WorksPhysics and Astronomy
Burgers vectors from Voronoi topology
Abstract
Automated dislocation analysis is standard in atomistic simulations, yet current tools are limited to a handful of crystal structures. We introduce an algorithm without adjustable parameters, based on Voronoi topology, that identifies dislocations in a broad class of crystals. Each face shared by two cells that match the ideal crystal is taken as a step in it, and summing these steps around closed circuits of such cells yields exact Burgers vectors, without geometric templates. Tested on dislocations inserted into all 9,610 Materials Project structures whose Voronoi cells are stable with a wide margin, the algorithm finds every dislocation with its correct Burgers vector. In raw molecular-dynamics snapshots, including eleven dislocation types in -AlCu and bcc iron up to its melting point, every Burgers vector it returns is correct, although near melting the dislocations merge into one cluster. Voronoi topology thus provides a natural language in which a crystal, its defects, and their Burgers vectors are all described.
Introduction
Atomistic simulations have become a primary tool for studying dislocation nucleation, mobility, and interactions in crystalline materials. Simulations of millions of atoms are now routine, and some reach hundreds of millions [1], making automated dislocation identification essential. The Dislocation Extraction Algorithm (DXA) developed by Stukowski and coworkers [2, 3, 4, 5] established automated analysis as standard practice. DXA first classifies each atom using a variant of common neighbor analysis [6, 7, 8], comparing the pattern of bonds among its neighbors with those of ideal reference structures to identify crystalline atoms. It then maps bonds between crystalline atoms onto ideal lattice vectors through a structure template and sums these vectors around closed circuits, yielding the Burgers vector . Other approaches estimate from the Nye dislocation density tensor, computed from local lattice distortions [9, 10], or from slip vectors, the changes in the separations of neighboring atoms relative to a reference configuration [11, 12]. A recent graph-based method reconstructs dislocation loops and their Burgers-vector families from local displacement patterns in body-centered cubic (bcc), face-centered cubic (fcc), and hexagonal close-packed (hcp) crystals [13].
These methods work for a limited set of crystal structures, each structure needing its own geometric template. DXA includes built-in templates for only five structures: fcc, bcc, hcp, and cubic and hexagonal diamond [4, 14]. Extending the analysis to additional structures requires dedicated effort. For example, adapting DXA to the C14 and C15 Laves phases required modifying the neighbor analysis and restricting the analysis to a single sublattice [15]. Other studies rely on manual workarounds: running DXA only on the Sr and O sublattices of SrTiO [16], or, in ordered fcc NiAl, using DXA for the Shockley partials and separately highlighting the atoms of the antiphase boundaries, planar faults across which the ordered arrangement of the two species is shifted [17]. Studies of other complex crystals analyze dislocations one sublattice at a time, as in perovskite [18], or combine DXA with a tool written for that structure, as in the Laves phase [19]. Consequently, many crystal structures found in intermetallics, complex alloys, and minerals remain outside existing automated Burgers-vector tools, although atomistic simulations increasingly target these materials [20, 21].
The Burgers vector is a topological invariant: it depends only on how atoms are connected, not on their exact positions, so in principle it can be computed from connectivity alone. The difficulty is deciding which atoms are connected. The usual choices, a fixed number of nearest neighbors or all atoms within a cutoff distance, need a parameter tuned to each structure, and in configurations perturbed by strain or thermal motion a single misassigned pair along a Burgers circuit can corrupt the result. Continuous measures of local order face a related limit: any continuous map from many coordinates to a few numbers gives the same value to some configurations that are arbitrarily far apart [22], so no cutoff on such a value can separate structures cleanly. A discrete classification, constant on whole regions of configuration space, has no such limit. What is needed is a notion of adjacency that needs no parameter, and whose robustness is a property of the crystal.
One such notion comes from the Voronoi tessellation, which divides space into cells, one per atom, each consisting of the region closer to that atom than to any other [23]. Two atoms are neighbors if their cells share a face. The topology of a cell is the way its faces, edges, and vertices fit together, regardless of their sizes and shapes: which faces meet along each edge, and which edges meet at each vertex. The tessellation is built from the atomic positions, but its topology is discrete and parameter-free [24, 25]: it is typically unchanged by small displacements, and changes only when atoms move far enough to create or remove a face or an edge [24]. In an ideal crystal, the cells of all atoms on one crystallographic (Wyckoff) site share the same topology, and in many crystals this topology is stable against small perturbations, a property called topological stability [26, 27]. A recently released package, VoroTypes, computes the cell topologies of general crystal structures, enumerating every topology that small perturbations can produce [26, 27]. For a topologically stable crystal, the only such topology is the ideal one, so its robustness is known before any sample is analyzed.
Here we show that Burgers vectors can be read from Voronoi topology in crystals whose cells are topologically stable. Each face shared by two cells whose topologies match those of the ideal crystal is read as the corresponding step in the ideal crystal. Atoms whose cells do not match form defect clusters, and closed loops through the layer of crystalline atoms around each cluster give its Burgers vector. The only input is the ideal crystal structure, and in these crystals the result is exact. We validate the method on all 9,610 Materials Project structures whose cells are stable with a wide margin and on raw molecular-dynamics snapshots up to the melting point of iron, and apply it to complex intermetallics under active plastic deformation. Crystals whose cells are not topologically stable are more complicated, and they include the most common metals, fcc and hcp. We show that the same ideas handle them in practice, with weaker guarantees, and that they extend to -titanium, a degenerate crystal for which DXA has no template. A general treatment of such crystals is left for future work.
Results
Burgers circuits in two dimensions
Although our method targets three-dimensional crystals, its ideas are most clearly introduced in two dimensions; we return to three dimensions in the next subsection. In two dimensions, a dislocation is a point defect (Figure 1a). Far from the defect, the crystal appears defect-free, yet it cannot be mapped continuously onto an ideal crystal everywhere: traversing a closed path around the defect reveals a relative displacement defined by the Burgers vector . The standard method to measure is the Burgers circuit (Figure 1). One steps along neighboring atoms around the defect until returning to the starting atom, recording each step as its corresponding vector in the ideal crystal. In defect-free crystal, these step vectors sum to zero; around a dislocation, they yield a nonzero sum equal to . Equivalently, if one step is left out so that the others would sum to zero, the walk around the dislocation fails to close, and the step needed to close it is (Figure 1a). We traverse circuits clockwise as seen with the dislocation line, normal to the plane, pointing toward the viewer. Summed from start to finish, the steps of such a circuit give the same as the finish-to-start, right-hand (FS/RH) convention [28], whose circuit runs the opposite way.

Figure 1. Defect clusters in a triangular crystal with thermal noise. Dots represent atoms and polygons their Voronoi cells; non-hexagons show their number of sides. Each defect cluster (orange) is separated from the crystal by its boundary (dark orange outline); the crystalline cells along it form the shell (blue), through which the circuit (black) runs clockwise from the open circle to the filled one. (a) A dislocation: the circuit fails to close, and the red step that closes it is the Burgers vector . (b) A dislocation of opposite sign whose core has merged with a thermal fluctuation; the circuit returns . (c) A bond flip, in which two neighboring atoms move apart and two of their common neighbors move together; . (d) Two adjacent bond flips. The cluster looks much like that in (b), but its circuit returns .
Our method rests on three topological properties of the Burgers vector. First, it is the same for every circuit that encloses the same dislocations and passes only through defect-free crystal, whatever the circuit’s size or shape. Second, it vanishes for circuits that enclose no dislocation. Third, the vectors of several enclosed dislocations add.
In two dimensions, automating the Burgers circuit requires two decisions for atoms displaced from their ideal positions: which atoms are neighbors, and the ideal direction of each step. In a triangular crystal, Voronoi topology provides both the neighbors and the directions [29]. Two atoms are neighbors if their Voronoi cells share a side. A polygon contains little information, so the topology of an atom in two dimensions is taken to include the numbers of sides of its neighbors, in order around the cell [29]. Unlike in three dimensions, it therefore depends on the neighboring cells as well as on the cell itself. Away from defects, nearly all cells remain hexagons even at high temperatures, while a dislocation core appears as a pentagon beside a heptagon [30] (Figure 1a). Each hexagon has its own frame, a numbering of its sides by the six directions of the crystal. The side two hexagons share fixes how their frames are related (Supplementary Fig. S1a); turning every step of a loop of hexagons into a common frame and summing the ideal steps gives .
In a two-dimensional triangular crystal, an atom is classified as crystalline if its cell and the cells of all its neighbors are hexagons [29]. Non-crystalline atoms are grouped into connected *defect clusters. Each cluster is separated from the surrounding crystal by its boundary, the closed curve of Voronoi sides between the cluster’s cells and crystalline ones, and the crystalline cells along the boundary form the cluster’s shell (Figure 1a). Following the boundary once around, stepping through the shell cells beside it, gives the Burgers circuit, which returns the cluster’s net Burgers vector. A cluster without holes has a single boundary, and so a single circuit: nothing is searched for or chosen. Small defect clusters also appear at finite temperature, such as the pentagon–heptagon–pentagon–heptagon cluster created by a single bond flip; their circuits return , so they are not counted as dislocations (Figure 1c). The circuit does not depend on what the cluster looks like: a dislocation core merged with a thermal fluctuation still returns its Burgers vector (Figure 1b; Supplementary Movie 1), while a cluster of fluctuations that appears disordered returns zero (Figure 1d).
Most of this framework extends directly to general crystal structures and to three dimensions: two atoms are neighbors if their Voronoi cells share a face, each shared face relates the frames of the two cells, and Burgers vectors are evaluated along the boundary between each defect cluster and the surrounding crystal, through the shell of crystalline cells beside it. Two aspects require generalization. First, whereas the triangular crystal has a single type of cell, with frames related by rotations of a hexagon, atoms on distinct Wyckoff sites generally have different cells; each site therefore requires its own reference cell, and frames are related by the symmetries of the crystal. Second, in three dimensions the boundary of a defect cluster is a surface rather than a curve, and a surface with several handles can enclose a dislocation network whose branches have different Burgers vectors.
Three dimensions: defects, shells, and Burgers vectors
In three dimensions, a dislocation is a line, and its type depends on the angle between the line and : edge if they are perpendicular, screw if parallel, and mixed at any angle in between. For a perfect dislocation, is a lattice vector, a translation that maps the ideal crystal onto itself. Perfect dislocations may dissociate into partial dislocations, whose Burgers vectors are not lattice vectors, joined by a ribbon of stacking fault, a planar defect that interrupts the stacking sequence of layers.
We now take the two changes in turn. First, the reference. For each Wyckoff site of the ideal crystal, the Voronoi cell of one atom serves as a reference cell, and each of its faces is labeled by the ideal vector to the neighbor across it [26]. A sample cell whose topology matches that of a reference cell receives these labels through a one-to-one correspondence of faces, edges, and vertices that preserves how they meet (a combinatorial isomorphism). We call this correspondence a labeling, and it gives the cell its local frame. A labeling assigns ideal positions to a cell’s atom and its neighbors. Two neighboring cells therefore both assign positions to themselves, to each other, and to their common neighbors. If one symmetry of the crystal maps the first cell’s positions of these atoms onto the second cell’s, the two labelings agree, and this symmetry is the cells’ frame relation. A cell is crystalline, as in earlier work, if its topology matches that of a reference cell [24, 25]. A Burgers circuit needs slightly more: each step from one crystalline cell to the next needs a frame relation, which exists only if the two cells’ labelings can agree. A crystalline cell with no labeling that agrees with those of its crystalline neighbors, for example because they are not arranged around it as in the crystal, cannot be joined to them by frame relations. For the purpose of constructing Burgers circuits, it is therefore treated as a defect. Composing frame relations around a closed loop of crystalline cells gives a net rotation and a net translation. Around a dislocation the net rotation is the identity, and the net translation is the Burgers vector, with the FS/RH convention above. A loop closes if its net rotation is the identity and its net translation is zero. In topologically stable crystals, where every Voronoi vertex is shared by exactly four cells, as in bcc metals, B2 alloys, and Laves and phases, every loop of crystalline cells that can be shrunk to a point through crystalline cells closes (Supplementary Section S1.3). Any two loops that can be deformed into one another through crystalline cells therefore give the same Burgers vector.
Second, the boundary. As in two dimensions, non-crystalline atoms are grouped, now through shared faces, into defect clusters. The boundary of a cluster, the Voronoi faces between its cells and crystalline ones, is a closed surface, and the crystalline cells along it form the cluster’s shell, one atom thick (Supplementary Figure S4). The boundary is still unique, but a surface has many closed curves, so the circuit is no longer determined by the cluster alone. What matters is how a curve winds around the surface, which depends on its genus , its number of handles. The boundary need not be connected, and for each of its connected pieces, follows from Euler’s formula, , where , , and are the numbers of vertices, edges, and faces of the surface. A spherical boundary (), such as those around point defects and thermal fluctuations, returns zero, since every curve on it can be shrunk to a point. A dislocation line or loop is encased by a torus (), on which a curve once around the tube returns and a curve along the tube that does not wind around it returns zero. For a straight line through the periodic box, a curve along the tube closes only through the box’s periodic boundary, so its net translation is not a Burgers vector, and it is not used (Methods). A torus can also return zero, for example around a ring-shaped thermal cluster. Surfaces of higher genus arise, for example, around dislocation networks, and a thermal fluctuation can add a handle to any cluster.
Figure S4. A defect cluster and its shell in three dimensions. A segment of an edge dislocation in bcc Fe: the Voronoi cells of its defect cluster (yellow) and, along the left half, its shell of crystalline cells (blue), each with the topology of the truncated octahedron, the Voronoi cell of bcc. The boundary between them, made of Voronoi faces, is a tube around the dislocation.
Each curve on the boundary defines a walk through the shell cells beside it. The shell therefore contains a circuit for every curve on the boundary, and no thicker layer is needed. A curve’s vector depends only on how it winds around the boundary surface, and only a few curves are needed: standard methods for surfaces give exactly two independent curves per handle [31] (Methods). The resulting net translations are the cluster’s Burgers vectors. At a junction Burgers vector is conserved, the vector of one branch being the sum of those of the others, because a curve around one branch can be deformed continuously on the surface into the combination of curves around the others.
This construction contains no adjustable parameters. Voronoi topology identifies the defect atoms, combinatorics determines the clusters, their boundary surfaces, and the curves on them, and exact frame relations yield the Burgers vectors. The vectors are exact under two conditions on the crystal: (1) it is topologically stable, and (2) it is not topologically pseudo-symmetric, that is, its Voronoi tessellation is not combinatorially more symmetric than the crystal. Crystals that fail the first are degenerate: some Voronoi vertices are shared by more than four cells, and the topology of the cells around them changes under arbitrarily small perturbations, as in fcc and hcp metals. We treat them below, with weaker guarantees. The simplest example of the second is an oblique two-dimensional lattice, whose cells are all hexagons, as in the more symmetric triangular lattice (Supplementary Figure S3). In such crystals topology alone can give more than one inequivalent Burgers vector, and the method chooses by geometry (Methods).

Figure S3. Topology alone cannot determine the Burgers vector in an oblique crystal. All cells are hexagons, so rotating every labeling by one face keeps them consistent. (a) An edge dislocation in a crystal with noise: its defect cluster (gray), shell (blue), and circuit, from the open circle. (b) The circuit’s steps read with the true labeling, drawn in the ideal crystal. Their sum, the red arrow from the open circle to the filled one, is . (c) Read with the rotated labeling, they give , also a lattice vector. Figure 1 and Figure 2 show the complementary picture: a circuit that is open in the sample but would close in the ideal crystal, with the red step closing it in the sample.
