research papers
Applications of the Clifford torus to material textures
^{a}Department of Materials Science and Engineering, Carnegie Mellon University, Pittsburgh, PA 152133890, USA
^{*}Correspondence email: mdg@andrew.cmu.edu
This paper introduces a new 2D representation of the orientation distribution function for an arbitrary material texture. The approach is based on the isometric square torus mapping of the Clifford torus, which allows for points on the unit quaternion hypersphere (each corresponding to a 3D orientation) to be represented in a periodic 2D square map. The combination of three such orthogonal mappings into a single RGB (red–green–blue) image provides a compact periodic representation of any set of orientations. Square torus representations of five different orientation sampling methods are compared and analyzed in terms of the Riesz s energies that quantify the uniformity of the samplings. The effect of on the square torus map is analyzed in terms of the Rodrigues fundamental zones for the rotational symmetry groups. The paper concludes with example representations of important texture components in cubic and hexagonal materials. The new RGB representation provides a convenient and compact way of generating training data for the automated analysis of material textures by means of neural networks.
Keywords: orientation distribution functions; texture; symmetry; quaternions; Clifford torus.
1. Introduction
The texture of a polycrystalline material is typically described by an orientation distribution function (ODF), which represents the φ_{1}, Φ, φ_{2}) corresponding to rotations around the Cartesian z, x′ and z′′ axes, respectively. There are several other frequently used rotation representations, including the set of neoEulerian representations (Rodrigues–Frank, homochoric or 3D stereographic vectors) and the quaternion representation; for details of each of these we refer the interested reader to Morawiec (2004). Details of the conversions between the representations can be found in the work of Rowenhorst et al. (2015).
of the sample with a particular orientation with respect to an external reference frame. In the materials field, this is usually the sample frame described by the RD–TD–ND directions (reference, transverse and normal directions, respectively), whereas the geological community typically selects a sample reference frame associated with the growth history of the sample. Traditionally, the orientations are parameterized by means of three Euler angles; in the Bunge Euler convention, those angles are represented by the triplet (Given the 3D nature of orientation space, the human brain's ability to interpret 3D structures correctly, and the widespread availability of highend computer graphics, several new orientation visualization techniques have been proposed in recent years (Berestova et al., 2018; Krakow et al., 2017; Callahan et al., 2017a,b). In these articles, several orientation representations, in particular the neoEulerian representations, are combined with graphical rendering software to produce 3D visualizations of textures as point clouds and clusters, density functions, and emission maps, either displaying the full orientation space or applying restrictions to the Rodrigues fundamental zone for textures and to the disorientation fundamental zone for multiphase textures (Callahan et al., 2017b).
Despite the elegance and widespread availability of these 3D renderings (e.g. in MTEX; Bachmann et al., 2010), they are typically restricted to interactive environments where the user can manipulate the viewing direction, zoom in on particular regions of orientation space or change the orientation representation mode, e.g. from Rodrigues–Frank vectors to homochoric or 3D stereographic vectors. Furthermore, none of these 3D visualizations lend themselves to being analyzed with machine learning techniques. The main goal of this paper is to propose a 2D visualization of a texture that can be formatted as a periodic RGB (red–green–blue) color image and is hence suitable as training data for a classification neural network to recognize texture components and/or texture fibers automatically.
From a numerical point of view, the quaternion representation provides an efficient and powerful way of working with 3D rotations, for instance in combining rotations or finding a geodesic path between two different orientations (Hanson, 2006). In the following section, we will thus start from the threesphere, , i.e. the unit quaternion sphere, to derive a 2D representation of 3D orientation space that can potentially be used for the training of neural networks.
2. Mathematical background
2.1. Definitions
Following the original description in Section 4.4 of Alexa (2022), we define the Clifford torus, also known as the Euclidean twotorus, as the of two unit circles which results in an object embedded in . Consider the unit circle parameterized by the angle θ and defined as the set of all points of the form with
Scaling the circle to a radius of ρ and taking the of two such circles results in the Clifford torus,
For the special choice , the Clifford torus becomes a submanifold of the unit threesphere, , because the norm of each point on the torus is equal to 1. Putting and , we have ; similarly, for and , we have , so that .
The Clifford torus has the special property that it is flat, i.e. there exists an from the torus to a 2D square with periodic boundaries; the edges of the square have length 2π and cover the interval [−π, π]. The isometric mapping, which can be shown to have a unit Jacobian, consists of taking the ratios
and inverting the relations to the coordinates (X, Y) = (θ, φ) in the square,
Here we use the traditional numerical twoargument version of the arctangent function which produces angles in the range [−π, π].
