betaFeedback

WorksPhysics and Astronomy

A Geometry-Guided Feature-Space Framework for Interpretable Unsupervised 4D-STEM Analysis of Complex Multiphase Materials

Abstract

Rapid, high-fidelity analysis of four-dimensional scanning transmission electron microscopy (4D-STEM) datasets from heterogeneous materials remains challenging because coexisting phases, orientation variants, local strain gradients, and intrinsic TEM thin-foil effects produce complex, overlapping diffraction signatures. Here, we present a machine-learning (ML) framework that combines physics-informed geometric descriptors with weighted high-dimensional embeddings to identify latent structural patterns with minimal prior knowledge. By constructing a regularized descriptor space before embedding and density-based clustering, the method improves cluster stability, suppresses noise-driven fragmentation, and reduces sensitivity to embedding hyperparameters. Spatial weighting produced the largest stability gain, increasing median cluster persistence from 0.04 to 0.55 while reducing the noise fraction from approximately 22% to 0%. Rotation weighting increased DBCV from 0.80 to 0.92, although the noise fraction rose from 0.2% to 1.0%, highlighting a trade-off between improved density separation and noise rejection. We validate the framework on two benchmark systems: Si/Six_xGe1−x_{1-x} epitaxial layers with continuous lattice variation and deformation, and multiphase Ti-Nb alloys containing distinct phases and orientation variants. Comparisons with a pretrained vanilla autoencoder and model-free pattern vectorization show that the physics-informed geometric descriptor provides higher cluster fidelity, greater interpretability, and substantially lower dimensionality.

Introduction

4D-STEM generates vast, information-rich datasets, driven by rapid advances in high-throughput detector architectures [1]. In this modality, the electron beam records a series of diffraction patterns that encode highly correlated information about a material’s phase, deformation, and orientation structure [2, 3, 4, 5, 6]. Data volume increases further in 5D-STEM with the addition of a temporal axis, resulting in terabytes of memory usage per user session. The key challenge is no longer data acquisition, but rapidly and reliably navigating these complex datasets and transforming them into physics-informed, readily interpretable visualizations and analyses necessary to derive deep materials inferences [1].

Current analysis pipelines remain highly parameterized, sequentially complex, computationally demanding, and domain-specific, and they often fail to resolve the latent relationships that govern material behavior. For example, the widely used template-matching approach for phase and orientation analysis requires prior crystallographic models and extensive template libraries, which become increasingly difficult to construct and search for complex multiphase systems, orientation variants, strained regions, or previously unknown phases [5, 7, 8]. Several 4D-STEM workflows use unsupervised ML and clustering to improve analysis while preserving the important details of large diffraction datasets [9, 10, 11, 12, 13, 14]. Prior studies have combined clustering with principal component analysis (PCA) for linear dimensionality reduction, while others employ uniform manifold approximation and projection (UMAP) for nonlinear embedding and visualization [9, 10, 11, 12, 13, 14, 15, 16, 17, 18].

Notable examples include latent-space mapping of atomically-resolved electric fields in Si-doped graphene from atomic-resolution STEM data [12] and a high-dimensional diffraction-vector workflow that uses guided multi-stage clustering to delineate phase, rotation, and bend contour effects in 2D materials and relatively thin TEM samples [11, 17, 18].

In more complex, thicker TEM samples, clustering robustness is often improved through preprocessing steps. Common strategies include masking to suppress the dominant central beam or Fourier-transform-based treatments to reduce tilt and thickness effects before flattening diffraction patterns into high-dimensional intensity descriptors [14]. More recently, energy-dispersive spectroscopy chemical maps have been incorporated to enable multimodal clustering approaches [19, 20, 21, 22].

In 4D-STEM, physically meaningful analysis requires a structured representation that preserves both diffraction-pattern signatures and their spatial context before applying neighborhood-based learning, manifold embedding, and clustering [9, 23, 24]. The goal is not merely dimensionality reduction but organizing diffraction-derived descriptors into a form that makes interpretation easier and robust.

A central challenge in analyzing these large, information-rich datasets is the curse of dimensionality. For example, a single 128x128-pixel diffraction pattern can be flattened into a 16,384-dimensional feature vector, where each dimension corresponds to the intensity recorded at a single detector pixel. As the dimensionality of the feature spaces expands, as in raw diffraction pattern representations, the data become increasingly sparse, local neighborhoods lose contrast, and the distinction between near and far points weakens [25, 26, 27]. These effects reduce the effectiveness of distance-based similarity measures and nearest-neighbor searches, increase sampling requirements, and make it more difficult to construct stable and physically interpretable low-dimensional representations of the data [19, 28]. At the same time, many high-dimensional datasets contain fingerprint information that can be represented in far fewer dimensions than their raw measurements suggest [29]. These limitations motivate an approach in which feature-space construction is treated as a primary part of the analysis rather than as a preprocessing detail.

In this work, we present a feature-engineering-focused machine learning framework for 4D-STEM analysis that emphasizes deliberate construction of the descriptor space before embedding and clustering. By using feature extraction to control dimensionality, the framework accelerates interpretation. It enables robust, physics-informed analysis of complex structural materials without labeled training data or prior phase assignments.

Our central premise is that the main challenge in unsupervised 4D-STEM cluster analysis is not only selecting an embedding or clustering algorithm but also designing and regularizing an interpretable feature space that preserves the underlying crystallographic relationships within the data. While existing workflows often emphasize diffraction-pattern preprocessing, black-box approaches to feature extraction, such as autoencoders, or downstream dimensionality reduction and clustering, comparatively less attention has been devoted to the physics-informed design and regularization of diffraction feature representations that preserve crystallographically meaningful information [9, 10, 11, 12, 13, 14, 15, 16, 19, 20, 21, 28, 30].

Feature-space construction and embedding

The workflow begins by constructing a physically interpretable feature space, referred to here as physics-informed descriptors (PIDs), from diffraction-disc centers detected by our neural-network object-detection model (Fig. 1) [31]. Before disc detection, each diffraction pattern is log-normalized and processed using Fourier-transform-based autocorrelation. This autocorrelation step reduces the influence of thickness- and tilt-related dynamical effects, similar to the robustness sought in cepstrum-based diffraction representations and improves the Friedel symmetry of the patterns [14, 31, 32].

The resulting descriptor representation is then passed to unsupervised embedding and clustering. As shown in Figure 1, each diffraction pattern is encoded using canonical triplet invariants, in-plane lattice rotation, and 2D scan coordinates. These descriptors are scaled and embedded before hierarchical density-based clustering assigns labels for 2D spatial mapping and 3D visualization. For 4D-STEM datasets, the latent space is defined by three coupled approaches: the selected diffraction features, their relative weighting, and the embedding used before density-based clustering.

Machine-learning workflow for high-dimensional 4D-STEM analysis

Figure 1. Machine-learning workflow for high-dimensional 4D-STEM analysis. Large 4D-STEM datasets are first processed by autocorrelation and neural-network-based diffraction-feature detection. Detected disc centers are converted into physics-informed descriptors (PIDs), including canonical triplet invariants, in-plane complex orientational order parameters, and scan-position coordinates. These descriptors are scaled and embedded into a weighted feature-vector space before UMAP and density-based clustering resolve phase, deformation, and orientation structure for downstream analysis. (example: 4D-STEM dataset from a Ti-50Nb sample).

Feature selection determines which physical distinctions the clustering can capture. While raw diffraction intensities preserve the full signal, they are dominated by the central beam and are sensitive to background variation and detector noise. In thicker samples, the background contribution is even larger, potentially overwhelming the clustering algorithm unless additional preprocessing is applied [1, 19, 24, 28].

In contrast, hand-crafted descriptors such as canonical triplets of Bragg vectors — together with explicit in-plane rotation and spatial-coordinate channels — compress each pattern into a compact fingerprint, rather than entangling them as in pixel-wise intensity representations.

Spatial context is widely used in computer vision segmentation tasks to preserve image structure and geometry, including in unsupervised ML frameworks such as SLIC superpixels and autoencoders [33, 34, 35]. However, their explicit integration with phase-, strain- and rotation-sensitive descriptors in unsupervised 4D-STEM analysis remains relatively unexplored. In our framework, they provide local spatial context alongside diffraction-derived descriptors.

Because feature descriptors combine heterogeneous quantities, including reciprocal-space lengths, angles, and scan positions, they must be brought onto a common scale before any distance metric can be interpreted. Per-channel zz-scoring removes unit imbalance, while explicit channel weights encode the relative importance of lattice geometry, rotation, and spatial coherence in the final metric. This weighted representation provides the foundation for the manifold embedding and clustering steps that follow.

To address this, the weighted feature space is embedded into a manifold that preserves local neighborhood structure while making high-density regions more accessible to the clustering algorithm. In this pipeline, feature design determines which physical attributes are represented, scaling controls the relative contribution of each channel, and embedding sets the effective resolution at which hierarchical density-based spatial clustering with noise (HDBSCAN) resolves structure [18]. Together, these steps enable pixel-resolved phase, rotation, and lattice deformation mapping from otherwise unlabeled diffraction data without requiring prior phase assignments.

Formulation of the geometric descriptor model

For each valid pixel ii at scan position (xi,yi)(x_i, y_i), we construct a geometric descriptor by concatenating three feature blocks: (1) canonical triplet invariants, (2) an in-plane complex orientational order parameter, and (3) spatial scan coordinates. The corresponding notation and dimensionality for each block are summarized in Table 1.

Each feature vector is normalized, zz-scored independently:

x~i,c=xi,c−μcσc,x∈{g,r,p}(1)\tilde{x}_{i,c} = \frac{x_{i,c} - \mu_c}{\sigma_c}, \quad x \in\{g,r,p\} \tag*{(1)}

Here, i=1,…,Ni = 1, \ldots, N indexes the valid scan pixels, cc the channel within each descriptor block, and gI,cg_{I,c}, rI,cr_{I,c}, and pI,cp_{I,c} the canonical triplet-geometry, in-plane rotation, and positional values at pixel ii, as shown in Fig. 1. The constants μc\mu_c and σc\sigma_c are the per-channel sample mean and standard deviation over the full scan, so the zz-scored channels g~I,c\tilde{g}_{I,c}, r~I,c\tilde{r}_{I,c}, and p~I,c\tilde{p}_{I,c} are dimensionless and of unit variance. The angle is the rotation-invariant interior angle α\alpha of the diffraction triplet, as seen in Fig. 2, calculated from a⃗\vec{a}, the shortest spot vector, and b⃗\vec{b}, the next non-collinear vector. When combined with the three magnitudes of these vectors, it yields a primitive reduced cell of the local lattice geometry in the Friedel-folded reciprocal space frame, independent of the crystal's rotation in that plane.

