{eq:coulomb_friction}
          145 \end{equation}
          146 The Coulomb-frictional coefficient $\mu$ may be parameterized as state 
          147 dependent, for example with different values for sliding and stationary 
          148 contacts.  However, for the following a single and constant value is used for 
          149 the frictional coefficient $\mu$.
          150 
          151 In the case of slip ($|\boldsymbol{f}_\text{t}| > \mu 
          152 |\boldsymbol{f}_\text{n}|$) the length of the contact-parallel travel path 
          153 $\boldsymbol{\delta}_\text{t}$ is reduced to be consistent to be consistent with 
          154 the Coulomb limit.  This accounts for the tangential contact plasticity and 
          155 irreversible work associated with sliding.
          156 
          157 Since the above model of tangential shear resistance is based on deformation 
          158 distance on the inter-floe contact plane, it requires solving for ice-floe 
          159 rotational kinematics of each ice floe and a bookkeeping algorithm for storing 
          160 contact histories.
          161 
          162 \subsection{Tensile contact strength}
          163 Cohesion is introduced by parameterizing resistance to extension beyond the 
          164 overlap distance between a pair of ice floes (i.e.\ $\delta_\text{n}^{ij} > 0$).  
          165 In real settings tensile strength may be provided by refreezing processes at the 
          166 ice-floe interface or mechanical ridging.  An ideal formulation of bond 
          167 deformation includes resistance to bond tension, shear, twist, and rolling 
          168 \citep[e.g.][]{Potyondy2004, Obermayr2013, Herman2016}.  However, for this study 
          169 we explore the possibility of using bond \emph{tension} alone as a mechanical 
          170 component contributing to bulk granular shear strength.
          171 
          172 Tensile strength is parameterized by applying Eq.~\ref{eq:f_n} for the extensive 
          173 regime ($\delta_\text{n} > 0$).  Eq.~\ref{eq:f_n} is applied until the tensile 
          174 stress exceeds the tensile strength $\sigma_\text{c}$ defined for the bonds.  
          175 The implementation allows for time-dependent tensile strengthening based on 
          176 contact age, but for the following we parameterize the bonds to obtain full 
          177 tensile strength as soon as a pair of ice floes first undergo compression 
          178 ($\delta_\text{n} < 0$).
          179 
          180 
          181 \subsection{Implementation}
          182 The intent of this work is to develop and apply an offline sea-ice dynamical 
          183 code in order to explore strengths and limitations of different methods related 
          184 to sea-ice mechanics.  Once the appropriate parameterization has been identified 
          185 it will be built in to the GFDL model of coupled ocean/atmosphere dynamics.
          186 
          187 The presented experiments are performed with the 
          188 \texttt{SeaIce.jl}\footnote{\url{https://github.com/anders-dc/SeaIce.jl}} Julia 
          189 package for offline simulations of sea-ice mechanics, which includes the above 
          190 parameterizations.  Furthermore, the model includes the option to include linear 
          191 (Newtonian) contact viscosity in the normal and tangential components, but an 
          192 analysis of these is not included here.  Simulation-specific scripts are 
          193 provided in the repository 
          194 \texttt{SeaIce-experiments}\footnote{\url{https://github.com/anders-dc/SeaIce-experiments}}.
          195 
          196 The offline model implementation identifies ice-floe contacts and placement in 
          197 the ocean/atmosphere grid through a cell-based spatial discretization, which 
          198 reduces the computational overhead significantly (Fig.~\ref{fig:performance}, 
          199 and additionally mirrors the current Lagrangian ice-berg implementation in the 
          200 GFDL ocean model.
          201 
          202 \begin{figure}
          203     \begin{center}
          204         \includegraphics[width=0.7\textwidth]{graphics/profiling-cpu.pdf}
          205     \end{center}
          206     \caption{\label{fig:performance} Single-threaded performance of the DEM 
          207     algorithm with all-to-all contact search, or a contact search using spatial 
          208     decomposition based on the ocean/atmosphere grid (Profiling script: 
          209     \url{https://github.com/anders-dc/SeaIce.jl/blob/master/test/profiling.jl}).}
          210 \end{figure}
          211 
          212 The experiments rely on pseudo-random number generation for generating ice-floe 
          213 particle size distributions (PSDs).  In order to assess the statistical 
          214 probability of granular jamming we seed the system with different values and 
          215 repeat each experiment ten times.
          216 
          217 For the presented runs the ice floes are forced with a uniform and constant wind 
          218 field oriented from north to south, with a simplified channel-like constriction 
          219 in the middle.  The ocean also flows from north to south but is constrained by 
          220 the basin geometry.  The spatial velocity pattern is determined by a stream 
          221 function under mass conservation, meaning that currents increase as
          222 the channel narrows.  Ice floes are initially placed in a pseudo-random 
          223 arrangement north of the channel.  All parameters are kept constant between the 
          224 experiment sets unless explicitly noted.
          225 
          226 \section{Results}
          227 
          228 \subsection{Experiment set 1}
          229 
          230 Increasing frictional coefficients, increasing tensile strength.  Constant 
          231 uniform size distribution.