Validation
We validated the method on samples with known dislocations, inserted into ideal crystals. Some samples were analyzed directly, after small random displacements of the atoms; others were first evolved by molecular dynamics (MD) under conditions in which the dislocations neither annihilated nor combined. In both cases the true Burgers vectors are the inserted ones. Neighboring dislocations may still share one cluster, when the non-crystalline regions around them touch, and such a cluster returns their vectors and their sums. A test counts as wrong if any vector found is none of these.
The Materials Project. We quantify how far a crystal’s Voronoi topology is from degenerate by its Voronoi stability gap , which VoroTypes reports as the minimum relative gap [26, 27]. In three dimensions, each Voronoi vertex is the center of a sphere that passes through four or more atoms and contains none inside (in two dimensions, Supplementary Fig. S2). In a degenerate crystal, some sphere passes through more than four atoms, and . Otherwise, at every vertex, the next nearest atom lies at least a fraction farther away than the four on the sphere, so atoms must move by a distance of order times the nearest-neighbor distance before any cell can change its topology. We computed for all 163,211 structures of the Materials Project [32]. Most (71%) have (above , the rounding of the stored coordinates), but their gaps are mostly tiny, with a median of ; the other 29% are degenerate. A subset of 9,610 structures (5.9%) has (Supplementary Table S1), so every Voronoi cell is stable by a wide margin, comparable to the stable crystals we test by MD below ( for bcc and B2, 0.50 for C15, and 0.29 for -AlCu at the potential’s lattice parameters). These 9,610 structures span all seven crystal systems and 103 space groups, have 1 to 57 Wyckoff sites, and fall into 58 families of cell topologies (distinct sets of reference-cell topologies) and 242 distinct geometries (structures alike up to scale when species are ignored); 2,302 are experimentally observed. About two thirds of them (6,353, in 29 geometries) are ordered compounds whose atoms, ignoring species, form a bcc arrangement, such as Heusler and CsCl-type phases; the other 3,257, in 97 space groups and 213 geometries, include 1,463 experimentally observed structures and such intermetallic types as Laves phases (C15, C14, C36), A15, CaCu, -AlCu (C16), and the and phases. The full list, with each structure’s identifier, gap, and results, is available with the data (Data availability).
Each structure was first snapped onto exact positions of its declared space group (Methods). We then inserted dislocation dipoles (pairs of dislocations with opposite Burgers vectors), each into its own fully periodic supercell (Methods). Two dipoles have their lines along the shortest lattice vector : a screw dipole, with , and a non-screw dipole, whose Burgers vector is the shortest lattice vector not parallel to ; the non-screw dipole is pure edge in 1,775 structures (18%) and mixed, at – to the line, in the rest. In 9,270 structures, each dipole gave two separate clusters, with and , in supercells about wide ( the nearest-neighbor distance; 2,028–12,312 atoms). The remaining 340 structures did so in supercells two or four times as wide. These have smaller gaps (median , against 0.61 for all 9,610), so noise and the strain around a dislocation change cell topologies farther from its core than in crystals with larger gaps (Discussion), and wider samples are needed. A third dipole, with and its line along the shortest lattice vector not parallel to , gave the same picture: 9,468 structures at and 142 in wider samples. Across all 29,360 analyses, no vector was wrong (Figure 2; Supplementary § 22). DXA ignores chemical species [14], so it can be applied to the bcc-based structures with its bcc template; in 20 of them chosen at random, it also correctly returned the inserted vectors of all 40 dipoles, two per structure. The other 3,257 lie outside its templates. Smaller gaps are discussed below.

Figure 2. Burgers circuits. Rows 1–3: crystals of the test set, with screw, mixed (30–60°), and edge dislocations chosen as in Supplementary Section S2.2, not the test’s own dipoles. Row 4: raw MD snapshots of edge dislocations in fcc Cu, where the defect cluster holds both Shockley partials and the stacking fault, and of screw dislocations in hcp Mg and -Ti. Each panel gives in the conventional cell and its length. Edge dislocations are viewed along the line, the others 30° off it. Atoms of the defect cluster are gray, and the others are colored by element. From the open circle to the filled one, the black steps run clockwise and would close in an ideal crystal. The red arrow that closes them is , up to the strain of the crystal.
Molecular dynamics. All MD samples are fully periodic supercells containing either four straight dislocations with zero net Burgers vector (a quadrupole) or a single prismatic dislocation loop in bcc Fe (Methods). Without quenching or time averaging, we analyzed 846 raw snapshots of seven crystals (Supplementary Table S2). In bcc Fe they contain edge and screw dislocations and a prismatic vacancy loop, from 300 to 1755 K (up to , with K the melting point of the potential [33], Methods). The others contain [111] screw superdislocations in B2 NiAl at 300 and 900 K, eleven combinations of Burgers vector and line in -AlCu at 300 and 500 K, and screw, edge, and mixed dislocations in C15 MgCu at 300 K. A superdislocation has a Burgers vector that preserves chemical order, here [111], twice the bcc vector [111]. In crystals with degenerate cells (next section), they contain edge dislocations in fcc Cu from 300 to 1200 K and [1] screw dislocations in hcp Mg and -Ti at 300 and 600 K.
No snapshot returned a wrong vector. In 703 of the 846, each dislocation forms its own cluster, which returns its inserted vector. In the other 143, the non-crystalline regions around neighboring
dislocations touch, so these dislocations share one defect cluster, which returns their vectors and their sums. Shared clusters occurred in 15 of the 110 snapshots at 500 K, all for the two largest Burgers vectors, ( Å) and , and in bcc Fe near melting: at in 2 of 42 snapshots, and at 0.93, 0.96, and in all 126. Near , thermal fluctuations leave a quarter to a third of the atoms classified as non-crystalline, and their clusters percolate; the box-spanning cluster returns only multiples of , from the dislocations it contains. In fcc Cu each edge dislocation dissociates into two Shockley partials, which, together with the stacking fault between them, form one cluster with the Burgers vector of the perfect dislocation, (next section); every Cu snapshot from 300 to 1200 K gave four such clusters. In Mg and -Ti every snapshot gave four separate clusters with , and in B2 NiAl four with . Analyzed instead with the species-blind bcc reference (next section), the same NiAl snapshots return the vector of the superpartials, the two dislocations into which a superdislocation dissociates, although in no snapshot were all of them in separate clusters. None was wrong.
We also ran DXA [4, 34] on the bcc Fe edge and screw samples from 300 to 1650 K (). In every snapshot it found segments with the inserted and no other vector. At it still traced separate line segments, although for the edge dislocations only about 60% of their length (Supplementary Section S2). DXA has no structure template for , C15, or -Ti.
Degenerate cells and partial dislocations
Not every crystal is topologically stable, and the most common metals are not. A general treatment of such crystals is beyond this paper. Here we show two things. First, a treatment built for fcc and hcp metals. Second, a route not specific to one structure, which we apply to -titanium.
In fcc and hcp metals, the ideal Voronoi cells are degenerate, as several of their vertices are shared by more than four cells. We call the twelve atoms whose cells share a face with an atom’s cell in the ideal crystal its full neighbors, and the six that touch it only at a degenerate vertex its partial neighbors [26]. Thermal fluctuations split the degenerate vertices randomly, so no cell keeps the ideal topology, and most gain small faces shared with partial neighbors, which a cell’s topology alone cannot always tell apart from faces shared with full neighbors. Because each degenerate vertex can split in only finitely many ways, the topologies that small perturbations can produce can be enumerated exhaustively: 2,815 for fcc and 9,556 for hcp [24, 27]. A given topology may arise from the ideal cell in more than one way, each with its own set of twelve faces belonging to full neighbors; we call each such set a topological alignment of the cell with the ideal one, or simply an alignment. The enumeration gives a table of every perturbed topology and its alignments. An alignment plays the part that a labeling plays in stable crystals, but fixes only which faces join full neighbors; the ideal vectors are assigned when these neighbors are matched to their ideal arrangement (Methods).
Evaluating cells individually leaves a choice of alignment for 10–12% of the atoms in fcc Cu and 62–64% in hcp Mg in our MD snapshots. Neighboring cells, however, must agree: a shared face joins full neighbors for both cells or for neither. Enforcing this face-by-face agreement reduces the ambiguous cells to at most 0.06% in Cu and 0.6% in Mg (Methods). Away from defects, agreement never removes the true alignment, since the true alignments of neighboring cells always agree. Near a defect core, however, a single remaining alignment can be wrong. A cell is therefore treated as a defect unless its twelve neighbors form their ideal arrangement and its frame relations with the crystalline cells beside it are crystal symmetries. Cells that fail these checks, or are left with several alignments or none, join the defect cluster. With 94–98% of the Cu atoms and 96–97% of the Mg atoms passing these checks, the method finds every dislocation in these samples with its correct Burgers vector (Figure 2; Supplementary Section S2).
Stacking-fault atoms have the full-neighbor arrangement of the other close-packed structure (hcp-type in fcc, fcc-type in hcp). Most of them fail these checks and join the defect cluster, so a dissociated dislocation and its stacking fault form a single cluster with the Burgers vector of the perfect dislocation. In fcc Cu, the and Shockley partials of an edge dislocation give one cluster with . Some atoms, 12–17% of the stacking-fault atoms and 16–37% of those at the partial cores, pass the checks with a wrong full neighbor, but these wrong alignments agree with one another and with the crystal around them, so every smallest loop still closes and every vector was correct (Supplementary Section S1.4.7). DXA, by contrast, recognizes stacking-fault atoms by pattern matching and traces partial dislocations individually [4]. Resolving partials in the present framework is left for future work.
Crystals with stable cells have their own counterpart of partial dislocations, which can be resolved by choosing the reference. In B2 NiAl, a [111] superdislocation dissociates into two superpartials joined by an antiphase boundary. The reference may include the species of the atoms or ignore them (Methods). With species, in the ordered B2 reference, a cell is treated as a defect if its neighbors, with their species, are not arranged around it as in B2. The cells along the antiphase boundary are then defects, and the two superpartials and the boundary form one cluster with [111]. Without species, in the bcc reference, the antiphase boundary is defect-free bcc, and a circuit can enclose a single superpartial and return its , although in our snapshots the superpartials were never all separated. The B2 reference cannot do this, since every circuit through defect-free B2 returns a lattice vector of B2, and , which joins a Ni site to an Al site, is not one.
The same enumeration works beyond close-packed metals. In the phase of titanium (P6/mmm, three atoms per unit cell), an atom on the Wyckoff 2d site has 11 full neighbors and six partial neighbors. Thermal motion opens each of the six contacts into a face or removes it, so at finite temperature its cell has 11 to 17 faces. Enumerating every combination gives 382 cell topologies at this site and 8 at the other, 388 in all, since two occur at both [26]; these serve directly as reference cells (Methods). In defect-free samples with noise of up to per component, 95–98% of the atoms are crystalline, and no cluster returns a vector. In all 27 dipoles inserted as in the Materials Project test, with or , both dislocations were found with their correct vectors, as were all four lines in each of the 38 MD snapshots (Figure 2; Supplementary Section S1.4.7). DXA has no template for this crystal.
Complex crystals in action
As an application, we simulated plastic deformation of two complex intermetallics: -AlCu, which lies outside the templates of existing automated tools, and C15 MgCu, for which the only automated analysis we know of is a tool written specifically for Laves phases [15]. Each sample contains atoms at 300 K in a fully periodic box, with seeded dislocation loops (Methods). Plastic flow is driven by straining the box at a constant rate, and the raw snapshots are analyzed directly (Figure 3); times are counted from the start of loading. Beyond the seeded loops, the evolution is not known in advance, so we rely on two consistency checks: each seeded loop must keep its vector until it meets another defect, and after merging, every branch of the resulting network must have the vector of the loop it came from or, where loops react, a sum of such vectors. In Figure 3, each core atom (atom of a defect cluster) is colored by the Burgers vector of its branch, read from the shortest circuits through the shell near it (Methods).
Figure 3. Dislocations in complex intermetallics during plastic flow. (a) C15 MgCu and (b) -AlCu at 0, 60, 120, and 180 ps, from periodic samples of atoms strained at 300 K. Only defect clusters with a nonzero Burgers vector are drawn, each atom colored by the vector of its branch (key at right).
In C15 MgCu, a Laves phase with two Wyckoff sites, four of the six seeded loops survive relaxation (minimization and 10 ps at 300 K before loading). For the first 151 ps each is a defect cluster of its own, whose boundary is a torus with a single Burgers vector; the four are distinct vectors of the seeded slip systems (Figure 3a). The loops then grow, and at 152 ps two of them meet: they form one cluster, on whose surface curves return the vectors of both branches and their sum, and by 167 ps all four form a single network of atoms (Figure 3a), which keeps growing. In all 181 snapshots, taken every 1 ps, the circuit around every branch returns one of the four seeded vectors. Applied with default parameters to the Mg sublattice, which forms a diamond lattice, as in LaCA [15], DXA finds all four vectors until the loops meet and three or four of them afterward, but it then traces only a fraction of the line. It also reports segments with other vectors, among them , which may be synchro-Shockley partials [35] that our analysis does not resolve (Supplementary Section S2.4). Before the loops meet, DXA’s lines follow our clusters but leave gaps and a few isolated segments with open ends. By about 180 ps the moving dislocations have left planar faults behind them. As with stacking faults in fcc metals, the faults join the defect clusters, which still return the exact total vectors of the dislocations they contain. Resolving the partial dislocations that bound the faults would require recognizing the atoms of the faults as crystal of their own structure, with frame relations to the host, so we follow C15 only up to this point (Supplementary Movie 2).
We follow -AlCu for 200 ps (Supplementary Movie 3). Of its six seeded loops, three have vectors, one has [100], and two have [1], on different planes (Figure 3b). For about 100 ps they remain six separate tori, apart from frequent brief contacts, in which two loops touch, share one cluster with the vectors of both and their sum, and separate again. They then merge, at first parting again from time to time. At 150 ps two clusters remain, and from 180 ps on all six loops form a single network (Figure 3b). Throughout, no branch with a new vector forms. In all 1,001 snapshots, taken every 0.2 ps, the circuit around every branch returns one of the seeded vectors.
To follow two loops whose vectors add, we placed two loops on intersecting planes, with vectors and that add to [100] (Methods). Where the two dislocation lines cross, the loops join during relaxation, and a circuit around the joined segment returns [100]. Circuits alone cannot tell whether the lines combine there into one core. The segment persists at 300 K throughout a 359 ps run (Supplementary Movie 4). A copy held at 150 K and then heated in steps to 600 K keeps the [100] segment in every snapshot, while its loops roughen but do not shrink (Figure 4; Supplementary Movie 5). In a second sample, each loop stops 8 Å short of the line where the two planes meet. The loops soon touch, and a segment with Burgers vector [100] forms where they touch, comes apart, and forms again. After about 80 ps it is present almost continuously (Supplementary Movie 6). In a third, with each loop 11 Å short of that line, no [100] segment formed within 50 ps.

Figure 4. Two loops in -AlCu joined along a segment with . Two loops whose dislocation lines cross, after energy minimization (0 ps) and at the end of holds at 150, 300, 450, and 600 K. Atoms of the defect cluster are colored by the Burgers vector of their branch, and crystal atoms on the two glide planes by the cell topology of their site (dark: neither ideal topology).