FeaturesDescriptorChannelsD
Geometricg~i\tilde{g}_is1s_1, s2s_2, s3s_3, α\alpha4
Rotationr~i\tilde{r}_iRe ΨnRe\ \Psi_n, Im ΨnIm\ \Psi_n2∣H∣2|H|
Spatialp~i\tilde{p}_ixix_i, yiy_i2

Table 1. Notation for the feature vectors, descriptors, channels, and their dimensions (D).

Ψn\Psi_n — complex orientational order parameter

α\alpha — the interior angle between s1s_1 and s2s_2 of the primitive reduced cell

s1s_1, s2s_2, s3s_3 — the three sorted vector magnitudes (lengths of ∣a⃗∣<∣b⃗∣<∣c⃗∣|\vec{a}| < |\vec{b}| < |\vec{c}|)

HH – denotes the set of retained harmonic orders

Figure 2. Schematic of the canonical triplet invariant.

We define two-channel weights that encode the relative contribution of the feature axes. The rotation weight wrw_r and spatial weight wpw_p are simple diagonal preconditioners applied after z-scoring. They propagate into the Euclidean metric, analogous to inverse-covariance conditioning in Mahalanobis distance [36, 37]. UMAP builds its fuzzy high-dimensional neighborhood graph from this weighted latent space.

The per-channel weights act as a diagonal preconditioner on the z-scored feature space, the variance of channel ii scales as wi2w_i^2, and the covariance between channel ii and jj scales as wiwjw_iw_j, while their Pearson correlation coefficient remains unchanged. Unlike the Mahalanobis metric, which uses inverse covariance matrix Σ−1\Sigma^{-1} to whiten the data and remove cross-channel correlations, this diagonal rescaling preserves the correlation structure of the descriptor and leaves the responsibility for capturing it to the subsequent UMAP embedding.

The full model vector (ZZ) that UMAP processes:

Zi=[ 1⋅g~i,c, wr⋅r~i,c, wp⋅p~i,c ](2)Z_i = [\,1 \cdot\tilde{g}_{i,c},\ w_r \cdot\tilde{r}_{i,c},\ w_p \cdot\tilde{p}_{i,c}\,] \tag*{(2)}

and Z∈RDZ \in\mathbb{R}^{D} where D is descriptor dimensionality. wrw_r controls sensitivity to in-plane orientation differences, while wpw_p controls the strength of spatial regularization by biasing the metric toward nearby scan positions.

The resulting squared weighed Euclidian distance between scan pixels ii and jj is defined as:

d2(i,j)=dg2(i,j)+wr2dr2(i,j)+wp2dp2(i,j)(3)d^2(i,j) = d_g^2(i,j) + w_r^2 d_r^2(i,j) + w_p^2 d_p^2(i,j) \tag*{(3)}

Here, dgd_g, drd_r and dpd_p denote the Euclidean distances computed from the standardized geometric, rotational, and spatial descriptor channels, respectively, as summarized in Table 1. A linear weight in feature space becomes a quadratic weight in the metric.

We adapt the complex orientational order parameter Ψn\Psi_n [38, 39] to the azimuthal distribution of detected Bragg vectors in each diffraction pattern. Angular harmonics of scattered intensity are an established probe of local orientational symmetry in X-ray cross-correlation analysis [40] and electron diffraction of amorphous solids [41] and pattern symmetry has been used as a per-pixel contrast mechanism in STEM [42]. Here, we instead evaluate Ψn\Psi_n on the discrete first-order Bragg-vector ring of crystalline patterns, which yields a spot-permutation-invariant orientation descriptor.

For a pixel with N first-ring spots at azimuthal angles φk\varphi_k, the order parameter is given by:

Ψn=1N∑k=1Ne(in⋅φk),n∈{2,4,6}(4)\Psi_n = \frac{1}{N}\sum_{k=1}^{N} e^{(i n \cdot\varphi_k)}, \qquad n \in\{2,4,6\} \tag*{(4)}

Where nn is the rotational symmetry order, and ii is the imaginary unit. Figure 3 shows the geometric interpretation of the in-plane complex orientational order parameter.

Argand diagram of the in-plane complex orientational order parameter

Figure 3. The in-plane complex orientational order parameter Ψn=∣Ψn∣einθ\Psi_n = |\Psi_n|e^{in\theta} on the complex plane- Argand diagram- (red vector). The magnitude ∣Ψn∣|\Psi_n| (dashed rings) is the nn-fold coherence, rr and the phase nθn\theta (blue arc) is the lattice orientation θ=arg⁡(Ψn)/n\theta= \arg(\Psi_n)/n. The components (Re⁡Ψn,Im⁡Ψn)(\operatorname{Re}\Psi_n,\operatorname{Im}\Psi_n) (dashed projections) are the rotation entries of the descriptor, weighted by wrw_r.

We decompose the complex order parameter into real and imaginary parts, which are the two rotation channels fed to clustering after normalization. The magnitude ∣Ψn∣|\Psi_n| measures the symmetry strength of nn-fold orientational coherence, r. The full clustering descriptor combines Re ΨnRe\ \Psi_n and Im ΨnIm\ \Psi_n (Cartesian components of Ψn\Psi_n) with the rotation-invariant geometric triplet [s1s_1, s2s_2, s3s_3, α\alpha].

Analogous azimuthal Fourier representations of Bragg-peak positions are used in template-based automated crystal orientation mapping to solve for the in-plane rotation angle [2]; here we use the first-ring angular harmonic directly, without a reference structure, as a rotation-equivariant feature for unsupervised phase segmentation, with the dominant symmetry order selected for every diffraction pattern.

The squared spatial contribution between two pixels separated by Δx\Delta x, Δy\Delta y scan-pixels is:

dp2(i,j)=(Δxσx)2+(Δyσy)2≈(12H2)(Δx2+Δy2)(5)d_p^2(i,j) = \left(\frac{\Delta x}{\sigma_x}\right)^2 + \left(\frac{\Delta y}{\sigma_y}\right)^2 \approx\left(\frac{12}{H^2}\right)(\Delta x^2+\Delta y^2) \tag*{(5)}

For an H×WH \times W square scan, the z-scored positions have unit variance, with σx≈H/12\sigma_x \approx H/\sqrt{12} for a uniform scan. The effect of dp2d_p^2 can be interpreted as a spatial coherence length. Small values impose only a weak spatial bias, with a coherence length much larger than the scan region H×WH \times W, so they influence only global organization rather than local cluster identity.

As wpw_p increases, the coherence length decreases, making spatial separation comparable to phase differences over shorter distances and causing clusters to fragment into spatially compact subregions. In this sense, the spatial weight acts as an inverse coherence length, setting how far apart two pixels can be in scan space while remaining candidate neighbors in the model metric. For each scan pixel, the descriptor channels are concatenated into a single composite feature vector in the weighted high-dimensional space, XmodelX_{\mathrm{model}}. UMAP first embeds this joint representation, after which HDBSCAN computes the mutual reachability distance:

dmr(i,j)=max⁡{d(i,j),dcoremax⁡(i,j)}(6)d_{\mathrm{mr}}(i,j) = \max\{d(i,j),d_{\mathrm{core}}^{\max}(i,j)\} \tag*{(6)}

Here, dcoremax⁡(i,j)d_{\mathrm{core}}^{\max}(i,j) denotes the distance from ii to its kk-th nearest neighbor. For 3D visualization, the high-dimensional embedding is linearly projected to 3D by PCA, retaining the three orthogonal directions of greatest variance. This projection serves only for visualization; HDBSCAN labels are assigned in the full UMAP embedding, and the 3D scatter plot renders the clustering structure in the higher-dimensional space.

Results and Discussion

Strain and lattice analysis of pseudomorphic Si/Six_xGe1−x_{1-x} epitaxial layers

To evaluate the fidelity of our structural analysis workflow, we used a Si/Six_xGe1−x_{1-x} epitaxial-multilayer TEM sample (dataset I) in 4D-STEM analysis. This traceable standard consists of alternating pure Si and Six_xGe1−x_{1-x} alloy layers deposited on a single-crystal Si substrate. It contains five Six_xGe1−x_{1-x} layers, each about 10 nm thick, separated by pure Si layers about 13 nm thick, for a total superlattice thickness of approximately 100 nm. The alternating Si and Six_xGe1−x_{1-x} layers produce the light and dark contrast seen in the virtual bright-field (vBF) image in Figure 4(a). The TEM foil thickness was approximately 70 nm, as determined by electron-energy-loss spectroscopy (EELS). We acquired XEDS-STEM profiles across the multilayer stack and measured the average Si-Ge composition using the Ge K-line at 9.874 keV and the Si K-line at 1.739 keV. The layers contained approximately 87 at. % Si (x = 87) and 13 at. % Ge. From this composition, we estimated a maximum expected out-of-plane (growth direction) strain, ε⊥\varepsilon_\perp, of about 0.85% using the approach of Munshi et al. [43].

Seven-panel characterization of the Si/Si$_{87}$Ge$_{13}$ multilayer, including vBF image, strain maps, spatial cluster maps, and 3D projections

Figure 4. (a) vBF image of the Si/Si87_{87}Ge13_{13} multilayer showing layer contrast and a bend contour in the thin TEM foil. (b, c) Strain maps ε⊥\varepsilon_{\perp} and ε∥\varepsilon_{\parallel} represent the out-of-plane and in-plane strain fields referenced to the ⟨100⟩\langle100\rangle and ⟨110⟩\langle110\rangle reciprocal-space directions used in the analysis, respectively. (d, e) Spatial cluster map and 3D projection highlighting foil bending and lattice variation in the Si substrate and Si87_{87}Ge13_{13} layers (wp=0.03w_{\mathrm{p}} = 0.03) (f, g) Spatial cluster map and 3D projection with spatial weight set to zero, without any spatial regularization, showing increased noise in the spatial cluster map and fragmented clusters in the 3D UMAP embedding (wp=0w_{\mathrm{p}} = 0).