          232 Generated by running \texttt{`make`} from 
          233 \url{https://github.com/anders-dc/SeaIce-experiments/tree/master/cohesion-friction-comparison}
          234 Animation with \texttt{seed=1} can be found at:
          235 \url{https://youtu.be/JmWvoi9Z6Zk}
          236 
          237 \begin{sidewaysfigure}
          238     \begin{center}
          239         \includegraphics[width=0.99\textheight]%
          240         {graphics/cohesion-friction-comparison.png}
          241         \includegraphics[width=0.99\textheight]%
          242         {{graphics/cohesion-friction-comparison-jam_fraction.png}}
          243     \end{center}
          244     \caption{\label{fig:set1}%
          245     \textbf{Experiment set 1}:
          246     Comparison between the frictional and cohesive model with different 
          247     Coulomb-frictional coefficient values and tensile strengths.
          248     First row: Ice fluxes over time with increasing friction ($\mu = [0.0, 0.1, 
          249     0.2, 0.3, 0.4]$).
          250     Second row: Ice fluxes over time with increasing tensile strength 
          251     ($\sigma_\text{c} = [10, 100, 200, 400, 800]$ kPa).
          252     Third row: Fraction of runs jammed over time with increasing friction.
          253     Fourth row: Fraction of runs jammed over time with increasing tensile 
          254     strength.
          255     }
          256 \end{sidewaysfigure}
          257 
          258 While the frictionless runs rarely jam, we observe that increasing either 
          259 friction or cohesion greatly increases the likelihood of jamming 
          260 (Fig.~\ref{fig:set1}).  Under cohesive runs with large tensile strengths the ice 
          261 floes north of the assemblage behave like a single rigid block, while the 
          262 corresponding frictional runs show more spread in the jam-fraction probability.  
          263 At intermediate to low frictional coefficients ($\mu \leq 0.3$ or 
          264 $\sigma_\text{c} \leq$ 200 kPa) there is good agreement between frictional and 
          265 cohesive models, where the runs with $\mu = 0.3$ most closely resemble the 
          266 jam-fraction distribution of $\sigma_\text{c} = 200$ kPa.  We base further runs 
          267 on comparing the differences between these two parameter sets under varying 
          268 circumstances.
          269 
          270 \subsection{Experiment set 2}
          271 Constant friction coefficient ($\mu = 0.3$) for frictional runs, constant 
          272 tensile strength ($\sigma_\text{c} = 200$ kPa) for cohesive runs.  Variable 
          273 channel width.  Monodisperse ice-floe size distribution.
          274 Generated by running \texttt{`make`} from 
          275 \url{https://github.com/anders-dc/SeaIce-experiments/tree/master/cohesion-friction-width-monodisperse-comparison}.  
          276 Animation with \texttt{seed=1} can be found at:
          277 \url{https://youtu.be/tPMSxdw1UZ8}.
          278 
          279 \begin{sidewaysfigure}
          280     \begin{center}
          281         \includegraphics[width=0.99\textheight]%
          282         {graphics/cohesion-friction-width-monodisperse-comparison.png}
          283         \includegraphics[width=0.99\textheight]%
          284         {{graphics/cohesion-friction-width-monodisperse-comparison-jam_fraction.png}}
          285     \end{center}
          286     \caption{\label{fig:set2}%
          287     \textbf{Experiment set 2}:
          288     Comparison between the frictional ($\mu = 0.3$ and $\sigma_\text{c}$ = 0 
          289     kPa) and cohesive ($\mu = 0$ and $\sigma_\text{c}$ = 200 kPa) models with a 
          290     monodisperse (single) ice-floe size and increasing channel width.
          291     First row: Ice fluxes over time with friction ($\mu = 0.3$) and increasing 
          292     channel width.
          293     Second row: Ice fluxes over time with cohesion ($\sigma_\text{c} = 200$ kPa) 
          294     and increasing channel width.
          295     Third row: Fraction of frictional runs jammed over time.
          296     Fourth row: Fraction of cohesive runs jammed over time.
          297     }
          298 \end{sidewaysfigure}
          299 
          300 The jamming probability of two-dimensional granular assemblages under gravity is 
          301 well described in the literature, especially for monodisperse (same-sized) 
          302 granular materials \citep[e.g.][]{Beverloo1961, To2001, Tang2009}.
          303 While the ocean/atmosphere forcing is different from the gravity drag in the 
          304 silo flows in these publications, we observe the same relationship between 
          305 conduit width and grain size in monodisperse granular assemblages as previously 
          306 reported.  The consistency between the frictional and cohesive model 
          307 (Fig.~\ref{fig:set2}) indicates that the simpler cohesive model is performing 
          308 well in this setting.
          309 
          310 \subsection{Experiment set 3}
          311 Constant friction coefficient ($\mu = 0.3$) for frictional runs, constant 
          312 tensile strength ($\sigma_\text{c} = 200$ kPa) for cohesive runs.  Variable 
          313 channel width.  Constant and uniformly-distributed ice-floe size distribution.
          314 Generated by running \texttt{`make`} from 
          315 \url{https://github.com/anders-dc/SeaIce-experiments/tree/master/cohesion-friction-width-comparison}.
