Message Boards Message Boards

[WSS18] Optimal Packing of Bidisperse Circles

My WSS project is to study packing of Bidisperse circles, i.e., circles of 2 different radii. To do so, I implemented the Lubachevsky-Stillinger algorithm with an Event Driven approach, based on the algorithmic details here and here, and the C++ implementation here.

As for a heuristic algorithmic description, check out this gif:

Cool Packing Gif

The disks are randomly initialized, and grow with time while colliding and bouncing around. Collisions are processed, new events are computed, and the simulation progresses until jamming (or a event error is flagged and the simulation aborts).

As for the implementation, I segmented it into 9 sections: Constraints, Initialization, Cell Method, Augmented MinHeap, Collisions, Transfers, Procedure, Visualization and Analysis. Each section has its own functions and description (43 functions total!), and I'll try to explain some of them here (I won't comment on Constraints because they are quite straightforward, nor the MinHeap, because it will get its own post).

Initialization

Here is where the initial radii, positions and velocities are defined. The initial positions need to satisfy the constraint functions, and to place another circle entails not overlapping any of the previously input circles, nor the square boundary. The function to do so is createDisk, below. It randomly tries circle coordinates, given a partial list of coordinates (circles) and a full list of radii (radii) which define the constraints.

createDisk[circles_List,radii_List, size_]:=Module[
{cnumber=circles//Length, position, riffled, count=0, insert=False, rc, r},

(*splice radii list, select next radius, riffle coordinates and radii for constraint functions*)
rc=radii[[;;cnumber]]; r=radii[[cnumber+1]]; riffled=Riffle[circles,rc]//Partition[#, 2]&;

(*testing loop*)
While[
(*stop when inserted or overcount*)
And[count<200, !insert ],

    (*randomly try next coordinate*)
    position=RandomReal[{0, size}, 2];

    (*if doesn't overlap border or overlap anyother circle*)
    If[And[
    borderC[position,r],
    And@@(overlapC[{position, r}, #]&/@riffled)
    ], insert=True];
count=count+1;
];
If[count==200, Print["fail"]];

position
]

createDisks

Nested application of createDisk.

Cell Method

Eventually, we will need to calculate the next collision time for some particle i . The easiest way to do so would be to calculate the times for all particles and take the minimum, but that would be computationally expensive for many body systems. Instead, I implemented the Cell Method, essentially segmenting the space into a grid, and placing each particle into bins (stored with a Matrix/Association pair). Now, instead of iterating over all particles to find the minimum collision time, just look at the particles in the neighboring bins. The functions below are 2 of the most important for this method: nearestCells and updateCell.

nearestCells

nearestCells takes a list of coordinates, number of rows and size and returns a matrix (rows, rows) of lists, where cellMatrix[[i, j]] stores the particle indices in that bin.

nearestCells[coordinates_, nrows_, size_]:= Module[ 
{nf=Nearest[coordinates->Range@Length@coordinates, DistanceFunction->ChessboardDistance]},
nf[binCenters[nrows, size], {Infinity, size/(2 nrows)}]
]

Where binCenters[nrows, size] returns a list of grid center coordinates. The visual representation: nearestCells in action

updateCell

As the disks move around, inevitably they will transfer between bins, and the correspondent cellMatrix (result of function above) and cellA (association mapping particle index to bin index pair) have to be updated. updateCell does exactly that:

updateCell[i_, newcell_] := Module[{currentk, currentj, oldk, oldj},
  (*initialize variables*)
  {currentk, currentj} = newcell;
  {oldk, oldj} = cellA[i];
  (*delete and reinsert element from cellMat*)
  cellMatrix[[oldk, oldj]] = Select[cellMatrix[[oldk, oldj]], # != i &]
  AppendTo[cellMatrix[[currentk, currentj]], i];
  (*reset element in association*)
  cellA[i] = {currentk, currentj};
  ]

Collisions

Here is where I implemented all collision-related functions: calculateCollision, predictCollision, findNextCollision, collisionCheck and processCollision. The first 3 are related to calculating the next collision for particle i, the 4^th enforces the collision is valid and symmetric (if i collides with j, j has to collide with i) and the last processes a collision event, through conservation of momentum. Here, events are parametrized by triples: {te, i, p}: te is the time of the event, i is the particle index, and p_Integer the event partner: if 0 < p < N+1, the event is a collision with particle p. Take a look at the 2 most important of them:

calculateCollision

This function calculates the collision time, if it exists (else, infinity). Additionally, this function aborts the program if there is considerable overlap between particles, and prints an error statement.

calculateCollision[i_, j_]:= Module[
{nRi, nRj, d, \[Sigma], \[CapitalDelta]R, \[CapitalDelta]V, A, B, C, t},
(*sets positions and radii to current time*)
nRi=positions[[i]]+(gtime-lastupdates[[i]])velocities[[i]];
nRj=positions[[j]]+(gtime-lastupdates[[j]])velocities[[j]];
\[Sigma]=(radii[[i]]+radii[[j]])^2;