The strain maps ε⊥\varepsilon_\perp and ε∥\varepsilon_\parallel in Figure 4(b) and (c), respectively, represent the out-of-plane and in-plane (parallel to interface) strain fields, respectively, referenced to the two reciprocal-space directions used in the analysis, corresponding to the ⟨100⟩\langle100\rangle and ⟨110⟩\langle110\rangle reflections. The strain fields were calculated by tracking diffraction-disc centers with the center-detection model and measuring their displacements relative to a strain-free region in the Si substrate using polar decomposition (Supplementary material: polar decomposition algorithm). Oscillations in the ε⊥\varepsilon_\perp map indicate tensile strain along the growth direction, with a mean variation of approximately 1%, in good agreement with the expected theoretical maximum for the Si87_{87}Ge13_{13} layers. The per-pixel clusters' labels are remapped to their (x, y) probe coordinates, rendering a real-space map.

These contiguous regions of identical color delineate the spatial extent of each structurally distinct phase. Figure 4(d) shows this real-space map, which we refer to as the spatial cluster map. Adjacent Figure 4(e) is the 3D projection of the 6D feature embedding composed of the canonical triplet invariants s1s_1, s2s_2, s3s_3, α\alpha, and the spatial coordinates xix_i, yiy_i.

The blue and orange regions in Figure 4(d) define a curved low-dimensional manifold in Figure 4(e) that distinguishes Si from Si87_{87}Ge13_{13}, while also capturing a continuous ribbon-like feature associated with local lattice-spacing variations and the bend contour components labeled I, II, and III in Figure 4(e). Collectively, these clusters reflect both Si/Si87_{87}Ge13_{13} contrast and intrinsic regional TEM foil bending. The manifold structure resolves the two sides of the bend contour within the Si substrate, labeled bending I and II. By contrast, it does not separate the bending III region from the Si87_{87}Ge13_{13} layers, both of which appear as the orange cluster in the spatial cluster map and 3D projection in Figure 4(d, Figure 4e). This behavior reflects an inherent limitation of the clustering approach, which assumes that clusters occupy high-density regions separated by lower-density gaps. In this case, the continuous deformation linking the bend contour to the Si87_{87}Ge13_{13} layers forms a smooth gradient without well-defined boundaries, thereby precluding distinct cluster partitioning.

Two principal UMAP hyperparameters are the number of neighbors and the minimum distance. The number of neighbors sets the local neighborhood size used to construct the fuzzy graph, whereas the minimum distance controls how tightly points may pack in the low-dimensional embedding. In this study, the number of neighbors was set to 100 and the minimum distance to 0.15 based on empirical optimization, as these values produced a stable manifold (Supplementary Fig. 1) with well-defined boundaries. For the weighted embedding, the spatial weight was set to 0.03, imposing a modest spatial prior.

The weighted clustering resolves the lattice difference between the Si and Si87_{87}Ge13_{13} layers and identifies the regions strained by bending. These bend contours are also visible in the virtual bright-field image as a broad alternating band of high and low intensities (Figure 4(a)) (Supplementary Figure 4–Supplementary Figure 4)). Bending of the thin TEM foil is also evident in the strain maps. In the ε∥\varepsilon_{\parallel} map, Figure 4(c), it produces continuous strain modulation of about 0.1–0.2% along the nominally strain-free ⟨110⟩\langle110 \rangle direction in the Si substrate. The spatial cluster map complements the strain analysis by providing structural context, reducing the risk of misinterpreting bending-related effects.

Schematic showing the vDF image contrast caused by the bend contour in panels (a) and (b), and the virtual mask positions used to generate the corresponding vDF images

Supplementary Figure 4. Schematic showing the vDF image contrast caused by the bend contour in panels (a) and (b). Panels (d) and (e) show the virtual mask positions used to generate the corresponding vDF images in (a) and (b).

Figure 4(d–Figure 4g) illustrates how feature-space weighting shapes the UMAP-fuzzy graph. With spatial weighting in the 6D embedding (Figure 4d, Figure 4e), the manifold remains well defined, and clustering resolves deformation and lattice variation. In the absence of this spatial prior in the 4D embedding (Figure 4(f, Figure 4g)), the structure fragments into noise, evident as gray over-fragmentation in the 3D projection and dark noise pixels in the spatial cluster map, leaving sparse, weakly connected regions that HDBSCAN predominantly classifies as noise. These results show that weighting reshapes the UMAP topology and facilitates k-nearest neighbors (kNN) graph construction by reorganizing local neighborhoods and rescaling feature axes in high-dimensional space. As a result, HDBSCAN identifies more stable clusters.

To assess embedding behavior across the UMAP hyperparameter space, we examined 3D projections generated without any spatial regularization; these showed greater noise, weaker cluster boundaries, and reduced stability (Supplementary Fig. 2). We then evaluated the effect of spatial channel weighting by varying it over four orders of magnitude for the Si/Si87_{87}Ge13_{13} epitaxial layers (dataset I; Table 2). Cluster stability was quantified using two complementary metrics: median per-cluster persistence (P~\widetilde{P}) from the HDBSCAN condensed-tree hierarchy and the density-based cluster validity (DBCV) score [44], both defined in the methods section.

Dataset I
wpw_pkkP~\widetilde{P}DBCV
0140.040.34
0.0350.550.44
0.160.440.72
0.330.540.89
130.470.78

Table 2. Unsupervised clustering metrics for dataset I across the spatial weight wpw_p from 0 to 1.

At wp=0w_p = 0, the geometric descriptor produced 14 candidate clusters with P~\widetilde{P} near the noise floor (0.04), indicating over-fragmentation from per-pixel descriptor noise, with approximately 22% of pixels classified as noise, consistent with Figure 4(f, Figure 4g). A modest regularization, wp=0.03w_p = 0.03, consolidated these into five density-supported clusters (P~\widetilde{P} 0.55, DBCV 0.44, noise 0%) while preserving fine continuous deformation features such as crystal bending, as shown in Figure 4(d, Figure 4e). As the spatial weight increased, these features were progressively absorbed into compact, spatially localized clusters, with a corresponding loss of manifold detail (Supplementary Fig. 3c–d). At wp=0.3w_p = 0.3, only three clusters remained, and DBCV reached its maximum value of 0.89—an apparent improvement driven by spatial-coherence smoothing rather than descriptor-based separation.

At still larger weights (wp=1w_p = 1), the spatial channel dominated and the embedding collapsed into larger clusters (Supplementary Figure 3(e)); both metrics declined (P~\widetilde{P} 0.47, DBCV 0.78). The two metrics peak at different operating points—P~\widetilde{P} at wp=0.03w_p = 0.03 and DBCV at wp=0.3w_p = 0.3—and their disagreement highlights the trade-off between descriptor fidelity and spatial regularization. We therefore adopt wp=0.03w_p = 0.03 as the default because it coincides with the P~\widetilde{P} peak, where descriptor information is used most effectively with the minimum regularization needed to suppress noise-driven fragmentation. The substantial gain in cluster persistence (P~\widetilde{P} increasing from 0.04 to 0.55 between wp=0w_p = 0 and wp=0.03w_p = 0.03) reflects the role of the spatial channel as a locality prior on the feature representation rather than a smoothing filter on the cluster map.

3D manifold morphology across increasing spatial regularization

Supplementary Figure 3. 3D manifold morphology across increasing spatial regularization for the Si/ Si87Ge13\mathrm{Si}_{87}\mathrm{Ge}_{13} epitaxial-layer sample. 3D maps (a-e) correspond to spatial weights (wpw_p) of 0, 0.03, 0.1, 0.3, and 1.0, respectively. As the spatial weight increases, the clusters become progressively more compact and spatially localized, with a corresponding loss of nonlinear manifold structure. This trend reflects a shift from lattice-sensitive organization toward spatially driven phase separation. The most stable and well-defined cluster structure is obtained with a spatial weight of 0.03.

The geometric descriptor encodes lattice geometry per-pixel. Still, it discards the spatial grid structure intrinsic to 4D-STEM data, so with wp=0w_p = 0, UMAP's kNN graph treats pixels at any two scan positions as interchangeable — analogous to applying a multilayer perceptron to flattened image data without the locality prior that convolutional architectures provide [44, 45]. Computational performance was benchmarked on a desktop workstation equipped with an 8-core Intel i7 CPU and a 12 GB NVIDIA RTX 3060 GPU.

For the Si/Si87Ge13\mathrm{Si}_{87}\mathrm{Ge}_{13} dataset, the workflow processed 9,702 diffraction patterns at 512×512512 \times512 pixel resolution, including geometric descriptor extraction and clustering, in approximately 55 s. Descriptor extraction required 45 s, corresponding to an effective throughput of about 200 frames per second, while clustering added another 10 s.

Phase and rotation analysis in multiphase metallic alloy

We next evaluated our workflow on the more challenging task of resolving complex heterostructure in a Ti-50Nb metallic alloy (dataset II). Refractory alloys often exhibit intricate chemical variations and associated multiphase distributions together with crystallographic variants and in-plane rotations [46].

The spatial cluster map in Figure 5(a) provides a phase-level interpretation of dataset II. Clustering is governed primarily by the rotation-invariant geometric descriptor and spatial coherence; therefore, variants with similar reduced-cell geometry remain grouped even when they differ in in-plane orientation. The orange background corresponds to the body-centered-cubic (BCC) β\beta-Ti matrix viewed along the [110110]β_\beta direction., while the red and blue precipitate regions correspond to hexagonal-closed-pack (α\alpha-HCP) and orthorhombic O/α′′\alpha^{\prime\prime} domains (distinguished by the level of tetragonal distortion of the BCC phase), respectively. In the 3D projection of the embedding in Figure 5(b), these precipitate-related clusters are separated mainly by differences in lattice metric in Ti-based alloys [47, 48, 49, 50]. The virtual dark-field (vDF) images in Figure 5(c) reveal the complex phase morphology of this alloy, where needle-like and globular domains are embedded on both sides of the twinned BCC matrix, forming a highly overlapping basket-weave-like structure.

