WorksComputer Science
Symmetry-aware super-resolution of crystal orientation maps via invariant latent-space learning
Abstract
Crystal-orientation maps are physical fields defined only up to crystal symmetry; electron backscatter diffraction (EBSD) resolves them experimentally, but acquisition-time constraints limit spatial resolution. Unlike conventional images, EBSD data lie on the quotient space , where is the crystal-symmetry group. Standard Euclidean interpolation can therefore mix symmetry-equivalent representations and blur grain boundaries. We introduce the Symmetry-Group-Aware Super-Resolution Attention Network (SG-SRAN), which incorporates crystal symmetry and boundary preservation by design. A frozen, locally isometric encoder maps equivalent orientations to a common latent representation in which Euclidean distance approximates misorientation. Super-resolution is performed in this space, with each high-resolution token restricted to a feature-consistent local support to prevent cross-boundary mixing. A dictionary-based decoder then recovers valid orientations. Across FCC and HCP benchmarks, SG-SRAN matches 15-16 million parameter backbones using only 27-49k trainable parameters, while achieving the lowest p68 errors, highest inverse-pole-figure fidelity, and zero-shot transfer to unseen alloys.
*Corresponding author(s). E-mail(s): [email protected]; Contributing authors: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; †These authors contributed equally to this work.
Electron backscatter diffraction (EBSD) is a widely used technique for resolving crystallographic orientation fields in structural materials, with applications ranging from grain-topology characterization to the study of deformation localization, recrystallization and damage initiation [1]. However, its spatial resolution comes at a substantial acquisition cost. Reducing the pixel pitch increases scan time and limits the area that can be mapped under a fixed measurement budget. A reliable digital super-resolution (SR) method could help relax this trade-off by reconstructing high-resolution microstructural information from more rapidly acquired low-resolution scans (Extended Data Fig. 1).
EBSD super-resolution is fundamentally different from conventional image super-resolution because each pixel represents a crystal orientation rather than a Euclidean intensity or color value (Fig. 1a). A crystal orientation is defined only up to the proper rotational symmetry group of the crystal. The physically meaningful orientation space is therefore the quotient manifold , where denotes the group of three-dimensional rotations and is the proper rotational crystal point group.
This geometry presents two distinct challenges. First, symmetry-equivalent orientations can have very different numerical representations. Interpolation that neglects crystal symmetry may therefore introduce artificial orientation gradients near fundamental-zone boundaries (Fig. 1b). Second, grain boundaries introduce discontinuities in the physical orientation field. Interpolation or feature aggregation across a grain boundary mixes information from distinct grains, producing nonphysical orientations and blurred, shifted, or smoothed boundaries. A naive upsampler can consequently generate visually smooth reconstructions while introducing speckled intra-grain errors and degraded grain boundaries (Fig. 1c). Common representations, including Euler angles, inverse-pole-figure color maps and raw quaternion coordinates, all retain ambiguities or discontinuities when processed using standard Euclidean operations.
Classical interpolation methods address parts of this problem. For example, Spherical linear interpolation (SLERP) preserves the unit-norm constraint of quaternions [2], and symmetry-aware variants can select shorter paths between symmetry-equivalent representations. These approaches nevertheless remain local, pairwise and non-learned. They cannot draw on mesoscale orientation structure or adapt their interpolation support to the surrounding grain topology. Learned EBSD super-resolution has also been explored using IPF-colour images and quaternion-valued maps [3, 4]. Related work has used serial-section learning to generate volumetric EBSD maps [4], for which accurate two-dimensional section-level super-resolution is a natural building block. Yet the underlying representation problem remains. When a network operates in a space whose distance metric does not agree with crystallographic misorientation, part of its capacity must be spent compensating for that geometric mismatch before it can learn the reconstruction task.
A learned backbone is also not automatically protected from mixing information across grains. Unless the architecture controls which neighbors contribute to a prediction, feature aggregation can still blur adjacent grains or bridge them with physically incorrect orientation gradients. An effective orientation-field super-resolution method must therefore distinguish between two different situations: orientations that appear different numerically but are equivalent under crystal symmetry, and orientations that are physically distinct because they lie on opposite sides of a grain boundary.
Two developments make this distinction tractable. First, locally isometric embeddings of the quotient space , constructed from Reynolds-projected Wigner- features, provide latent coordinates that are invariant to the crystal point-group action while remaining organized into irreducible-representation channels [5, 6, 7]. Symmetry-equivalent orientations therefore map to the same latent representation, while small Euclidean distances approximate small crystallographic misorientations. Second, -equivariant operations preserve the transformation laws of these latent features throughout the network rather than flattening them into generic scalar channels [8, 9, 10, 11].
Here we introduce SG-SRAN, a symmetry-aware and locally routed architecture for orientation-map super-resolution. A frozen encoder maps each input quaternion to a locally isometric latent representation constructed from Reynolds-projected Wigner- features [7]. The resulting features are invariant to the proper rotational crystal point group while transforming as structured irreducible-representation channels under global rotations. The trainable SG-SRAN backbone operates entirely on these latent features using -equivariant operations implemented in e3nn, preserving their transformation structure throughout the network. Feature-distance-masked latent convolutions restrict local support and reduce information exchange between crystallographically incompatible orientations. During upsampling, neighboring features are grouped into a small set of feature-consistent slots, and each high-resolution (HR) token is routed to its most compatible slot. Finally, a dictionary-based decoder maps the predicted latent features back to unit quaternions using an orientation dictionary generated by cubochoric sampling [12, 13].
We evaluate SG-SRAN through three sets of experiments. First, the frozen encoder–dictionary-based decoder reconstructs held-out orientations with milliradian accuracy for both face-centered cubic (FCC) IN718 and hexagonal close-packed (HCP) Ti-6Al-4V data, showing that the latent space provides a faithful computational interface across distinct crystal symmetries. Second, on the IN718 benchmark, evaluated over five seeds against four classical interpolants and seven learned baselines, the leading methods are not materially separated by pooled mean misorientation. SG-SRAN nevertheless achieves the lowest p68 and p95 misorientation errors and the highest IPF-space PSNR, while using approximately fewer parameters than the attention-based backbones. This comparison refers to trainable backbone size; end-to-end inference also includes the fixed dictionary and dictionary-based decoder, whose runtime cost is reported in Methods. Third, the same architecture family supports a second crystal symmetry using symmetry-specific invariant encoders and configuration choices. SG-SRAN is the strongest compact model on the Ti-6Al-4V HCP benchmark under the proper symmetry protocol and supports zero-shot transfer to out-of-distribution FCC CoNi and HCP Ti-Al targets (Table 3). Together, these results show that quotient geometry, equivariant feature processing and local routing provide useful inductive biases for orientation-valued super-resolution, rather than corrections applied after training a conventional Euclidean image backbone.
Table 3 Zero-shot transfer to out-of-distribution FCC and HCP targets. All entries use source-domain checkpoints without target retraining. Composition F1 is the boundary-window composition score; its spurious-rate and recall components are reported in Extended Data Table 2. Lower is better for angular metrics; higher is better for Boundary F1, Composition F1, PSNR and SSIM. Bold marks the best available method in each column. SG-SRAN uses 49K trainable parameters for the IN718-source FCC transfer and 27K for the Ti-6Al-4V-source HCP transfer.
a. Orientation-accuracy and boundary metrics.
| Method | Params (K) | Mean () | Median () | Boundary F1 | Composition F1 | Interior () | Boundary band () |
| IN718 CoNi (FCC): zero-shot test split | |||||||
| Nearest | – | 2.85 | 0.27 | 0.405 | 0.741 | 0.35 | 8.30 |
| Bicubic | – | 3.94 | 0.40 | 0.446 | 0.387 | 0.47 | 11.53 |
| SLERP | – | 3.49 | 0.27 | 0.452 | 0.364 | 0.33 | 10.40 |
| Symm-SLERP | – | 3.12 | 0.27 | 0.455 | 0.401 | 0.32 | 9.22 |
| Atindama inpainting [15] | 25,784 | 7.87 | 2.49 | 0.260 | 0.240 | 5.80 | 12.40 |
| Q-RBSA-adapted [4] | 6,175 | 2.35 | 0.88 | 0.666 | 0.636 | 0.91 | 5.49 |
| QEDSR [14] | 1,815 | 1.66 | 0.23 | 0.628 | 0.801 | 0.29 | 4.65 |
| EDSR [14] | 7,241 | 1.65 | 0.25 | 0.644 | 0.798 | 0.30 | 4.59 |
| RCAN [16] | 15,594 | 1.66 | 0.24 | 0.640 | 0.804 | 0.29 | 4.65 |
| SAN [17] | 15,862 | 1.62 | 0.26 | 0.631 | 0.844 | 0.31 | 4.48 |
| HAN [18] | 16,073 | 1.68 | 0.32 | 0.675 | 0.760 | 0.34 | 4.61 |
| SG-SRAN (ours) | 49 | 1.59 | 0.23 | 0.670 | 0.834 | 0.29 | 4.42 |
| Ti-6Al-4V Ti-Al (HCP): zero-shot test split | |||||||
| Nearest | – | 2.21 | 0.36 | 0.393 | 0.818 | 0.38 | 7.55 |
| Bicubic | – | 4.08 | 0.43 | 0.364 | 0.395 | 0.66 | 14.04 |
| SLERP | – | 3.78 | 0.33 | 0.404 | 0.411 | 0.53 | 13.26 |
| Symm-SLERP | – | 2.48 | 0.33 | 0.415 | 0.462 | 0.35 | 8.68 |
| Atindama inpainting [15] | 25,784 | 12.01 | 6.47 | 0.115 | 0.243 | 9.79 | 18.51 |
| Q-RBSA-adapted [4] | 6,175 | 2.05 | 0.79 | 0.533 | 0.867 | 0.80 | 5.69 |
| QEDSR [14] | 1,815 | 1.68 | 0.38 | 0.536 | 0.850 | 0.42 | 5.37 |
| EDSR [14] | 7,241 | 1.72 | 0.43 | 0.554 | 0.823 | 0.48 | 5.36 |
| RCAN [16] | 15,594 | 5.15 | 2.72 | 0.456 | 0.456 | 3.76 | 9.20 |
| SAN [17] | 15,862 | 1.99 | 0.67 | 0.514 | 0.900 | 0.76 | 5.60 |
| HAN [18] | 16,073 | 3.07 | 1.45 | 0.470 | 0.711 | 1.85 | 6.62 |
| SG-SRAN (ours) | 27 | 1.21 | 0.30 | 0.630 | 0.915 | 0.33 | 3.78 |
Table 3.