In materials texture analysis, it is common practice to perform computations involving 3D orientations by means of unit quaternions, , with q_{0} the scalar part and q the vector part of the quaternion. Since unit quaternions reside on and the Clifford torus is a submanifold of the threesphere, the following question arises naturally: What role can the Clifford torus play in the description of sets of 3D orientations/rotations and, thus, the description of 3D materials textures?
2.2. Projection of unit quaternions onto the Clifford torus
For a unit quaternion q with components (q_{0}, q_{1}, q_{2}, q_{3}), the projection on the Clifford torus results in the point x with coordinates
The projected coordinates in the square torus are then readily shown to be given by
The reason for the use of the symbol Z_{Y} will become clear below. For a 3D rotation with unit rotation axis and rotation angle ω, the unit quaternion is given by
Substitution of the quaternion components q_{i} into equation (4) then results in
This can be converted to the Rodrigues–Frank representation by noting that R = = , so that
The notation Z_{Y} reflects the fact that the ratio R_{z}/R_{y} is used to determine the second square torus coordinate. Note that this relation is not invertible, since there are only two on the square torus and 3D rotations require three The representation in terms of the Rodrigues–Frank vector R suggests that the loss of information during the projection onto the square torus occurs due to the ratio of the R_{z} and R_{y} components in the Z_{Y} coordinate. This, in turn, suggests that two other projections can be defined by cyclic permutation of the vector components (q_{1}, q_{2}, q_{3}) of the quaternion in equation (3); starting from this equation, we can derive two additional square torus projections via the relations
We can think of the three coordinate pairs as three different isometric projections of an orientation onto three orthogonal square tori. We will label the square tori by their coordinate symbols; when no coordinate label is present, the (X, Z_{Y}) projection will be assumed. In terms of the Rodrigues–Frank vector components, the cyclic permutations correspond to 120° rotations about the principal diagonal axis of the Rodrigues reference frame.
Fig. 1 shows the square torus (X, Z_{Y}) in the range [−π, π] along both horizontal and vertical axes; the left and right vertical edges connect to each other, as do the top and bottom edges. The shaded areas correspond to different sign combinations of the quaternion components, with regions I–IV corresponding to quaternions with a positive scalar part (the default convention for 3D rotations) and a negative scalar part for the outer regions. Rectangles with the same gray level are exact copies of each other due to the doublecover nature of .
2.3. Relation between the square torus map and the Euler angle representation
The (Z, Y_{X}) square torus map is related to a projection of Euler space along the Φ axis. The transformation relations from the Rodrigues vector components to the Bunge Euler angles are given by (Callahan et al., 2017a)
For the (Z, Y_{X}) map, we have and and, using the sum and difference formulas for the tangent function, it is easy to show that
so that
This means that the (Z, Y_{X}) square torus map is identical to a projection of Euler space along the Φ axis followed by a 45° rotation, bringing the φ_{1} = φ_{2} diagonal parallel to the Y_{X} axis of the square torus map. The two other maps, (X, Z_{Y}) and (Y, X_{Z}), do not appear to have simple interpretations in terms of linear projections through Euler space; they are more complicated nonlinear projections.
2.4. Zoneplate function representation
For a given set of N unit quaternions, one can represent each quaternion by a narrow symmetric normalized 2D Gaussian function at the position (X, Z_{Y}) and add all Gaussians together to obtain a 2D `intensity' landscape representing the set . Several versions of this landscape can be generated:
(i) . All orientations in the set are represented visually, without application of symmetry or conversion to quaternions with a positive scalar part.
(ii) . All orientations are reduced to the Rodrigues fundamental zone (RFZ) for the rotational
corresponding to the (this implies that they have a positive scalar part).(iii) or . All symmetrically equivalent orientations are computed but only those with either a positive or a negative scalar part are visualized on the square torus.
(iv) . All symmetrically equivalent orientations with both positive and negative scalar quaternion parts are represented.
For each of these cases, one can apply a zone plate to the square torus, i.e. the intensity at each point is multiplied by a modulation function with spatially varying frequency content (Öztireli, 2020). For orientation sets that are supposed to be uniform in SO(3), the application of a zone plate sometimes facilitates the interpretation of the intensity distribution and makes it easier to spot nonuniformities. When a zoneplate function is applied, we represent this by a subscript on the intensity landscape, e.g. or . In the following section we will explore what these representations look like for uniform samples of orientations, as well as for a number of well known textures. It should be noted that application of a zoneplate function in general destroys the equivalence of regions with equal shading in Fig. 1.