Clustering analysis of the multiphase Ti–50Nb alloy, dataset II

Figure 5. Clustering analysis of the multiphase Ti–50Nb alloy, dataset II. (a) Spatial cluster map, showing the BCC β\beta-Ti matrix (orange), α\alpha-HCP (red) and Ti orthorhombic α′′\alpha^{\prime\prime} domains (blue). (b) Corresponding 3D projection from 6D embedding (c) vDF images of the precipitates and twin matrix.

The asynchronous object-detection workflow processed 10,000 diffraction patterns in 150 s, yielding an effective throughput of 150 fps with a batch size of 64. As in the Si/Si87_{87}Ge13_{13} case, clustering required about 10 s, bringing the total analysis time to about 2.5 min.

The spatial cluster maps were post-processed with spatial kernel averaging (3x3) to remove noise and unassigned pixels from the UMAP embedding. This process reduces isolated wrong assignments within uniform domains while preserving grain boundaries and features, and it is applied purely as a visualization step — all reported cluster-stability and quality metrics (P~\tilde{P}, DBCV) were computed on the raw, pre-filtered HDBSCAN labels.

Incorporating the coherence-gated multi-order channels (Re Ψn\Psi_n, Im Ψn\Psi_n; n=2,6n = 2, 6) into the 10-D embedding resolves three coexisting phases in a Ti-50Nb powder alloy in a single unsupervised pass (Figure 6a,Figure 6b). First, the β\beta-Ti matrix (Clusters 1, 2): exhibits an identical BCC cell (d≈0.239d \approx0.239, 0.170, 0.139 nm — {110}, {200}, {211}; α≈90∘\alpha\approx90^\circ). The two clusters represent twin domains viewed along the [110]β_\beta zone axis, distinguished only by in-plane orientation, with a 72° rotation and Ψ2\Psi_2 coherence r≈1r \approx1 (Figure 7b). Second, an orthorhombic martensite phase, denoted α′′\alpha'' oriented along [1]α′′_{\alpha''} (Clusters 0, 13, 6, 7): a common large cell (d≈0.41d \approx0.41, 0.27, 0.25 nm; α≈83∘\alpha\approx83^\circ) appearing as four Ψ2\Psi_2-distinct orientation variants (0/13 and 6/7, each pair ≈60∘\approx60^\circ apart) that tile the basket-weave according to the adjoining matrix twin. The large variations in lattice spacing are likely associated with double diffraction in the TEM specimen. Third, α\alpha-HCP precipitates (Cluster 3): near-equal lattice spacings (~0.27 nm) with Ψ2\Psi_2 incoherent (r=0.03r = 0.03) but strong six-fold coherence (∣Ψ6∣=0.76|\Psi_6| = 0.76), the [1]α-HCP_{\alpha\text{-HCP}} projection of a hexagonal grain. Note that this is a phase that a single-order Ψ2\Psi_{2} embedding cannot represent and that is recovered here only because the descriptor is influenced by Ψ6\Psi_{6}. Cluster 4 traces the β/α′′\beta/\alpha^{\prime\prime} phase boundaries: it inherits the Ψ2\Psi_{2} orientation of Cluster 0 but carries a distorted metric (α=46.6∘\alpha= 46.6^\circ) arising from the double diffraction effect at the overlapping interfaces and is not a crystallographic distinct phase. The observed phase geometry is consistent with the orientation relationship [110]β//[001]α′′//[0001]α-HCP[110]_{\beta} // [001]_{\alpha^{\prime\prime}} // [0001]_{\alpha\text{-HCP}}, which places the β\beta matrix, orthorhombic α′′\alpha^{\prime\prime} martensite, and α\alpha-HCP phase along aligned zone axes and provides a crystallographic basis for the variant selection observed in the cluster map. α\alpha-HCP (Cluster 3) appears as elongated laths localized almost entirely within the lower β\beta-twin domain (Cluster 2), indicating preferential nucleation on one twin variant rather than uniform distribution — a variant-selection effect consistent with a fixed β→α\beta\to\alpha orientation relationship. Here, UMAP parameters matched dataset I (100 neighbors, minimum distance 0.15); rotation and spatial weights were wr=0.25w_r = 0.25 and wp=0.03w_p = 0.03.

Geometric-descriptor clustering of dataset II

Figure 6. Geometric-descriptor clustering of dataset II. (a) The spatial cluster map is shown with each cluster and (b) medoid patterns color-coded by cluster ID and annotated (green s1s_1, red s2s_2, cyan s3s_3) triplet vectors. The descriptor is robust for simple single-lattice patterns, as demonstrated by clusters corresponding to the matrix twin (Clusters 1 and 2) and precipitate variants (Clusters 0, 13, 6, 7, and 3). (c) One-way ANOVA η2\eta^2 explained (d) Channel-cluster heat map

Diffraction ring panel and orientation–coherence polar summary

Figure 7. (a) Medoid diffraction rings (30% Ψn\Psi_n ring width) for β\beta matrix twins (1, 2) and orthorhombic α′′\alpha^{\prime\prime} domain rotational variants (0, 13, 6, 7) and α\alpha-HCP (3); filled spots feed Ψn\Psi_n. (b) Cluster orientation θ\theta vs coherence rr polar summary (2-fold, 180° period): matrix twin ~72° separation; precipitate pairs 0/13 and 6/7 share metric but differ in orientation—colors as in Figure 6(a).

A one-way ANOVA across clusters ranks the descriptor channels by discriminability (Figure 6 c,Figure 6d). The lattice-metric edges and the Ψ2\Psi_{2} channels dominate — s1s_1 (F = 24640, η2=0.97\eta^2 = 0.97), s3s_3 (F = 24131, η2=0.97\eta^2 = 0.97), Im Ψ2\Psi_{2} (F = 21962, η2=0.97\eta^2 = 0.97), s2s_2 (F = 12914, η2=0.95\eta^2 = 0.95) and Re Ψ2\Psi_{2} (F = 11964, η2=0.95\eta^2 = 0.95) — confirming that metric and dominant two-fold orientation separate the β\beta and α′′\alpha^{\prime\prime} families. The Ψ6\Psi_{6} channels rank lowest overall (Re Ψ6\Psi_{6}: F = 1350, η2=0.67\eta^2 = 0.67; Im Ψ6\Psi_{6}: F = 942, η2=0.58\eta^2 = 0.58) because only one cluster is six-fold coherent: they are the sole channels that isolate the α\alpha-HCP grain (Cluster 3), which is otherwise metric- and Ψ2\Psi_{2}-ambiguous. Global channel importance and phase-resolving power are therefore distinct — a low-variance channel can still be indispensable for a minority phase, which is precisely the advantage of retaining multiple coherence-gated orders.

Complex orientational order provides a direct check that the Ψ2\Psi_{2} and Ψ6\Psi_{6} channels encode meaningful in-plane symmetries. Figure 7 summarizes representative medoid patterns for the matrix twin, the dominant precipitate variants, and the six-fold family (Cluster 3). In the first-order ring plots (Figure 7(a)), detected Bragg vectors are shown in reciprocal space with the direct beam at the origin; the inner solid circle marks min⁡∣g∣\min|g| and the outer dashed circle the Ψn\Psi_n ring cutoff (30% width). Spots inside this band (filled symbols) are exactly those summed in the complex orientational order parameter — the same first-ring spots enter both the Ψ2\Psi_{2} and Ψ6\Psi_{6} sums — whereas excluded spots (hollow grey) lie on higher-order shells. The matrix twins (Clusters 1 and 2) occupy the same ring radius

— confirming an identical rectangular reciprocal cell — but differ in the azimuth of their first-ring spots, i.e., in orientation only. The large-cell precipitate variants differ from the matrix by metric (a larger reciprocal cell, smaller ∣g∣|g|; d≈0.40d \approx0.40–0.42, 0.27, 0.25 nm, also evident in the channel-cluster heatmap of Figure 6(d)) and from one another by orientation. Indeed, clusters 0 and 13 share this ring but differ in spot alignment, as do Clusters 6 and 7. Cluster 3, shown on its n=6n = 6 ring, presents an approximately six-fold arrangement of first-ring spots with high coherence (r6=∣Ψ6∣≈0.76r_6 = |\Psi_6| \approx0.76); this is the direct-space confirmation that its separation reflects genuine six-fold orientational order rather than residual noise in the two-fold channel. The cluster-average orientation θ\theta and coherence rr (Figure 7b) make the variant structure explicit. The β\beta twins (Clusters 1, 2) and all four α′′\alpha^{\prime\prime} variants (0, 13, 6, 7) sit at r≈0.96r \approx0.96–0.99, i.e. near-perfect Ψ2\Psi_2 coherence, but at distinct angles: the β\beta pair at θ=60.5∘\theta= 60.5^\circ and 168.6∘168.6^\circ (72∘72^\circ apart), and the α′′\alpha^{\prime\prime} variants spread across θ=47.7∘\theta= 47.7^\circ, 107.9∘107.9^\circ, 121∘121^\circ and 180∘180^\circ in the ≈60∘\approx60^\circ-spaced pairs that build the basket-weave. Cluster 3 is the diagnostic outlier — it collapses to r=0.03r = 0.03 on the Ψ2\Psi_2 dial (no two-fold orientation), and its coherence instead resides on the Ψ6\Psi_6 order, marking it unambiguously as the hexagonal α\alpha-HCP grain rather than a poorly-oriented matrix or martensite variant.

To assess the effect of the rotation on the clustering, we varied the rotation weight wrw_r from a rotation-descriptor free baseline (wr=0w_r = 0) across wr=0.03w_r = 0.03, 0.1, 0.25, 0.3, and 1 for dataset II, which contains both twin variants and rotation-distinct precipitates. Joint analysis of the stability metrics — cluster count kk, mean cluster persistence P~\widetilde{P}, DBCV, and noise fraction — together with the spatial cluster maps reveals two regimes (Table 3; Supplementary Figure 5–Supplementary Figure 5).

Dataset II
wrw_rkkP~\widetilde{P}DBCVNoise (%)
090.630.800.2
0.03100.620.921
0.1120.580.880.9
0.25140.560.842.3
0.3140.610.874.1
1170.490.843.8

Table 3.