          316 
          317 \begin{sidewaysfigure}
          318     \begin{center}
          319         \includegraphics[width=0.99\textheight]%
          320         {graphics/cohesion-friction-width-comparison.png}
          321         \includegraphics[width=0.99\textheight]%
          322         {{graphics/cohesion-friction-width-comparison-jam_fraction.png}}
          323     \end{center}
          324     \caption{\label{fig:set3}%
          325     \textbf{Experiment set 3}:
          326     Comparison between the frictional ($\mu = 0.3$ and $\sigma_\text{c}$ = 0 
          327     kPa) and cohesive ($\mu = 0$ and $\sigma_\text{c}$ = 200 kPa) models with a 
          328     uniform ice-floe size distribution and increasing channel width.
          329     First row: Ice fluxes over time with friction ($\mu = 0.3$) and increasing 
          330     channel width.
          331     Second row: Ice fluxes over time with cohesion ($\sigma_\text{c} = 200$ kPa) 
          332     and increasing channel width.
          333     Third row: Fraction of frictional runs jammed over time.
          334     Fourth row: Fraction of cohesive runs jammed over time.
          335     }
          336 \end{sidewaysfigure}
          337 
          338 While the monodisperse ice-floe size used in Set 2 provides a comparative basis 
          339 to the literature on granular jamming, we have for Set 3 used a wide grain size 
          340 distribution and varied the channel width as before.  We observe that the width 
          341 of ice-floe size distribution decreases the likelihood of jamming, and that 
          342 these runs are less likely to jam than the corresponding monodisperse ensemble 
          343 runs.  Again we see a good correspondence between the frictional and cohesive 
          344 model runs (Fig.~\ref{fig:set3}).
          345 
          346 \subsection{Experiment set 4}
          347 Constant friction coefficient ($\mu = 0.3$) for frictional runs, constant 
          348 tensile strength ($\sigma_\text{c} = 200$ kPa) for cohesive runs.  Constant 
          349 channel width.  Different widths in power-law distributed ice-floe size 
          350 distribution.
          351 Generated by running \texttt{`make`} from 
          352 \url{https://github.com/anders-dc/SeaIce-experiments/tree/master/cohesion-friction-psd-comparison}.
          353 Animation with \texttt{seed=1} can be found at: 
          354 \url{https://youtu.be/I9P8XsA1S0I}
          355 
          356 \begin{sidewaysfigure}
          357     \begin{center}
          358         \includegraphics[width=0.99\textheight]%
          359         {graphics/cohesion-friction-psd-comparison.png}
          360         \includegraphics[width=0.99\textheight]%
          361         {{graphics/cohesion-friction-psd-comparison-jam_fraction.png}}
          362     \end{center}
          363     \caption{\label{fig:set4}%
          364     \textbf{Experiment set 4}:
          365     Comparison between the frictional ($\mu = 0.3$ and $\sigma_\text{c}$ = 0 
          366     kPa) and cohesive ($\mu = 0$ and $\sigma_\text{c}$ = 200 kPa) models with a 
          367     increasingly wide power-law ice-floe size distribution and constant channel 
          368     width.
          369     First row: Ice fluxes over time with friction ($\mu = 0.3$) and increasing 
          370     ice-floe size span.
          371     Second row: Ice fluxes over time with cohesion ($\sigma_\text{c} = 200$ kPa) 
          372     and increasing ice-floe size span.
          373     Third row: Fraction of frictional runs jammed over time.
          374     Fourth row: Fraction of cohesive runs jammed over time.
          375     }
          376 \end{sidewaysfigure}
          377 
          378 Ice-floe distributions in natural sea-ice settings have been observed to follow 
          379 a power-law probability distribution with an exponent of $\alpha \approx -1.8$ 
          380 \citep[e.g.][]{Steer2008, Herman2013}.  With constant channel size and an 
          381 increasing width of the ice-floe size distribution, we observe that the 
          382 probability of jamming decreases, and observe good agreement between the 
          383 frictional and cohesive model runs.
          384 
          385 \section{Summary}
          386 We construct a flexible discrete-element framework for simulating Lagrangian 
          387 sea-ice dynamics at the ice-floe scale, forced by ocean and atmosphere velocity 
          388 fields.  While frictionless contact models based on tensile stiffness alone are 
          389 very unlikely to jam, we describe two different approaches based on friction and 
          390 tensile strength, both resulting in increased bulk shear strength of the 
          391 granular assemblage.  We demonstrate that the discrete-element approach is able 
          392 to undergo dynamic jamming when forced through an idealized confinement, where 
          393 the probability of jamming is determined by the channel width, ice-floe size, 
          394 and ice-floe size variability.  The frictionless but cohesive contact model can 
          395 with certain tensile strength values display jamming behavior which on the large 
          396 scale is highly similar to a model with contact friction and ice-floe rotation.  
          397 The frictionless and cohesive model is likely an ideal candidate for coupled 
          398 computations due to its reduced algorithmic complexity.
          399 
          400 \printbibliography{}
          401 
          402 \end{document}