Alexa (2022) suggested the following zoneplate function:
where κ is a constant that determines the number of oscillations of the zoneplate function in the interval [−π, π] and d(p, q) is the natural metric on ,
with 〈p, q〉 the standard dot product between two quaternions projected onto the Clifford torus. q_{f} is an arbitrary point on the torus, so that the zoneplate function uses the geodesic distance between q and q_{f} along the surface of the torus. In this paper, we select the reference point
which is clearly located on the Clifford torus and causes z(q) to be symmetric with respect to the point (X, Z_{Y}) = (0, 0), as shown in Fig. 2; in this figure, we have set κ = 40 and subdivided the interval [−π, π] into an equidistant grid of 1001 points along both axes.
3. Uniform and random samplings of SO(3)
In this section, we explore a number of different orientation sampling approaches and their representation on the square torus using a zoneplate function. The following sampling approaches are used to generate orientation sets:
(i) . Each quaternion is composed of four components, each uniformly sampled on the interval [−1, 1] using the Mersenne twister algorithm (Matsumoto & Nishimura, 1998), and the quaternion is subsequently normalized.
(ii) . Each unit quaternion is generated using the Marsaglia sampling approach (Marsaglia, 1972). Draw two uniform random numbers x_{1} and y_{1} from [−1, 1] until s_{1} = . Repeat for x_{2}, y_{2} until s_{2} = . Then replace s_{2} by and form the unit quaternion q = [x_{1}, y_{1}, x_{2}s_{2}, y_{2}s_{2}].
(iii) . In the Shoemake algorithm (Shoemake, 1992), three random numbers are generated using the Mersenne twister algorithm: u_{1} is selected uniformly from the interval [0, 1] and u_{2} and u_{3} are selected uniformly from the interval [0, 2π]. Setting a = and b = , the random unit quaternion is then generated as q = .
(iv) . Cubochoric sampling (Roşca et al., 2014) is used to generate a uniform 3D grid of points inside a cube of edge length . Each of these points is then mapped using an equalvolume mapping onto the northern hemisphere of (the northern hemisphere corresponds to quaternions with positive scalar part), resulting in a uniform sampling of SO(3).
(v) . SuperFibonacci sampling (Alexa, 2022) is a relatively new sampling approach that relies on two irrational numbers (ϕ, ψ). One potential choice is related to the golden ratio (positive root of ϕ^{2} = ϕ + 1) and the supergolden ratio (positive root of ψ^{3} = ψ^{2} + 1). Another choice, producing a uniform set of quaternions with a lower dispersion, sets and ψ as the positive root of ψ^{4} = ψ + 4 → ψ = 1.53375117. To generate N uniformly distributed quaternions, the algorithm is as follows: for i ∈ [0, …, N − 1], set s = i + 1/2, t = s/n, d = 2πs, , , α = d/ϕ and β = d/ψ. The quaternion q_{i} is then formed as q_{i} = .
For each of these five sampling approaches, an orientation set of 10^{6} samples was generated and represented using the zoneplate function approach. For each orientation in the set, the quaternion was projected onto the Clifford torus and, at the corresponding point (X, Y) in the square torus, a narrow unitamplitude 2D Gaussian kernel was added to the intensity plot and multiplied by the zoneplate function value for that point. The resulting intensity plots are shown in Fig. 3. Due to the discrete grid nature of cubochoric sampling, the number of sampling points for the set in general cannot be set arbitrarily; however, in the absence of crystal symmetry, a sampling of the cubochoric grid with 100 equidistant points on the side results in a sampling of orientation space with precisely 10^{6} points.
For the set , the zoneplate function shown in Fig. 3(a) shows a regular pattern of excess intensities corresponding to (X, Y) values close to π/4 and 3π/4, indicating that the uniform sampling approach based on the Mersenne twister does not produce a uniform orientation sampling. For the Marsaglia generator, the function shown in Fig. 3(b) shows a better uniformity than the uniform sampling approach, but there is an excess intensity for orientations near X = ±π/2. The Shoemake algorithm produces a smoothly varying zoneplate representation [Fig. 3(c)], indicating that the sampling is uniform. The cubochoric sampling approach shows a relatively smooth zoneplate function [Fig. 3(d)], but there are highly localized excess intensities (arrowed) for Y = ±π/4 and Y = ±3π/4. These are probably due to the fact that the cubochoric sampling algorithm relies on sampling of six pyramidal volumes that together make up the cube [see Fig. 1 in the report by Roşca et al. (2014)]; where the pyramids meet (along the body diagonals of the cube) the sampling is apparently not as uniform as elsewhere inside the cube. However, the overall zone plate for the cubochoric sampling approach is significantly smoother than that for the uniform and Marsaglia sampling methods, indicating that the orientation set represents a more uniform sampling of SO(3). Finally, the zoneplate function in Fig. 3(e) shows that the superFibonacci sampling approach produces a smooth intensity profile with no obvious nonuniformities.
The uniformity of a sampling of points on the sphere can be quantified using the concept of Riesz energy (Hardin & Saff, 2004). For a set of N orientations represented by unit quaternions, the Riesz energy is defined as
This expression can be interpreted as a generalized Coulomb energy for a collection of points on the sphere and reduces to the standard Coulomb energy for s = 1. The sum covers the entire threesphere, not just the northern hemisphere . The optimal Riesz energies for uniform coverage of any hypersphere are well known (Hardin & Saff, 2004); the values for the threesphere for an orientation set with N elements are given by
Table 1 lists the optimal values for an orientation data set of 2 × 10^{6} unit quaternions (counting q and −q as distinct points) along with the relative values for each of the orientation sets of Fig. 3; for an optimal sampling, all three ratios should be equal to unity. Both uniform and Marsaglia sampling have ratios that are significantly different from unity, in particular the r_{3} value, which is more sensitive than the other two in terms of local sampling nonuniformities. The Shoemake algorithm produces r_{1} and r_{2} ratios that are very close to unity, but has an r_{3} value nearly twice the optimal value; this indicates that globally this sampling has excellent uniformity but on a local scale the sampling is not as optimal as the cubochoric and superFibonacci sampling approaches. The latter two are comparable in their r_{i} values, which are all relatively close to unity, indicating that there is no significant oversampling of subregions of orientation space. The visualization of the orientation data sets in terms of the zoneplate function is qualitatively consistent with the Riesz energy ratios, so the zoneplate images can be used as a visual substitute for the more accurate computation of the Riesz energies.

4. Material textures and the square torus representation
4.1. Fundamental zone representations
In this section we review how orientation information restricted to an RFZ is represented in the square torus. The example orientation data sets used in Sections 4.2 and 4.3 are taken from the supplementary material of Callahan et al. (2017b). In the context of material textures, the two additional square tori (Y, X_{Z}) and (Z, Y_{X}) become relevant since they project the orientation data along different directions. One can think of a texture representation as an intensity distribution on three adjacent faces of a cube with edge length 2π, or as a periodic RGB color image.
4.1.1. Cyclic pointgroup symmetry
For the cyclic rotational point groups 2 (C_{2}), 3 (C_{3}), 4 (C_{4}) and 6 (C_{6}), the RFZ corresponds to the region between two parallel planes normal to the R_{z} axis [for the monoclinic 2 (C_{2}), with b as the unique axis, the planes are normal to the R_{y} axis] at a distance of from the origin, where n is the order of the rotation axis. Points inside this region are of the form
with for n = 3, 4 and 6, and of the form
with for n = 2. Producing a uniform sampling of these RFZs results in nonuniform distributions in the square torus because there are more points far away from the origin than nearby.
The zoneplate functions for the cyclic rotational point groups are shown in the top row in Fig. 4. The orientation sets were generated by the superFibonacci algorithm and subsequently reduced to the Rodrigues fundamental zone, i.e. the intensity plots are of the type . Note that for 2 (C_{2}), the denser (brighter) regions are shifted up by half a unit due to the selection of the b axis as the unique monoclinic axis. Along the horizontal axis, orientations span the range [−π/2, π/2], whereas in the vertical direction the entire range [−π, π] is used. As the order of the rotation axis increases, the intensity becomes more focused near the points (±π/2, 0).
4.1.2. Dihedral, tetrahedral and octahedral pointgroup symmetry
For the dihedral point groups 222 (D_{2}), 32 (D_{3}), 422 (D_{4}) and 622 (D_{6}) and for the two cubic groups 23 (T) and 432 (O), the RFZs are finite and bounded by planar facets. Fig. 5 shows the projections of the RFZ edges onto the square torus along with a volume rendering of the RFZ; all inset images have the same scale and the R_{z} axis is vertical in all cases. The corresponding zone plates for these point groups are shown in the bottom row of Fig. 4.