A series of spatial cluster maps showing changes in phase structure with increasing rotation regularization for Dataset II

Supplementary Figure 5. A series of spatial cluster maps illustrates changes in phase structure with increasing rotation regularization for Dataset II. Panels (a–f) correspond to rotation weights wrw_r of 0, 0.03, 0.1, 0.25, 0.3, and 1, respectively. Increasing the rotation-descriptor weight progressively increases the number of clusters, cluster fragmentation, and noise, while promoting the formation of low-density clusters.

Table 3. Unsupervised clustering metrics for dataset II as a function of the rotation weight wrw_r (0–1): cluster count kk, mean cluster persistence P~\widetilde{P}, DBCV, and noise fraction.

At wr=0w_r = 0, embedding yields only k=9k = 9 clusters (P~=0.63\widetilde{P} = 0.63, DBCV = 0.80, noise = 0.2%): the metric channels alone still separate the β\beta, α′′\alpha^{\prime\prime} and α\alpha-HCP phases along the characteristic two-domain architecture (Figure 5a), but the two β\beta twins and the four α′′\alpha^{\prime\prime} orientation variants each collapse into a single metric class, since variants that share a reduced cell are indistinguishable without orientation information. A small weight makes the embedding orientation-sensitive, and DBCV peaks at 0.92 at wr=0.03w_r = 0.03 (k=10k = 10), where the dominant ≈72∘\approx72^\circ β\beta twin is first resolved.

Increasing wrw_r progressively resolves finer orientation variants — k=12k = 12 (0.1) then 14 (0.25) — until the full set of β\beta, α′′\alpha^{\prime\prime} and α\alpha-HCP variants analyzed in Figure 6(a, Figure 6b) and Figure 7 is captured at wr=0.25w_r = 0.25 (P~=0.56\widetilde{P} = 0.56, DBCV = 0.84, noise = 2.3%). Beyond this the partition saturates: wr=0.3w_r = 0.3 yields the same 14 clusters but nearly doubles the noise (4.1%), and only at wr=1w_r = 1 does over-weighting fragment the map further (k=17k = 17, P~=0.49\widetilde{P} = 0.49, noise = 3.8%). We therefore adopt wr=0.25w_r = 0.25 — the smallest weight at which the complete variant set is resolved, and at the lowest noise for that partition. The marginally higher persistence and DBCV at wr=0.3w_r = 0.3 do not indicate better structure but more aggressive noise rejection: with the same 14 clusters, the metric gain comes entirely from discarding ∼1.8%\sim1.8\% more (predominantly boundary) pixels as noise. We interpret wrw_r as an angular-resolution control, with the weight required to resolve two metric-identical variants scaling approximately as wr∝1/Δθw_r \propto1/\Delta\theta. The dominant β\beta twin ($\Delta\theta \approx 72∘72^\circ in-plane) is resolved at the lowest weight, where DBCV peaks (wr=0.03w_r = 0.03). Small precipitate rotations require more weight: Clusters 0 and 7 are both α′′\alpha'' with an identical cell (s≈0.41s \approx0.41, 0.27, 0.25 nm; α≈83∘\alpha\approx83^\circ) and differ by only ≈13∘\approx13^\circ in Ψ2\Psi_2 orientation, so they merge into a single mode for wr<0.25w_r < 0.25 and separate into distinct density peaks only at wr=0.25w_r = 0.25. The rotation weight should therefore be chosen according to the angular scale of the feature of interest — coarse textures at low weight, few-degree orientation variants at higher weight — up to the point where noise degrades the partition (wr≳0.3w_r \gtrsim0.3).

The two datasets, dataset I (Si/Si87_{87}Ge13_{13}) and dataset II (Ti-Nb multiphase alloy), pose fundamentally different clustering problems due to their intrinsic structures. Dataset I is dominated by continuous, non-linear lattice deformation: the per-pixel descriptors form a single, connected manifold with no intrinsic density gaps, so there is no discrete set of classes to be recovered. Clustering here is therefore better understood as a coarse-graining of a continuum than as class discovery, where the spatial weight wpw_p acts as a granularity control, trading a finer, manifold-preserving partition for a coarser set of spatially compact regions, with the operating point chosen to match the length scale of interest. dataset II, by contrast, contains discrete phases together with twin and precipitate variants that occupy genuinely distinct, well-separated density modes. Here, clustering is well-posed and recovers physically meaningful classes: a small rotation weight (wr=0.03w_r = 0.03) suffices to separate the dominant twin variant (∼72∘\sim72^\circ misorientation) while keeping the cluster density stable, whereas excessive weighting over-fragments the map. The same descriptor-and-HDBSCAN pipeline thus behaves very differently across datasets: in dataset I, it segments a continuous deformation field, while in dataset II, it resolves discrete phase and rotation variants.

Comparison of descriptors

Using dataset II, we benchmarked the geometric descriptor against two data-driven alternatives—a pretrained convolutional-autoencoder and a model-free flattened descriptor—using the same UMAP→HDBSCAN pipeline. The geometric descriptor produced a well-separated, spatial cluster map (6 dimensions; DBCV = 0.80, P~\tilde{P} = 0.63), as seen in Figure 5(a) and Table 4, whereas both data-driven descriptors failed: the 16,262-dimensional flattened descriptor was essentially unstructured (DBCV = −0.61, P~\tilde{P} = 0.13), generating a continuous manifold of a single large cluster, as seen in Figure 8(a). The 130-dimensional autoencoder resolved the dominant twin orientation by splitting the cluster into two low-density clusters, as shown in the 3D projection in Figure 8(c) (DBCV = −0.46, P~\tilde{P} = 0.31) and in Table 4.

DescriptorDimDBCVP~\tilde{P}Noise (%)
Geometric60.800.630.2
Autoencoder130−0.460.311.1
Model-free16262−0.610.130.7

Table 4.

Comparison of descriptor representations for clustering the dataset II

Figure 8. Comparison of descriptor representations for clustering the dataset II. (a–b) Model-free flattened diffraction pattern descriptor results, showing primarily coarse continuous manifold clustering, like an autoencoder. (c–d) Pretrained autoencoder descriptor results show weaker separation of large clusters and formation of fewer stable, low-density clusters with lower persistence.

The HDBSCAN condensed trees in Figure 8(b) and Figure 8(d) highlight this distinction: the autoencoder and model-free descriptors produce only a few low-density modes within an otherwise continuous manifold. In contrast, the geometric descriptor yields 14 smaller clusters mostly corresponding to distinct phases, as shown in Figure 5(a).

Table 4. Comparison of geometric, pretrained autoencoder, and model-free flattened descriptors for dataset II. Metrics include feature dimensionality, DBCV, median HDBSCAN cluster persistence, and noise (%). The geometric descriptor provides the highest cluster quality and stability while using the lowest-dimensional, physically interpretable representation.

Together, these results show that the compact, physics-informed descriptor approach substantially outperforms both learned and generic high-dimensional representations. Specifically, its built-in rotation, translation, and permutation invariances isolate the discriminative lattice signal and yield genuinely discrete clusters, whereas the model-free descriptor and the autoencoder latent — optimized for reconstruction rather than class separability and pretrained on a narrow image set — encodes a relatively large continuous manifold of low cluster density. Beyond accuracy, the geometric descriptor is two-to-three orders of magnitude lower-dimensional (6 versus 130–16,262 features), making it faster to cluster and, unlike the learned and pixel-based representations, directly physically interpretable. We repeated the same descriptor comparison using dataset I, which contains simpler diffraction patterns than the multiphase Ti–50Nb alloy. At 512×512 detector resolution, the clustering still followed a similar trend: the geometric descriptor produced denser and more physically meaningful clusters with higher persistence, as shown in Supplementary Figure 6(a–Figure 6i).

Within the explored UMAP hyperparameter space, both data-driven descriptors more clearly separated deformation associated with bend contour and lattice spacing variations between the Si substrate and Si87_{87}Ge13_{13} layers in dataset I. However, whether these latent-space separations reflect meaningful crystallographic variables rather than factors coupled with sample thickness, diffraction intensity, or beam-condition variations—remains unclear, particularly considering the strain maps in Figure 4(b, Figure 4c) showing a clear strain relaxation region in Si adjacent to the top of the Si87_{87}Ge13_{13} layer stack in the TEM sample.

A limitation arises for high-symmetry diffraction patterns when s1≈s2s_1 \approx s_2 and α\alpha remains nearly constant. Over-fragmentation can occur because the clustering becomes sensitive to small numerical fluctuations rather than meaningful structural differences. Symmetry-aware canonicalization or feature weighting could reduce this instability.

In future work, we plan to expand the autoencoder architecture with residual connections to preserve spatial resolution. Additionally, we plan to curate data to expand the training datasets to include diffraction patterns acquired over a broader range of camera lengths, spot spacing, pattern geometry, and intensity distributions for both autoencoder and neural network object detection models.

Conclusion

In this work, we present a parallelized post-processing workflow for large-scale 4D-STEM datasets that combines a neural network-based object detection model for feature extraction with feature-regularized unsupervised ML clustering. The framework enables rapid classification of diffraction features and supports accurate crystallographic analysis (assigning both phase and crystal orientation) without labeled training data or prior phase assignments. This approach proved to be successful in disentangling the various components of a complex and multiphase alloy with nanoscale features and double-diffraction effects. By accelerating interpretation across varied experimental conditions, this approach can help streamline materials discovery and understanding in systems ranging from semiconductors to metallic alloys.

Methods

Materials

The MAG*I*CAL® specimen was acquired from Ted Pella Inc. and is widely used to benchmark strain and lattice measurements because the lattice mismatch between epitaxial Si and Six_xGe1−x_{1-x} produces a well-defined reference system. The lattice parameters are aGe=0.5658a_{\mathrm{Ge}} = 0.5658 nm for Ge and aSi=0.5431a_{\mathrm{Si}} = 0.5431 nm for Si. A cross-section TEM sample was conventionally prepared by dimpling and Ar-ion milling.

The Ti-50Nb alloy was prepared by arc melting, followed by splat quenching to produce foils about 250 μ\mum thick, and then homogenized to ensure a uniform bulk composition of 50 at.% Nb. Oxygen enrichment was achieved by wrapping the alloy in Nb foil, sealing it in fused silica ampoules under an Ar/O2_2 gas mixture, and heat-treating at 800°C for 2 h to produce an oxygen concentration of about 1 at.%. TEM specimens were prepared by focused ion beam milling at 30 kV on a Thermo Fisher Scientific Helios NanoLab 600, followed by final thinning at 5 kV to minimize Ga+^+ ion-beam damage.