(*define helper variables*)
\[CapitalDelta]R=nRi-nRj;
\[CapitalDelta]V=velocities[[i]]-velocities[[j]];
A=\[CapitalDelta]V.\[CapitalDelta]V-rate^2 \[Sigma];
B=\[CapitalDelta]V.\[CapitalDelta]R-rate(1+rate gtime)\[Sigma];
C=\[CapitalDelta]R.\[CapitalDelta]R-\[Sigma] (1+rate gtime)^2;
d=B^2-A C;

(*debugging tests*)
If[C<-10^(-2)(radii[[i]]+radii[[j]])(1+gtime rate), Print["After ", totalevents, " events, overlap."]; Abort[]];

(*calculate time*)
Which[
    C <-10^(-4)(radii[[i]]+radii[[j]])(1+gtime rate) && B < 0, Print["spheres are already overlapping ", C]; Infinity,

     d <= 0, Infinity,  

    True, -(B+Sqrt[d])/(A)
    ]
]

collisionCheck

collisionCheck enforces collisions are symmetric, as well as updating p's event partner's events if necessary (see Procedure section for further details):

collisionCheck[collision_] := Module[
  {tc, i, j, symcol, k, n = nextevents // Length, ek},
  {tc, i, j} = collision;
  symcol = {tc, j, i};
  k = nextevents[[j]][[3]];
  (*j cannot have event before colliding with i*)
  If[tc > nextevents[[j]][[1]], Print["big mistake"]; Abort[]];

  (*if j's next event was collision, give k a check*)
  If[(k <= n) && (k != i) && (k >= 1), 
   ek = nextevents[[k]];
   nextevents[[k]] = {ek[[1]], k, Infinity};];

  (*set symmetric collision as j's next event*)
  nextevents[[j]] = symcol;
  (*update Heap*)
  decreaseKey[j, tc];
  ]

Ignore the heap stuff for now, I'll explain in a bit.

Transfers

This section is related specifically to transfers: events corresponding to movement between cell walls. Each transfer event can be parametrized using the same triples, now, with N < p < N+5, where p-N indicates the specific cell wall: 1 for the right wall, 2 for the left, 3 for the top, 4 for the bottom. There are 2 functions in this section: findNextTransfer and processTransfer. The second is quite straightforward, and is basically a set of Which cases and updateCell calls. findNextTransfer, however, is just slightly more elaborated.

findNextTransfer

This function is also a set of Which cases, for each direction (1,2, 3 or 4), now considering radii growth if colliding with a hard boundary. See below:

findNextTransfer[i_]:= Module[
{r, rx, ry, vx, vy, wallindex, binindex, bincenter, step=size/nrows, times=Table[Infinity, {4}]},

(*set positions, set velocities, set radius*)
r = radii[[i]];
{rx, ry} = positions[[i]] + velocities[[i]](gtime-lastupdates[[i]]);
{vx, vy} = velocities[[i]];

(*obtain bin position*)
binindex = cellA[i];
bincenter = step(binindex - {1,1}/2);

(*consider x axis*)
Which[
    vx>0 && binindex[[1]]!=nrows, 
       times[[1]]=(bincenter[[1]]+step/2-rx)/vx;, 
    vx >0 && binindex[[1]] == nrows,
       times[[1]]=(bincenter[[1]]+step/2-rx-r(1 + gtime rate))/(vx+rate radii[[i]]);,
    vx<0 && binindex[[1]]!=1,
       times[[2]]=(bincenter[[1]]-step/2-rx)/vx;,
    vx <0 && binindex[[1]]==1, 
       times[[2]]=(bincenter[[1]]-step/2-rx+r(1+gtime rate))/(vx-rate radii[[i]]);
];

(*consider y axis*)
Which[
    vy>0 && binindex[[2]]!=nrows, 
       times[[3]]=(bincenter[[2]]+step/2-ry)/vy;, 
    vy>0 && binindex[[2]] == nrows,
       times[[3]]=(bincenter[[2]]+step/2-ry-r(1+gtime rate))/(vy+rate radii[[i]]);, 
    vy<0 && binindex[[2]]!=1, 
       times[[4]]=(bincenter[[2]]-step/2-ry)/vy;,
    vy<0 && binindex[[2]]==1,
       times[[4]]=(bincenter[[2]]-step/2-ry+r(1+gtime rate))/(vy-rate radii[[i]]);
];
wallindex = Ordering[times][[1]];

(*debugging statement*)
If[!(And@@Positive@times), Print["negative time", times, binindex]; Print@borderC[{rx, ry}, r(1+gtime rate)]; Abort[]];

(*return event*)
{Min[times]+gtime, i, n+ wallindex}
]

Procedure

This section is where the previous sections come together. This section includes findNextEvent, ProcessEvent, and synchronize, as well as Process, which puts everything together. The first calls findNextTransfer and findNextCollision, and returns the event of lowest start time (with some complications). The second extracts the next event from the heap, processes the event, and calls findNextEvent on the particle in question and reinserts the new event into the heap.