A few general trends can be observed in the zone plates for the four dihedral groups:
(i) The projected RFZ outlines in Fig. 5 have the same horizontal width and stretch across the entire vertical dimension for all four dihedral groups. One can think of the projected outline as a distorted 2D net corresponding to the polyhedral RFZ shape. For instance, for 422 (D_{4}) the central square in the outline corresponds to the square face normal to the x axis in Rodrigues space and the entire polyhedron is `unfolded' in the vertical (z) direction to produce the distorted octagons above and below (labeled `top' and `bottom' in Fig. 5). The other vertical facets give rise to the curved areas at the Y = 0 line as well as the top and bottom Y = ±π regions of the square torus. The RFZ facets perpendicular to the y axis are projected onto the vertical X = ±π/2 edges of the outline.
(ii) The zoneplate plots in Fig. 4 show that the intensity becomes more concentrated along the Y = 0 and Y = ±π lines as the rotational order increases. This is because the `thickness' of the RFZ along the z axis (Fig. 5) decreases [recall that the top and bottom facets are at a distance from the origin] and this results in fewer projected orientations near the Y = ±π/2 lines.
For the cubic rotational groups 23 (T) and 432 (O), the width of the zone plate corresponding to the RFZs is narrower than that for the dihedral groups. For the tetrahedral group 23 (T), the vertical edges of the region with nonzero intensity correspond to the single intersection points of the octahedral RFZ (Fig. 5) with the y axis; those orientations are completely degenerate in the square torus representation. For both cubic groups, the zoneplate intensity shown in Fig. 4 reaches a maximum near the horizontal projections of the RFZ edges.
4.2. Basic texturetype representations
In this section, we use two basic texture components, the cube or {100}〈001〉 texture and the Goss texture {110}〈001〉, to demonstrate how these orientation sets are represented using the square torus and/or the zoneplate intensity map. Synthetic cube and Goss textures were generated using the EMsampleRFZ program which is part of the EMsoft open source package for electron scattering simulations (Singh et al., 2017). Each orientation set contains 1 000 000 unique orientations clustered around the respective texture component mean orientations, i.e. the origin of Rodrigues space for the cube texture and the point for the Goss texture {i.e. texture component (100)[011]}. The orientations were generated using von Mises–Fisher sampling with halfwidths of 5° for the Goss texture and 10° for the cube texture.
Fig. 6 shows the two orientation distributions represented as point clouds in the cubic RFZ (leftmost column); note that the cluster spans the RFZ boundaries for the Goss texture. Figs. 6(a) and 6(c) show the square torus (ST) representations for both textures, and , along with the RFZ outlines. All orientations were reduced to the fundamental zone before being projected onto the square torus. In both cases, the representations contain vertical lines due to the `unfolding' of the fundamental zone along the vertical direction of the ST map. In Figs. 6(b) and 6(d) the cubic rotational symmetry operations were applied to the orientation set (including the equivalence of quaternions q and −q) before projection; the resulting ST maps correspond to and . Note that the Goss map is identical to the cube map, but is shifted diagonally by a vector (X, Y) = (π/8, −π/8).
Figs. 6(e) and 6(f) show RGB representations of the three square torus maps, with (X, Z_{Y}) mapped onto the red channel, (Y, X_{Z}) onto green and (Z, Y_{X}) onto blue. For the cube texture in Fig. 6(e), the three maps are identical so that the RGB image becomes a grayscale image. For the Goss texture in Fig. 6(f), on the other hand, the three projections differ from each other and the resulting RGB image shows distinct clusters in each color.
Because of the asymmetric way in which the R_{x} and (R_{y}, R_{z}) components of a Rodrigues vector contribute to the coordinates (X, Y) in the square torus [equation (6)], different symmetrically equivalent texture components will have different ST representations. Figs. 7(a) and 7(b) show the ST representations of the Goss texture components centered on the points and , respectively, and Fig. 7(c) shows the sum of all three Goss maps for equal weights.
Fig. 8 shows the square torus representations for several common rolling texture components for facecentered cubic (f.c.c.), bodycentered cubic (b.c.c.) and hexagonal closepacked (h.c.p.) crystal structures, as indicated in the figure caption. For cubic symmetry, both f.c.c. and b.c.c., the intensity distributions show three different basic components: a nearly circular peak [colored red in Fig. 8(a)] or a circular peak [red in Fig. 8(d)], a horizontally elongated peak (horizontal ellipse, green) and a vertically elongated peak (vertical ellipse, yellow). The texture components differ in the relative positioning of these three basic elements. In Fig. 8(d), the ellipses overlap, giving rise to a crosslike appearance (blue).
The size of the intensity peaks increases nonlinearly with the angular range Δθ (or, equivalently, the concentration parameter κ of the Watson distribution used to generate the samples). This is illustrated in Fig. 9, which shows the f.c.c. (copper) texture component for Δθ = 5°, 10° and 20°; the corresponding concentration values κ for the Watson distribution are shown below the intensity maps.
4.3. Experimental texture representations
In this section we consider several experimental data sets obtained using the electron backscatter diffraction (EBSD) technique. The first data set consists of three separate EBSD scans of a polycrystalline sample of the orthorhombic mineral forsterite (Mg_{2}SiO_{4}, No. 62, Pbnm, also known as olivine). The three data sets (courtesy of Dr K. Marquardt, University of Oxford, UK) were acquired at a beam energy of 20 keV and have a combined total of 7 039 154 sampling points covering an area of about 7 mm^{2} with a step size of 1 µm. The zoneplate and square torus representations of this orientation data set are shown in Figs. 10(a) and 10(b), respectively. The intensity distributions are very nearly uniform with a few higherintensity clusters (one is circled in the square torus map), indicating that the texture of this sample is nearly random. There are eight equivalent higherintensity cluster regions due to the order of the rotational (4) and the equivalence of q and −q (2).
To quantify the slight nonuniformity of this orientation set, a μ and concentration parameter κ for the Watson distribution on the threesphere , under application of the rotational symmetry group 222 and the doublecover property; the details of the fitting algorithm are described by Chen et al. (2015). The fitted mean orientation is given by the unit quaternion,
estimation was performed of the mean orientation quaternioncorresponding to the Rodrigues vector
and square torus coordinates (X, Y) = (0.75129592, 2.3204532); this point is indicated by a small white dot in Fig. 10(b). This corresponds to a rotation of 105.56° around the [0.70946461, −0.48020196, 0.51581597] axis in the (RD, TD, ND) sample reference frame. The Watson concentration parameter is given by κ = 2.1 (or 58.4°), indicating a very weak and broad clustered texture. This is in agreement with a Tmatrix analysis (Mardia & Jupp, 2009), which results in the eigenvalues λ_{i} = [0.613515, 0.142532, 0.126213, 0.117740]; the largest eigenvalue is only about four times larger than the others, which indicates a weakly clustered distribution.
For the synthetic orientation sets from Section 4.2 (and also in Fig. 9), the concentration parameters are one to two orders of magnitude larger than for the forsterite data set and the clustering around the mean orientation is clearly observed in the square torus representation. For the cube texture of Fig. 6(a), the eigenvalues of the Tmatrix are λ_{i} = [0.983452, 0.005506, 0.005516, 0.005525]; the ratio of the first two eigenvalues is 178.6, indicating a strongly clustered texture, in agreement with the Watson distribution κ value of 91.2. These results indicate that simple visual observation of the square torus representation can reveal even very small nonuniformities in the orientation distribution function.
The second experimental example is based on a Ti–6Al–4V rolledplate textured sample containing both α (h.c.p.) and β (b.c.c.) phases (Callahan et al., 2017b). An EBSD scan of a 2.25 mm^{2} region with step size 0.5 µm acquired at 20 keV beam energy resulted in an orientation data set of 9 × 10^{6} sampling points. As reported by Callahan et al. (2017b), the microstructure shows microtextured regions along the rolling direction with similar grain orientations, with a strong preference for [11.0] directions to be aligned along the sample normal direction. Fig. 11 shows the square torus RFZ representations for the α [panel (a)] and β [panel (d)] phases, as well as the fully symmetrized versions of the orientation data sets [panels (b) and (e)]. Note the two red lines superimposed on features of the maps; the lines are perpendicular to each other, which suggests that there may be a relation between the two phases. This is indeed the case and Callahan et al. (2017b) showed that the two phases have the Burgers orientation relation between them. Figs. 11(c) and 11(f) show the RGB torus maps for the two phases.
4.4. Fiber textures
4.4.1. F.c.c. fibers
Consider the α fiber in an f.c.c. material. Its orientations are located around the line (φ_{1}, π/4, π/2) in Euler space, with φ_{1} ∈ [0, π/2]. The corresponding unit quaternions are obtained by setting
and then forming the quaternion as q = , resulting in
with τ = (2φ_{1} + π)/4. Conversion to the square torus coordinates (X, Y) using equation (4) results in
The square torus coordinates for the Goss point (0, 45°, 90°) are then (X, Y) = (−0.392699, −1.1781), and for the point where the β fiber branches off (φ_{1} = 35.26°) we have (X, Y) = (−0.674817, −1.35956); these points are indicated on the schematic diagram in Fig. 12(a).
For the β fiber, the S and C texture components are and , respectively, leading to the square torus coordinates
and
Fig. 12(a) shows these points, along with curved line segments corresponding to quaternion spherical linear interpolation (SLERP) between pairwise end points B–S and S–C. Note that the line (φ_{1}, π/4, π/2) in Euler space maps onto the sigmoidal line from lower left to upper right shown in Fig. 12(a). Superimposed on the line are the locations for which φ_{1} is a multiple of π/2. Note that the range of φ_{1} is equivalent to [0, 4π], in agreement with the fact that the true periodicity of Euler space in the context of a quaternion mapping has all three Euler angles in the range [0, 4π] (Callahan et al., 2017b).
Application of the cubic symmetry elements and the (q, −q) doublecover property to the α and β fiber segments results in the square torus map shown in inverted contrast in Fig. 12(b); there are 48 equivalent locations for the fiber segments. Note that in some cases the α fiber maps onto a single point instead of a line segment.
4.4.2. B.c.c. fibers
Consider the α, γ and ε fibers in a b.c.c. material. In Euler space, all orientations lie along the following lines:
After conversion to the square torus coordinates, we find that the α fiber is represented by the curve
between the points (X, Y) = (0, −π/2) for Φ = 0 and (−π/4, −π/4) for Φ = π/2.
For the ε fiber we find
This curve intersects the X = 0 axis at the point (0, −π/2) and curves downwards towards the point for Φ = π/2.
The γ fiber sits in between the two curves and is represented by
where and τ_{±} = (π ± 4φ_{1})/8.
The intersection points of the γ fiber with the α and ε fibers have coordinates
for the α fiber and
for the ε fiber. All the relevant points for the α, γ and ε b.c.c. fibers are shown in Fig. 13(a), which displays the lower left quadrant of the square torus map. The full square torus map for these b.c.c. fiber textures is shown in Fig. 13(b). In these plots, the free angle, φ_{1} or Φ, is swept through the entire range [0, 4π] to cover both q and −q quaternions.
4.5. Experimental fiber texture example
As an example of a square torus representation of an experimental fiber texture, we consider a strong γ fiber in an electric steel; the data set consists of 28 306 orientations. When projected onto the (Φ, φ_{2}) plane, the resulting distribution is a sharp nearly symmetric Gaussian with an FWHM of 2.1° around the point . Fig. 14(a) shows the RGB square torus map for this data set after application of the (q, −q) doublecover property; the intensity along the curves is nearly constant, in agreement with the nearly uniform distribution of the φ_{1} angle in the interval [0, 2π]. After application of the cubic rotational symmetry elements, the resulting RGB map is shown in Fig. 14(b).
5. Discussion and conclusions
In this paper, we have introduced a new 2D representation of material textures in terms of three isometric projections from the Clifford torus , a submanifold of the threesphere , onto square torus maps which are subsequently arranged in the red, green and blue channels of a periodic RGB image. Two of the maps are nonlinear projections involving the components of the Rodrigues–Frank vector. The third projection, which also involves Rodrigues–Frank vector components, was shown to be equivalent to a linear projection through Bunge Euler space along the Φ direction onto the (φ_{1}, φ_{2}) plane. Each 3D orientation, which has three is projected onto three periodic square torus maps and is represented in these maps by a narrow Gaussian peak; the superposition of all these peaks generates an intensity map that can optionally be modulated by a zoneplate function.
The zoneplate function representation is particularly useful to determine visually whether or not a set of orientations uniformly covers orientation space. For the orientation sampling algorithms of Section 3, the analysis in terms of the Riesz s energies provides some insight into the quality of the samples, i.e. how close to optimal the orientation sample is. For the Shoemake sampling, which has r_{1} and r_{2} values that are closer to unity than any of the other sampling methods, the value of r_{3}, for s = 3, is nearly twice the optimal value. For increasing values of s, the Riesz energies become increasingly more sensitive to the local arrangements of sampling points, with only the nearest neighbors contributing in the limit s → ∞. For the orientation sampling , the Riesz energies indicate that, on a global level, the sampling is very close to optimal, but at the local level there are some deviations from optimal. Whether or not these deviations are important depends on the application of the sampling. If the orientation set is to be used as an initial sampling for dictionary indexing, for instance (Singh & De Graef, 2016), then the local nonoptimality is unimportant, since the indexing will typically be followed by an orientation step where the orientations are allowed to wander away from the initial orientations.
Different from more conventional 3D representations of material textures, the RGB square torus map representation opens a unique path to the use of neural networks to automate the analysis of material textures, in particular to determine the mixture of texture components that are present in the orientation distribution. The use of ST maps in this context is the topic of ongoing investigations.
Acknowledgements
The author would like to acknowledge stimulating discussions with A. D. Rollet, M. P. Echlin, T. M. Pollock, S. Wright, W. Lenthe, D. Rowenhorst, B. Hutchinson, C. Lafond, G. Austin and S. Niezgoda.
Funding information
The author acknowledges financial support from the National Science Foundation, Directorate for Mathematical and Physical Sciences (grant No. DMR2203378), and the use of the computational resources of the Materials Characterization Facility at Carnegie Mellon University (grant No. MCF677785). The author also acknowledges support from the John and Claire Bertucci Distinguished Professorship in Engineering.
References
Alexa, M. (2022). 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 18–24 June 2022, New Orleans, Louisiana, USA, pp. 8281–8290. New York: IEEE. Google Scholar
Bachmann, F., Hielscher, R. & Schaeben, H. (2010). Solid State Phenom. 160, 63–68. CrossRef CAS Google Scholar
Berestova, S. A., Kopytov, N. P., Misura, N. E. & Mityushov, E. A. (2018). AIP Conf. Proc. 2053, 030005. Google Scholar
Callahan, P. G., Echlin, M., Pollock, T. M., Singh, S. & De Graef, M. (2017a). J. Appl. Cryst. 50, 430–440. Web of Science CrossRef CAS IUCr Journals Google Scholar
Callahan, P. G., Echlin, M. P., Stinville, J. C., Pollock, T. M., Singh, S., Ram, F. & De Graef, M. (2017b). J. Appl. Cryst. 50, 1267–1279. Web of Science CrossRef CAS IUCr Journals Google Scholar
Chen, Y., Wei, D., Newstadt, G., DeGraef, M., Simmons, J. & Hero, A. (2015). IEEE Signal Process. Lett. 22, 1152–1155. Web of Science CrossRef Google Scholar
Hanson, A. (2006). Visualizing Quaternions, ch. 18, pp. 167–171. Amsterdam: Elsevier. Google Scholar
Hardin, D. & Saff, E. (2004). Not. AMS, 51, 1186–1194. Google Scholar
Krakow, R., Bennett, R. J., Johnstone, D. N., Vukmanovic, Z., SolanoAlvarez, W., Lainé, S. J., Einsle, J. F., Midgley, P. A., Rae, C. M. F. & Hielscher, R. (2017). Proc. R. Soc. A, 473, 20170274. Web of Science CrossRef PubMed Google Scholar
Mardia, K. V. & Jupp, P. E. (2009). Editors. Directional Statistics. Chichester: John Wiley & Sons. Google Scholar
Marsaglia, G. (1972). Ann. Math. Stat. 43, 645–646. CrossRef Web of Science Google Scholar
Matsumoto, M. & Nishimura, T. (1998). ACM Trans. Model. Comput. Simul. 8, 3–30. CrossRef Google Scholar
Morawiec, A. (2004). Orientations and Rotations: Computations in Crystallographic Textures. New York: Springer Verlag. Google Scholar
Öztireli, A. C. (2020). Comput. Graph. Forum, 39, 133–148. Google Scholar
Roşca, D., Morawiec, A. & De Graef, M. (2014). Modell. Simul. Mater. Sci. Eng. 22, 075013. Google Scholar
Rowenhorst, D., Rollett, A., Rohrer, G. S., Groeber, M., Jackson, M., Konijnenberg, P. & De Graef, M. (2015). Modell. Simul. Mater. Sci. Eng. 23, 083501. Web of Science CrossRef Google Scholar
Shoemake, K. (1992). Graphics Gems III (IBM Version), edited by D. Kirk, pp. 124–132. Cambridge: Academic Press. Google Scholar
Singh, S. & De Graef, M. (2016). Modell. Simul. Mater. Sci. Eng. 24, 085013. Web of Science CrossRef Google Scholar
Singh, S., Ram, F. & De Graef, M. (2017). Microsc. Microanal. 23(S1), 212–213. CrossRef Google Scholar
This is an openaccess article distributed under the terms of the Creative Commons Attribution (CCBY) Licence, which permits unrestricted use, distribution, and reproduction in any medium, provided the original authors and source are cited.