4D-STEM data acquisition

The dataset I acquired from the Si/Six_xGe1−x_{1-x} epitaxial-layer sample using a GATAN Stela hybrid-pixel detector operated in energy-filtered mode at 200 kV with a probe convergence angle of 0.6 mrad. The scan covered a 147 ×\times 66 pixel region, yielding 9,702 scan positions, each recorded at 512 ×\times 512 pixel diffraction resolution.

Dataset II was acquired from the Ti–50Nb alloy sample using a Thermo Fisher Scientific Ceta S CMOS detector and Spectra 200 STEM operated at 200 kV with a probe convergence angle of 0.6 mrad. The scan covered a 100 ×\times 100-pixel region, with each diffraction pattern recorded at 256 ×\times 256-pixel resolution.

Clustering metrics

Cluster persistence. For cluster kk, we define the average persistence as

Pers⁡(k)=(1k)∑xi∈k(λdeath(k)−λbirth(i,k))(7)\operatorname{Pers}(k) = \left(\frac{1}{k}\right) \sum_{x_i \in k} \left(\lambda_{\mathrm{death}}^{(k)} - \lambda_{\mathrm{birth}}^{(i,k)}\right) \tag*{(7)}

and report the summary statistic across clusters as

P~=median⁡{Pers⁡(ki)}i=1K(8)\widetilde{P} = \operatorname{median}\{\operatorname{Pers}(k_i)\}_{i=1}^{K} \tag*{(8)}
  • P~\widetilde{P}: median persistence

  • kk: a cluster selected from the HDBSCAN condensed tree

  • ∣k∣\lvert k\rvert: number of points assigned to cluster kk

  • xix_i: a point belonging to cluster kk

  • λ=1/ε\lambda= 1 / \varepsilon: inverse distance scale used in the HDBSCAN density hierarchy

  • λbirth(i,k)\lambda_{\mathrm{birth}}^{(i,k)}: scale at which point xix_i first becomes part of cluster kk

  • λdeath(k)\lambda_{\mathrm{death}}^{(k)}: scale at which cluster kk disappears or splits

  • KK: total number of selected clusters

Density-based clustering validation (DBCV). We compute the global validity score as

DBCV=∑k=1K(∣k∣N)V(k)(9)DBCV = \sum_{k=1}^{K} \left(\frac{\lvert k\rvert}{N}\right)V(k) \tag*{(9)}

with cluster validity term

V(k)=(DSPC(k)−DSC(k)max⁡(DSPC(k),DSC(k)))(10)V(k) = \left(\frac{DSPC(k) - DSC(k)}{\max(DSPC(k), DSC(k))}\right) \tag*{(10)}
  • NN: total number of clustered points, excluding noise points

  • V(k)V(k): validity of cluster kk

  • DSC(k)DSC(k): density sparseness of cluster kk, measuring the weakest internal density connectivity

  • DSPC(k)DSPC(k): density separation between cluster kk and its nearest neighboring cluster

In this formulation, higher persistence indicates that clusters remain stable across a broader range of density thresholds. In contrast, higher DBCV indicates a cluster that is both internally dense and well separated from neighboring clusters.

HDBSCAN condense tree visualization

Phase clusters were identified with HDBSCAN and visualized using its condensed cluster tree. In this representation, the vertical axis is the density parameter λ=1/dmr\lambda= 1/d_{\mathrm{mr}}, where dd is the mutual-reachability distance; larger λ\lambda corresponds to higher density (smaller distance scale). Each branch ("icicle") denotes a candidate cluster, and its width represents the number of constituent points at the given λ\lambda. As λ\lambda increases, points progressively drop out of a cluster until it either splits into child clusters (when both children meet the minimum cluster size) or dissolves into noise; the persistence of a branch over a range of λ\lambda defines its stability. Branch color encodes cluster size as log (number of points), a logarithmic scale used so that clusters spanning several orders of magnitude in population remain distinguishable. The clusters retained by HDBSCAN's excess-of-mass selection are outlined and assigned the phase color palette used throughout in the spatial cluster maps.

Channel importance (one-way ANOVA)

To determine which descriptor channels drive clustering, we performed a per-channel one-way analysis of variance (ANOVA) across HDBSCAN cluster labels, treating each cluster as a group and excluding noise pixels (label =−1= -1) and the scan-position channels. For channel cc with NN clustered pixels partitioned into kk clusters of sizes ngn_g, we computed the between-cluster and total sums of squares:

SSbetween,c=∑g=1kng(xˉg,c−xˉc)2(11)SS_{\mathrm{between},c} = \sum_{g=1}^{k} n_g(\bar{x}_{g,c} - \bar{x}_c)^2 \tag*{(11)}
SStotal,c=∑i=1N(xˉi,c−xˉc)2(12)SS_{\mathrm{total},c} = \sum_{i=1}^{N}(\bar{x}_{i,c} - \bar{x}_c)^2 \tag*{(12)}

where xˉg,c\bar{x}_{g,c} is the mean of channel cc in cluster gg and xˉc\bar{x}_c is the grand mean.

SSwithin,c=SStotal,c−SSbetween,c(13)SS_{\mathrm{within},c} = SS_{\mathrm{total},c} - SS_{\mathrm{between},c} \tag*{(13)}

From these, we report the F-ratio and effect size,

Fc=SSbetween,c/(k−1)SSwithin,c/(N−k)(14)F_c = \frac{SS_{\mathrm{between},c}/(k-1)}{SS_{\mathrm{within},c}/(N-k)} \tag*{(14)}
ηc2=SSbetween,cSStotal,c(15)\eta_c^2 = \frac{SS_{\mathrm{between},c}}{SS_{\mathrm{total},c}} \tag*{(15)}

Here ηc2\eta_c^2 is the fraction of each channel’s variance explained by the cluster partition (0 = no separation, 1 = fully separated). Because FcF_c and ηc2\eta_c^2 are invariant to constant per-channel rescaling, they are unaffected by the z-scoring and channel weighting applied before embedding; channels are ranked by FcF_c to order their contribution to the clustering (Figure 4(c)).

Channel cluster heatmap

To visualize how each cluster differs, we tabulated the mean descriptor of every cluster from the raw (unscaled) feature values. Reciprocal-vector magnitudes were converted to real-space d-spacing using the diffraction pixel size pp (nm−1^{-1} per pixel)

d=1s⋅p(16)d = \frac{1}{s \cdot p} \tag*{(16)}

With the interior angle α\alpha reported in degrees, and the complex orientational channels Re Ψn\mathrm{Re}\,\Psi_n and Im Ψn\mathrm{Im}\,\Psi_n were left dimensionless. In parallel, we computed a standardized effect for each cluster–channel pair, where σc\sigma_c is the global standard deviation of channel cc over all clustered pixels; this expresses each cluster mean as a deviation from the population mean in units of standard deviation, so that phase- and orientation-distinguishing channels are directly comparable across the descriptor. The resulting cluster-mean and standardized-effect tables are rendered as the channel–cluster heatmap in Figure 4(d).

Pretrained autoencoder descriptor.

As a learned alternative to the hand-crafted descriptors, we used a convolutional autoencoder as a fixed feature extractor. The network is a four-stage stride-convolution encoder (32→64→128→256 channels, each Conv–BatchNorm–ReLU) feeding a 128D-channel bottleneck, mirrored by a transposed-convolution decoder; it was pretrained self-supervised (reconstruction loss) on 99 diffraction patterns — the same image domain used to train the diffraction peak center detector (neural network object detection model)—. The convolutional autoencoder was trained to minimize MSE reconstruction loss using the AdamW optimizer with an initial learning rate of 0.001, a weight decay of 0.0001, and a cosine annealing learning-rate scheduler. The model was trained for a maximum of 60 epochs with a batch size of 8. A 5% validation split and early stopping with a patience of 20 epochs were used to mitigate overfitting. The checkpoint with the lowest validation loss was retained for downstream analysis. Its weights were held fixed during clustering. At inference, each scan position's diffraction pattern was converted to its log-autocorrelation, per-image min–max normalized to [0,1], and contrast-inverted to match the polarity of the training images (bright background, dark peaks), then passed through the encoder. The bottleneck feature map was globally average-pooled and L2-normalized to yield a 128-dimensional embedding per pixel. As with the geometric descriptor, the two scan-position channels were then concatenated, yielding a 130-dimensional feature vector that was fed into the common pipeline (per-channel z-scoring, UMAP, HDBSCAN).

Model-free descriptor

As a model-free baseline to the geometric descriptor, each scan position was represented by its diffraction pattern autocorrelation, computed as the inverse Fourier transform of the power spectrum (Wiener–Khinchin), normalized by its central peak, and log-compressed. Each autocorrelation was area-downsampled to N×NN \times N pixels (N=128N = 128), its central self-correlation peak excised with a circular mask of radius f⋅(N/2)f \cdot(N/2) (f=0.1f = 0.1), and the retained pixels were max-normalized per pattern and flattened into a feature vector. These vectors were then fed to the common pipeline (per-channel z-scoring, UMAP, HDBSCAN). This descriptor is translation-invariant but rotation-variant and is therefore sensitive to in-plane lattice rotation (twin/orientation variants) as well as lattice spacing.

Data and Code Availability

The datasets and source code generated and/or analyzed during the current study are not publicly available but are available from the corresponding author upon reasonable request, subject to any applicable institutional or collaborator agreements.

Acknowledgements

This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, as part of the Genesis Mission National Science and Technology Challenges under Award Number DE-SC0026808 AWD20795.

Author Contributions

A.G. and R.S. acquired the 4D-STEM data. A.G. developed the machine-learning algorithm and data-analysis workflow. D.B. analyzed the Ti–50Nb alloy data, and A.N. developed and trained the autoencoder model. A.G., R.S., and D.G. edited the manuscript. D.G. supervised the experiments and analyses.

Competing Interests

Authors declare no competing interests.