Here, it is worth commenting on how the events are stored overall: a augmented min-heap stores tuples {te, i}, indicating the time te of particle i's next event. A list nextevents stores these events, where nextevents[[i]] is a event triple . This allows events to be inserted and the minimum time event obtained in a relatively fast time (O(log n) complexity for those operations, as well as lookup in O(1) with a association on top).

Now, let us take a look at these functions:

findNextEvent

In addition to the explanation above, findNextEvent has to account for and return an extra possibilities: In case particle i has outgrown it's own bin size, or will do so before the next collision or transfer, a p=-Infinity flag is set as the event partner to be processed later.

findNextEvent[i_] := Module[
{outgrowtime, step = size/(nrows), t, c, event,r},

r= radii[[i]];

outgrowtime = gtime+(step/2 - r(1+gtime rate))/(rate r);
If[outgrowtime < 0, outgrowtime = 0];

(*obtain transfer and collision events*)
t = findNextTransfer[i];
c = findNextCollision[i];

Which[

    (*outgrow event*)
    (outgrowtime < c[[1]]) && (outgrowtime < t[[1]]) && (nrows > 1),
       event = {outgrowtime, i, -Infinity};,

    (*next event is collision*)
    (c[[1]] < t[[1]]),
       collisionCheck[c];
       event = c;,

    (*next event is transfer*)
    (t[[1]] <= c[[1]]),
       event = t;
];

event
]

processEvent

This function processes exactly 1 event, outputting it from the heap and accessing it using nextevents. Then, the event is processed, and the next event for that particle is calculated, and inserted into the heap. Counters are set to sample data as the simulation progresses. If the event is a check (parametrized by p=Infinity), there is nothing to process and just obtain the next event. Moreover, if the event is an outgrowth (p=-Infinity), rescale the system, update the Cells (ChangeBins) and recommence the simulation.

processEvent := Module[ {te, i, j, nextevent, cevent},

(*increment counter, store data*)
nevents++;
If[Divisible[nevents, Ceiling[steps/20]], dataStore[]];

(*extract event from heap, initialize event variables*)
{te, i} = obtainMin[heap];
j = nextevents[[i]][[3]];
cevent = nextevents[[i]];

(*Divide into event cases*)
Which[
    (*event is a collision*)
    (j >= 1) && (j <= n),

       (*process collision*)
       ncollisions = ncollisions + 1;
       processCollision[{te, i, j}];

       (*make sure event was symmetric and check j*)
       If[ (nextevents[[j]][[3]] != i) || (nextevents[[j]][[1]] != gtime),
         Print["Events not symmetric", cevent, nextevents, gtime];
         Abort[],
         nextevents[[j]][[3]] = Infinity;
       ];

       (*set next event*)
       nextevent = findNextEvent[i];
       nextevents[[i]] = nextevent;
       increaseKey[i, nextevent[[1]]];

       If[te >= nextevent[[1]], Print["collision event time error"]];
    ,

    (*event is a transfer*)
    (j > n) && (j < n + 5),

       (*process transfer*)
       ntrans = ntrans + 1;
       processTransfer[{te, i, j}];

       (*set next event*)
       nextevent = findNextEvent[i];
       nextevents[[i]] = nextevent;
       increaseKey[i, nextevent[[1]]];

       If[te >= nextevent[[1]], Print["transfer event error", cevent, nextevent]; Abort[]];
    ,

    (*event is a check*)
    j == Infinity,

       (*process check*)
       nchecks = nchecks + 1;

       (*set next event*)
       nextevent = findNextEvent[i];
       nextevents[[i]] = nextevent;
       increaseKey[i, nextevent[[1]]];

       If[te > nextevent[[1]] && nextevent[[3]]!= -Infinity, 
       Print["check event error", cevent, nextevent];
       ];
    ,

    (*disk outgrew cell, decrease nrows*)  
    j == - Infinity,

       (*synchronize system*)
       gtime = te;
       synchronize[False];
       nrows = nrows - 1;

       (*need to resize system and update parameters*)  
       changeBins[nrows];
       setInitialEvents[n];

       Process[steps];
    ];
];

step 1 step2

Look at the bottom particle colliding with the wall!

Visualization

Here, I built a visualization function to display circles of different radii using Riffle. A couple results have already been shown above.

displayParticles

displayParticles[positions_, radii_, size_]:= Module[
{rifflelist, border, borderg},
border = Line@{{0,0}, {0, size}, {size, size}, {size, 0}, {0 ,0}};
borderg =  Graphics[{Thick, Red, border}];
rifflelist = Riffle[positions, radii]//Partition[#, 2]&;
Show[Graphics[{Disk[#[[1]], #[[2]]]&/@rifflelist}], borderg]
]

Analysis

Finally, a couple functions to analyze the simulation results.

packingFraction

packingFraction := radii.radii Pi (1 + rate gtime)^2/size^2

dataStore

Appends position and current radii, for visualization purposes, as well as the current packing fraction.

dataStore:=(
AppendTo[pf, packingFraction];
    nestpositions = AppendTo[nestpositions, positions];
    nestradii = AppendTo[nestradii, radii];
)

Packing Fraction as Events Progress

I attached my full implementation below, with a couple more simulations and plots, and additional documentation for the functions not explained above. There still are precision errors, and I still plan to add further analysis functions and simulations of different systems (different particle number, radii ratio and particle ratio). To optimize the pack, do play around a lot with the initialization constants (initial packing fraction, initial velocities and growth rate). If any optimizations are found, please send me them!

Attachments:
6 Replies

Another idea is to use Compile, which should give much better performance than plain Mathematica code, in particular when the code is procedural (as here) and it doesn't use any high-level functions (i.e. it's basically loops, arithmetic, and list manipulation). I might have missed something but it seemed to me that the only high-level function you used was Nearest.

I'd still be curious about how your current code performs—I just can't run it for the reason I mentioned earlier.

POSTED BY: Szabolcs Horvát

Hmn, I was unaware of Compile's better performance, I'll use it in the future. Thanks!

It looks like an interesting project!

I was hoping to play with your implementation and see how it performs. I downloaded the notebook and evaluated the "Development" section, which seemed to contain the definitions.

Then I evaluated "Numerical Variable Definition" and finally "Global Variables", which gave the error:

Nearest::near1: createDisks[{1,1,1,0.5,0.5,0.5,0.5}[0.5,1,300,0.5,0.9],1]->{1,2} is neither a list of real points nor a valid list of rules.

So I was wondering if the notebook is complete or is missing some bits that are necessary to play with the algorithm.


I was also wondering about the decision to implement the entire algorithm in Mathematica. This looks like a complex and highly procedural algorithm that relies on data structures that Mathematica does not have built-in (min-heap). It seems like exactly that thing that Mathematica does not do well with. Of course it is going to work, but probably very slowly.

Thus, I am curious about what motivated the decision to re-do everything in Mathematica instead of leaving only the data processing and visualization to it (the things it does the best), and what advantages you were able to derive from this choice. How does performance compare to the C++ version? I would expect a difference of orders of magnitude.

Typical reasons one would use Mathematica:

  • You don't know C++. I assume this does not apply since you based it on an existing C++ implementation.
  • Mathematica is good for prototyping. But you went as far as implementing a min heap, which is merely a performance optimization and goes far beyond prototyping.
  • Finally, Mathematica is very good at visualization and data analysis. However, this does not imply that the simulation needs to be in Mathematica.

I will usually write such simulations in C++ and connect them to Mathematica through LibraryLink. Thus I get the best of both worlds. My LTemplate package takes most of the pain out of LibraryLink development. It comes with detailed documentation and many examples. One of the examples shows how to implement an Ising model in C++ and visualize it in Mathematica in real-time.

I use Mathematica daily and I love it. However, I also believe in using the right tool for each job. This is why I advocate for developing more links between Mathematica and other systems.

POSTED BY: Szabolcs Horvát

Hey Szabolcs!

I hadn't obtained that error personally, so I played around until I was able to repeat it. It seems createDisks wasn't evaluated in your nb, it should just be a matter of evaluating the initialization cells.

The intention was part scientific experiment and part learning experience, to try to keep it 100% in WL (otherwise I would indeed do it in C++). Thanks for the package suggestion though, I'll definitely use it.

As for the performance, the average simulation takes a couple minutes, but is highly sensitive on initialization parameters (play around, I want to know how to make it better).

Great project and post! I'm curious about the geometry (fractal?) of the 'fault lines' your algorithm produces large-scale. I would really be interested to see any real materials that mimic those 'fault lines'. Is 3D necessary for that? (wikipedia doesn't mention a 3D algorithm -- is it feasible?)

POSTED BY: Peter Barendse

Thanks Peter! I imagine the fault lines should be strongly dependent on the region format (for now only squares, easily extendable to rectangles). The code should be generalizable to 3D without difficulty, I should try that too.

Reply to this discussion
Community posts can be styled and formatted using the Markdown syntax.
Reply Preview
Attachments
Remove
or Discard

Group Abstract Group Abstract