Discussion
We have shown that Burgers vectors can be derived from the Voronoi topology of an atomic configuration. In topologically stable crystals without pseudo-symmetry the result is exact, with no adjustable parameter. In practice, the method needs only the ideal crystal structure, as given in a crystallographic file, from which everything it uses is derived. Across the 9,610 structures tested, it returned no wrong vector and found every inserted dislocation separately, some only in wider samples. A third of them, in 97 space groups across all seven crystal systems, lie outside DXA’s templates; among them are Laves phases, A15 compounds, and phases, and CaCu- and -AlCu-type structures.
Three further differences from template methods follow from the topology. First, because circuits run on the boundary surface of a defect cluster, whatever its size, no maximum circuit length is needed. DXA, in contrast, does not find dislocations whose cores are too wide for its longest trial circuit [14]. Second, the core needs no separate criterion: we take it to be the dislocation’s defect cluster, the atoms whose cells are not crystalline or are treated as defects. DXA defines the
core by a separate criterion [14]. Nor does the core’s structure need to be recognized: two clusters that look alike can differ in what they hold (Figure 1b,Figure 1d), since that is decided by the crystal around them. Third, the output always conserves Burgers vector. Because every defect cluster is bounded by a closed surface and every vector comes from a closed curve on it, no dislocation can end inside crystal, and Burgers vector is conserved at every junction by construction. DXA’s traced lines can have open ends, as in the C15 sample above (Supplementary Fig. S9a). DXA does, however, do three things our method does not. It traces lines, with their directions and lengths, whereas we report the Burgers vectors of clusters and their atoms. It traces the two partial dislocations of a dissociated dislocation in fcc and hcp metals separately, and near melting it still traces separate lines where our clusters have merged (Supplementary § 22). The second could be added to our approach, in which stacking-fault atoms are recognized as crystal of the other stacking, with frame relations between the two, and we plan to address it in future work. The third reflects what the two methods report, as discussed below.
Degenerate crystals are handled with weaker guarantees. For fcc and hcp metals we built a specific treatment, in which each cell’s alignment is fixed by agreement between neighbors and a check of its twelve full neighbors. For -titanium, every topology that small perturbations can produce becomes a reference cell. VoroTypes enumerates these topologies automatically, without code specific to the structure, so we expect this route to extend to many, perhaps most, of the degenerate and small-gap crystals, though this remains to be tested. In both, exactness is no longer guaranteed near defects. Instead, the software checks that the smallest circuits there return zero, and every vector in our tests was correct. For fcc and hcp metals the method is built for thermal snapshots, in which thermal motion splits the degenerate vertices randomly, and does not apply to energy-minimized samples. A general treatment, tested beyond these three crystals, is left for future work.
Beyond degenerate cells, the method has three limitations. First, our guarantees are stated for fully periodic samples, the usual setting of bulk simulations, in which every defect cluster is enclosed by a closed surface in the crystal. Where a free surface cuts the crystal, it is no longer clear what a Burgers circuit is when part of it would lie outside the sample; we leave this case to future work. Second, the method requires the ideal cell topology to survive around the defects. Dislocation strain, about at distance from the line, destroys it where it exceeds , within a distance of order , so in crystals with small gaps the topological core, the region of cells classified as non-crystalline, extends far beyond the core itself, and clusters grow and merge rather than return wrong vectors. This affects many crystals: 71% of Materials Project structures have , but in most of them the gap is so small that thermal vibrations destroy the ideal cell topology. We have therefore restricted our study of stable-cell crystals to . Treating small-gap crystals as degenerate, by enumerating perturbed cell topologies and enforcing neighbor agreement, is one possible remedy, though it requires a table for each structure; we leave it, with a general treatment of degenerate cells, to a separate study. That study has a wide scope: about 93% of Materials Project structures match no fcc, hcp, bcc, or diamond arrangement even when species are ignored, most of them with small positive gaps. Third, in topologically pseudo-symmetric crystals the Voronoi tessellation is more symmetric than the crystal, as in any oblique two-dimensional lattice, whose cells are all hexagons (Supplementary Fig. S3). Topology alone then leaves one discrete choice of labeling per region of crystalline cells, which the method makes from the fit of the whole region’s geometry, reporting its margin (Methods). Pseudo-symmetry is a property of the ideal crystal, can be detected before any sample is analyzed, and affects at most about 4% of the structures we tested (Supplementary § 21).
The method needs no adjustable parameters because the three ingredients of a Burgers circuit come from a single object, the Voronoi tessellation. In particular, a circuit needs to know which atoms are neighbors, how their local frames are related, and which smallest loops must close. Template methods obtain these by separate procedures. In our framework, Voronoi neighbors share a face, so adjacency is mutual. Both cells take their frame relation from that same face, so it is consistent in both directions. And the smallest loops, around Voronoi edges, provably close in stable crystals (Supplementary § 21.3). In such crystals the three cannot disagree.
At finite temperature, “crystalline” and “defect” have no unique definition, and existing tools supply operational ones through their criteria: the neighbor patterns accepted by common neighbor analysis [6, 7] and polyhedral template matching [36], the maximum circuit length of DXA [4], or cutoffs on distances and order parameters. The Burgers vectors are fixed by how a sample was made, and every method should return them. But which atoms a method counts as core, and whether nearby dislocations appear as one cluster or as separate lines, depend on its definitions, and on these points methods can be compared only for consistency. Near melting this distinction matters. In bcc Fe at 0.93–0.99 , Voronoi topology classifies a quarter to a third of the atoms as not crystalline, and these regions join the dislocations into one cluster. The common neighbor analysis in DXA finds even more, 46% at 0.93 , yet DXA still traces separate lines there. Voronoi topology gives a definition without adjustable parameters, under which every quantity we computed is consistent: the Burgers vectors are exact, and clusters that enclose no dislocation return zero.
Conceptually, the construction can be seen as a discrete version of the classical geometric theory of defects, in which the crystal is treated as a continuum whose lattice directions define a local frame at every point, and a dislocation appears where steps along these directions fail to close around a loop [37, 38]. Here the frames are the cells’ labelings, and the relations between neighboring frames are obtained exactly from Voronoi topology. Because these relations include rotations as well as translations, the construction measures more than Burgers vectors: a net rotation around a loop would mark a disclination.
Because the method analyzes raw thermal configurations directly, without minimization or time averaging, it can be applied snapshot by snapshot to entire molecular-dynamics trajectories. Each analysis also partitions the sample into connected regions of crystalline cells, each with its own crystal frame, and the non-crystalline cells can themselves be classified by their Voronoi topology, which distinguishes point defects and core structures [24, 25]. Voronoi topology thus offers a single description of crystal structure, defects, and their Burgers vectors.
Methods
Overview
The analysis takes the atomic positions, the periodic box, and the ideal crystal structure, and proceeds in four steps. (1) The Voronoi tessellation of the sample is computed. (2) Each atom is classified as crystalline if the topology of its Voronoi cell matches that of a reference cell, and as a defect atom otherwise. Further, a crystalline atom that cannot be joined to its crystalline neighbors by frame relations is treated as a defect atom for the Burgers analysis (below). (3) Defect atoms whose cells share a face are grouped into defect clusters, and the boundary of each cluster is extracted as a closed surface, with its genus. (4) Curves on each boundary, two per handle (none on a sphere), are evaluated through the shell as compositions of frame relations, giving the cluster’s Burgers vectors. Each step works cell by cell or cluster by cluster, so the work grows in proportion to the number of atoms, except when one cluster holds much of the sample (below).
Voronoi tessellation
Voronoi tessellations are computed with voro++ [39], using its multithreaded extension [40]. For every cell we record its faces as oriented cycles of vertices, and the atom across each face. A Voronoi vertex is identified by the set of cells that meet there, so two cells sharing a face agree on its vertices without reference to coordinates.
Reference cells and labelings
The reference is built from the ideal crystal structure alone: its unit-cell vectors, fractional coordinates, and, optionally, species. For each Wyckoff site (an orbit of the space group, found with spglib [41]) the Voronoi cell of one atom serves as a reference. It stores the ideal vector to the neighbor across each face and the ideal position of each vertex. Labelings are found as combinatorial isomorphisms onto the reference cell of the same topology [42, 43]. When a cell’s topology has more symmetries than its site has in the crystal, it has several labelings. All are kept, and those incompatible with the labelings of its neighbors are pruned by arc consistency [44], a constraint-propagation method that repeatedly removes each labeling with which no remaining labeling of some neighbor agrees. A cell left with no labeling cannot be joined to some neighbor by a frame relation, so it is treated as a defect, and propagation restarts without it. The result does not depend on the order in which cells are processed (Supplementary § 21). Neighbors usually rule such extra labelings out, because their own cells do not share the extra symmetry. Only when the whole tessellation shares it, under topological pseudo-symmetry (Results), does a choice remain that topology cannot make (below). If species are given, each atom must have its site’s species.
Frame relations
The frame relation of two labeled cells and that share a face is the crystal symmetry , with in the point group of the crystal and the ideal vector from to , that maps the ideal positions of the face’s vertices given by one cell’s labeling onto those given by the other’s:
where is the ideal position that the labeling of cell gives vertex . Only ideal positions enter, so the frame relation is fixed by the two labelings, and the sample’s coordinates play no part (its computation, including a tolerance for rounding of the input structure, is described in Supplementary § 21). Loops never cross a face whose two labelings no crystal symmetry relates in this way. In a two-dimensional triangular crystal, the frame relation reduces to a rotation fixed by the numbers that two hexagons assign to their shared side (Supplementary Fig. S1a).
Choosing labelings
Where a cell still has several labelings, they are chosen by growing regions of crystalline cells from a seed, the unlabeled cell with the fewest remaining labelings. The seed’s labeling fixes the region’s frame: a different choice would rotate every frame of the region alike and leave every Burgers vector unchanged. A neighboring cell is labeled once topology decides it, that is, once all of its labelings consistent with its labeled neighbors are equivalent (related by symmetries of its site). These differ only by a symmetry of the cell’s site, which changes no Burgers vector, so we take the first. A cell whose consistent labelings are not all equivalent waits until more of its neighbors are labeled. Regions are grown through all crystalline cells, not only those around defect clusters, because agreement may have to spread from farther away to settle a cell near a defect, and no fixed distance would always suffice.
Without pseudo-symmetry, this labels a large defect-free region correctly, since the only labeling that all its cells accept is the true one, up to the frame. In practice, choices remain open only near defects, where information is missing. If every remaining cell is waiting, the one with the most labeled neighbors is decided by geometry, taking the labeling with the smallest misfit (below), and the software counts and reports such decisions. In the defect-free samples of all 9,610 structures of the Materials Project test, no such choice occurred, except in one structure that cannot be snapped onto its declared space group (Supplementary Section S1).
Under topological pseudo-symmetry the seed itself has several inequivalent labelings, which topology cannot tell apart. The region is then grown from each in turn, and the labeling whose cells fit best overall is kept, with its margin over the runner-up reported (Supplementary Section S1, with a two-dimensional example). The fit is measured by the misfit: the root-mean-square deviation of the actual neighbor vectors from the ideal ones after the best orthogonal map and uniform scaling. Of the defect-free samples of the 9,610 structures in the Materials Project test, each snapped onto its declared space group where possible, 390 (4.1%) have several such labelings (an upper bound, Supplementary Section S1). In 73 of them the runner-up’s misfit exceeds the best by less than 0.1 Å, and in 3 by less than 0.01 Å. All 9,610 were nevertheless analyzed correctly (Results).
Loops and Burgers vectors
Composing the frame relations around a closed loop of cells gives a net rotation and a net translation , , with the frame relation of Eq. (1), composed as . Around a dislocation the net rotation is the identity (no rotation), and for a loop that runs clockwise when the line points toward the viewer, is the Burgers vector of the FS/RH convention (Results). The software checks that every smallest loop of crystalline cells in the sample, around a single Voronoi edge, closes. In topologically stable crystals this is guaranteed (Supplementary Section S1.3), so a loop that does not close signals a problem with the input, such as a periodic box too short along a dislocation line, and is reported. A loop’s Burgers vector then depends only on the defects it encloses, and it is automatically a lattice vector. For topologically stable crystals it is the true Burgers vector whenever the loop can be deformed through crystalline cells into crystal far from all defects, provided the labeling agrees with the true crystal at one cell. Topology guarantees this agreement except under topological pseudo-symmetry, where the region-level fit decides it (above). Rotations are transported along frame relations, so all Burgers vectors are expressed in one crystal frame per crystalline region (connected region of crystalline cells), and vectors around different clusters can be compared directly. Evaluated in separate frames, two dislocations with the same Burgers vector could return it as and as , and a junction would not visibly conserve it. No region is singled out as the bulk. Each connected piece of a boundary surface faces a single crystalline region in stable crystals, and its curves are evaluated in that region’s frame; vectors are reported together with their region.
Defect clusters, boundary surfaces, and loops
Clusters and boundary. A defect cluster is a connected component, through shared Voronoi faces, of the atoms that are not crystalline or are treated as defects, taken over the whole sample. Its boundary surface is made of the Voronoi faces, edges, and vertices shared by atoms both in and outside it: mixed pairs (faces), triples (edges), and quadruples (vertices) of atoms. In a generic configuration (three cells at every Voronoi edge, four at every vertex), as thermal snapshots are, each mixed edge lies on exactly two mixed faces, so is a closed surface. Each connected piece of has Euler characteristic , with , , and the numbers of vertices, edges, and faces of that piece, and genus ( is always even).
Curves. On each piece, a standard construction gives curves, two per handle, from which every closed curve on the surface can be composed [31] (Supplementary Section S1). Curves that can be shrunk to a point through the crystal outside the cluster return zero. A curve, a cyclic sequence of mixed pairs with , becomes the walk through the shell atoms , with consecutive repeats removed. Consecutive shell atoms share a face, so every step has a frame relation (for close-packed crystals, see below).
Burgers vectors. Each region of crystalline cells has its own frame, so curves are combined only within one region. A curve’s winding around the periodic box is the sum of its minimum-image steps in fractional coordinates of the (possibly triclinic) box. Curves that wind around the box enter only through integer combinations with zero total winding, found by exact integer elimination. These combinations give the cluster’s Burgers vectors, which are lattice vectors (Supplementary Section S1). We report the vectors of the cluster’s shortest closed curves, up to sign, and generators of the lattice they span. The sign is left open, since a closed curve, like a dislocation line, has no preferred direction. In the MD tests with straight lines it was read from each curve’s winding around the inserted lines.
Burgers vector of each core atom. A cluster that contains a dislocation network has one vector per branch, so each atom takes the vector of the branch it lies on. If the cluster has a single vector, every atom takes it. Otherwise, each shell atom takes the net translation, up to sign, of the shortest closed walk through within the shell whose net translation is not zero and which does not wind around the box. Such a walk runs once around the branch on which lies. Where walks of the same least length give different vectors, takes all of them. Each atom of takes the vectors of the shell atoms nearest to it, with distances counted in steps between face-sharing atoms through the cluster, and again keeps all vectors that tie. In our samples, ties occur only near junctions. In drawings, an atom with several vectors takes one of them. The assignment does not depend on the choice of starting atom or the order of search. It can be computed exactly because the frames of the shell atoms can be fixed once, along a spanning tree, so that the net translation of a walk is a sum of integer vectors, one per step (Supplementary Section S1).
Circuits for figures. The Burgers circuits drawn in Figure 1 and Figure 2 are open walks through crystalline atoms whose steps, taken as ideal interatomic vectors, sum to zero; one step equal to closes each (Supplementary Section S1). They illustrate the result; every reported vector comes from the boundary surface.