Additional Information

Supplementary information: the online version contains supplemental material available at

Supplementary information for A Geometry-Guided Feature-Space Framework for Interpretable Unsupervised 4D-STEM Analysis of Complex Multiphase Materials

Arda Genc; Daniel A. Bolden; Albert Nguyen; Daniel S. Gianola; Ravit Silverstein

3D projections of manifold morphology across embedding hyperparameter space

Figure S1. A series of 3D projections illustrate manifold morphology across the explored high-dimensional embedding hyperparameter space for the Si/Si87Ge13\mathrm{Si}/\mathrm{Si}_{87}\mathrm{Ge}_{13} epitaxial-layer sample. Rows correspond to neighborhood sizes of 50, 100, and 200, and columns correspond to minimum-distance values of 0.05, 0.15, and 0.30. When spatial coordinates are included with a regularization weight (wp=0.03w_p = 0.03), the cluster structure remains largely stable across these parameter variations. Reducing the minimum distance to 0.05 at 100 neighbors decouples the bending-related contribution (green cluster) from the Si substrate (blue cluster). The most stable and well-defined cluster structure is obtained with 100 neighbors and a minimum distance of 0.15.

Nine 3D projections showing manifold morphology across neighborhood sizes and minimum-distance values

Figure S2. A series of 3D projections illustrates manifold morphology across the explored high-dimensional embedding hyperparameter space in the absence of spatial regularization (wp=0w_{p}=0) for the Si/Si87Ge13\mathrm{Si}/\mathrm{Si}_{87}\mathrm{Ge}_{13} epitaxial-layer sample. Rows correspond to neighborhood sizes of 50, 100, and 200, and columns correspond to minimum-distance values of 0.05, 0.15, and 0.30. Without regularization, the embedding becomes noisier and less stable across parameter variations, with weaker cluster boundaries and greater fragmentations.

Comparison of descriptor representations for clustering the Si/Si$_{87}$Ge$_{13}$ epitaxial layer 4D-STEM dataset

Supplementary Figure 6. Comparison of descriptor representations for clustering the Si/Si87Ge13\mathrm{Si}_{87}\mathrm{Ge}_{13} epitaxial layer 4D-STEM dataset. (a–c) Geometric descriptor results, including the spatial cluster map, 3D projection, and HDBSCAN condensed tree, showing continuous manifold variation from lattice and bending deformation. (d–f) Pretrained autoencoder descriptor results from 128×128128 \times128 cropped diffraction patterns, showing weaker separation and low-density large clusters which assigned as noise by HDBSCAN. (g–i) Model-free flattened diffraction pattern descriptor results, showing several low-density clusters and noise but identify the Si87_{87}Ge13_{13} layers over Si background and the bend contours.

Polar decomposition algorithm

For each scan position, the algorithm matches measured reciprocal-lattice vectors to a reference set, estimates the reciprocal-space transformation by least squares, converts that transformation to the real-space deformation gradient, and applies polar decomposition to separate rigid-body rotation from elastic stretch. Strain is then calculated from the symmetric stretch tensor and reported relative to the selected reference region.

Step 1: Build the reference (GrefG_{\mathrm{ref}}) matrix

Pick a template pixel from the ROI center. Extract its Bragg vectors — these are 2D positions of diffraction spots relative to the central beam:

Gref={g1,g2,…,gN},gi∈R2G_{\mathrm{ref}} = \{g_1, g_2, \ldots, g_N\},\quad g_i \in\mathbb{R}^2

These are the "unstrained" reference vectors. Then filter them across the ROI to reduce noise and eliminate outliers.

Step 2: Match vectors at each pixel

At every scan pixel, the detected Bragg vectors (xx, yy in pixels) can be in a different order or have sign flips in Friedel pairs. At each pixel and Bragg vector, gjmeasg_j^{\mathrm{meas}}, compute the Euclidean distance (L2 norm) matrix against girefg_i^{\mathrm{ref}} considering both signs. Take the minimum and do greedy assignment (closest pair first) to establish which measured vector corresponds to which reference vector.

d(i,j)=∥giref−gjmeas∥2d(i,j)=\left\lVert g_i^{\mathrm{ref}}-g_j^{\mathrm{meas}}\right\rVert_2

ii and jj are the reference and measured Bragg vectors, respectively.

Step 3: Fit the affine transformation matrix (A)

With matched Bragg vector pairs (grefg^{\mathrm{ref}}, gmeasg^{\mathrm{meas}}), solve the least-squares problem (minimum 8 pairs from high-speed machine learning (ML) diffraction spot center detection):

gmeas=A⋅grefg^{\mathrm{meas}} = A \cdot g^{\mathrm{ref}}

This is a 2×22 \times2 matrix A that best maps reference vectors to measured vectors in the detector coordinate system.

A=(a11a12a21a22)(xmeasymeas)=(a11a12a21a22)(xrefyref)\begin{aligned} A &= \begin{pmatrix} a_{11} & a_{12} \\ a_{21} & a_{22} \end{pmatrix} \\ \begin{pmatrix} x_{\mathrm{meas}} \\ y_{\mathrm{meas}} \end{pmatrix} &= \begin{pmatrix} a_{11} & a_{12} \\ a_{21} & a_{22} \end{pmatrix} \begin{pmatrix} x_{\mathrm{ref}} \\ y_{\mathrm{ref}} \end{pmatrix} \end{aligned}

8 pairs x 2 equations each = 16 equations for 4 unknowns (overdetermined and robust).

The fit uses at least eight matched Bragg-vector pairs, yielding at least 16 scalar equations for the four unknown entries of A. This redundancy improves numerical stability, although robustness still depends on vector geometry, localization uncertainty, correspondence quality, and outlier rejection.

Deformation gradient (F)

The affine map A operates in reciprocal space. The real-space deformation gradient, FF, is:

F=(A−1)TF = (A^{-1})^{T}

Inversion converts ‘reciprocal space mapping” to “real space mapping”. Transpose accounts for the difference between how row vectors and column vectors transform.

Polar decomposition

Decompose FF into rotation, RR and stretch, UU:

F=R⋅UF = R \cdot U

where:

  • R is rotation matrix (det = 1, orthogonal)

  • U is a stretch tensor

RR and UU are calculated through the Numpy Singular Value Decomposition (SVD) function. The output is rotation1, stretch, and rotation2.

  1. Usvd,S,Vt=np.linalg.svd(F)U_{\mathrm{svd}}, S, V_t = \text{np.linalg.svd}(F)

  2. R=Usvd@VtR = U_{\mathrm{svd}} @ V_t

  3. U=VtT@np.diag(S)@VtU = V_t^{T} @ \text{np.diag}(S) @ V_t

Linear strain tensor (Biot strain tensor) EE,

E=U−I=(εxxεxyεyxεyy)(17)E = U - I = \begin{pmatrix} \varepsilon_{xx} & \varepsilon_{xy} \\ \varepsilon_{yx} & \varepsilon_{yy} \end{pmatrix} \tag*{(17)}

where :

I=(1001)I = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}
  • εxx\varepsilon_{xx}: normal strain in detector x-direction

  • εyy\varepsilon_{yy}: normal strain in detector y-direction

  • εxy,εyx\varepsilon_{xy}, \varepsilon_{yx}: shear strain (symmetric, off-diagonal)

Step 7: Crystal frame projection of strain tensor EE

Detector frame:

Detector frame diagram

Figure 1 (PDF p. 29).

E=(εxxεxyεyxεyy)E = \begin{pmatrix} \varepsilon_{xx} & \varepsilon_{xy} \\ \varepsilon_{yx} & \varepsilon_{yy} \end{pmatrix}

εxy\varepsilon_{xy}, shear strain representing the change in angle between x and y detector directions.

Crystal frame:

Crystal frame diagram

Figure 2 (PDF p. 29).

E=(εhεhkεkhεk)E = \begin{pmatrix} \varepsilon_h & \varepsilon_{hk} \\ \varepsilon_{kh} & \varepsilon_k \end{pmatrix}

If h and k directions are defined by the lattice directions in the diffraction pattern, to rotate the strain tensor in the detector frame to the crystal frame, we do the following:

εn=n^TEn^\varepsilon_n = \hat{n}^{T} E \hat{n}
εh=nx2εxx+2nxnyεxy+ny2εyy\varepsilon_h = n_x^2 \varepsilon_{xx} + 2n_x n_y \varepsilon_{xy} + n_y^2 \varepsilon_{yy}

nin_i, unit vector along the lattice direction.

εh\varepsilon_h, εk\varepsilon_k normal strain along the lattice direction.α\alpha

εhk\varepsilon_{hk}, shear strain representing the change in angle between basis vectors.

εxypred=∣ε1−ε22χsin⁡2α∣\varepsilon_{xy}^{pred} = \left|\frac{\varepsilon_1-\varepsilon_2}{2\chi\sin2\alpha}\right|

α\alpha is the angle in degrees between the principal strain (no shear) and the crystal frame coordinates. ε1\varepsilon_1 and ε2\varepsilon_2 are rotation-free principal strain values, defining maximum and minimum strains in orthogonal directions.

Step 8: Rotation

R=F⋅U−1R = F \cdot U^{-1}
R=(R00R01R10R11)=(cos⁡θ−sin⁡θsin⁡θcos⁡θ)R = \begin{pmatrix} R_{00} & R_{01} \\ R_{10} & R_{11} \end{pmatrix} = \begin{pmatrix} \cos\theta& -\sin\theta\\ \sin\theta& \cos\theta\end{pmatrix}
θ=arctan⁡2(sin⁡θ,cos⁡θ)\theta= \arctan2(\sin\theta,\cos\theta)

This gives the in-plane rigid-body rotation at each pixel, relative to the ROI average orientation. Reported in degrees.

Summary of outputs:

MapMeaning
Strain A (εxx,εh\varepsilon_{xx}, \varepsilon_h)Normal strain along the x detector or h crystal directions
Strain B (εyy,εk\varepsilon_{yy}, \varepsilon_k)Normal strain along the y detector or k crystal directions
Strain C (εxy,εhk\varepsilon_{xy}, \varepsilon_{hk})Shear strain along detector and crystal directions
RotationRigid-body in-plane rotation (degrees) in polar coordinates

Table 1 (PDF p. 30).