Figure 1. Geometric and microstructural challenges in orientation-map super-resolution. a, A low-resolution (LR) orientation map and its high-resolution (HR) target, rendered in IPF-Z color. b, Crystal orientations are points on the quotient manifold : pairwise spherical-linear interpolation (SLERP) on the quaternion sphere follows a geodesic but is local and symmetry-blind; the color bar reports the residual misorientation such an interpolant leaves across an interface. c, A “naive” generative upsampler (SLERP here) that treats orientations as Euclidean image channels; the magnified insets reveal its characteristic, physically invalid failure modes—speckled intra-grain noise and blurred or bridged grain boundaries—even where the output looks visually smooth. These artifacts motivate operating in a quotient-aware latent representation with explicit, boundary-aware routing rather than on rendered RGB maps. Scale bar, ; in the task, one LR sample corresponds to a block of HR pixels.
Results
Pipeline overview
The SG-SRAN pipeline separates the EBSD SR problem into a frozen geometric representation and a trainable latent-space upsampler (Figure 2). The encoder collapses symmetry-equivalent orientations exactly and makes small latent Euclidean distances correspond to small crystallographic misorientations (see Methods). Therefore, the trainable backbone never compares quaternions directly: local context aggregation, support-bank clustering and routed patch synthesis all happen in the latent feature space. The routing operation is symmetry-aware: SG-SRAN clusters a local window-support bank (for instance, the bank shown in Figure 2) by Euclidean distance in the locally isometric latent feature embedding, and the router selects among those symmetry-collapsed feature slots. Thus, upsampling preserves crystal-symmetry equivalence while discouraging support from distinct grains being merged across a boundary.

Figure 2. Symmetry-aware latent EBSD super-resolution pipeline and local routed synthesis in SG-SRAN. The top row summarizes the end-to-end model. Input orientations are mapped by the frozen /Reynolds encoder into a symmetry-invariant latent stack, refined once at LR, upsampled by SG-SRAN, refined at HR according to the material-specific configuration, and decoded back to crystallographically valid quaternions. The small decoded crops above the latent stacks provide an interpretable view of the orientation field before and after latent upsampling. The lower row expands one local SG-SRAN decision at a boundary-rich site. The displayed IN718 example uses a support bank around the marked LR pixel, which is partitioned into candidate orientation slots in the locally isometric latent embedding; the proposal generator builds token-conditioned candidate features for each slot, while a router uses the complete binary slot-composition masks and learned subpixel position embeddings to assign each HR token to one slot. Each routed HR token is therefore initialized from one observed local orientation branch before HR refinement. Detailed layer definitions and equations are given in Methods.

Figure 2. Frozen orientation interface and encoder–dictionary-based decoder reconstruction fidelity. a, Each pixel orientation is mapped by the Reynolds-projected encoder to crystal-symmetry-invariant latent features. Predicted HR features are matched against the dictionary, generated by cubochoric sampling, and decoded to unit quaternions, with optional local iterative refinement. The annotation denotes the surrounding SR task: spatial upsampling is performed by the SG-SRAN backbone (Figure 2), whereas the frozen encoder and dictionary-based decoder act pointwise on the LR and HR orientation fields. b, Round-trip encodedecode reconstructions on full IPF-Z scan slices for FCC IN718 using the proper rotational subgroup (left pair) and Ti-6Al-4V HCP test data using the proper rotational subgroup (right pair). In each pair, the ground-truth input and dictionary-based decoder output are visually indistinguishable, including at grain boundaries. The dense-lookup round-trip misorientation is approximately () for both systems (Table 2a), showing that the frozen latent interface round-trips orientation information across the two symmetry classes. Panel b isolates pointwise encoder–dictionary-based decoder reconstruction at unchanged spatial sampling.
Figure 2 presents the pipeline at two levels. The top row (a) gives the end-to-end latent SR path: frozen encoding, one LR latent refinement, SG-SRAN upsampling, configuration-dependent HR latent refinement and decoding. The lower row (b) depicts one typical local SG-SRAN decision: a boundary-centered local window-support bank is partitioned into latent feature slots, a proposal generator forms token-conditioned candidate features for each slot, and a MLP router uses the complete binary slot-composition masks together with learned subpixel position to choose the local orientation branch for each HR token in the final patch. Detailed layer definitions are given in Methods.
The frozen encoder–dictionary-based decoder round-trips orientations across FCC and HCP
We first isolate the latent interface from the SR task by measuring round-trip encode–decode fidelity on held-out test orientations: with no spatial upsampling, how accurately do the frozen encoder and dictionary-based decoder recover the input orientation? Dense dictionary lookup achieves milliradian-level accuracy on both crystal systems (Table 2a): the mean per-sample symmetry-aware error is 0.0076 rad on the FCC IN718 test split and 0.0079 rad on the Ti-6Al-4V HCP test split. These values measure the end-to-end round-trip fidelity of the frozen Reynolds encoder and dictionary-based decoder independently of the spatial super-resolution network. Their close agreement shows that the latent interface attains comparable round-trip accuracy across the two crystal-symmetry classes. The two geometric properties that make this latent space a locally faithful interface—exact crystal-symmetry invariance and local isometry with respect to misorientation—are verified directly in Figure 3; representative round-trip reconstructions are shown in Extended Data Figure 2.
| Method | Params (K) | Mean (°) | Median (°) | Boundary F1 | Composition F1 | Interior / Boundary mean (°) |
| IN718 (FCC, held-out test patches; learned methods over five seeds) | ||||||
| Nearest | – | 4.45 ± 0.60 | 1.00 | 0.425 | 0.807 | 0.69 / 9.36 |
| Bicubic | – | 6.71 ± 0.99 | 1.22 | 0.432 | 0.353 | 1.57 / 13.42 |
| SLERP | – | 6.16 ± 1.00 | 0.88 | 0.442 | 0.325 | 1.38 / 12.39 |
| Symm-SLERP | – | 4.92 ± 0.65 | 0.86 | 0.477 | 0.381 | 0.67 / 10.46 |
| Atindama inpainting [15] | 25,784 | 7.14 ± 0.42 | 2.82 | 0.400 | 0.245 | 3.57 / 11.79 |
| Q-RBSA-adapted [4] | 6,175 | 3.05 ± 0.12 | 0.76 | 0.665 | 0.665 | 0.64 / 6.19 |
| QEDSR [14] | 1,815 | 2.68 ± 0.01 | 0.60 | 0.626 | 0.853 | 0.56 / 5.44 |
| EDSR [14] | 7,241 | 2.59 ± 0.04 | 0.61 | 0.647 | 0.849 | 0.58 / 5.22 |
| RCAN [16] | 15,594 | 2.52 ± 0.04 | 0.65 | 0.678 | 0.866 | 0.58 / 5.05 |
| SAN [17] | 15,862 | 2.62 ± 0.03 | 0.61 | 0.640 | 0.899 | 0.58 / 5.27 |
| HAN [18] | 16,073 | 2.63 ± 0.05 | 0.62 | 0.643 | 0.812 | 0.57 / 5.32 |
| SG-SRAN (ours) | 49 | 2.60 ± 0.06 | 0.61 | 0.658 | 0.891 | 0.61 / 5.20 |
| Ti-6Al-4V (HCP, held-out test patches; learned methods over five seeds) | ||||||
| Nearest | – | 13.85 | 0.68 | 0.544 | 0.689 | 0.67 / 13.98 |
| Bicubic | – | 20.96 | 11.68 | 0.611 | 0.293 | 2.71 / 21.14 |
| SLERP | – | 20.72 | 11.39 | 0.605 | 0.311 | 1.51 / 20.91 |
| Symm-SLERP | – | 14.94 | 7.22 | 0.561 | 0.406 | 0.73 / 15.08 |
| Atindama inpainting [15] | 25,784 | 20.76 ± 0.47 | 13.03 | 0.618 | 0.263 | 8.50 / 20.88 |
| Q-RBSA-adapted [4] | 6,175 | 10.36 ± 0.09 | 1.00 | 0.620 | 0.773 | 0.92 / 10.46 |
| QEDSR [14] | 1,815 | 10.09 ± 0.03 | 0.66 | 0.619 | 0.766 | 0.59 / 10.18 |
| EDSR [14] | 7,241 | 11.72 ± 3.55 | 3.34 | 0.609 | 0.739 | 2.72 / 11.81 |
| RCAN [16] | 15,594 | 10.74 ± 0.69 | 1.88 | 0.613 | 0.511 | 1.27 / 10.83 |
| SAN [17] | 15,862 | 10.27 ± 0.15 | 1.18 | 0.621 | 0.804 | 0.95 / 10.36 |
| HAN [18] | 16,073 | 10.64 ± 0.64 | 1.79 | 0.624 | 0.671 | 1.26 / 10.73 |
| SG-SRAN (ours) | 27 | 9.54 ± 0.12 | 0.48 | 0.639 | 0.782 | 0.51 / 9.63 |
Table 2a.

Figure 3. The frozen Reynolds embedding is crystal-symmetry-invariant and locally isometric for both crystal systems. a,b, Scalar field over the sphere of rotation axes , rendered in 3D for the FCC/ (a) and HCP/ (b) embeddings; the encoder uses the proper rotational subgroups, while these scalar norm fields also display the corresponding cubic and hexagonal Laue-class visual symmetries ( and ). c,d, The same fields unrolled on the polar and azimuthal angles , exposing the symmetry-repeating fundamental tiles (FCC, c; HCP, d). e, Locally isometric embedding distance versus symmetry-reduced crystallographic misorientation for low-discrepancy (Fibonacci-lattice sampling) orientation-pair separations per system (line, median; band, interquartile range). Both systems are linear near the origin (dashed small-angle reference); the first-order calibration holds to rad (, red dashed) and saturates beyond. f, Numerical verification of crystal-symmetry invariance. For each low-discrepancy sampled orientation, we evaluate all 24 FCC or 12 HCP symmetry-equivalent copies and report the maximum deviation from their orbit-mean embedding, normalized by the mean norm. Relative deviations remain below , confirming invariance to single-precision accuracy. The retained Reynolds degrees are for FCC and for HCP.

Figure 3. Expanded learned-baseline comparison on IN718 and Ti-6Al-4V held-out 4 × 4 examples. a, IN718 comparison with the full Open718 block [20] labeled Test 0, followed by held-out test samples 1–3, shown at a common extent and aspect ratio. b, Ti-6Al-4V trained-baseline comparison on held-out test samples 0–3 from the HCP split. In both panels, columns show the display-upsampled LR input, Atindama inpainting, EDSR, HAN, RCAN, SAN, Q-RBSA-adapted, QEDSR, SG-SRAN and the HR target, with the material-specific IPF-Z key at right. One displayed LR pixel corresponds to a 4 × 4 HR block.
Benchmark on held-out IN718 and Ti-6Al-4V at 4 × 4
We evaluate SG-SRAN at scale factor against four classical interpolants (nearest-neighbor, bicubic with renormalization, SLERP and symmetry-aware SLERP) and seven learned baselines trained independently on the matched splits: Q-RBSA-adapted [4], EDSR [14], a quaternion EDSR (QEDSR) [14], an Atindama-style partial-convolution inpainting network [15], RCAN [16], SAN [17] and HAN [18]. Every learned method is trained over five seeds per material and evaluated under a single fixed protocol (symmetry-aware error , a boundary mask, the same held-out test splits of IN718 and Ti-6Al-4V patches; Methods), so all rows of tbl:2b are directly comparable. The Ti-6Al-4V metrics use the 12-element proper rotational subgroup.
On IN718, the attention backbones, the quaternion baselines and SG-SRAN form a closely grouped tier on pooled mean misorientation: RCAN records the lowest mean (), with SG-SRAN (), SAN () and HAN () separated by only about a tenth of a degree. The complementary metrics resolve distinct strengths: RCAN leads Boundary F1 (0.678) and boundary-band error, QEDSR leads median and interior error, SAN leads boundary-window Composition F1, and SG-SRAN attains the lowest p68 and p95 error percentiles (p68 and p95 ; the next-best learned p68 is EDSR at and the next-best learned p95 is SAN at , with the remaining learned rows spanning –; Table 1) and the highest IPF-space PSNR and SSIM. SG-SRAN reaches this with a ∼49k-parameter trainable backbone—∼320× smaller than RCAN/SAN/HAN, ∼126× smaller than Q-RBSA-adapted and ∼530× smaller than Atindama inpainting. The IN718 comparison therefore establishes SG-SRAN’s compactness–accuracy advantage: it leads p68/p95 and parameter efficiency while maintaining competitive accuracy among other metrics.
| Method | Params (K) | p68 (°) | p95 (°) | p99 (°) | PSNR (dB) | SSIM |
| IN718 4 × 4 (FCC, ) | ||||||
| Nearest | – | 1.00 | 46.83 | 59.76 | 17.58 ± 0.73 | 0.615 ± 0.021 |
| Bicubic | – | 3.99 | 39.33 | 57.73 | 14.64 ± 0.90 | 0.474 ± 0.044 |
| SLERP | – | 1.24 | 38.32 | 56.40 | 14.74 ± 0.99 | 0.527 ± 0.042 |
| Symm-SLERP | – | 1.15 | 32.40 | 53.04 | 15.89 ± 0.65 | 0.558 ± 0.031 |
| Atindama inpainting [15] | 25,784 | 4.60 | 34.62 | 50.13 | 13.90 ± 0.26 | 0.411 ± 0.033 |
| Q-RBSA-adapted [4] | 6,175 | 1.03 | 14.64 | 56.23 | 19.01 ± 0.31 | 0.652 ± 0.012 |
| QEDSR [14] | 1,815 | 0.86 | 5.23 | 59.21 | 19.91 ± 0.04 | 0.706 ± 0.004 |
| EDSR [14] | 7,241 | 0.83 | 3.39 | 59.04 | 20.12 ± 0.11 | 0.711 ± 0.004 |
| RCAN [16] | 15,594 | 0.89 | 3.88 | 58.30 | 20.29 ± 0.05 | 0.706 ± 0.013 |
| SAN [17] | 15,862 | 0.83 | 2.70 | 59.12 | 20.13 ± 0.06 | 0.710 ± 0.004 |
| HAN [18] | 16,073 | 0.85 | 3.17 | 59.11 | 20.10 ± 0.14 | 0.709 ± 0.009 |
| SG-SRAN (ours) | 49 | 0.81 | 1.73 | 59.55 | 20.36 ± 0.15 | 0.715 ± 0.006 |
| Ti-6Al-4V 4 × 4 (HCP, proper D, ) | ||||||
| Nearest | – | 7.93 | 73.56 | 88.08 | 12.90 | 0.449 |
| Bicubic | – | 22.45 | 72.26 | 87.23 | 10.59 | 0.262 |
| SLERP | – | 24.19 | 72.66 | 87.14 | 10.44 | 0.280 |
| Symm-SLERP | – | 15.58 | 56.12 | 79.57 | 12.76 | 0.405 |
| Atindama inpainting [15] | 25,784 | 22.79 | 69.53 | 86.87 | 10.15 ± 0.10 | 0.275 ± 0.012 |
| Q-RBSA-adapted [4] | 6,175 | 1.93 | 66.00 | 86.61 | 14.29 ± 0.03 | 0.557 ± 0.001 |
| QEDSR [14] | 1,815 | 1.99 | 65.58 | 86.49 | 14.36 ± 0.03 | 0.562 ± 0.002 |
| EDSR [14] | 7,241 | 6.66 | 61.35 | 84.99 | 13.80 ± 1.32 | 0.503 ± 0.105 |
| RCAN [16] | 15,594 | 3.64 | 63.57 | 85.78 | 14.18 ± 0.23 | 0.534 ± 0.032 |
| SAN [17] | 15,862 | 2.89 | 64.48 | 86.15 | 14.36 ± 0.05 | 0.547 ± 0.009 |
| HAN [18] | 16,073 | 4.00 | 63.34 | 85.79 | 14.22 ± 0.21 | 0.529 ± 0.029 |
| SG-SRAN (ours) | 27 | 1.36 | 65.16 | 86.51 | 14.60 ± 0.06 | 0.575 ± 0.005 |
Table 1. Model compactness, orientation accuracy and image fidelity (all baselines). Per-method trainable parameter count (thousands, K), error percentiles (p68/p95/p99) of the per-pixel symmetry-aware misorientation, and IPF-space PSNR/SSIM, on the held-out IN718 () and Ti-6Al-4V () test splits. SG-SRAN parameter counts are configuration-specific: 49K for the IN718 support-window setting and 27K for the Ti-6Al-4V setting. Deterministic interpolants are marked “–” and learned rows are five-seed means; ± reports seed-level standard deviation for learned rows and per-patch test-set standard deviation for deterministic IN718 rows. Pooled mean and median misorientation are in tbl:2b. Lower is better for misorientation, higher for PSNR/SSIM; bold marks the best per column. Most IN718 learned methods cluster near 59° at p99, whereas the boundary-rich Ti-6Al-4V benchmark yields larger extreme errors across methods.
| Method | IN718 4 × 4 | Ti-6Al-4V 4 × 4 | ||||
| Spurious | Recall | F1 | Spurious | Recall | F1 | |
| Nearest | 0.030 | 0.691 | 0.807 | 0.149 | 0.579 | 0.689 |
| Bicubic | 0.691 | 0.410 | 0.353 | 0.734 | 0.327 | 0.293 |
| SLERP | 0.673 | 0.323 | 0.325 | 0.679 | 0.301 | 0.311 |
| Symm-SLERP | 0.613 | 0.374 | 0.381 | 0.566 | 0.382 | 0.406 |
| Atindama inpainting [15] | 0.813 | 0.355 | 0.245 | 0.792 | 0.358 | 0.263 |
| Q-RBSA-adapted [4] | 0.335 | 0.665 | 0.665 | 0.076 | 0.664 | 0.773 |
| QEDSR [14] | 0.080 | 0.796 | 0.853 | 0.104 | 0.669 | 0.766 |
| EDSR [14] | 0.083 | 0.791 | 0.849 | 0.147 | 0.652 | 0.739 |
| RCAN [16] | 0.070 | 0.810 | 0.866 | 0.476 | 0.499 | 0.511 |
| SAN [17] | 0.036 | 0.843 | 0.899 | 0.106 | 0.731 | 0.804 |
| HAN [18] | 0.151 | 0.777 | 0.812 | 0.279 | 0.627 | 0.671 |
| SG-SRAN (ours) | 0.038 | 0.830 | 0.891 | 0.093 | 0.687 | 0.782 |
Table 1. Extended Data Table 1 Boundary-window composition analysis on in-distribution test splits. This pooled diagnostic is evaluated on the selected prediction summaries used for the composition analysis. For each HR boundary pixel, the HR and SR windows centered at that pixel are compared under crystal symmetry with a tolerance. The spurious rate is the fraction of SR window observations not matching any GT orientation group in the corresponding HR window (lower is better). Recall is the fraction of GT orientation groups recovered by at least one SR observation, and Composition F1 uses precision spurious rate together with this recall. Boundary centers use the HR 8-neighbor boundary mask; overlapping boundary windows are counted separately.
On Ti-6Al-4V, the task is substantially harder because the held-out HCP patches contain a dense network of fine boundaries. The 27k-parameter SG-SRAN configuration nevertheless gives the best pooled mean (), median (), interior mean (0.51°), boundary-band mean (9.63°), Boundary F1 (0.639), PSNR and SSIM. SG-SRAN also leads p68; EDSR leads learned-method p95, and symmetry-aware SLERP leads overall p95 and p99. The Ti-6Al-4V comparison therefore separates boundary-aware reconstruction from extreme-tail ranking: SG-SRAN gives the most accurate and most faithful average reconstruction while SAN narrowly gives the best boundary-window Composition F1 (0.804 versus 0.782 for SG-SRAN).
Qualitative comparisons support the same picture: SG-SRAN and Q-RBSA-adapted recover coherent grain interiors while Atindama leaves periodic stripe artifacts from the sparse observation mask (Extended Data Figure 3); matched held-out test examples for both materials are shown in Extended Data Figure 4. IPF renderings are used only as qualitative visualizations; quantitative claims rest on and and the boundary diagnostics. Scale dependence across the , and tasks on both materials is summarized in Extended Data Figure 7.

Figure 4. Held-out test samples on both crystal systems. Within each material panel, from top to bottom, rows show the display-upsampled low-resolution input, the SG-SRAN super-resolved output and the high-resolution target; the first three columns show IPF-X, IPF-Y and IPF-Z, and the fourth visual column shows the IPF-Z residual with brighter pixels indicating larger residual. The LR input is displayed at the HR grid via nearest-neighbor upsampling only for visual alignment; the SG-SRAN SR and HR fields are at the native HR resolution. (a) IN718 held-out test sample. (b) Ti-6Al-4V held-out test sample from the HCP test split. One LR pixel corresponds to a HR block; the displayed HR fields are pixels for IN718 and pixels for Ti-6Al-4V.

Figure 7. Scale dependence across crystal systems and upsampling factors. Seed-42 performance is shown for deterministic interpolants, learned baselines and SG-SRAN as grouped normalized score bars at , and . a, IN718 (FCC, ). b, Ti-6Al-4V (HCP, proper ). Each scale subplot reports median symmetry-aware misorientation, IPF-space PSNR, IPF-space SSIM, Boundary F1 and boundary-window Composition F1. Because these metrics have different units, values are normalised to against fixed physical anchors rather than subplotwise min–max ranges, after orienting all metrics so higher bars indicate better performance: median misorientation is inverted on a logarithmic scale between and the crystal-symmetry covering radius ( for cubic and for hexagonal ), PSNR is linear on dB, and SSIM, Boundary F1 and Composition F1 are linear on . Each bar is evaluated from the corresponding scale-specific prediction summary; SG-SRAN uses the reported scale-specific invariant-embedding router configuration.
Grain Boundary analysis
The boundary diagnostics measure complementary reconstruction properties: strict Boundary F1 scores edge localization, boundary-band misorientation scores angular accuracy, and boundary-window Composition F1 scores recovery of the distinct local orientation branches present within an interface window. RCAN leads IN718 Boundary F1, while SG-SRAN leads the IN718 p68/p95 errors and the Ti-6Al-4V mean, median, and boundary-band metrics. tbl:2b reports Composition F1, and Extended Data Table 1 separates its spurious-rate and recall components. Spurious rate is the fraction of SR observations in boundary windows that do not match any HR orientation group within the prescribed misorientation threshold. Nearest-neighbor copying produces few spurious observations but misses minority local orientation branches; SG-SRAN reaches 0.891 on IN718 and 0.782 on Ti-6Al-4V, close to SAN’s respective leading values of 0.899 and 0.804.
The evidence for the mechanism is therefore local, matching the scale at which SG-SRAN forms routed patch assignments. The walkthrough in Extended Data Figure 6 shows that crop-level owner changes and reduced router confidence follow grain-boundary neighborhoods, while support-bank clustering and decoded candidate features remain consistent with the local orientation branches in the same locally isometric latent feature space. This supports the interpretation that SG-SRAN selects among local orientation branches rather than smoothing across the interface. These owner, confidence, and slot-usage maps also make the learned upsampler inspectable: failure modes can be examined in the same latent feature space used for routing, rather than only through the final decoded IPF maps.

Figure 6. Extended Data Fig. 6 SG-SRAN local-routing walkthrough on a boundary-rich IN718 test crop. a, Boundary-rich crop context: the four tiles show the sparse LR input, the decoded SG-SRAN SR map, the HR target and the SR–HR misorientation map, localizing residual error to grain-boundary neighborhoods. b, Crop-scale routing analysis: owner-slot assignments, router confidence, the HR boundary mask and the owner-boundary overlay show that route changes and confidence drops align with grain-boundary neighborhoods. c, Local support-bank decomposition for the marked LR site: the configured local window-support bank, shown here as a IN718 neighborhood, is grouped into orientation-consistent slots in the isometric latent feature space, and the rightmost bars report the center-site slot masses. They summarize local support composition for interpretation; the router itself consumes the complete binary slot masks and token position. d, Patch-level routing decision for the selected HR block: crop context, per-pixel owner slots, selected owners and router margins show which support branch is selected at each HR location. e, Decoded candidate features: each candidate slot is mapped back to an IPF-colored orientation branch, making residual errors interpretable as branch-selection errors among discrete local candidate features rather than unconstrained quaternion averaging. The displayed LR support footprint spans HR pixels at upsampling, and the selected routed output block is HR pixels.
The stagewise decoded-error analysis provides a complementary aggregate check on both crystal systems. The routed upsampler accounts for nearly all of the reduction in decoded error from the encoded input to the final output: 98.0% on IN718 (147 held-out test samples) and 99.7% on Ti-6Al-4V (72 held-out test samples), with subsequent HR refinement changing the mean error by at most in either material (Extended Data Figure 5). The no-routing control isolates the contribution of the routed-patch upsampler: replacing only this component with latent bicubic or nearest-neighbor interpolation, while holding every other component fixed, reduces IN718 Boundary F1 to 0.445 and 0.425, respectively, compared with 0.656 for the full pipeline (Methods).

Figure 5. Stagewise decoded error and sample-level interpretability for SG-SRAN. a, Mean decoded symmetry-aware misorientation for intermediate latent features from the seed-42 SG-SRAN checkpoints on IN718 and Ti-6Al-4V. The routed-upsampler decrease accounts for 98.0% of the total encoded-LR-to-final decrease on IN718 and 99.7% on Ti-6Al-4V; error bars show the standard deviation across per-sample means. b, IN718 Test sample 0 decoded through the SG-SRAN pipeline. From top to bottom, the rows show the observed LR samples, decoded LR embedding, context aggregation, routed patch synthesis, final SG-SRAN output and HR target. Columns show IPF-X, IPF-Y, IPF-Z and the boundary map. LR-space rows are displayed on the HR canvas only at sampled sites; unsampled pixels are black. The IN718 test canvas is HR pixels, with LR rows sampled every fourth HR pixel.
Zero-shot transfer to out-of-distribution FCC and HCP datasets
Because the pipeline is built around a frozen symmetry-aware representation, trained checkpoints can be applied unchanged to new materials within the same symmetry class. Table 3 reports zero-shot transfer to two out-of-distribution targets: a CoNi alloy (FCC) and a Ti-Al alloy (HCP). SG-SRAN remains strongest on CoNi mean, median, p68, p95 and IPF fidelity, while HAN has the highest CoNi Boundary F1 and SAN has the highest CoNi Composition F1. On Ti-Al, the Ti-6Al-4V-trained SG-SRAN gives the best mean, boundary-band mean, Boundary F1, composition F1, p68, p95, p99, PSNR and SSIM among the reported methods. The zero-shot boundary-Composition F1 values are integrated into tbl:3a, with the full spurious-rate and recall decomposition in Extended Data Table 2. This analysis separates the two transfer regimes: SAN narrowly gives the highest CoNi Composition F1 (0.844 versus 0.834 for SG-SRAN), whereas SG-SRAN gives the highest Ti-Al Composition F1 (0.915). Representative IPF-Z maps for two CoNi and two Ti-Al zero-shot samples, including all deterministic and learned baselines, are shown in Extended Data Figure 8.
| Variant | IN718 (FCC, ) | Ti-6Al-4V (HCP, ) |
| latent features + dictionary lookup | 0.0076 ± 0.0004 | 0.0079 ± 0.0001 |
Table 2. Encoder fidelity and orientation-accuracy breakdown. (a) Frozen Reynolds encoder–dictionary-based decoder round-trip fidelity on held-out test samples, as mean and standard deviation of per-sample mean symmetry-aware error in radians. (b) IN718 and Ti-6Al-4V super-resolution accuracy breakdown on the task: trainable parameter count (K), pooled mean and median per-pixel symmetry-aware misorientation, Boundary F1, boundary-window Composition F1, and the interior / boundary-band split of mean misorientation. SG-SRAN parameter counts are configuration-specific: 49K for IN718 and 27K for Ti-6Al-4V. Error-tail percentiles and IPF-space PSNR/SSIM are in Table 1. The composition-F1 column is the pooled boundary-window analysis on the selected prediction summaries, with its spurious-rate and recall components reported in Extended Data Table 1. Lower is better except Boundary F1 and Composition F1; classical interpolants are deterministic and marked “–”, and the Ti-6Al-4V rows use the 12 proper rotations. Except for the composition-F1 analysis column, learned values are five-seed means and the Mean-column ± is the seed-level standard deviation. Classical-row ± values in the Mean column are per-patch test-set standard deviations where available, not seed spreads.

Figure 8. Zero-shot IPF-Z outputs across all baselines for FCC and HCP transfer targets. a, IN718-trained models evaluated without target retraining on two CoNi samples. b, Ti-6Al-4V-trained models evaluated without target retraining on two Ti-Al samples. Within each material panel, two seven-column rows show the display-upsampled LR input, deterministic interpolants, learned quaternion or image-space baselines, SG-SRAN and the HR target. The LR input is nearest-neighbor expanded only for visual alignment; all SR and HR panels are rendered at the native HR grid from saved quaternion arrays, with SLERP and Symm-SLERP recomputed from the same LR samples at render time. Each subfigure has one material-appropriate Laue IPF-Z key, while the misorientation/evaluation protocol uses the rotational subgroups and .
| Method | IN718 → CoNi | Ti-6Al-4V → Ti-Al | ||||
| Spurious ↓ | Recall ↑ | F1 ↑ | Spurious ↓ | Recall ↑ | F1 ↑ | |
| Nearest | 0.054 | 0.609 | 0.741 | 0.014 | 0.699 | 0.818 |
| Bicubic | 0.620 | 0.393 | 0.387 | 0.626 | 0.419 | 0.395 |
| SLERP | 0.598 | 0.332 | 0.364 | 0.569 | 0.393 | 0.411 |
| Symm-SLERP | 0.554 | 0.364 | 0.401 | 0.510 | 0.437 | 0.462 |
| Atindama inpainting [15] | 0.813 | 0.335 | 0.240 | 0.820 | 0.374 | 0.243 |
| Q-RBSA-adapted [4] | 0.323 | 0.600 | 0.636 | 0.007 | 0.769 | 0.867 |
| QEDSR [14] | 0.090 | 0.716 | 0.801 | 0.049 | 0.769 | 0.850 |
| EDSR [14] | 0.090 | 0.711 | 0.798 | 0.092 | 0.753 | 0.823 |
| RCAN [16] | 0.091 | 0.721 | 0.804 | 0.578 | 0.496 | 0.456 |
| SAN [17] | 0.045 | 0.756 | 0.844 | 0.046 | 0.852 | 0.900 |
| HAN [18] | 0.160 | 0.693 | 0.760 | 0.283 | 0.705 | 0.711 |
| SG-SRAN (ours) | 0.054 | 0.745 | 0.834 | 0.005 | 0.847 | 0.915 |
Table 2.
| Method | Params (K) | p68 () | p95 () | p99 () | PSNR (dB) | SSIM |
| (FCC): zero-shot test split n = 20 | ||||||
| Nearest | – | 0.42 | 24.04 | 59.76 | 19.69 | 0.770 |
| Bicubic | – | 0.95 | 26.12 | 52.14 | 17.61 | 0.680 |
| SLERP | – | 0.46 | 26.42 | 50.57 | 17.89 | 0.728 |
| Symm-SLERP | – | 0.46 | 22.37 | 47.61 | 18.36 | 0.733 |
| Atindama inpainting [15] | 25,784 | 5.47 | 40.48 | 48.59 | 14.03 | 0.402 |
| Q-RBSA-adapted [4] | 6,175 | 1.16 | 5.62 | 46.23 | 21.04 | 0.779 |
| QEDSR [14] | 1,815 | 0.34 | 1.62 | 53.84 | 22.28 | 0.831 |
| EDSR [14] | 7,241 | 0.37 | 1.67 | 53.05 | 22.30 | 0.831 |
| RCAN [16] | 15,594 | 0.36 | 1.62 | 53.69 | 22.25 | 0.828 |
| SAN [17] | 15,862 | 0.37 | 1.43 | 54.71 | 22.51 | 0.837 |
| HAN [18] | 16,073 | 0.49 | 2.64 | 50.14 | 22.44 | 0.813 |
| SG-SRAN (ours) | 49 | 0.32 | 1.20 | 54.97 | 22.66 | 0.842 |
| --- (HCP): zero-shot test split n = 133 | ||||||
| Nearest | – | 0.49 | 1.31 | 64.92 | 22.15 | 0.828 |
| Bicubic | – | 0.72 | 22.50 | 71.33 | 18.75 | 0.732 |
| SLERP | – | 0.48 | 25.55 | 73.02 | 18.81 | 0.771 |
| Symm-SLERP | – | 0.47 | 16.13 | 46.46 | 21.50 | 0.814 |
| Atindama inpainting [15] | 25,784 | 11.10 | 51.19 | 80.47 | 13.49 | 0.334 |
| Q-RBSA-adapted [4] | 6,175 | 0.96 | 1.60 | 56.70 | 23.48 | 0.828 |
| QEDSR [14] | 1,815 | 0.53 | 1.44 | 54.82 | 23.69 | 0.833 |
| EDSR [14] | 7,241 | 0.59 | 1.67 | 51.04 | 23.49 | 0.824 |
| RCAN [16] | 15,594 | 3.79 | 17.65 | 50.42 | 18.25 | 0.599 |
| SAN [17] | 15,862 | 0.89 | 2.64 | 54.47 | 23.31 | 0.763 |
| HAN [18] | 16,073 | 2.16 | 6.71 | 47.56 | 21.05 | 0.687 |
| SG-SRAN (ours) | 27 | 0.40 | 0.88 | 42.61 | 25.31 | 0.886 |
Table 3b. Error-tail percentiles and IPF image fidelity.
Discussion
Orientation super-resolution should be treated as reconstruction on the symmetry quotient , rather than as interpolation of generic image channels. By combining this geometry with local routing at grain boundaries, SG-SRAN produces accurate, physically valid outputs through an inspectable routed upsampling process. Across five seeds, the compact 27–49k-parameter model performs strongly on tail-error and boundary-sensitive metrics across both crystal systems, although competing methods lead individual measures. Orientation-field super-resolution should therefore be assessed jointly through angular accuracy, boundary localization, local orientation-branch recovery, and compactness.
The boundary performance arises from the geometry-aware representation and routed upsampler acting together. The right--invariant, locally isometric encoder gives nearby Euclidean latent distances a first-order interpretation as crystallographic misorientation. Routing changes and reduced confidence concentrate near grain boundaries, where the upsampler forms distinct candidate features and assigns them to the high-resolution output grid. The slot assignments are made during upsampling; subsequent convolution layers refine the latent features without re-clustering the local support, while boundary-window Composition F1 measures whether the corresponding local branches survive in the final output. SG-SRAN therefore preserves interfaces by assigning among local orientation branches rather than averaging across them.
The principal limitations are difficult orientation selection, decoding cost, and validation breadth. On the boundary-rich Ti-6Al-4V benchmark, the most extreme – errors remain comparable to those of other learned methods; because dictionary-based decoding always returns a canonical unit quaternion, these errors are valid but incorrect orientation choices rather than invalid states. Dictionary search and refinement dominate inference, requiring approximately 4.63 s per patch compared with 28.5 ms for encoding and latent super-resolution. The CoNi and Ti–Al transfers provide initial evidence within FCC and HCP symmetry classes, but broader cross-specimen validation remains necessary.
These findings point directly to router calibration, learnable decoding, and broader cross-specimen validation. More broadly, they support a representation-first strategy: quotient geometry defines physical equivalence, local-isometry normalization gives nearby latent distances a crystallographic meaning, and routed upsampling preserves distinct local branches at interfaces. Building these properties into both the representation and the upsampling process, rather than correcting a Euclidean image model after the fact, may provide a useful strategy for reconstructing fields defined only up to a physical symmetry.
Methods
Global notation.
We write and for the low- and high-resolution quaternion fields, and for the frozen encoder and dictionary-based decoder, and for the symmetry-aware misorientation metric of (2). The spatial upsampling factor is and the task studied here is . Latent feature fields after successive backbone stages are denoted as in equation (8).
Crystal orientations and quotient geometry
In this work, crystal orientation is represented as an active crystal-to-specimen rotation , defined by , where and are the coordinates of the same physical direction in right-handed crystal and specimen frames, respectively. We write and represent it by a scalar-first Hamilton unit quaternion , acting as , where denotes the pure quaternion with zero scalar part and vector part , and extracts the vector part of a quaternion. Under this convention, composition satisfies . The antipodal quaternions and represent the same rotation, so . When an EBSD tool supplies a passive specimen-to-crystal orientation, we denote it by ; after expressing it in our crystal and specimen frames, we convert it once at import as and . Thereafter, and always denote active crystal-to-specimen orientations, with conversion back to the source convention performed only for visualization. Our frame, handedness, and active/passive conventions follow Britton et al. [19].
The crystal lattice has a finite proper rotational crystal point group , whose elements act actively in the crystal frame. For each , let be a unit-quaternion lift satisfying . Because maps crystal coordinates to specimen coordinates, crystal symmetry acts from the right: , equivalently . Thus, the physical orientation space is the quotient manifold . For face-centered cubic (FCC) metals, is the cubic rotation group with ; for hexagonal close-packed (HCP) metals, it is the dihedral rotation group with .
The fundamental zone (FZ) selects one representative from each symmetry orbit, up to boundary ties. Enumerating , we form
where is the scalar part of . We define and choose its antipodal sign such that . This selects the symmetry-equivalent rotation closest to the identity; remaining boundary ties are resolved deterministically. The symmetry-aware angular error is
which is zero for symmetry-equivalent orientations and reduces to when .
Locally isometric orientation encoder
The encoder maps each pixel quaternion to a finite-dimensional Euclidean vector that is (i) exactly invariant under and (ii) locally isometric such that small latent distances track small geodesic distances on near the identity equivalence class.
Symmetry-invariant feature blocks.
For each angular degree , the real (tesseral) Wigner matrix is the degree- irreducible representation of , obtained from the complex Wigner- matrices by the unitary change of basis to real tesseral harmonics; it preserves the homomorphism and keeps all features and weights real. The Reynolds projector
projects onto the -fixed subspace . With and an orthonormal basis of , the crystal-invariant feature block at degree is
where is the active rotation of . Because for all , invariance under follows from the homomorphism property and :
Two group actions must be distinguished: the encoder is invariant under the right crystal-symmetry action (equation (5)), which collapses symmetry-equivalent orientations, and equivariant under the left global-rotation action , with (i.e. ), for which . The backbone operates on these right--invariant feature blocks as left--equivariant e3nn representations—processing orientation as tensors—while the routing and feature-matching loss introduced in Section 3.3 are constructed from invariant contractions of these equivariant features. The raw descriptor is the concatenation over the selected invariant degrees .
Constructing the invariant bases.
The invariant bases are built once, offline, and stored as non-trainable buffers. For each degree in the truncation set, we form the Reynolds projector of equation (3) in the real tesseral Wigner basis, symmetrize it numerically, and take the eigenvectors with eigenvalue one as an orthonormal basis for the fixed subspace. The scalar block is mathematically invariant but is not emitted by the orientation encoder because it carries no orientation information. The emitted descriptor therefore contains only non-scalar invariant blocks.
For FCC (), the descriptor is truncated at . Since and , the first non-trivial cubic invariant is the 9-dimensional block. For HCP (), the descriptor is truncated at and keeps every non-scalar projector component below that cutoff: , , and . The HCP encoder output is therefore , i.e. dimensions, and serves as both the encoder output and the backbone working space.
Local-isometry normalization.
The Reynolds projector fixes the invariant subspace , but it does not fix the Euclidean scale assigned to each emitted feature block. We therefore calibrate the scale of the actual encoder feature map at the identity. For each retained block , let , with , denote the Jacobian of with respect to infinitesimal axis-angle perturbations at . The normalized encoder is
where the scalar block weights are chosen so that
These scalar weights change only the metric scale; they do not change the invariant subspaces or the right- invariance established in (5).
For FCC (), only the block is retained through . Its unnormalized local metric is to numerical precision, so we use
For HCP (), the retained blocks through are . Their unnormalized local metric contributions are , , . With equal weights, the combined local metric would be , corresponding to the small anisotropy . The implemented normalization instead uses
which makes the combined local metric equal to up to numerical precision.
Thus, for a sufficiently small axis-angle vector , let denote its skew-symmetric matrix, and let be a unit-quaternion lift satisfying . Let , so that . Then
Consequently, mean-squared error in the normalized latent space tracks squared geodesic orientation error to leading order. This local-isometry construction follows [5]. In our numerical analysis, the approximation remains tight for misorientations up to approximately rad, or , beyond which higher-order Taylor terms become non-negligible.
Completeness and truncation.
In the untruncated harmonic construction, all nonconstant degrees with together separate quotient orientations in . The implemented encoders retain finite degree sets, for FCC and for HCP, yielding finite-dimensional representations. The selected degrees determine the harmonic content, latent dimension and computational cost of each encoder. The local-isometry analysis characterizes the infinitesimal geometry of these representations, while the encoder–dictionary round-trip error in Table 2 measures orientation recovery by the implemented encoder–dictionary pair on the orientations used in the experiments.
Orientation-cluster routed-patch super-resolution backbone
The SG-SRAN super-resolution stage uses an orientation-cluster routed-patch upsampler that acts directly on the locally isometric latent field, predicting one high-resolution (HR) latent patch per low-resolution (LR) pixel from a local, orientation-conditioned support bank. The emitted patch is for the SR task. Every stage operates on the crystal-invariant latent of the encoder (the 9-dimensional block for FCC, the 40-dimensional descriptor for HCP), so crystal invariance of the predicted orientation is preserved end to end.
Let be the encoded LR feature map, with batch size , flattened LR grid , and latent dimension . The pipeline is
where and are feature-similarity-masked equivariant convolutions, is the SG-SRAN upsampler with used for clustering and used for token synthesis, and denotes the optional second HR refinement pass. SG-SRAN uses a configured odd local support window (e.g. for IN718 and for Ti-6Al-4V), up to candidate slots per LR pixel, and an 8-neighbor feature-space clustering graph. The router is a small MLP (hidden = 64) over the complete binary slot-composition image and the learned subpixel position embedding, with a center-prior weight of 5.0, and the cross-attention proposal uses a 54-dimensional hidden MLP and a 32-dimensional query/key space.
Feature-similarity-masked equivariant convolutions.
Each convolution forms local context only from neighbors whose latent direction is aligned with the center pixel, using a cosine-similarity threshold between the locally isometric feature vectors. This cosine gate is used as a scale-invariant feature-direction mask; the calibrated angular neighborhoods are the L2 thresholds in equation 10. The masked neighborhood average is combined with the center feature through an e3nn fully connected tensor product [11] followed by a residual connection, keeping the update exactly equivariant while suppressing crystallographically dissimilar neighbors. The IN718 configuration uses one LR convolution with residual weight 1.0, followed by two HR refinements with residual weight 0.3; the Ti-6Al-4V configuration uses one LR convolution with residual weight 1.0, followed by one HR refinement with residual weight 0.2, avoiding broader post-routing smoothing on the finer HCP boundary network.
Local support banks and feature-space clustering.
For each LR pixel , SG-SRAN extracts co-registered clustering and value banks
in the reported configurations,
with replicate padding at boundaries. The frozen encoded features determine cluster membership, whereas the context-refined equivariant features provide the values later synthesized into HR tokens. At upsampling this LR window spans roughly HR pixels per side (for example, 36 HR pixels when and 20 when ). The nodes form an 8-neighbor graph in which two adjacent nodes are linked when
corresponding under the local isometry to in the IN718 configuration and in the Ti-6Al-4V configuration. Connected components of this graph define local orientation regions; the largest components are packed into deterministic slots, and for each slot SG-SRAN stores a binary membership mask and a compact metadata vector (slot validity, slot-rank one-hot code, relative mass, centroid and spatial dispersion).
Token-conditioned cross-attention proposal.
SG-SRAN synthesises each candidate HR-token feature using a single token-conditioned cross-attention operation over the members of a local orientation slot. Let the emitted patch be indexed by , with for the task, and let denote the learned subpixel position embedding for token . For slot at LR site , define its member set as . A query constructed from the slot metadata and the learned subpixel position embedding attends over these members; keys are derived from member window coordinates and invariant norms of the context-refined member features, while values are the context-refined member features :
where is a masked softmax over . This directly produces an anchor-less, token-conditioned proposal for each slot. Since the attention weights are invariant scalars and the values are equivariant features, is equivariant. This single attention operation performs both pooling and proposal generation, allowing different tokens within the same HR patch to attend to different members of the same local orientation cluster.
Geometric routing and owner selection.
For each token the router emits logits over slots from the complete binary slot-composition image and the token position,
where is the learned subpixel position embedding for token , when slot contains the window center and is zero otherwise, and biases the decision toward the center slot when occupancy evidence is weak. The reported MLP router jointly processes all mask entries and the token-position embedding, making routing geometric and separate from orientation-feature proposal synthesis; logits for empty slots are masked before selection. A hard owner selects the slot for token , implemented with a straight-through estimator so gradients flow through the logits, and the emitted token is the cross-attention proposal of that slot,
The selected tokens are reshaped into the HR patch for site and used directly as the HR latent prediction; residual connections appear only within the downstream HR refinement convolutions, after routed token selection.
Dictionary-based decoder and fundamental-zone canonicalization
The dictionary-based decoder maps a predicted latent feature back to a unit quaternion in two steps. First, a precomputed lookup table of fundamental-zone quaternions sampled by cubochoric sampling [12, 13] is encoded once, , and the nearest table entries to are retrieved by chunked squared-Euclidean search. Second, starting from these seeds, quaternion candidates are refined in parallel by Adam (60 steps, learning rate 0.03) to minimize , with renormalized to the unit sphere after each step, and the candidate with the lowest residual is selected and canonicalized to the fundamental zone by (1). The finite cubochoric dictionary sets the small frozen-interface floor reported in Table 2a; that floor is below the downstream SR errors and is separated from the trainable upsampler when reporting representation fidelity. The encoder and dictionary-based decoder are frozen throughout training; only the SG-SRAN upsampler and the refinement convolutions are optimized. The standalone representation-fidelity result in Table 2a reports the dense lookup-only floor, isolating the frozen encoder–dictionary-based decoder interface.
On an NVIDIA A100-PCIE-40GB, end-to-end IN718 inference takes 4.65 s per HR patch. The dictionary-based decoder takes s (70.6 s per HR pixel), compared with 0.70 ms for the encoder and 27.8 ms for the SR backbone. The dictionary search and refinement therefore set the current inference latency.
Training objective and implementation
The SG-SRAN backbone is trained with a feature-space mean-squared-error loss,
where is the predicted HR latent from the full pipeline (equation (8)), is the frozen encoded HR target (computed under torch.no_grad() and detached), and is the number of HR pixels per sample. By equation (7), minimizing is a first-order surrogate for the squared geodesic orientation error,
with the per-pixel misorientation angle; the correspondence is exact to first order in the locally isometric small-angle regime and remains an effective objective empirically at larger angles. The backbone is trained with this loss alone, and the geometric router and its straight-through owner assignment are learned end to end from the feature-matching gradient.
The model is implemented in PyTorch with e3nn [11] for the equivariant tensor products. At run time, Reynolds invariant features are evaluated from precomputed projector bases using spherical-harmonic synthesis of , entirely on GPU; the Wigner-D eigensolve and angular-momentum-generator machinery enters only in the offline construction of the projectors and bases. Optimization uses AdamW (weight decay ) at peak learning rate with cosine annealing (2 warmup epochs, minimum learning rate ), gradient clipping at norm 1.0, and bfloat16 mixed precision with TensorFloat-32, for 150 epochs at batch size 2 on IN718 and 5 on Ti-6Al-4V. The aggregate benchmark reports five independent SG-SRAN seeds (42–46); seed 42 is used for the representative qualitative figures and mechanism analysis.
Datasets and evaluation protocol
Materials.
Two EBSD datasets are used for in-domain training and testing. The IN718 source map is an EBSD orientation field from the multi-modal 3D microstructure dataset of Stinville et al. [20]; the Ti-6Al-4V source data are drawn from the Ti-6Al-4V 3D microstructure dataset of Jangid et al. [21]. The IN718 dataset (nickel superalloy, FCC, ) provides 1174/147/147 train/validation/test paired patches at HR size ; the Ti-6Al-4V dataset (titanium alloy, HCP, ) provides 581/73/72 train/validation/test paired patches at HR size . SG-SRAN is trained and evaluated end to end on both materials under the task. The frozen encoder–dictionary-based decoder round-trip baseline in Table 2a is reported on the IN718 and Ti-6Al-4V test splits as an FCC/HCP representation analysis rather than as a trained SR result. The dataset manifests record patch counts, sizes, and crystal symmetry; train, validation, and test splits are disjoint at the patch-file level. The zero-shot transfer table uses two held-out targets: CoNi (FCC) and Ti-Al (HCP). The CoNi scan is a held-out FCC transfer map prepared with the same manifest workflow, and the Ti-Al target uses an HCP Ti-Al split with 1064/133/133 train/validation/test paired patches; only its test split is used for zero-shot evaluation.
Training-pair construction.
For each source map, quaternions are normalized, mapped to a consistent hemisphere, and reduced to the fundamental zone. Random aligned HR patches are extracted, and the matching LR patches are formed by regular stride- subsampling along each axis (), so LR and HR remain spatially registered without an extra interpolation step. This construction models the spatial-resolution component of the LR/HR trade-off.
Baselines.
We compare against four classical interpolants. Nearest: nearest-neighbor upsampling with quaternion renormalization. Bicubic: per-channel bicubic interpolation, renormalized to unit norm. SLERP: pairwise spherical linear interpolation [2] on the quaternion sphere. Symmetry-aware SLERP: bilinear SLERP in which each pair is first symmetry-resolved to the shortest crystallographic path. All four are generated from the same LR test patches used by SG-SRAN. Seven learned baselines are trained independently on the matched IN718 and Ti-6Al-4V train/validation splits and evaluated on the corresponding and held-out test patches at , each over five training seeds (42–46). Q-RBSA-adapted [4] retains the released 10-block, 128-feature quaternion residual self-attention head and body. Its local adapter converts the passive scalar-first dataset quaternions to the active scalar-last convention expected by the released implementation, selects the 24-element proper subgroup for IN718 or the 12-element proper subgroup for Ti-6Al-4V from the dataset point-group metadata for the rotational loss, and exports predictions through the same fundamental-zone and held-out evaluation protocol as SG-SRAN. The model replaces the released one-dimensional pixel-shuffle tail with a dimension-matched two-dimensional quaternion pixel-shuffle tail. It is trained from scratch for 300 epochs and contains 6.17 million parameters, using AdamW with cosine decay (peak learning rate , weight decay , two warmup epochs, seeds 42–46).
QEDSR is a quaternion adaptation of EDSR [14]: the standard EDSR residual body operates on four quaternion channels (active scalar-last, with the passive scalar-first convention restored after inference), followed by a two-dimensional quaternion pixel-shuffle tail. It contains 1.82M trainable parameters and is trained from scratch for 150 epochs with AdamW (peak learning rate , weight decay , two warmup epochs, five seeds 42–46) under the same dataset-selected or rotational loss and held-out evaluator as SG-SRAN. The corresponding real-valued EDSR baseline uses the same residual-block schedule and two-dimensional pixel-shuffle tail on four scalar quaternion channels, without quaternion convolutions or quaternion attention. It contains 7.24M trainable parameters and is trained and evaluated with the same five-seed held-out protocol as the other learned baselines.
RCAN [16], SAN [17] and HAN [18] retain their published residual-channel, second-order and holistic attention bodies, respectively. RGB mean-shift layers are omitted and active scalar-last quaternions are supplied as four channels; the passive scalar-first convention is restored after inference. RCAN uses 10 residual groups with 20 blocks each, SAN uses 20 groups with 10 blocks each, and HAN uses 10 groups with 20 blocks each, all with 64 features. The models use a two-dimensional pixel-shuffle tail; parameter counts are 15.59M (RCAN), 15.86M (SAN) and 16.07M (HAN). Each dataset is trained independently for 300 epochs with AdamW, cosine decay, weight decay , two warmup epochs and five seeds (42–46). SAN uses its published base rate on both materials; RCAN/HAN use on IN718 and the authors’ rate on Ti-6Al-4V after the larger rate proved unstable under the proper- loss. SAN covariance pooling is evaluated by the exactly equivalent centered product rather than materializing an centering matrix, and its trace-normalized Newton–Schulz square root uses native autograd with a machine-precision diagonal regularizer. All three models use the same dataset-selected or rotational loss and held-out evaluator as SG-SRAN.
The Atindama baseline adapts the partial-convolution U-Net from the EBSD restoration pipeline of Atindama et al. [15]. HR orientations are converted to normalized intrinsic ZXZ Euler channels, the observed stride- samples define a periodic known-pixel mask, and the network predicts the missing pixels while known samples are copied exactly into the output. The model is trained for 150 epochs with Adam (learning rate , five seeds 42–46) using squared geodesic error on unknown pixels without crystal-symmetry minimization. Training is therefore symmetry-agnostic, but all exported Ti predictions are reduced and evaluated with the same 12 proper rotations used by the other methods. Each model has 25.78 million trainable parameters. The published Criminisi refinement is not applied: the regular SR masks contain no fully known source patch, so its exemplar-search precondition is unsatisfied.
An architecture-matched no-routing control isolates the routing mechanism: the encoder, refinement convolutions and dictionary-based decoder are fixed, and only the routed-patch upsampler is replaced by bicubic or nearest interpolation in the latent space. The control was run for the IN718 Reynolds checkpoint under a single-seed protocol; its results are reported alongside the boundary metrics in Results.
Evaluation metrics.
Angular accuracy is measured by the symmetry-aware misorientation (equation (2)). The FCC metric minimizes over the 24 proper rotations of , and the HCP metric minimizes over the 12 proper rotations of , so each material is evaluated with its own rotational crystal symmetry. For the frozen encoder–dictionary-based decoder round trip, we report the mean and standard deviation of the per-sample mean in radians on the held-out IN718 and Ti-6Al-4V test splits. For the super-resolution benchmark, angular errors are reported in degrees and pooled over all HR pixels in the 147 held-out IN718 patches or the 72 held-out Ti-6Al-4V patches. The mean, median, p68, p95 and p99 summarize this pooled per-pixel distribution; per-patch or seed-level standard deviations are reported where uncertainty is shown.
To evaluate boundary recovery, we also split into grain-interior and grain-boundary-band pixels. The HR boundary mask is obtained by thresholding 4-connected nearest-neighbor crystallographic misorientation at , and the boundary band contains pixels within five pixels of this HR mask. Boundary F1 is computed from predicted and HR boundary masks using the same rule, with TP, FP and FN accumulated over all patches: .
Inverse-pole-figure (IPF) maps color each pixel by the crystal direction parallel to a selected sample axis (, or ), making grains and sharp orientation changes visually legible; we render them with orix [22, 23]. IPF-space PSNR and SSIM are computed on the IPF-X, Y and Z renderings of the prediction against the HR reference and averaged over the three axes. All angular metrics are computed directly on the quaternion orientation field.
For the boundary-window composition analysis, an 8-neighbor symmetry-aware rule defines HR boundary centers. Each center contributes clipped HR and SR windows. Transitive pairwise matches within partition the HR window into orientation groups. The spurious rate is the fraction of SR observations farther than from every HR group; composition recall is the fraction of HR groups recovered by at least one SR observation; and Composition F1 is the harmonic mean of recall and precision, with precision . Overlapping boundary windows are counted separately.
Data availability
EBSD datasets and encoded lookup tables will be deposited in a public repository upon publication. Requests for data prior to publication should be directed to the corresponding author.
Code availability
Code implementing the SG-SRAN-based latent super-resolution model, the locally isometric encoder, and the dictionary-based decoder will be released under the Apache 2.0 license on a public GitHub repository upon publication. The release will include trained checkpoints for the IN718 and Ti-6Al-4V models, the precomputed Reynolds basis files, and the cubochoric-sampled dictionary lookup tables ( per symmetry class for FCC and HCP), which would otherwise be non-trivial to recompute.
Acknowledgements
The authors acknowledge Neal Brodnik (Postdoc, UCSB) for his intellectual contributions and early discussions on this problem. We also thank Joaquin Giorgio (M.S., UCSB) for his prior work on SLERP-based orientation upsampling, which informed the baseline comparison in this study. This research was supported in part by the NSF CSSI Award #2411453. TMP and MPE also acknowledge the support of Army Research Office (ARO) MURI Grant W911NF-25-2-0164.
Competing interests
The authors declare no competing interests.
Author contributions
U.G. developed the direct Reynolds-projection embedding used in the final HCP and FCC representations, along with its calibration, and developed the anchorless routing architecture. U.G. led the experimental validation and conducted the experiments, and wrote the manuscript. W.Z. proposed applying the Reynolds operator to enforce symmetry-compliant network outputs and developed the original locally isometric embedding and its calibration, using rank-tensor decomposition. U.G. and W.Z. jointly developed and tested early implementations of this approach. Both authors revised the manuscript; U.G. verified the manuscript’s claims against the implementation. M.P.E. acquired and provided EBSD datasets, contributed to experimental design, and reviewed the manuscript. S.H.D. provided guidance on data interpretation, and reviewed the manuscript. T.M.P. provided guidance on materials characterization and EBSD methodology, and reviewed the manuscript. B.S.M. supervised the overall project, provided guidance on the manuscript, and reviewed and edited the manuscript.
Extended Data

Figure 1. Electron backscatter diffraction (EBSD) and the inverse pole figure (IPF) representation. a, A focused electron beam is scanned across a specimen tilted to 70° in a scanning electron microscope, and the backscatter diffraction pattern collected at each scan point is indexed to a crystallographic orientation; the resulting orientation field is rendered as an inverse pole-figure (IPF-Z) map. b, In the IPF representation, color encodes the crystal direction aligned with the sample Z axis according to the IPF color key (colored fundamental triangle). Grains appear as contiguous regions of slowly varying color separated by sharp orientation boundaries; recovering these boundaries at high spatial resolution, under a fixed acquisition budget, is the objective of this work.
Extended Data Table 2 Zero-shot boundary-window composition analysis. The same HR-boundary-window composition diagnostic as Extended Data Table 1 is applied to the two headline out-of-distribution targets. Spurious rate is lower better; recall and Composition F1 are higher better.
References
- [1]Humphreys, F. J. Review of grain and subgrain characterisation by electron backscatter diffraction. Journal of Materials Science 36, 3833–3854 (2001).DOI
- [2]Shoemake, K. Animating rotation with quaternion curves. ACM SIGGRAPH Computer Graphics 19, 245–254 (1985).
- [3]Seret, A., Moussa, C. & Jonas, J. J. Application of super-resolution convolutional neural networks to EBSD maps. Materials Characterization 153, 1–9 (2019).
- [4]Jangid, D. K. et al. Q-RBSA: High-resolution 3d EBSD map generation using an efficient quaternion transformer network. npj Computational Materials 10, 27 (2024).arxiv.org/abs/2303.10722
- [5]Hielscher, R. & Lippert, L. Locally isometric embeddings of quotients of the rotation group modulo finite symmetries. Journal of Multivariate Analysis 185, 104764 (2021).arxiv.org/abs/2007.09664
- [6]Arnold, R., Jupp, P. E. & Schaeben, H. Statistics of ambiguous rotations. Journal of Multivariate Analysis 165, 73–85 (2018).arxiv.org/abs/1701.01579
- [7]Lenthe, W., Singh, S. & Graef, M. D. A spherical harmonic transform approach to the indexing of electron back-scattered diffraction patterns. Ultramicroscopy 207, 112841 (2019).DOI
- [8]Cohen, T. S. & Welling, M. Group equivariant convolutional networks. Proceedings of ICML (2016).arxiv.org/abs/1602.07576
- [9]Weiler, M. & Cesa, G. General E(2)-equivariant steerable CNNs. Advances in NeurIPS (2019).arxiv.org/abs/1911.08251
- [10]Thomas, N. et al. Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds. arXiv:1802.08219 (2018).arxiv.org/abs/1802.08219
- [11]Geiger, M. & Smidt, T. e3nn: Euclidean neural networks. arXiv:2207.09453 (2022).arxiv.org/abs/2207.09453
- [12]Rosca, D., Morawiec, A. & De Graef, M. A new method of constructing a grid in the space of 3d rotations and its applications to texture analysis. Modelling and Simulation in Materials Science and Engineering 22, 075013 (2014).DOI
- [13]Rowenhorst, D. et al. Consistent representations of and conversions between 3d rotations. Modelling and Simulation in Materials Science and Engineering 23, 083501 (2015).DOI
- [14]Lim, B., Son, S., Kim, H., Nah, S. & Lee, K. M. Enhanced deep residual networks for single image super-resolution. CVPR Workshops (2017).arxiv.org/abs/1707.02921
- [15]Atindama, E., Lef, P., Doğan, G. & Athavale, P. Restoration of noisy orientation maps from electron backscatter diffraction imaging. Integrating Materials and Manufacturing Innovation 12, 251–266 (2023).DOI
- [16]Zhang, Y. et al. Image super-resolution using very deep residual channel attention networks. European Conference on Computer Vision 286–301 (2018).arxiv.org/abs/1807.02758
- [17]Dai, T., Cai, J., Zhang, Y., Xia, S.-T. & Zhang, L. Second-order attention network for single image super-resolution. IEEE/CVF Conference on Computer Vision and Pattern Recognition 11065–11074 (2019).DOI
- [18]Niu, B. et al. Single image super-resolution via a holistic attention network. European Conference on Computer Vision 191–207 (2020).arxiv.org/abs/2008.08767
- [19]Britton, T. B. et al. Tutorial: Crystal orientations and EBSD—or which way is up? Materials Characterization 117, 113–126 (2016).DOI
- [20]Stinville, J. C. et al. Multi-modal dataset of a polycrystalline metallic material: 3D microstructure and deformation fields. Scientific Data 9, 460 (2022).DOI
- [21]Jangid, D. et al. Titanium 3d microstructure for physics-based generative models: A dataset and primer. ICML 2023 Workshop on the Synergy of Scientific and Machine Learning Modelling (2023).
- [22]Johnstone, D. N., Martineau, B. H., Crout, P., Midgley, P. A. & Eggeman, A. S. Density-based clustering of crystal (mis)orientations and the orix Python library. Journal of Applied Crystallography 53, 1293–1298 (2020).DOI
- [23]orix developers. orix: Analysing crystal orientations and symmetry in Python, version 0.13.0 (2024). Software.