Close-packed crystals
For fcc and hcp crystals, a table of cell topologies tells which faces of a cell join its full neighbors. A small perturbation of the ideal crystal changes a cell’s topology only at its degenerate vertices (the vertices shared by more than four cells, each of which can split in multiple ways; Supplementary Section S1), so we enumerate the splittings of these vertices around one atom, and all compatible combinations up to the site symmetry, as in VoroTypes [27]. For each resulting topology the table records the sets of its faces that join the twelve full neighbors of the ideal crystal, over all isomorphisms with the enumerated cells. The fcc table has 2,815 topologies and the hcp table 9,556. Supplementary Section S1.4.7 shows the same construction in two dimensions. Agreement. The table gives each cell a list of possible alignments. Two neighboring cells must agree on their shared face: it joins full neighbors for both cells or for neither. If an alignment of a cell disagrees with every remaining alignment of one of its neighbors, we discard it. We repeat this until nothing more can be discarded (arc consistency [44]). A cell whose topology is not in the table, or that has no alignment left, is treated as a defect. Its neighbors then ignore it, and agreement starts again without it. Checks. A cell left with a single alignment is treated as a defect unless its twelve full neighbors form their ideal arrangement. Each pair of the twelve that should be full neighbors of each other must be confirmed as such by at least one of the two that itself has a single alignment. Matching the twelve to this arrangement gives each its ideal vector, and hence the cell’s labeling. The cell is also treated as a defect unless its frame relation with every crystalline cell it shares a face with is a crystal symmetry. Atoms treated as defects join the defect clusters. Consecutive atoms of a boundary curve that are only partial neighbors are related through crystalline atoms that neighbor both (Supplementary Section S1).
Other degenerate crystals
For -Ti, unlike fcc and hcp, the perturbed cells serve directly as reference cells. For each site we list with VoroTypes every way its degenerate vertices can split, up to the site symmetry, and keep the resulting cells, each with the ideal neighbor vector on every face. A topology can occur with different assignments of neighbor vectors to its faces, not related by a symmetry of the site, and each is kept as a separate labeling. Of the 388 topologies (8 at the 1a site and 382 at the 2d site, two of them at both), 384 have a single labeling, three have two, and one has nine, for 399 labelings in all. Each cell of a sample is then labeled by matching it to one of these, as in a stable crystal. One difference remains. Near a partial neighbor, the labeling does not determine the ideal positions of a perturbed cell’s vertices, so frame relations are obtained not from the vertices of the shared face but from the ideal positions both labelings give to the two atoms of the face and their common neighbors.
Crystal structures
For each Voronoi vertex of the ideal crystal in three dimensions, let and be its distances to the fourth and fifth nearest atoms. The Voronoi stability gap of the crystal is the minimum of over all vertices. If , every vertex is shared by exactly four cells (the crystal is topologically stable), so every cell has a single topology, stable under small displacements. We computed from the stored coordinates (Materials Project database version 2026.09.28, all structures ordered). VoroTypes [26, 27], an independent implementation, found a single cell topology per Wyckoff site and the same gap, to three decimals, in all 9,610 structures of the test. Topology families are sets of distinct reference cell topologies; geometries were grouped with pymatgen’s structure matcher [45], ignoring species.
Each structure was first snapped onto exact positions of its Materials Project space group, at the smallest tolerance ( to 0.2 Å) at which spglib [41] finds that group, as in VoroTypes [26]. For 9 structures, no tolerance in this range recovers the declared group, and their stored coordinates were used unchanged. Snapping changes the primitive cell of 34 structures, whose files list a supercell of the crystal (for example, a seven-atom triclinic cell of a one-atom rhombohedral Nb). The analysis then sees the crystal as the database classifies it, and its results can depend on that classification, as they depend on the choice between an ordered reference and a species-blind one (Results).
Each reference was built from the snapped structure alone: its lattice was LLL-reduced [46] (reduced to a basis of short, nearly orthogonal vectors), with the shortest lattice vector as the cell vector , and written in restricted-triclinic form (the LAMMPS convention). The supercells are fully periodic, at least wide in cross-section ( or where needed; Results) and at least long, with the nearest-neighbor distance. Whether a curve wraps the box is read from its steps, each taken to the nearest periodic image, so the sample must be longer along the line than twice the longest neighbor distance. A length of satisfies this in every one of the 9,610 structures. Two dipoles were inserted: a screw dipole with , and a non-screw dipole with the shortest lattice vector not parallel to . The two lines of the non-screw dipole are separated along the edge component of , so that the cut between them contains (a cut across it would add or remove a sheet of atoms). A third dipole had its lines along that vector and . The displacement is the textbook elastic field of a straight dislocation in an isotropic medium [28], summed for the dipole over its periodic images in the cross-section. It uses a Poisson ratio of 0.3 for every structure, since only the Burgers vector it inserts matters for the test. The residual uniform gradient of the finite sum is removed, so that the sample is periodic to within 0.002 Å. Gaussian noise of standard deviation per Cartesian component was then added, to mimic thermal displacements.
Each Burgers vector is computed exactly in the frame of the crystal (Supplementary § 21). To compare it with the inserted vector, it is rotated into the laboratory frame, using the crystal’s orientation, which the software fits to the neighbor vectors, and replaced by the nearest lattice vector. This is correct as long as the rotated vector is off by less than half the shortest lattice vector, whose length is at least . The largest error in the whole test was , well within this margin. A dipole was counted as correct if some cluster returns a vector and every vector found is , as wrong if any other vector appears, and as undecided if no cluster returns a vector. Undecided dipoles, and dipoles whose two lines merged into one cluster, were analyzed again in cross-sections twice and four times as wide.
Molecular dynamics
Simulations used LAMMPS [47] with embedded-atom potentials for Fe [33], Cu [48], Mg [49], NiAl [50], and MgCu (from the database of Zhou, Johnson, and Wadley [51]), the angular-dependent potential of Apostol and Mishin for Al–Cu [52], and the spline-based modified embedded-atom potential of Hennig et al. for Ti [53]. Validation samples. Screw dislocations form a checkerboard, and edge and mixed ones the standard quadrupole, with partners on different glide planes. The displacement is the same elastic field as in the Materials Project test, summed over periodic images and superposed for the four lines. Where a cut has a climb component, the atoms it doubles are removed by an exact rule decided on the ideal crystal: atom is removed when the segment from to crosses the cut. This removes exactly the slab volume times the density of each species. The prismatic loop was built by removing a disc of three layers and closing the gap with the solid-angle displacement field [28], which jumps by exactly across the disc. Cross-sections are 59–285 Å. Samples were minimized with the box relaxed, run for 15 ps at constant temperature and zero pressure (NPT), and then, with the box fixed at its average, at constant temperature (NVT, Nosé–Hoover thermostat). After 5 ps, raw snapshots were taken every 1 ps (1.5 ps for , 3 ps for Mg). Wherever the lines of a quadrupole annihilated or merged during the run, the cell was doubled. In -Ti the lines annihilated after about 18 ps of this protocol, so we analyzed a run at constant pressure started from the minimized sample, every 1 ps from 1 to 19 ps (Supplementary Section S1.4.7). The close-packed analysis currently requires orthogonal boxes, so Cu and Mg were relaxed with the box kept orthogonal. DXA analyses used OVITO 3.17 [34] with default parameters unless stated otherwise. The Fe potential’s melting temperature, K, was found by solid–liquid coexistence in a box of atoms (Supplementary § 22).
Plastic flow (Figure 3). The samples of -AlCu (Apostol–Mishin potential) and C15 MgCu (embedded-atom potential [51]) were seeded with circular glide loops. For C15, they lie on the four planes and on , all with vectors. For , they cover the five most favorable systems identified by Andre et al. [54]: (two loops), , , , and . Each loop was inserted by the solid-angle displacement field. After energy minimization and 10 ps at 300 K (Nosé–Hoover thermostat [55, 56]), the box was deformed at a constant strain rate of for 200 ps (normal components applied to the box lengths, shear components as tilts, with atoms remapped), a total strain of 10%. The strain tensor, with normal and shear components, and the sign of each loop were chosen so that every seeded loop expands. Snapshots were written every 0.2 ps () and 1 ps (C15) and analyzed directly, without minimization or averaging, in their triclinic boxes.
Junction (Fig. 4). Two circular loops of radius 60 Å, on with and on with , were inserted in a fully periodic -AlCu box of 2,274,480 atoms by the solid-angle displacement field, tapered to zero far from the loops so that the field is periodic. The planes meet along , the line of a junction. In one sample both loops pass through the same point of this line, at the center of the box; in the others each loop stops 8 or 11 Å short of it, within its own plane. After energy minimization, the samples were run at 300 K (Nosé–Hoover) with no applied strain for 200 ps, the 11 Å sample for 50 ps. The sample with crossing lines was continued at 300 K to 359 ps. A copy of the same initial sample, minimized again, was held at 150 K for 50 ps and then heated to 300, 450, and 600 K, with 25 ps ramps and 50 ps holds. All runs kept the box fixed. Snapshots were written every 1 ps, except in the 8 and 11 Å samples (every 0.2 ps for the first 50 ps, and every 0.5 ps afterward) and in the copy heated in steps (every 0.5 ps, and every 0.25 ps during heating).
Implementation
The analysis, for crystals with stable cells and for fcc and hcp, is implemented in C++ on top of voro++, with OpenMP through the multithreaded voro++ containers [40]. A Python package provides the defect clusters, their surfaces and loops, approximate line positions for display, and a command-line tool. For degenerate crystals such as -Ti, VoroTypes [27] enumerates the perturbed cells.
Run time and memory
The stages applied to every atom take time proportional to the number of atoms , up to sorting: each cell has a bounded number of faces and labelings, and arc consistency revisits a cell only when a neighbor loses a labeling. The cluster analysis involves only the atoms in and around defect clusters. We timed the analysis, from the positions in memory to the Burgers vectors of all clusters, on molecular-dynamics snapshots tiled periodically to 1.2–79 million atoms, on four types of node, with DXA (OVITO 3.17, default parameters, with no template for -AlCu or C15) on the same samples, node, and threads (Table 1; Supplementary Section S2.5). With 16 threads the analysis takes 6–10 s per million atoms for bcc Fe and -AlCu, 8–11 s for fcc Cu, and 13–18 s for hcp Mg. For bcc Fe from 1.2 to 79 million atoms, the time grows about linearly in (fitted exponents 0.88–1.05 on the four nodes), and peak memory is 1.4–1.5 GB per million atoms, against 1.1–1.3 GB for DXA. On one thread DXA is 1.2–2.3 times faster, but our analysis gains more from threads (5–7 times from 1 to 16 threads on a Gold 6348 node, against 1.5 times for DXA), and with 16 threads it takes 0.4–0.8 times the time of DXA on bcc Fe and fcc Cu, and 0.9–1.1 times on hcp Mg. Since DXA also traces and smooths lines, these comparisons show only that the analysis is practical at the sizes where DXA is used. It is slowest when one defect cluster holds much of the sample: in the C15 network (12% of the atoms not crystalline), the cluster analysis, in Python, takes about half the time, and the time per atom about doubles from 1.2 to 9.7 million atoms on Gold nodes.
| sample | atoms () | time (s) ours | time (s) DXA | memory (GB) ours | memory (GB) DXA |
| bcc Fe, prismatic loops | 1.24 | 7.8 | 15.1 | 1.9 | 1.6 |
| 9.90 | 66.5 | 113.4 | 13.7 | 11.1 | |
| 79.17 | 591.8 | 915.3 | 108.7 | 84.9 | |
| bcc Fe, screw quadrupole | 1.55 | 9.7 | 18.6 | 2.3 | 2.0 |
| bcc Fe, edge quadrupole | 2.12 | 14.1 | 25.7 | 3.0 | 2.6 |
| -AlCu, junction | 2.27 | 14.7 | – | 3.4 | – |
| 18.20 | 129.0 | – | 24.7 | – | |
| C15 MgCu, network | 1.22 | 15.0 | – | 2.0 | – |
| 9.73 | 264.1 | – | 19.7 | – | |
| fcc Cu, dissociated edges | 1.94 | 17.9 | 27.1 | 3.3 | 2.5 |
| 15.53 | 145.6 | 198.8 | 22.5 | 17.7 | |
| hcp Mg, [1] screws | 1.32 | 18.1 | 18.0 | 2.8 | 1.7 |
| 10.58 | 150.6 | 135.1 | 16.6 | 12.1 |
Table 1. Run time and peak memory for one snapshot, with 16 threads on a Xeon Gold 6348 node.
Use of AI tools
Anthropic’s Claude Code, a coding assistant based on a large language model, was used extensively throughout this work, in algorithm development, code implementation, computational verification, and manuscript drafting. The author directed the work, checked the results, and takes full responsibility for the content.
Code availability
The analysis is part of VoroTop 1.2 (BSD 3-clause license, https://github.com/VoroTop/VoroTop). The scripts used for this paper are archived with the data.
Data availability
The 9,610 tested structures (Materials Project identifiers, gaps, families, and results), all molecular-dynamics inputs and initial configurations, the analysis outputs of every snapshot, and the snapshots shown in the figures are on Zenodo (https://doi.org/10.5281/zenodo.24302060).
Acknowledgments
This paper is dedicated to the memory of Robert V. Kohn (1953–2026) and Frank H. Lutz (1968–2023).
The author acknowledges support from the Israel Science Foundation under grant 202/24.
Author contributions
E.A.L. conceived the study, carried out the research, and wrote the manuscript.
Competing interests
The author declares no competing interests.
Supplementary Information
Burgers vectors from Voronoi topology
Algorithmic details
Frame relations in two dimensions
In a triangular crystal, each hexagon numbers its six sides from 1 to 6, starting from any side and turning in the same sense in every cell (clockwise in Figure S1). This numbering is the cell’s frame. The neighbor across side lies in direction , so the frame also numbers the six directions to the neighbors. Suppose cells and share a side that numbers and numbers . Then direction of points back along direction of , so it is ’s direction , and adding (mod 6) to the side numbers of gives those of . This rotation, together with the ideal step in direction of , is the frame relation of the two cells. Supplementary Movie 1 follows these steps around a whole circuit. The three-dimensional frame relations generalize this rule: faces take the place of sides, and crystal symmetries take the place of rotations by sixths of a turn (Figure S1).

Figure S1. Frame relations in two and three dimensions. (a) Two neighboring hexagons and in a triangular crystal with thermal noise number their sides from arbitrary starts. Their shared side (red) is side of and side of , so adding (mod 6) to the side numbers of gives those of . (b) The same in three dimensions, for two neighboring cells of a bcc crystal, truncated octahedra drawn apart. Each cell’s frame is three numbered axes, and the two cells number the same directions differently. Their shared hexagonal face (red) fixes the crystal symmetry relating the two frames, here a third of a turn about the line through the atoms.
The Voronoi stability gap
Figure S2 illustrates the stability gap (Methods) in two dimensions, where each Voronoi vertex is the center of a circle through three or more atoms, with none inside, and and take the place of and . The triangular crystal has . The square crystal, whose vertices are each shared by four cells (degenerate), has .

Figure S2. The Voronoi stability gap in two dimensions. Each panel shows a Voronoi vertex (open circle), its nearest atoms (blue), and the circle through them. (a) Triangular crystal: three atoms lie on the circle, and the next ones (red) are twice as far, so . (b) Square crystal: four atoms lie on the circle, so four cells meet at the vertex and . An arbitrarily small displacement changes which cells meet there.
Why the vectors are exact
In a topologically stable crystal without topological pseudo-symmetry, four facts make the Burgers vectors exact. All four rest on one elementary fact: a crystal symmetry, like any isometry, is fixed by where it sends four points not all in one plane, so the only crystal symmetry that fixes four such points is the identity.
Frame relations are unique. Let cells and share a face. Each edge of that face is shared with a third cell, whose atom is a neighbor of both and , so both labelings give it an ideal position. These atoms, at least three, surround the line through and , so together with and they include four points not all in one plane. At most one crystal symmetry therefore carries the positions given by onto those given by . Once both cells are labeled, their frame relation is fixed, and the frame relation from to is its inverse.
Loops around a Voronoi edge close. Let crystalline cells , , and share a Voronoi edge, whose two ends are vertices shared with a fourth cell, at one end and at the other. In a topologically stable crystal, any two of the four cells at a vertex share a face, so each of , , and gives ideal positions to all five atoms. Each frame relation carries one cell’s positions of these atoms onto the next cell’s, so the composition of the three around the edge carries the positions given by back onto themselves. Since and lie on opposite sides of the plane through , , and , the five atoms are not all in one plane, and the composition is the identity: no rotation and no translation. The argument treats the five atoms as distinct. In a periodic box shorter than about three neighbor distances, a triangle of neighbors can instead wrap around the box, and the software reports such loops (Methods).
A loop’s vector depends only on how it winds. A step from one cell to a neighbor and straight back contributes nothing, since the two frame relations are inverses. Moving a loop across a Voronoi edge whose three cells are crystalline inserts or removes the smallest loop around that edge, which also contributes nothing. Two loops that can be deformed into each other through crystalline cells therefore return the same vector, and a loop that can be shrunk to a point returns zero. When the net rotations are the identity, as around dislocations, the net translation of two loops joined at a common cell is the sum of their net translations. On the boundary of a defect cluster, a closed surface with handles, every closed curve can be deformed into an integer combination of basic curves, two for each handle [31], so their vectors determine those of all the others.
Vectors add at junctions. Where dislocation lines meet, a curve around one branch can be deformed on the boundary into the sum of curves around the other branches, all oriented the same way along the lines. By the additivity above, its vector is the sum of theirs, so Burgers vector is conserved without being imposed.
Far from defects, without pseudo-symmetry, every labeling is the true one up to one symmetry for the whole region. A loop that can be enlarged through crystalline cells into such crystal therefore returns the true Burgers vector.
Implementation details
Reference cells
Reference cells (Methods) are taken from the Voronoi tessellation of a large finite cluster of the ideal crystal, from one atom near its center for each Wyckoff site. spglib also gives the point group , so the input need not declare its symmetry, and its coordinates may be rounded. Relative to that atom, each face stores the ideal vector to the neighbor across it, and each vertex its ideal position , the center of the sphere through the four atoms whose cells meet there.
For each site we compare two counts: , the number of elements of that map the atom’s neighbors, with their species, onto themselves, and , the number of combinatorial isomorphisms of the reference cell onto itself or onto its mirror image (all face cycles reversed). Each site symmetry gives such an isomorphism, so . We call a site rigid if the two counts are equal and no other site has a cell with the same number of faces with each number of edges. Any two labelings of a cell onto the reference of a rigid site then differ by a site symmetry, so the cell has only one labeling, up to symmetry. A site that is not rigid does not imply pseudo-symmetry: its cell alone allows several labelings, but its neighbors usually leave one (§ 21.4.4).
Labelings
A cell with a vertex shared by more than four cells receives no labeling here (degenerate cells are treated in Section S1.4.7). Each sample cell is compared with every reference cell and its mirror image, and each combinatorial isomorphism found is a labeling of the cell. To save time, only references with the same number of faces with each number of edges are tried. Isomorphisms are found by a standard search: one directed edge of the cell is sent to each directed edge of the reference in turn, and the map is extended face by face [42]. If species are given, the atom must have the species of the reference site. For a rigid site, all labelings are equivalent, so the search stops at the first one. For any other site, each face of the reference is also tagged with the number of faces with each number of edges of the neighbor’s cell, and with the neighbor’s species if species are given, and an isomorphism must preserve these tags. This removes some isomorphisms that are symmetries of the polyhedron but not of the atom’s surroundings in the crystal. If the tags leave only site symmetries, one labeling again suffices. Otherwise all labelings are kept, and agreement with neighbors removes the rest (§ 21.4.4). Through a labeling, each face of a sample cell inherits an ideal vector , and each vertex an ideal position . We write for when is the face that cell shares with cell .
Frame relations
To find the frame relation of Eq. (1) (Methods), with , we solve that equation at three vertices of the shared face, take the nearest element of , and accept it only if the equation holds at every vertex, in both directions. Only ideal positions enter, so the tolerance needs only to absorb the rounding of the input: of , the cube root of the average volume per atom, for a structure snapped onto its space group (Methods), and 0.2% otherwise.
Fixing the labelings
Arc consistency, seeds, waiting cells, and the misfit are as in Methods. Arc consistency runs in rounds, each deciding every cell from the labelings left at its start, so of two neighbors that cannot both be right, neither is kept, and the result does not depend on the order in which cells are processed. In a growing region, cells are taken in order of how many labeled neighbors they have.
Topological pseudo-symmetry. When the Voronoi tessellation of the crystal is more symmetric than the crystal (Results), the seed has several inequivalent labelings that each extend over the region, and topology cannot decide between them. The region is then grown from each, and we keep the one with the smallest mean misfit over the region’s cells.
The simplest example is two-dimensional (Figure S3). Every Voronoi cell of a generic oblique lattice is a hexagon, so its tessellation has the topology of a honeycomb. That topology has 12 symmetries, six times the 2 of the oblique lattice. Labeling one cell determines the labelings of all others, so a region has six labelings. They differ by which actual neighbor vector each ideal one is assigned to. For example, the labeling rotated by one face reads each step as , and as . A loop then returns a lattice vector, but not the right one (Figure S3). Geometry then tells the labelings apart, since only the true one maps the ideal neighbor vectors close to the actual ones. How clearly it does so depends on the lattice, and here pseudo-symmetry and the gap pull in opposite directions. A lattice close to rectangular has a short pair of sides and a small gap, but its labelings are easily told apart. A lattice close to triangular has a large gap, but its labelings fit almost equally well, and at exactly triangular they become equivalent. Without geometry, only a known orientation of the crystal could tell them apart. The space group, for example from a CIF file, does not help: it constrains only the frame relations, and these are the same in every labeling. In three dimensions the same happens in a slightly sheared bcc crystal, whose cells are still truncated octahedra.
Pseudo-symmetry is a property of the ideal crystal and can be detected from it alone, by counting the inequivalent labelings of a crystalline region and comparing their misfits. We did this for all 9,610 structures of the Materials Project test, each snapped onto its declared space group as in the test (all but 9, Methods). Each sample was a periodic block of unit cells without noise, with chosen to give at least about 600 atoms. In every sample all atoms were crystalline and formed a single region. In 9,220 structures (95.9%) the seed has a single labeling up to symmetry. In the other 390 (4.1%) it has several, most often two to four. These counts are upper bounds, because arc consistency, which we use to list the labelings, also admits some that do not extend over the whole region. Without noise the best labeling fits exactly. The runner-up has a misfit below 0.01 Å in 3 structures, 0.01–0.1 Å in 70, 0.1–0.3 Å in 128, and above 0.3 Å in 189. For each sample, the software reports the smallest margin, its fit, between the best labeling and the runner-up over its regions, and warns when it is below 0.01 Å.
The one structure in which the misfit had to decide a cell (Methods), CaCeHg, was analyzed as given, because no tolerance reproduces its declared space group (Supplementary § 22.1). Its reference therefore has a lower symmetry than the declared one, and atoms that the declared symmetry makes equivalent have nearly identical cells, which the frame relations cannot tell apart.
Loops and their net translations
Loops, their net rotation and net translation , and the sign convention for are as in Methods. For every Voronoi edge whose three cells are crystalline, the software checks that the loop around it closes, as it must in topologically stable crystals (Results).
Each cell’s frame is related to that of its region’s seed by composing frame relations along any path of crystalline cells. When no loop of the region has a net rotation, as only a disclination would, the result does not depend on the path. All Burgers vectors of a region are thus expressed in one crystal frame, that of the seed up to a symmetry of the crystal. Among these equivalent settings, the software reports the one whose axes lie closest to an orientation supplied by the user, by default the laboratory axes.
Boundary surfaces, loops, and per-atom vectors
As in Methods, a defect cluster is a connected set of atoms that are not crystalline or are treated as defects, and its boundary surface is made of the Voronoi faces it shares with the crystal (Figure S4). Each handle of carries two closed curves, each curve becomes a walk through the shell atoms beside it, and the net translations of these walks give the cluster’s Burgers vectors. This section adds details of each step.
Surface. In a configuration that is not generic, as in some samples without noise, need not be a closed surface. The cluster is then flagged in the output.
Curves. The curves come from the tree–cotree construction [31]. We take a spanning tree of the faces, joining two faces that share an edge, and then a spanning tree of the vertices through the edges that the first tree does not use. Exactly edges are in neither tree. Each joins two faces, and with the path between them in the first tree it forms one curve. A sphere has no such curves, so a cluster enclosed only by spheres, such as a thermal fluctuation or a point defect, has no Burgers vector.
Walks. Consecutive faces of a curve often border the same shell atom, so a walk can be shorter than its curve. A walk that visits only one or two atoms returns the identity. For example, a ring-shaped cluster around a single crystalline atom has a curve that runs around the inside of the ring, beside that atom alone. In close-packed crystals, consecutive shell atoms that are partial neighbors, at opposite corners of a split octahedral hole, are related through the hole’s four other atoms. A step that still has no frame relation is replaced by a two-step detour through a common full neighbor outside the cluster, and all such detours must give the same frame relation, or the curve is not used.
Content. In close-packed crystals, regions grow only through full neighbors, while consecutive shell atoms may be partial neighbors, so a component of could face more than one region. Its curves would then not be combined, and the software reports it. None did in our samples. The integer combinations of Methods are found by row operations on the curves’ windings, and each is rounded to the nearest translation of the crystal.
Per-atom Burgers vectors. The shortest walk through each shell atom (Methods) is found exactly by a breadth-first search that tracks, for each atom reached, the translation accumulated on the way. This search is needed only at some atoms. If atoms and are steps apart, a walk through becomes one through with the same net translation and more steps, by walking from to and back. So if at the shortest walk with one vector is steps shorter than that with any other, every atom fewer than steps from has that vector alone. The remaining atoms, mostly near junctions, are searched one by one.
Circuits for figures. The circuits drawn in Figure 2 and Supplementary Figs. S6–S8 (Methods) run once clockwise around the core through crystalline atoms, from a start (open circle) to an end (filled circle). By convention, the end is an atom of the shell walk that returned , which also fixes the sign of , and the start is the crystalline atom that lies from the end in the ideal crystal. Of all such circuits, over every choice of end, we draw a shortest one in total length.
Degenerate cells
Why degenerate cells cannot be read one at a time. In a degenerate crystal (such as fcc and hcp, but not bcc), a small displacement changes the topology of the cells, so a cell in a finite-temperature sample need not have the ideal topology. The square crystal shows this in its simplest form (Figure S5, Figure S5). Each ideal cell is a square, and four cells meet at each corner. Its full neighbors are the four atoms across its sides, and its partial neighbors the diagonal ones. A small displacement splits each corner into two vertices joined by a short side, a partial side, which makes one diagonal pair of the four cells neighbors. A cell gains such a side at each corner where it belongs to that pair, so a cell at finite temperature has four to eight sides, and its polygon does not tell which four are the sides of the ideal square. In the limit of small noise only about 4.4% of the cells remain squares [57]. The statistics of chains, blocks, and clusters of such cells have also been derived [58]. Geometry separates the two kinds of sides: a partial side is about as long as the noise, and a side shared with a full neighbor about . A rule based on lengths would need a threshold, however, and we use topology instead. In fcc and hcp the problem is worse. Each ideal cell has vertices where four of its own faces meet (a generic cell has three faces at every vertex), so these vertices always split and no cell keeps the ideal topology. In our samples with thermal-like noise only 5–7% of the cells have exactly twelve faces, one per full neighbor, and many admit more than one choice of full-neighbor faces.

Figure S5. Two degenerate crystals in the plane, ideal and with Gaussian noise of 0.03 per component, each atom colored by the number of sides of its cell (key at right). (a, b) The square crystal. Every vertex is shared by four cells, and with noise each splits into two vertices joined by a short partial side. (c, d) The Lieb crystal, with square and hexagonal cells. Only the vertices shared by four hexagons are degenerate, so the square cells remain squares.
A table of cell topologies. The table of Methods is built as follows for fcc. The fcc cell has six degenerate (octahedral) vertices. Seen from the central cell, each can split into an edge (in one of two orientations), a triangular face (in one of four positions), or a quadrilateral face (in one way): three kinds of splitting and seven splittings in all. The enumeration of these splittings over all six degenerate vertices gives 2,815 topologies for fcc, and the same enumeration gives 9,556 for hcp. Eight fcc and 223 hcp topologies have more than one alignment. Read one cell at a time, the table leaves more than one alignment for 10–12% of the cells in our Cu snapshots and 62–64% in Mg (all 84 snapshots of Supplementary Section S2).
Neighbor agreement. Away from defects the true alignments of neighboring cells always agree, so agreement (Methods) never removes them, and a cell left with a single alignment has its true full neighbors. In the Cu and Mg snapshots, agreement leaves more than one alignment for at most 0.06% of the Cu cells and 0.6% of the Mg cells. Cells left with no alignment make up 0.5–1.0%, and cells whose topology is not in the table 0.2–1.3%, rising with temperature. Near a dislocation, agreement may leave a strip of cells along the glide plane undecided, because alternative alignments there place the core at different positions along the plane. These cells join the defect cluster, which therefore extends along the glide plane. The Burgers vector, returned by a circuit through the crystalline cells around the cluster, is unaffected.
Two-dimensional illustration. The same construction applies in the plane. As in the main text, the topology of a two-dimensional cell is its polygon together with the numbers of sides of its neighbors, in a canonical cyclic order [29]. An alignment is then a set of sides that join full neighbors. In the square crystal, the topology of an atom’s cell depends on the 16 vertices of the block of cells centered on it, all of them degenerate. All splittings of these vertices occur under small perturbations, and they give 1,261 topologies, 322 of which allow more than one alignment. The Lieb crystal (Figure S5, Figure S5) has atoms at the corners and edge midpoints of the squares of a square lattice. The centers of these squares, where no atom sits, are its only degenerate vertices, each shared by the cells of four edge-midpoint atoms, and the same enumeration gives 55 topologies, none of them ambiguous. The two crystals mirror fcc and hcp. In the Lieb crystal, as for most fcc cells, the table alone determines the full neighbors of every cell. In the square crystal, as in hcp Mg, it leaves about two thirds of the cells with several alignments, and agreement decides all of them. We tested both in periodic samples of unit cells, five per noise level, with Gaussian noise of – per component, where is the side of the square unit cell ( in the square crystal and in the Lieb crystal). In the square crystal, the table and agreement gave the true full neighbors of every cell. In the Lieb crystal they did so for every cell up to . At , 3.5% of the Lieb cells had also changed topology at generic vertices, so their topologies were not in the table and they were treated as defects. In neither crystal was a cell given a wrong alignment.
Labeling by the full-neighbor arrangement. In fcc and hcp, the twelve full neighbors of a cell with a single alignment, with the full-neighbor bonds among them, are matched to their ideal arrangement in fcc or hcp stacking, written in the frame of the crystal, the hcp one in both of its orientations, which differ by a half turn about the stacking axis. Matches must preserve orientation, taken from the oriented Delaunay tetrahedra. Matching gives each full neighbor its ideal vector, and hence the cell’s labeling. The hcp arrangement has more combinatorial than geometric symmetry (its basal mirror), so it can admit a match that is combinatorially allowed but geometrically wrong. Such arrangements are therefore labeled by consistency with their labeled neighbors, starting from the cells with the most labeled neighbors and growing outward. Frame relations are obtained from the common full neighbors of two cells and , , with in the point group of the crystal.
Stacking faults. Stacking-fault atoms have the full-neighbor arrangement of the other stacking. Their topologies cannot exclude them: 880 of the 2,815 fcc topologies also arise in hcp, among them the most common ones, so excluding these would remove most of the crystal. Most stacking-fault atoms fail the checks (Methods), however, so a dissociated dislocation and its stacking fault form one cluster, with the Burgers vector of the perfect dislocation. Resolving each partial would require recognizing stacking-fault atoms as crystal of the other stacking, with frame relations between the two, which we leave to future work.
Wrong alignments near cores. The checks compare arrangements, not distances. Near the core of a partial dislocation the crystal is strongly sheared, and a partial neighbor, at in the ideal crystal, can come almost as close as a full neighbor. A cell can then pass the checks with that partial neighbor in place of one of its full neighbors. To measure how often, we quenched all 80 snapshots of the Cu runs of Supplementary § 22.3 (16 each at 300, 600, 900, 1100, and 1200 K, including the two per temperature before 7 ps that Table S2 does not count) to 0 K by energy minimization. In the quenched positions, each atom’s twelve nearest atoms are its true full neighbors, and their arrangement gives its true stacking. None of the about atoms of fcc stacking that passed the checks had a wrong full neighbor. Of the stacking-fault atoms, about 6,600 per temperature, 12–17% passed the checks, all with a wrong full neighbor, with no trend in temperature. So did 637–1,511 of the 4,096 atoms per temperature at the partial cores, of neither stacking, fewer at higher temperature. In all but one of these atoms exactly one of the twelve was wrong, a partial neighbor at a median distance of 1.3 in the quenched positions. The wrongly aligned stacking-fault atoms lie mostly beside the cores: 53% have a core atom as a nearest neighbor, against 31% of all stacking-fault atoms. These wrong alignments are still consistent with one another and with the crystal around them. In all 80 snapshots every smallest loop, a triangle of crystalline atoms that are full neighbors of one another, closed: loops, of them through a wrongly aligned cell. Every Burgers vector was correct.
| crystal | dislocations | (K) | atoms | snaps. | resolved | merged | lost | wrong |
| bcc Fe | edge , 59×121 Å | 300, 900, 1200 | 16,920 | 63 | 63 | 0 | 0 | 0 |
| edge, 119×242 Å | 300, 900, 1200 | 68,400 | 63 | 63 | 0 | 0 | 0 | |
| 1500 | 68,400 | 21 | 20 | 1 | 0 | 0 | ||
| 1650 | 68,400 | 21 | 0 | 21 | 0 | 0 | ||
| 1700, 1755 | 68,400 | 42 | 0 | 42 | 0 | 0 | ||
| screw , 168×170 Å | 300, 900, 1200 | 96,768 | 63 | 63 | 0 | 0 | 0 | |
| 1500 | 96,768 | 21 | 20 | 1 | 0 | 0 | ||
| 1650 | 96,768 | 21 | 0 | 21 | 0 | 0 | ||
| 1700, 1755 | 96,768 | 42 | 0 | 42 | 0 | 0 | ||
| prismatic vacancy loop | 900, 1200 | 45,815 | 42 | 42 | 0 | 0 | 0 | |
| B2 NiAl | [111] screw, B2 reference | 300, 900 | 96,768 | 42 | 42 | 0 | 0 | 0 |
| C15 MgCu | screw | 300 | 15,120 | 21 | 21 | 0 | 0 | 0 |
| edge | 300 | 30,240 | 21 | 21 | 0 | 0 | 0 | |
| 60° mixed | 300 | 33,696 | 21 | 21 | 0 | 0 | 0 | |
| -AlCu | 11 (b, line) combinations | 300 | 32–81k | 110 | 110 | 0 | 0 | 0 |
| 500 | 32–81k | 110 | 95 | 15 | 0 | 0 | ||
| fcc Cu | edge , 102×200 Å | 300, 600, 900, 1100, 1200 | 30,336 | 70 | 70 | 0 | 0 | 0 |
| hcp Mg | screw [1], 192×188 Å | 300, 600 | 48,960 | 14 | 14 | 0 | 0 | 0 |
| -Ti | screw [1], 193×191 Å | 300, 600 | 60,480 | 38 | 38 | 0 | 0 | 0 |
| total | 846 | 703 | 143 | 0 | 0 |
Table S2. Molecular-dynamics validation in fully periodic samples, analyzed whole, with each raw snapshot counted once. Cu and Mg were analyzed with the table of cell topologies and neighbor agreement, and -Ti with its enumerated reference cells (Supplementary Section S1.4.7).
Flat faults and defect-free crystal. Faults without dislocations gave no wrong alignments. In fcc and hcp crystals of about 5,000 atoms with two flat stacking faults and Gaussian noise of 0.005–0.08 times the full-neighbor distance per component (21 samples of each), no fault atom passed the checks in fcc, and 11 of 21,168 did in hcp, all with their true full neighbors. No other atom had a wrong full neighbor. In defect-free periodic fcc and hcp crystals of 4,000 atoms with Gaussian noise up to 0.12 times the full-neighbor distance per component (three samples per level), no atom passing the checks had a wrong full neighbor, and no cluster returned a Burgers vector. In fcc every atom passed the checks up to 0.06, and at least 99.5% at 0.08. In hcp at least 99.3% passed up to 0.06, and 97.5–99.5% at 0.08.
The phase. -Ti (P6/mmm) has one atom on the 1a site and two on the 2d site, and is treated as in Methods. With the potential of Hennig et al. [53], its lattice constants are Å and Å. The cell of a 2d atom has 11 faces of positive area and touches its six partial neighbors at single points, while the cell of a 1a atom has 14 faces, which it keeps under perturbation. Combinations of splittings that are compatible pair by pair need not close into a cell, so we keep only those in which every Delaunay triangle containing the atom lies in exactly two Delaunay tetrahedra. This gives the 388 topologies [26] and 399 labelings of Methods. The degenerate 2d cells of the exactly ideal crystal match none of them and are not crystalline.
The defect-free samples of Results had 3,360 atoms, with Gaussian noise of 0.005, 0.01, 0.02, 0.03, 0.04, and 0.05 per component, and their crystalline atoms formed a single region. The three dipoles were inserted in supercells 28 wide, with nine noise realizations each, and 95–98% of the atoms were crystalline. In molecular dynamics, a [1] screw quadrupole (b = c) annihilated during equilibration with the protocol of the other crystals, after about 18 ps at 300 K, hence the early snapshots of Methods. An independent count, the winding of the atomic displacements around each line, found all four lines in every one of the 38 snapshots, in which 91.4–94.1% of the atoms were crystalline.
Validation protocol and benchmarks
Materials Project test
Selection. The gaps of Results were computed by examining every Voronoi vertex of each stored structure, so they are exact up to the rounding of the stored coordinates, about . Of the 163,211 structures, 25%, 11%, and 8% have above 0.008, 0.05, and 0.1, respectively. The 93% that match no fcc, hcp, bcc, or diamond arrangement (Discussion) were found with spglib (tolerance 0.25 Å), with all species replaced by one: 11,708 structures match one of these arrangements, and 104,187 of the other 151,503 have . Table S1 lists the 9,610 structures with by crystal system. They comprise 128 elemental structures, 2,801 binary and 6,514 ternary compounds, and 167 compounds of four or five elements. The 2,302 that are experimentally observed fall into 48 of the 58 topology families and 135 of the 242 geometries.
| crystal system | structures | observed | space groups | atoms per cell | median atoms | Wyckoff sites |
| cubic | 6,556 | 1,211 | 16 | 1–156 | 4 | 1–20 |
| tetragonal | 663 | 270 | 25 | 1–34 | 4 | 1–14 |
| orthorhombic | 903 | 133 | 29 | 1–100 | 8 | 1–32 |
| hexagonal | 817 | 518 | 10 | 3–96 | 12 | 1–30 |
| trigonal | 537 | 157 | 11 | 1–171 | 12 | 1–57 |
| monoclinic | 107 | 13 | 10 | 1–55 | 12 | 1–46 |
| triclinic | 27 | 0 | 2 | 4–55 | 18 | 3–55 |
| all | 9,610 | 2,302 | 103 | 1–171 | 4 | 1–57 |
Table S1. The 9,610 Materials Project structures of the test (), by crystal system: the numbers of structures, of those observed experimentally, and of space groups, and the numbers of atoms and of Wyckoff sites in the primitive cell as tested. All but 9 structures were snapped onto their declared space group (Methods).
Declared symmetry. Snapping onto the declared space group (Methods) mostly removes rounding, but it matters for the 34 structures whose files list a supercell of the declared crystal. Analyzed as given, such a structure asks the topology to distinguish atoms that the declared symmetry makes equivalent. The Nb of Methods (mp-aaafzwer), seven atoms in a triclinic cell, is within 0.1 Å of the declared rhombohedral crystal () with one atom per cell. As given, 16% of its atoms were not crystalline at noise 0.02 , and one dipole was never clean. Snapped, every atom was crystalline at the same noise, and every dipole was clean at 28. The 9 structures for which no tolerance up to 0.2 Å reproduces the declared space group (Methods) are DyYIr, CaCeHg, two polymorphs of MgSi, BaSrNp, YbGdZn, SrMg, and two polymorphs of Rb.
Thermal stability. Defect-free samples of the 28 screw supercell were analyzed with Gaussian noise of 0.02 and 0.03 per component. At 0.02 , at least 95% of the atoms are crystalline in every structure, and at least 99% in 9,600. At 0.03 , 9,500 structures reach 95% and 9,323 reach 99%. In no sample do the crystalline atoms form more than one crystalline region.
Dislocation character. The screw dipoles are pure screw by construction. The Burgers vector of the non-screw dipole is the shortest lattice vector not parallel to the line , so it makes an angle of at least 60° with the line: at a smaller angle, or would be a shorter lattice vector, also not parallel to . Hence the non-screw dipole is pure edge or mixed at 60–89° to the line (Results). Dislocations at about 45°, and lines along longer lattice vectors, are tested below.
Test. Each dipole was inserted into a fully periodic, restricted-triclinic supercell, and the whole sample was analyzed (Methods). We call a dipole clean if its two lines come out as two separate defect clusters with . At the first width, 28 (2,028 to 12,312 atoms, median 7,056; Results), the screw dipole was clean in 9,560 structures, the non-screw dipole in 9,281, and both in the 9,270 of Results. Of the 50 screw and 329 non-screw dipoles that were not clean, 41 and 173 had a single cluster that returned , typically because both cores lay in one cluster whose boundary surface has a handle around each line. In the other 9 and 156, no cluster returned a Burgers vector, as when both cores lie in a single tube whose curves enclose both lines and return zero. These dipoles were analyzed again in cross-sections 56 and 112 wide (8,640 to 49,248 and 42,336 to 165,888 atoms). At 56, 49 of these screw and 325 non-screw dipoles were clean, and at 112 the remaining screw dipole and 4 non-screw dipoles were clean. Both were thus clean in all 9,610 structures.
The third dipole (Methods), edge or mixed at 60–90° to , was clean in 9,468 structures at 28, in 138 more at 56, and in the last 4 at 112. Counting the wider samples, the three dipoles required the 29,360 analyses of Results: 28,830 were clean, one per dipole, and none of the other returned a wrong vector. The samples were at least long, which suffices in every structure tested (Methods). In a molecular crystal, where is a short intramolecular bond, may be too short.
Noise or strain? A dipole may need a wider sample because the noise destroys the ideal cell topology near its cores, or because the elastic strain does. To separate the two, the 340 structures whose screw or non-screw dipole needed a wider sample were analyzed again with noise of and , the samples being otherwise unchanged. Of their 680 dipoles, 301 were clean at with noise , 462 with , and 502 with . Noise thus accounts for part of the widening. Even at a quarter of the noise, however, 178 of these dipoles (26%) still needed wider samples, which suggests that strain sets their topological core: it extends to a distance of order from the line, large when is small (Discussion). At every noise level every dipole was clean at some width, and none returned a wrong vector.
Further Burgers vectors and lines. For one structure of each of the 242 geometries, we took up to three Burgers vectors, the shortest lattice vector of each distinct length up to twice the shortest (–). Each was inserted along up to three lines: along (screw), along the short lattice vector closest to perpendicular to , and along the one closest to from it. Each combination of vector and line was inserted as a dipole in a fully periodic cell whose third vector is the line, with the protocol of the main test: wide, then and while not clean, at least long, and noise . For 69 of the 2,127 combinations, 66 of them screw, no such cell could be built from short lattice vectors. Each of the other 2,058 returned the inserted vector and no other. In 2,055 the two lines formed separate clusters, at (1,585), (398), or (72). In the other three they still shared one cluster at , which returned . In two samples voro++ could not tessellate a very thin, strongly tilted box, and the same sample, twice as long along the line, was used instead.
Gallery of Burgers circuits
Supplementary Figs. S6–S8 show 16 screw, 16 mixed, and 16 edge dislocations in crystals of the Materials Project test, drawn as in Figure 2. Within each figure every panel shows a different crystal, and none of them appears in the matching row of Figure 2. These dislocations are not those of the test itself. For each crystal, we took as Burgers vectors the shortest lattice vector of each distinct length up to twice the shortest, and as lines short lattice vectors along (screw), at – to it (mixed), or at or more (edge; in these figures). Each candidate was inserted as a dipole and analyzed as in the test ( samples, widened while not clean, noise ), and every one returned its inserted vector. For each crystal we show the candidate whose circuit is roundest and smoothest seen along the line, and within a figure the panels are ordered by the width of the core.
Molecular-dynamics validation
This section gives in full the protocol summarized in Methods.
Samples. Every sample is a fully periodic box, and the analysis receives only the atoms and the box. The supercell is spanned by three lattice vectors, with the lines along the third, and is written in restricted-triclinic form. Each sample holds four lines with zero total Burgers vector, so that the box is not sheared. The displacement is the same elastic field as in the Materials Project test (Methods), summed over periodic images, superposed for the four lines, and periodic to within Å. Screw dislocations form a checkerboard, with at and and at and , in fractions of the two cell vectors that span the cross-section. Edge and mixed dislocations form the standard quadrupole, with at and and at and . Here

Figure S6. Screw dislocations in 16 crystals, viewed 30° off the line, drawn as in Figure 2: atoms of the defect cluster are gray, and the others are colored by element.
Figure S7. Mixed dislocations in 16 crystals (30–60° between and the line), viewed 30° off the line, drawn as in Figure 2: atoms of the defect cluster are gray, and the others are colored by element.

Figure S8. Edge dislocations in 16 crystals, viewed along the line, drawn as in Figure 2: atoms of the defect cluster are gray, and the others are colored by element.
the second coordinate runs along the glide-plane normal, and the cell is doubled in that direction, so that partners of opposite sign lie on different glide planes. Every line lies midway between two atomic glide planes. A cut between lines on different glide planes has a climb component and doubles a slab of atoms. The extra atoms are removed by a rule decided on the ideal crystal: atom is removed when the segment from to crosses the cut. This removes exactly the slab volume times the density of each species (in the C15 edge sample, 288 Mg and 576 Cu atoms), which a rule based on distances after displacement fails to do near cores with large and two species. The prismatic vacancy loop in Fe (, radius 20 Å) was built by removing a disc of three (111) layers and closing the gap with the solid-angle displacement field.
The cells must be wide enough that the lines neither annihilate nor merge during the run. In narrower cells the screw quadrupoles of Fe and NiAl annihilated at 900 K and above, and the lines shared clusters, so these cells of Table S2 are twice as wide in both directions of the cross-section. In Mg, screw quadrupoles annihilated in every arrangement we tried, in cells up to 200 Å wide. The winding of the atomic coordinates around each line confirms that the lines are physically gone, and the method correctly reports nothing, so Table S2 has no result for Mg.
Molecular dynamics. Samples were minimized with the box relaxed, kept orthogonal for Fe, NiAl, Cu, and Mg and allowed to tilt for and C15. The close-packed analysis requires an orthogonal box, which is why Cu and Mg keep theirs. Each sample was then run for 15 ps at constant temperature and zero pressure (NPT), and then, with the box fixed at its average over the last 5 ps, at constant temperature (NVT), with a time step of 1 fs. After 5 ps, raw snapshots were taken every 1 ps (1.5 ps for ) and analyzed without quenching or averaging, in Mg every third one and in Cu the 14 from 7 ps on. The -Ti samples followed a different protocol (Supplementary Section S1.4.7). The minimized 0 K configurations of the crystals with stable cells were also analyzed, and every one was resolved. They are not counted in Table S2. Minimized Cu, Mg, and -Ti configurations are not counted, because the cells of a minimized degenerate crystal are exactly degenerate, outside the generic geometry the construction assumes.
Classification. Each snapshot is placed in the first of four classes that applies. It is wrong if any cluster returns a vector that is not an integer multiple of the inserted . It is resolved if each inserted line is a defect cluster of its own with exactly (in Cu, a dissociated dislocation and its fault form one cluster, Section S1.4.7). It is merged if some cluster returns more than one vector, counting and as different, or a multiple of other than , and so must contain several lines. It is lost if fewer clusters return a vector than lines were inserted. The classes are read from the vectors alone: a cluster that holds two lines with the same vector returns only , and counts as one line.
Results. No vector was wrong in any of the 846 snapshots, and every snapshot was resolved except 15 of at 500 K and 128 of Fe at 1500–1755 K, which are merged (Table S2). In at 500 K these are 5 of the 10 snapshots of the [100] edge dislocation along [10] and all 10 of the [110] screw dislocation, whose Burgers vector is the largest ( Å). Every other type is resolved in all 10 snapshots at both temperatures. The 42 NiAl snapshots were also analyzed with the species-blind bcc reference, which can separate the superpartials. The two superpartials of each superdislocation often share one cluster, which returns the of each, and no snapshot was resolved: 26 were merged, 16 lost, and none wrong. In the 16 lost snapshots, fewer than eight clusters return a vector, but their boundary surfaces have at least eight handles in all, so superpartials of the same sign share clusters and are read as single lines. In none of these snapshots are all eight superpartials separate clusters.
Close to melting. At 1500 K (0.85) the lines are resolved in 40 of 42 snapshots. At 1650 K (0.93) a quarter of the atoms are not classified as crystalline even in defect-free crystal (24% in a defect-free sample, 25% in the samples with dislocations), and their clusters percolate. In all 42 snapshots the four lines then lie in one box-spanning cluster of – atoms, which returns nothing but the vectors of the dislocations it contains: its curves return (in two snapshots also ), or in one snapshot alone, with all curves of the same sign. That snapshot would read as lost, but its cluster holds all four lines, and we count it as merged. The same holds at 1700 and 1755 K (0.96 and 0.99 ), where 28–35% of the atoms are not classified as crystalline: in all 84 snapshots the four lines lie in one cluster of – atoms, whose curves return , in 14 snapshots also , and nothing else. No sample melted during the run, since the potential energy and the crystalline fraction stay constant. At these temperatures a sample analyzed whole does not resolve the lines individually, and a larger sample would not help, since the thermal clusters percolate. The melting temperature of the potential, K, was found by solid–liquid coexistence. A bcc box of atoms (20 × 20 × 80 cells), half melted, was run at constant enthalpy and zero pressure for 600 ps, and its temperature settled at 1767–1770 K with both phases present.
Close-packed crystals. Cu and Mg were analyzed with the table of cell topologies and neighbor agreement (Supplementary Section S1.4.7). Most stacking-fault atoms fail the checks of the host crystal, so each dissociated Cu dislocation, with its two Shockley partials and the fault between them, forms one cluster with the Burgers vector of the perfect dislocation, . Every Cu snapshot gave four such clusters, with with zero sum, and 94–98% of the atoms passed the checks (97–98% at 300 K, 94–96% at 1200 K). Every Mg snapshot gave four clusters with , and 96–97% of the atoms passed the checks. The other atoms form small clusters that return no vector.
Comparison with DXA
bcc Fe. We ran DXA (OVITO 3.17, default parameters, bcc input structure) on the larger bcc Fe edge quadrupole and on the screw quadrupole of Table S2, 21 snapshots at each of 300, 900, 1200, 1500, and 1650 K. In every snapshot DXA found segments with and no other Burgers vector. Up to 1500 K the total length it traced stayed close to that of the four lines, about 115 Å for the edge and 160 Å for the screw sample. At 1650 K (0.93 ) it fell to 68 Å on average for the edge sample, so parts of those lines were not traced, while the screw lines were traced almost fully (154 Å). DXA classified 99.6–99.8% of the atoms as bcc at 300 K, 89% at 1200 K, and 54% at 1650 K. At 1650 K, DXA thus still traces the four lines separately, though with gaps, while our analysis joins them into one cluster spanning the box, which returns only the vectors of the lines it contains (§ 22.3). Neither method returned a wrong vector.
C15 MgCu. In C15 the Mg atoms form a diamond lattice, and LaCA identifies dislocations by running DXA on this sublattice [15]. We followed this procedure with OVITO’s DXA, not with LaCA’s own software. DXA (OVITO 3.17, default parameters, cubic-diamond input structure) was run on the Mg atoms alone in every snapshot of the C15 plastic-flow run of the main text, every 1 ps from 0 to 180 ps (181 snapshots). Burgers vectors are given in units of the cubic cell, which C15 and its Mg sublattice share (Figure S9).
Figure S9. DXA on the Mg sublattice of C15 MgCu during plastic flow. (a) At 140 ps, before the loops meet: atoms of our defect clusters with a Burgers vector (gray), and the lines DXA traces (blue, ; green, ; red, other vectors). (b) Total length of DXA’s lines by Burgers vector: (dark), (light), and others (gray). Dotted line: the loops first meet (152 ps).
Up to 150 ps, DXA traced all four loops with their vectors, the same loops and vectors as our analysis, with 1,210–1,370 Å of line in all. In some snapshots it also reported short segments with other vectors, up to about 120 Å in all. Our clusters at this stage are closed rings, one around each loop. DXA’s lines lie inside them but have gaps (Figure S9).
The loops first meet at 152 ps. From then on DXA still finds three or four of the four vectors in every snapshot, but the total length of line with these vectors shrinks to about 260 Å at 180 ps, a fifth of its earlier length, although the loops have grown. DXA also reports segments with , from about 100 ps on, the Burgers vector of synchro-Shockley partials in Laves phases [35], which may be real partials that our analysis does not resolve. It also reports segments with vectors that are not lattice vectors of the crystal, such as , , , and , some of them from the first snapshots on. Over the same snapshots, every branch in our analysis returns one of the four seeded vectors (main text).
A dislocation line cannot end inside crystal, so open ends mark lines that were not fully traced or segments that are not dislocations. Our analysis cannot produce them (Discussion).
Run time and memory
Each sample is a periodic molecular-dynamics snapshot tiled times: bcc Fe with prismatic loops at 900 K (45,815 atoms per tile), screw and edge quadrupoles in bcc Fe, dislocated edges in fcc Cu, and [1] screws in hcp Mg, all at 300 K, the -AlCu junction at 200 ps, and the C15 network at 180 ps. Every case ran alone on a whole node, one after another (on the Platinum 8558 node, 46 of its 96 physical cores, the most its queue allows). Times run from the positions in memory to the Burgers vectors of all clusters, for ours and for DXA (OVITO 3.17, default parameters, with the bcc, fcc, or hcp input structure), and exclude reading the file. Our analysis runs as a Python process and C++ child processes, so its peak memory is the largest sum, over its processes, of their proportional set sizes, sampled every 0.1 s (for the 79-million-atom sample, their resident memory, larger by under 1%). For DXA it is the peak resident memory. The four node types are two 28-core Xeon Gold 6348 processors (2.6 GHz), one 32-core Xeon Gold 6338 (2.0 GHz), two 18-core Xeon Gold 6240 (2.6 GHz), and two 48-core Xeon Platinum 8558.
Across the four node types, times differ by up to a factor of 1.7 (Table S3), and the ratios to DXA given in Methods hold on every one. On a Gold 6348 node, from 1 to 16 threads our analysis runs 4.8–6.8 times faster, and DXA 1.5 times (Table S4). The gain is similar on the Gold 6338 and Gold 6240 nodes and smaller on the Platinum 8558 node, 2.7–4.2 times (Table S3). Beyond 16 threads the tessellation, computed with the multithreaded voro++ containers [40], became slower: for 9.9 million atoms of bcc Fe it took 14 s on 16 threads and 64 s on 56, while the other stages together became slightly faster.
| sample | atoms () | Gold 6348 ours | Gold 6348 DXA | Gold 6338 ours | Gold 6338 DXA | Gold 6240 ours | Gold 6240 DXA | Platinum 8558 ours | Platinum 8558 DXA |
| bcc Fe, prismatic loops | 1.24 | 7.8 | 15.1 | 7.9 | 19.2 | 9.7 | 16.8 | 11.7 | 17.2 |
| 2.93 | 18.8 | 34.3 | 19.3 | 43.3 | 22.9 | 38.3 | 25.3 | 35.1 | |
| 9.90 | 66.5 | 113.4 | 65.2 | 139.8 | 80.6 | 124.0 | 67.0 | 94.5 | |
| 23.46 | 163.5 | 265.0 | 162.7 | 333.8 | 196.4 | 292.0 | 138.9 | 218.0 | |
| 45.82 | 328.2 | 519.9 | 318.6 | 643.5 | 410.3 | 622.0 | 270.8 | 418.3 | |
| 79.17 | 591.8 | 915.3 | 573.9 | 1108.2 | 809.4 | 1017.3 | 483.4 | 738.8 | |
| bcc Fe, screw quadrupole | 1.55 | 9.7 | 18.6 | 9.7 | 22.8 | 11.5 | 20.4 | 12.8 | 21.8 |
| bcc Fe, edge quadrupole | 2.12 | 14.1 | 25.7 | 14.7 | 30.9 | 16.9 | 28.2 | 17.9 | 21.5 |
| -AlCu, junction | 2.27 | 14.7 | – | 15.0 | – | 19.4 | – | 18.9 | – |
| 18.20 | 129.0 | – | 139.8 | – | 162.5 | – | 114.7 | – | |
| C15 MgCu, network | 1.22 | 15.0 | – | 15.5 | – | 17.8 | – | 20.1 | – |
| 9.73 | 264.1 | – | 280.3 | – | 312.0 | – | 219.8 | – | |
| fcc Cu, dissociated edges | 1.94 | 17.9 | 27.1 | 18.6 | 32.8 | 20.8 | 29.2 | 22.2 | 28.9 |
| 15.53 | 145.6 | 198.8 | 146.6 | 246.2 | 169.1 | 218.6 | 123.8 | 166.1 | |
| hcp Mg, [1] screws | 1.32 | 18.1 | 18.0 | 18.9 | 21.8 | 20.7 | 19.9 | 23.2 | 21.2 |
| 10.58 | 150.6 | 135.1 | 157.2 | 162.2 | 172.0 | 152.5 | 133.0 | 117.6 | |
| one thread | |||||||||
| bcc Fe, prismatic loops | 1.24 | 39.6 | 22.4 | 43.7 | 27.0 | 44.8 | 24.6 | 31.6 | 27.3 |
| -AlCu, junction | 2.27 | 101.4 | – | 107.6 | – | 114.8 | – | 78.6 | – |
| fcc Cu, dissociated edges | 1.94 | 89.7 | 41.5 | 96.2 | 48.0 | 96.3 | 44.5 | 77.9 | 33.4 |
Table S3. Run time (s) for one snapshot on four node types, with 16 threads and with one thread.
| threads | ||||||||
| 1 | 2 | 4 | 8 | 16 | 28 | 56 | ||
| bcc Fe | ours | 321.4 | 189.2 | 120.0 | 83.5 | 67.6 | 76.4 | 117.4 |
| DXA | 174.4 | 141.4 | 125.2 | 120.0 | 113.1 | 111.5 | 110.1 | |
| fcc Cu | ours | 89.7 | 52.2 | 33.6 | 23.1 | 17.9 | 15.7 | 14.2 |
| DXA | 41.4 | 33.5 | 29.5 | 27.9 | 26.8 | 26.2 | 26.1 | |
| -AlCu | ours | 101.6 | 53.7 | 33.0 | 20.3 | 14.9 | 17.7 | 21.8 |
Table S4. Run time (s) against the number of threads on a second Gold 6348 node, so that times at 16 threads differ slightly from Table S3, for bcc Fe with prismatic loops (9.90 million atoms), fcc Cu with dissociated edges (1.94 million), and the -AlCu junction (2.27 million).
Supplementary Movies
The movies can be viewed and downloaded at https://www.vorotop.org/burgers-movies.html.
Supplementary Movie 1. Burgers vectors from Voronoi topology in two dimensions. A two-dimensional Lennard-Jones crystal of 440 atoms in a periodic box, heated to and held. Atoms are colored by the number of sides of their Voronoi cells, as in [29]. At the first bond flip the movie pauses and builds, step by step, its defect cluster, boundary, shell, and circuit. Each step of the circuit is drawn in the ideal crystal on the right, and the steps close. The crystal is then slowly sheared, and a bond flip splits into two dislocations. To show the pair close together, the run is continued from 0.15 time units after the split with velocities redrawn at the same temperature, and the movie shows one of 1,200 such continuations, in which one dislocation has merged with a thermal fluctuation. The circuits of the two clusters give and , and a circuit around both gives zero.
Supplementary Movie 2. C15 MgCu during plastic flow at 300 K. The sample of Figure 3a, every 1 ps from the start of straining to 180 ps. Atoms of the defect clusters with a nonzero Burgers vector are colored by the vector of their branch, as in Figure 3. Each branch is drawn whole across the periodic boundaries unless it is comparable in size to the box.
Supplementary Movie 3. -AlCu during plastic flow at 300 K. The sample of Figure 3b, every 1 ps from the start of straining to 200 ps, drawn as in Supplementary Movie 2.
Supplementary Movie 4. Crossing loops in -AlCu at 300 K, seen from all sides. The crossing loops of Figure 4, every 1 ps from 0 to 359 ps, while the camera circles the sample once, one degree per snapshot. Atoms of the defect cluster are colored by the Burgers vector of their branch, as in Figure 4, and crystal atoms are not drawn. The [100] segment is present in every snapshot.
Supplementary Movie 5. Crossing loops in -AlCu heated in steps. The sample of Figure 4, held at 150 K for 50 ps and then heated to 300, 450, and 600 K, with 25 ps ramps and 50 ps holds, every 0.5 ps (every 0.25 ps during heating). The view and colors are those of Figure 4: atoms of the defect cluster are colored by the Burgers vector of their branch, and one fixed set of crystal atoms on the two glide planes by cell topology. The [100] segment is present in every snapshot. The dark cells, matching neither ideal cell, are thermal fluctuations and return no Burgers vector.
Supplementary Movie 6. Two loops in -AlCu that join along a [100] segment at 300 K. The sample in which each loop stops 8 Å short of the line where the two glide planes meet, from 0 to 200 ps. The dislocations are drawn as closed curves inside their defect cluster, smoothed for display and colored by Burgers vector.
References
- [1]L. A. Zepeda-Ruiz, A. Stukowski, T. Oppelstrup, N. Bertin, N. R. Barton, R. Freitas, and V. V. Bulatov. Atomistic insights into metal hardening. Nat. Mater., 20:315–320, 2021. doi: 10.1038/s41563-020-00815-1.
- [2]A. Stukowski and K. Albe. Dislocation detection algorithm for atomistic simulations. Modelling Simul. Mater. Sci. Eng., 18:025016, 2010. doi: 10.1088/0965-0393/18/2/025016.DOI
- [3]A. Stukowski and K. Albe. Extracting dislocations and non-dislocation crystal defects from atomistic simulation data. Modelling Simul. Mater. Sci. Eng., 18:085001, 2010. doi: 10.1088/0965-0393/18/8/085001.DOI
- [4]A. Stukowski, V. V. Bulatov, and A. Arsenlis. Automated identification and indexing of dislocations in crystal interfaces. Modelling Simul. Mater. Sci. Eng., 20:085007, 2012. doi: 10.1088/0965-0393/20/8/085007.DOI
- [5]A. Stukowski. A triangulation-based method to identify dislocations in atomistic models. J. Mech. Phys. Solids, 70:314–319, 2014. doi: 10.1016/j.jmps.2014.06.009.DOI
- [6]J. D. Honeycutt and H. C. Andersen. Molecular dynamics study of melting and freezing of small Lennard-Jones clusters. J. Phys. Chem., 91:4950–4963, 1987. doi: 10.1021/j100303a014.DOI
- [7]D. Faken and H. Jónsson. Systematic analysis of local atomic structure combined with 3D computer graphics. Comput. Mater. Sci., 2:279–286, 1994. doi: 10.1016/0927-0256(94)90109-0.DOI
- [8]A. Stukowski. Structure identification methods for atomistic simulations of crystalline materials. Modelling Simul. Mater. Sci. Eng., 20:045021, 2012. doi: 10.1088/0965-0393/20/4/045021.DOI
- [9]C. S. Hartley and Y. Mishin. Characterization and visualization of the lattice misfit associated with dislocation cores. Acta Mater., 53:1313–1321, 2005. doi: 10.1016/j.actamat.2004.11.027.DOI
- [10]F. Z. Dai and W. Z. Zhang. An automatic and simple method for specifying dislocation features in atomistic simulations. Comput. Phys. Commun., 188:103–109, 2015. doi: 10.1016/j.cpc.2014.11.014.DOI
- [11]J. A. Zimmerman, C. L. Kelchner, P. A. Klein, J. C. Hamilton, and S. M. Foiles. Surface step effects on nanoindentation. Phys. Rev. Lett., 87:165507, 2001. doi: 10.1103/PhysRevLett.87.165507.DOI
- [12]J. Hua and A. Hartmaier. Determining Burgers vectors and geometrically necessary dislocation densities from atomistic data. Modelling Simul. Mater. Sci. Eng., 18:045007, 2010. doi: 10.1088/0965-0393/18/4/045007.
- [13]U. Bhardwaj. Savi-Bhransha: Graph-theoretic dislocation-loop characterization in crystals. arXiv:2607.29328, 2026. doi: 10.48550/arXiv.2607.29328.DOI
- [14]Dislocation analysis (DXA). OVITO 3 user manual, https://www.ovito.org/docs/current/reference/pipelines/modifiers/dislocation_analysis.html, 2026.
- [15]Z. Xie, D. Chauraud, E. Bitzek, S. Korte-Kerzel, and J. Guénolé. Laves phase crystal analysis (LaCA): Atomistic identification of lattice defects in C14 and C15 topologically close-packed phases. J. Mater. Res., 36:2010–2024, 2021. doi: 10.1557/s43578-021-00237-y.DOI
- [16]C. Okafor, A. Sayyadi-Shahraki, S. Bruns, T. Frömling, P. Hirel, P. Carrez, K. Durst, and X. Fang. Coupled electromigration-nanoindentation study on dislocation nucleation in SrTiO₃. J. Am. Ceram. Soc., 108(10):e70015, 2025. doi: 10.1111/jace.70015.DOI
- [17]X. Xu, X. Zhang, E. Bitzek, S. Schmauder, and B. Grabowski. Origin of the yield stress anomaly in L12 intermetallics unveiled with physically informed machine-learning potentials. Acta Mater., 281:120423, 2024. doi: 10.1016/j.actamat.2024.120423.
- [18]P. Hirel, A. Kraych, P. Carrez, and P. Cordier. Atomic core structure and mobility of [100](010) and [010](100) dislocations in MgSiO3 perovskite. Acta Mater., 79:117–125, 2014. doi: 10.1016/j.actamat.2014.07.001.
- [19]M. Freund, J. Spille, P.-L. Sun, M. Lipińska-Chwałek, J. Mayer, Z. Xie, and S. Korte-Kerzel. Redefining plasticity in a cubic intermetallic: Dislocation dynamics in the CaAl₂ Laves phase. Acta Mater., 303:121604, 2026. doi: 10.1016/j.actamat.2025.121604.DOI
- [20]S. Zhang, Y. Li, S. Suzuki, A. Nakamura, and S. Ogata. Neural network potential for dislocation plasticity in ceramics. npj Comput. Mater., 10:266, 2024. doi: 10.1038/s41524-024-01456-7.DOI
- [21]Q. Zhou, J. Wang, A. Misra, P. Huang, F. Wang, and K. Xu. Dislocations interaction induced structural instability in intermetallic Al₂Cu. npj Comput. Mater., 3:24, 2017. doi: 10.1038/s41524-017-0030-2.DOI
- [22]P. S. Landweber, E. A. Lazar, and N. Patel. On fiber diameters of continuous maps. Am. Math. Mon., 123(4):392–397, 2016. doi: 10.4169/amer.math.monthly.123.4.392.DOI
- [23]E. A. Lazar, J. Lu, and C. H. Rycroft. Voronoi cell analysis: The shapes of particle systems. Am. J. Phys., 90(6):469–480, 2022. doi: 10.1119/5.0087591.DOI
- [24]E. A. Lazar, J. Han, and D. J. Srolovitz. Topological framework for local structure analysis in condensed matter. Proc. Natl. Acad. Sci. USA, 112:E5769–E5776, 2015. doi: 10.1073/pnas.1505788112.DOI
- [25]E. A. Lazar. VoroTop: Voronoi cell topology visualization and analysis toolkit. Modelling Simul. Mater. Sci. Eng., 26:015011, 2018. doi: 10.1088/1361-651X/aa9a01.DOI
- [26]E. A. Lazar and C. H. Rycroft. Structural identification in general crystalline systems via Voronoi cell topology. arXiv:2610.20498, 2026. doi: 10.48550/arXiv.2610.20498.DOI
- [27]E. A. Lazar and C. H. Rycroft. VoroTypes: automated enumeration of Voronoi cell topologies for structural identification of complex crystals in atomistic simulations. arXiv:2610.21911, 2026. doi: 10.48550/arXiv.2610.21911.DOI
- [28]J. P. Hirth and J. Lothe. Theory of Dislocations. Wiley, 2nd edition, 1982.
- [29]E. A. Lazar, J. Lu, C. H. Rycroft, and D. Schwarcz. Characterizing structural features of two-dimensional particle systems through Voronoi topology. Modelling Simul. Mater. Sci. Eng., 32:085022, 2024. doi: 10.1088/1361-651X/ad8ad9.DOI
- [30]D. R. Nelson and B. I. Halperin. Dislocation-mediated melting in two dimensions. Phys. Rev. B, 19:2457–2484, 1979. doi: 10.1103/PhysRevB.19.2457.DOI
- [31]D. Eppstein. Dynamic generators of topologically embedded graphs. In Proceedings of the 14th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 599–608, 2003.
- [32]A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. A. Persson. Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. APL Mater., 1:011002, 2013. doi: 10.1063/1.4812323.DOI
- [33]M. I. Mendelev, S. Han, D. J. Srolovitz, G. J. Ackland, D. Y. Sun, and M. Asta. Development of new interatomic potentials appropriate for crystalline and liquid iron. Philos. Mag., 83:3977–3994, 2003. doi: 10.1080/14786430310001613264.DOI
- [34]A. Stukowski. Visualization and analysis of atomistic simulation data with OVITO–the Open Visualization Tool. Modelling Simul. Mater. Sci. Eng., 18:015012, 2010. doi: 10.1088/0965-0393/18/1/015012.DOI
- [35]Z. Xie, D. Chauraud, A. Atila, E. Bitzek, S. Korte-Kerzel, and J. Guénolé. Unveiling the mechanisms of motion of synchro-Shockley dislocations in Laves phases. Phys. Rev. Mater., 7:053605, 2023. doi: 10.1103/PhysRevMaterials.7.053605.DOI
- [36]P. M. Larsen, S. Schmidt, and J. Schiøtz. Robust structural identification via polyhedral template matching. Modelling Simul. Mater. Sci. Eng., 24:055007, 2016. doi: 10.1088/0965-0393/24/5/055007.DOI
- [37]B. A. Bilby, R. Bullough, and E. Smith. Continuous distributions of dislocations: a new application of the methods of non-Riemannian geometry. Proc. R. Soc. Lond. A, 231:263–273, 1955. doi: 10.1098/rspa.1955.0171.DOI
- [38]M. Kléman and J. Friedel. Disclinations, dislocations, and continuous defects: A reappraisal. Rev. Mod. Phys., 80:61–115, 2008. doi: 10.1103/RevModPhys.80.61.DOI
- [39]C. H. Rycroft. VORO++: A three-dimensional Voronoi cell library in C++. Chaos, 19:041111, 2009. doi: 10.1063/1.3215722.DOI
- [40]J. Lu, E. A. Lazar, and C. H. Rycroft. An extension to VORO++ for multithreaded computation of Voronoi cells. Comput. Phys. Commun., 291:108832, 2023. doi: 10.1016/j.cpc.2023.108832.DOI
- [41]A. Togo, K. Shinohara, and I. Tanaka. Spglib: a software library for crystal symmetry search. Sci. Technol. Adv. Mater. Methods, 4:2384822, 2024. doi: 10.1080/27660400.2024.2384822.DOI
- [42]L. Weinberg. A simple and efficient algorithm for determining isomorphism of planar triply connected graphs. IEEE Trans. Circuit Theory, 13:142–148, 1966. doi: 10.1109/TCT.1966.1082573.DOI
- [43]E. A. Lazar, J. K. Mason, R. D. MacPherson, and D. J. Srolovitz. Complete topology of cells, grains, and bubbles in three-dimensional microstructures. Phys. Rev. Lett., 109:095505, 2012. doi: 10.1103/PhysRevLett.109.095505.DOI
- [44]A. K. Mackworth. Consistency in networks of relations. Artif. Intell., 8:99–118, 1977. doi: 10.1016/0004-3702(77)90007-8.DOI
- [45]S. P. Ong, W. D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia, D. Gunter, V. L. Chevrier, K. A. Persson, and G. Ceder. Python Materials Genomics (pymatgen): A robust, open-source Python library for materials analysis. Comput. Mater. Sci., 68:314–319, 2013. doi: 10.1016/j.commatsci.2012.10.028.DOI
- [46]A. K. Lenstra, H. W. Lenstra, and L. Lovász. Factoring polynomials with rational coefficients. Math. Ann., 261:515–534, 1982. doi: 10.1007/BF01457454.DOI
- [47]A. P. Thompson, H. M. Aktulga, R. Berger, D. S. Bolintineanu, W. M. Brown, P. S. Crozier, P. J. in ’t Veld, A. Kohlmeyer, S. G. Moore, T. D. Nguyen, R. Shan, M. J. Stevens, J. Tranchida, C. Trott, and S. J. Plimpton. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Comput. Phys. Commun., 271:108171, 2022. doi: 10.1016/j.cpc.2021.108171.DOI
- [48]Y. Mishin, M. J. Mehl, D. A. Papaconstantopoulos, A. F. Voter, and J. D. Kress. Structural stability and lattice defects in copper: Ab initio, tight-binding, and embedded-atom calculations. Phys. Rev. B, 63:224106, 2001. doi: 10.1103/PhysRevB.63.224106.DOI
- [49]D. Y. Sun, M. I. Mendelev, C. A. Becker, K. Kudin, T. Haxhimali, M. Asta, J. J. Hoyt, A. Karma, and D. J. Srolovitz. Crystal-melt interfacial free energies in hcp metals: A molecular dynamics study of Mg. Phys. Rev. B, 73:024116, 2006. doi: 10.1103/PhysRevB.73.024116.DOI
- [50]Y. Mishin, M. J. Mehl, and D. A. Papaconstantopoulos. Embedded-atom potential for B2-NiAl. Phys. Rev. B, 65:224114, 2002. doi: 10.1103/PhysRevB.65.224114.DOI
- [51]X. W. Zhou, R. A. Johnson, and H. N. G. Wadley. Misfit-energy-increasing dislocations in vapor-deposited CoFe/NiFe multilayers. Phys. Rev. B, 69:144113, 2004. doi: 10.1103/PhysRevB.69.144113.DOI
- [52]F. Apostol and Y. Mishin. Interatomic potential for the Al-Cu system. Phys. Rev. B, 83:054116, 2011. doi: 10.1103/PhysRevB.83.054116.DOI
- [53]R. G. Hennig, T. J. Lenosky, D. R. Trinkle, S. P. Rudin, and J. W. Wilkins. Classical potential describes martensitic phase transformations between the α, β, and ω titanium phases. Phys. Rev. B, 78:054121, 2008. doi: 10.1103/PhysRevB.78.054121.DOI
- [54]D. Andre, Z. Xie, F. Ott, J. T. Pürstl, N. Lohrey, W. J. Clegg, S. Sandlöbes-Haut, and S. Korte-Kerzel. Dislocation-mediated plasticity in the Al₂Cu θ-phase. Acta Mater., 209:116748, 2021. doi: 10.1016/j.actamat.2021.116748.DOI
- [55]S. Nosé. A unified formulation of the constant temperature molecular dynamics methods. J. Chem. Phys., 81:511–519, 1984. doi: 10.1063/1.447334.DOI
- [56]W. G. Hoover. Canonical dynamics: Equilibrium phase-space distributions. Phys. Rev. A, 31:1695–1697, 1985. doi: 10.1103/PhysRevA.31.1695.DOI
- [57]H. Leipold, E. A. Lazar, K. A. Brakke, and D. J. Srolovitz. Statistical topology of perturbed two-dimensional lattices. J. Stat. Mech., page 043103, 2016. doi: 10.1088/1742-5468/2016/04/043103.DOI
- [58]E. A. Lazar and A. Shoan. Voronoi chains, blocks, and clusters in perturbed square lattices. J. Stat. Mech., page 103204, 2020. doi: 10.1088/1742-5468/ab b6e3.DOI