References

  1. [1]Spurgeon, S. R. et al. Towards data-driven next-generation electron microscopy. Nat. Mater. 20, 274–279 (2021).DOI
  2. [2]Ophus, C. et al. Automated Crystal Orientation Mapping in py4DSTEM using Sparse Correlation Matching. Microscopy and Microanalysis 28, 390–403 (2022).DOI
  3. [3]Gammer, C. et al. Local and transient nanoscale strain mapping during in situ deformation. Appl. Phys. Lett. 109, (2016).DOI
  4. [4]Ozdol, V. B. et al. Strain mapping at nanometer resolution using advanced nano-beam electron diffraction. Appl. Phys. Lett. 106, (2015).DOI
  5. [5]Savitzky, B. H. et al. Py4DSTEM: A Software Package for Four-Dimensional Scanning Transmission Electron Microscopy Data Analysis. Microscopy and Microanalysis 27, 712–743 (2021).
  6. [6]Ophus, C. Four-Dimensional Scanning Transmission Electron Microscopy (4D-STEM): From Scanning Nanodiffraction to Ptychography and Beyond. Microscopy and Microanalysis vol. 25 563–582 Preprint at https://doi.org/10.1017/S1431927619000497 (2019).
  7. [7]Cautaearts, N. et al. Free, flexible and fast: Orientation mapping using the multi-core and GPU-accelerated template matching capabilities in the Python-based open source 4D-STEM analysis toolbox Pyxem. Ultramicroscopy 237, (2022).
  8. [8]Clausen, A. et al. LiberTEM: Software platform for scalable multidimensional data processing in transmission electron microscopy. J. Open Source Softw. 5, 2006 (2020).DOI
  9. [9]Bruefach, A., Ophus, C. & Scott, M. C. Robust design of semi-automated clustering models for 4D-STEM datasets. APL Machine Learning 1, (2023).DOI
  10. [10]Lee, S. et al. Unsupervised segmentation and clustering workflow for efficient processing of 4D-STEM and 5D-STEM data. http://arxiv.org/abs/2601.17262 (2026).
  11. [11]Shi, C. et al. Uncovering material deformations via machine learning combined with four-dimensional scanning transmission electron microscopy. NPJ Comput. Mater. 8, (2022).arxiv.org/abs/2111.06496
  12. [12]Li, X. et al. Manifold learning of four-dimensional scanning transmission electron microscopy. NPJ Comput. Mater. 5, (2019).DOI
  13. [13]Martineau, B. H., Johnstone, D. N., van Helvoort, A. T. J., Midgley, P. A. & Eggeman, A. S. Unsupervised machine learning applied to scanning precession electron diffraction data. Adv. Struct. Chem. Imaging 5, (2019).DOI
  14. [14]Yoo, T. et al. Unsupervised machine learning and cepstral analysis with 4D-STEM for characterizing complex microstructures of metallic alloys. NPJ Comput. Mater. 10, (2024).DOI
  15. [15]Kimoto, K. et al. Unsupervised machine learning combined with 4D scanning transmission electron microscopy for bimodal nanostructural analysis. Sci. Rep. 14, (2024).DOI
  16. [16]Bergh, T. et al. Nanocrystal segmentation in scanning precession electron diffraction data. J. Microsc. 279, 158–167 (2020).DOI
  17. [17]McInnes, L., Healy, J. & Melville, J. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. http://arxiv.org/abs/1802.03426 (2020).
  18. [18]McInnes, L., Healy, J. & Astels, S. hdbscan: Hierarchical density based clustering. The Journal of Open Source Software 2, 205 (2017).DOI
  19. [19]Kho, Z. et al. On the use of clustering workflows for automated microstructure segmentation of analytical STEM datasets. APL Mater. 13, (2025).
  20. [20]Duran, E. C. et al. Correlated electron diffraction and energy-dispersive X-ray for automated microstructure analysis. Comput. Mater. Sci. 228, (2023).DOI
  21. [21]Thronsen, E. et al. Scanning precession electron diffraction data analysis approaches for phase mapping of precipitates in aluminium alloys. Ultramicroscopy 255, 304–3991 (2024).DOI
  22. [22]Kumar, M. & Rakita, Y. Prediction of EDS Maps from 4DSTEM Diffraction Patterns Using Convolutional Neural Networks. http://arxiv.org/abs/2508.20657 (2025).
  23. [23]Bruefach, A., Ophus, C. & Scott, M. C. Analysis of Interpretable Data Representations for 4D-STEM Using Unsupervised Learning. Microscopy and Microanalysis 28, 1998–2008 (2022).DOI
  24. [24]Kacher, J. et al. In Situ Transmission Electron Microscopy: Signal processing challenges and examples. IEEE Signal Process. Mag. 39, 89–103 (2022).DOI
  25. [25]Aggarwal, C. C., Hinneburg, A. & Keim, D. A. On the Surprising Behavior of Distance Metrics in High Dimensional Spaces. in Proceedings of the 8th International Conference on Database Theory 420–434 (Springer-Verlag, Berlin, Heidelberg, 2001).
  26. [26]Beyer Kevin and Goldstein, J. and R. R. and S. U. When Is “Nearest Neighbor” Meaningful? in Database Theory — ICDT’99 (ed. Beeri Catriel and Buneman, P.) 217–235 (Springer Berlin Heidelberg, Berlin, Heidelberg, 1999).DOI
  27. [27]Donoho, D. High-Dimensional Data Analysis: The Curses and Blessings of Dimensionality. AMS Math Challenges Lecture 1–32 (2000).
  28. [28]Cao, J. et al. Unsupervised multi-clustering and decision-making strategies for 4D-STEM orientation mapping. Digital Discovery 4, 3610–3622 (2025).DOI
  29. [29]Hinton, G. E. & Roweis, S. Stochastic Neighbor Embedding. in Advances in Neural Information Processing Systems (eds. Becker, S., Thrun, S. & Obermayer, K.) vol. 15 (MIT Press, 2002).
  30. [30]Bridger, A., David, W. I. F., Wood, T. J., Danaie, M. & Butler, K. T. Versatile domain mapping of scanning electron nanobeam diffraction datasets utilising variational autoencoders. NPJ Comput. Mater. 9, (2023).DOI
  31. [31]Genc, A. & Silverstein, R. Neural Object Detection for 4D-STEM: High-Throughput Sub-Pixel Electron Diffraction Pattern Recognition. in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops 3624–3634 (2025).DOI
  32. [32]Padgett, E. et al. The exit-wave power-spectrum transform for scanning nanobeam electron diffraction: robust strain mapping at subnanometer resolution and subpicometer precision. Ultramicroscopy 214, (2020).DOI
  33. [33]Watters, N., Matthey, L., Burgess, C. P. & Lerchner, A. Spatial Broadcast Decoder: A Simple Architecture for Learning Disentangled Representations in VAEs. http://arxiv.org/abs/1901.07017 (2019).
  34. [34]Dong, K. & Zhang, S. Deciphering spatial domains from spatially resolved transcriptomics with an adaptive graph attention auto-encoder. Nat. Commun. 13, (2022).
  35. [35]Achanta, R. et al. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach. Intell. 34, 2274–2281 (2012).pubmed.ncbi.nlm.nih.gov/22641706
  36. [36]Etherington, T. R. Mahalanobis distances and ecological niche modelling: correcting a chi-squared probability error. PeerJ 7, e6678 (2019).DOI
  37. [37]Bishop, C. M. . Pattern Recognition and Machine Learning. (Springer Science + Business Media, 2009).
  38. [38]Steinhardt, P. J., Nelson, D. R. & Ronchetti, M. Bond-Orientational Order in Liquids and Glasses. PHYSICAL REVIEW B vol. 28 (1983).DOI
  39. [39]Haeberle, J., Sperl, M. & Born, P. Distinguishing noisy crystalline structures using bond orientational order parameters. European Physical Journal E 42, (2019).arxiv.org/abs/1906.08111
  40. [40]Altarelli, M., Kurta, R. P. & Vartanyants, I. A. X-ray cross-correlation analysis and local symmetries of disordered systems: General theory. Phys. Rev. B Condens. Matter Phys. 82, (2010).DOI
  41. [41]Huang, S., Francis, C., Ketkaew, J., Schroers, J. & Voyles, P. M. Correlation symmetry analysis of electron nanodiffraction from amorphous materials. Ultramicroscopy 232, (2022).DOI
  42. [42]Krajnak, M. & Etheridge, J. A symmetry-derived mechanism for atomic resolution imaging. Proceedings of the National Academy of Sciences 117, 27805–27810 (2020).DOI
  43. [43]Munshi, J. et al. Disentangling multiple scattering with deep learning: application to strain mapping from electron diffraction patterns. NPJ Comput. Mater. 8, (2022).arxiv.org/abs/2202.00204
  44. [44]Moulavi, D., Jaskowiak, P. A., Campello, R. J. G. B., Zimek, A. & Sander, J. Density-Based Clustering Validation. in Proceedings of the 2014 SIAM International Conference on Data Mining (SDM) 839–847 (2014). doi:10.1137/1.9781611973440.96.DOI
  45. [45]Goodfellow, I., Bengio, Y. & Courville, A. Deep Learning. (MIT Press, 2016).
  46. [46]Lecun, Y., Bottou, L., Bengio, Y. & Haffner, P. Gradient-based learning applied to document recognition. Proceedings of the IEEE 86, 2278–2324 (1998).
  47. [47]Silverstein, R. et al. Interstitial-mediated spinal decomposition pathways leading to strengthening in TiNb. Acta Mater. 304, (2026).DOI
  48. [48]Gao, Y. et al. Unique twinning mode and extended twin boundary core structure associated with symmetry breaking in a multifunctional Ti-Nb-based alloy. Acta Mater. 286, (2025).DOI
  49. [49]Dong, R. et al. Stress-induced α″ martensite and its variant selection in a metastable β titanium alloy under an uniaxial compression. Journal of Materials Research and Technology 30, 1795–1799 (2024).
  50. [50]Marczewski, M. et al. Microstructure and mechanical properties of Ti-Nb alloys: comparing conventional powder metallurgy, mechanical alloying, and high power impulse magnetron sputtering processes for supporting materials screening. J. Mater. Sci. 59, 9107–9125 (2024).DOI

Paper details

vlnc:W_uzy8zjfv
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents