Application of machine learning in grain-related clustering of Laue spots in a polycrystalline energy dispersive Laue pattern

Amir Tosson; Christian Gutt; Eduard Mikayelyan; Mahmoud Al Humaidi; Mohammad Shokr; Ulrich Pietsch

REVIEW 4 major objections 6 minor 2 references

Application of machine learning in grain-related clustering of Laue spots in a polycrystalline energy dispersive Laue pattern

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Unsupervised clustering of Laue spots recovers grain membership and grain count in polycrystalline energy-dispersive Laue diffraction, with perfect simulated accuracy below 200 grains.

desk verdict The AT-map representation is a plausible feature, but the paper never specifies how experimental reflections get paired with simulated twins, so the reported clustering success presupposes the indexing problem it claims to solve. read the letter →

arxiv 2412.12224 v1 pith:EFZYRMFS submitted 2024-12-16 cond-mat.mtrl-sci cs.LGphysics.app-phphysics.data-an

Amir Tosson , Mohammad Shokr , Mahmoud Al Humaidi , Eduard Mikayelyan , Christian Gutt , Ulrich Pietsch This is my paper · ORCID

classification cond-mat.mtrl-scics.LGphysics.app-phphysics.data-an

keywords machinelearningLauediffractionX-rayhierarchicalclusteringK-meanscrystallographyenergydispersivedetectiongrainidentification

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the longstanding problem of telling which Laue spots in a polycrystalline energy-dispersive Laue diffraction pattern come from which grain can be solved accurately by ordinary unsupervised machine learning. Its recipe is to convert each measured spot into an 'orientation stamp' that records the 3D rotation angle between that spot's crystal plane and the corresponding plane in a simulated reference grain, forming a feature space called the AT-map, and then cluster those stamps. The authors report that a two-stage scheme — hierarchical clustering to propose the number of grains, K-means with the elbow method to confirm it — achieves an adjusted Rand index of 100% on simulated patterns with fewer than 200 grains and correctly recovers nine grains in a real polycrystalline nickel wire. The appeal is practical: the method needs no high-performance computing and, unlike earlier trial-and-error indexing, does not require at least three reflections per grain.

What carries the argument

The load-bearing object is the 'orientation stamp' and its collective representation, the AT-map. For every experimental reflection, the algorithm finds the identical reflection in a simulated reference grain, computes the 3D rotation between their plane normals, and records the three planar components of that rotation (projections onto x–y, y–z, and z–x). This turns raw mixing of many grains' reflections into separated clouds of points, one cloud per grain, which hierarchical clustering and K-means then partition. The orientation stamp is what makes the clustering claim work: it is a similarity feature that is identical for all spots of one grain and different for spots of different grains.

What would settle it

Take a polycrystalline sample whose grain count and orientations are established independently by electron backscatter diffraction, feed its EDLD pattern through the full pipeline including automatic index-twin pairing, and compare the recovered cluster count and adjusted Rand index with the EBSD ground truth; disagreement would falsify the central claim.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that grain-related clustering of Laue spots becomes a textbook clustering problem once each spot is represented not by its detector coordinates or reciprocal-space vector but by its orientation angle relative to a reference grain. Each Laue spot is paired with its indexing-twin in a simulated reference grain; the rotation needed to bring the two into coincidence defines the spot's orientation stamp, and the stamp's three planar projections make up the AT-map. Spots from the same grain share the same stamp, so clusters of stamps are grains: the number of clusters equals the number of grains. The twofold clustering algorithm — HC for an initial cluster count, then elbow-validated K-means — is shown to return perfect cluster assignments on simulated datasets up to about 200 grains, to tolerate as few as one reflection per grain, to survive detector-window constraints on reciprocal space, and to reproduce the previously known nine-grain structure of a measured nickel wire.

Load-bearing premise

Every experimental reflection must have a known 'indexing-twin' in the simulated reference grain, and the paper gives no algorithm or error analysis for pairing; if that correspondence is not already known, orientation stamps cannot be computed.

Editorial extensions

If this is right

Grain counting reduces to counting clusters: the k returned by the HC-plus-elbow-K-means pipeline is the paper's estimate of the number of grains in the illuminated volume.
Reflections-per-grain is no longer a barrier: the non-homogeneous 50-grain simulation, with one to six reflections per grain, still clusters perfectly, unlike trial-and-error methods that need three reflections.
The pipeline runs in near real time: reference-grain simulation stays under a second for practical HKL ranges and the clustering step adds negligible latency, so on-the-fly grain analysis at synchrotron beamlines becomes plausible.
Detector-window constraints do not break the method: when reciprocal space is restricted to the detector's active area in the 50-grain simulation, the algorithm still identifies 47 clusters, matching the constrained set's effective grain count.
Practical operating limits are explicit: simulated accuracy stays at or above 98% for fewer than about 200 grains and for orientation differences of 2° or more, with performance degrading beyond that.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The method's autonomy rests on a step the paper does not specify: pairing every experimental spot with its indexing-twin in the simulated reference grain. If that pairing is automated and robust, the whole pipeline is automatic; if not, the clustering result inherits whatever errors the pairing step makes.
Because the AT-map is just a rotation-invariant encoding, the same recipe could be tried on other multi-entity pattern problems where a reference 'twin' can be defined, such as classifying objects in images by relative pose; the paper itself gestures at these domains but supplies no demonstration.
The reported accuracy is for simulated ground truth; the only experimental validation compares cluster count (nine) with an earlier manual analysis. A stronger test would inject synthetic grains with known orientations into a measured background pattern and measure ARI on the mixture.
The drop beyond 200 grains and below 2° separation suggests the limitation is cluster overlap in the AT-map's angular projections; clustering on the full 3D rotation group (for example quaternion distances) rather than its planar projections might extend the range.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know this paper is a mixed bag. The AT-map idea—encoding each Laue spot's orientation relative to a reference grain and then clustering those stamps—is new in the EDLD context and worth thinking about. But the central step, pairing experimental reflections with their simulated reference twins, is never specified. That pairing is essentially the indexing problem, so the paper validates clustering on data where the hard part is already done.

The paper does well to motivate the problem and to show, in simulation, that HC + elbow K-means on AT-map features cleanly recovers 3, 20, and 50 grains, with ARI 100% below 200 grains. The latency analysis is useful: the clustering itself is cheap. The experimental demonstration on GaAs and Ni wires is at least a sanity check.

The soft spot is load-bearing. Section 3.2 lists "Pairing each exp indexed reflection with its ref indexing-twin" as a step, with no algorithm, criterion, or tolerance. In a real EDLD experiment you don't know which measured q-vector corresponds to which simulated hkl. If you did, you'd already know the grain orientation. So the AT-map cannot be computed from raw data as described. The simulated validation is self-generated, with no error bars, no sensitivity analysis, and no code or data. The experimental section says reflections were "carefully selected and processed" without revealing how. There are also no baseline comparisons against existing indexing methods. Minor: the equation formatting in Section 2 is garbled.

I disagree with the reader's verdict in one respect: the paper is not worthless. The AT-map is a plausible feature for a post-indexing clustering step, and the simulation results suggest the representation is discriminative. But the paper's stated claim—unsupervised grain clustering directly from raw EDLD—is unsupported.

If this came across my desk, I'd send it to peer review: the problem is real, and a revision that specifies the pairing step and validates on data with known ground truth could turn it into a useful contribution. As is, I would not accept it. For a reading group, maybe—it is a decent case study in how a feature representation can smuggle in the answer.

Referee Report

4 major / 6 minor

Summary. The paper proposes an unsupervised machine-learning pipeline for grouping Laue reflections according to their originating grain in energy-dispersive Laue diffraction (EDLD) experiments. The core idea is to transform each reflection into an 'AT-map' of orientation stamps computed with respect to a simulated reference grain, and then to cluster these stamps using hierarchical clustering (HC) combined with K-means and the elbow method. The method is tested on simulated GaAs datasets (3, 20, 50, and constrained 50-grain cases) with reported adjusted Rand index (ARI) values up to 100% for fewer than 200 grains, and on two experimental datasets: a GaAs single crystal (1 grain) and a polycrystalline Ni wire previously analysed as containing 9 grains. The paper claims that the number of clusters equals the number of grains and that the approach is accurate, fast, and free of the three-reflection-per-grain limitation of prior methods.

Significance. If the AT-map representation could be computed directly from raw EDLD data, the proposed pipeline would offer a computationally inexpensive route to grain clustering that avoids exhaustive indexing. The paper is commendable for formulating the problem as a clustering task and for providing pseudocode (Algorithms 1 and 2), latency measurements, and an explicit accuracy metric (ARI) over several simulated scenarios. However, the central contribution hinges on a preparation step that is asserted but never specified: pairing each experimental reflection with a simulated reference twin (Section 3.2). Without an algorithm for that pairing, the orientation-stamp features cannot be produced from an actual unindexed Laue pattern, and therefore the reported validation exercises only the clustering step on data whose labels are already known by construction. The experimental demonstration likewise does not reveal how the pairing was performed. These gaps make the central claim unsupported as stated.

major comments (4)

[Section 3.2] The AT-map calculation begins with 'Pairing each exp indexed reflection with its ref indexing-twin,' but the paper provides no algorithm, matching criterion, tolerance, or error analysis for this pairing. In a real EDLD experiment, reflections arrive as unindexed detector positions and energies; establishing which measured reflection corresponds to which simulated hkl in the reference grain is itself the indexing problem. If the pairing is already known, the grain orientation can be derived directly from a few reflections, making the subsequent clustering redundant. If the pairing is not known, the AT-map cannot be computed and the claimed unsupervised pipeline cannot be applied to raw data. Section 7.2 states only that reflections were 'carefully selected and processed' without revealing any pairing procedure, and no code or data is provided that would allow an independent check of the correspondence.
[Section 6.1] The ARI evaluation is performed exclusively on simulated datasets generated by the same simulation (Algorithm 1) used to construct the reference grain, so the necessary pairing of experimental reflections with reference twins is trivially known by construction. No noise, misindexing, missed reflections, or detector geometric distortions are introduced, no error bars or repeated trials are reported, and no independent experimental ground truth is used for the polycrystalline Ni dataset; the nine-grain result in Section 7.2 is compared only to the same authors' prior analysis. The reported 100% ARI therefore does not establish accuracy on real, unindexed EDLD data and is not a test of the full proposed pipeline.
[Section 5.1] The HC clustering uses 'a threshold of 3° per cluster representing the angular resolution of the detector,' but this threshold is a free parameter. The paper does not justify equating the HC threshold with the detector angular resolution, and no sensitivity analysis shows how the recovered number of clusters and the ARI vary with the chosen threshold. Because the resolvable orientation-stamp difference directly determines which grains can be separated (Section 6.1, Figure 8), the method's stated operating range depends on this unexamined parameter.
[Section 5.4] In the constrained reciprocal-space scenario, 47 clusters are recovered for 50 grains, but the paper does not analyse whether the three missing grains result from merged clusters, from grains with too few reflections, or from reflections outside the constrained detector region. Without this analysis, the paper's statement in Section 2 that 'the number of clusters (k) indicating the number of grains present in the sample' is not supported under constrained detector geometries, which are the typical experimental situation.

minor comments (6)

[Section 1.2] The sentence 'This technique proves particularly valuable in scenarios where quantifying the composition of complex and locally inhomogeneous materials is essential...' is repeated verbatim in consecutive paragraphs, and the phrase 'Here, we demonstrate a new machine-learning approach...' is also duplicated.
[Section 2] The equations for the orientation stamp ε2 and the angles α and β contain garbled symbols and missing typesetting in the provided text; the definitions of 'norms' versus 'normal vectors' are also confusing and should be clarified with consistent notation.
[Section 6.1] The wording '1 < the number of reflections per grain > 6' and 'falls between 20 and 10' should be 'between 1 and 6' and 'between 10 and 20', respectively.
[Figure 11] The caption of Figure 11 lists an extra '(b)' label directly after '(b)', so the subfigure labels do not match the referenced panels.
[Section 8] The concluding paragraph refers to 'electron diffraction analysis,' but the paper concerns X-ray (Laue) diffraction; this appears to be a terminology error.
[Section 1.2] The claim that the method 'has no limitation regarding the number of collected reflections per grain' is in tension with Section 6.1, where accuracy degrades for samples with sparse reflections and for orientation differences below 2°; this should be qualified in the introduction.

Circularity Check

2 steps flagged · score 7.0 of 10

AT-map feature requires pre-indexed, pre-paired reflection data, so the claimed unsupervised grain clustering is applied after the grain-assignment problem has already been solved.

self definitional [Section 3.2, 'AT-map calculation', first bullet]
"Pairing each exp indexed reflection with its ref indexing-twin."

The AT-map is the only similarity feature fed to the clustering algorithms, and its first construction step is to pair every experimental reflection with a simulated reference twin. In a polycrystalline EDLD experiment, identifying which measured spot corresponds to which simulated hkl reflection of the reference grain is precisely the indexing and grain-assignment problem the paper claims to solve. The paper provides no algorithm, tolerance, or error analysis for this pairing; Algorithm 1 only generates a simulated reference grain, not the exp-ref correspondence.
self citation load bearing [Section 7.2, 'Polycrystalline material']
"after carefully selecting and processing several reflections, we visualised the Laue pattern along with the corresponding assigned reflections in Figure 11(c). In a previous analysis (Shokr, 2019), it was determined that this reflection pattern was generated by nine different grains."

The experimental validation rests on the same group's earlier thesis (Shokr, 2019) for both the ground-truth grain count and, implicitly, for the reflection assignments needed to build the AT-map. The paper says the reflections were 'assigned' and that a previous analysis determined nine grains, but it never describes an independent exp-ref pairing procedure. The clustering outcome (k = 9) is therefore compared against the very assignment that must have been used, or is at least required, to prepare the input. This makes the experimental test a validation loop rather than a prediction from raw data.

full rationale

The simulation studies (Figures 4-8) are internally coherent: for synthetic data the exp-ref pairing is known by construction because each rotated grain is generated from the reference grain, so the ARI measures the clustering step on a feature whose computation is artificially supplied by ground truth. The circularity is in the paper's central claim of solving grain clustering from a polycrystalline EDLD pattern. The AT-map is defined in Section 3.2 from 'exp indexed reflections' paired with 'ref indexing-twins', which presupposes that the indexing problem is already solved; no algorithm is given to obtain this pairing from raw detector data. For the single experimental polycrystalline case, Section 7.2 relies on a previous analysis from the same group (Shokr, 2019) to establish the number of grains and the assigned reflections, so the validation is not independent. These issues do not make the clustering math itself incorrect, but they mean the paper's headline result reduces, at the input level, to information that already contains the grain assignment it claims to predict. Score 7 reflects this partial but central circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The method adds two free parameters (the HC threshold and the undefined elbow threshold) and relies on several domain assumptions about grain ideality and the completeness of the simulated reference. The most serious entry is the indexing-twin pairing axiom, which is the unproven load-bearing premise of the entire AT-map construction.

free parameters (2)

HC clustering threshold = 3 degrees
Section 5.1 sets a 3 degree per-cluster threshold, described as the detector angular resolution; this directly determines the HC cluster count and no sensitivity analysis is given.
Elbow convergence threshold = not specified
Algorithm 2 stops when the elbow 'satisfies a specific threshold condition', but that condition is never defined, making the stopping rule an uncalibrated free parameter.

assumptions (4)

domain assumption Every grain is a single-crystal domain whose lattice is related to the reference grain by one fixed 3D rotation.
Section 2 states that all crystallographic planes in a grain are inclined by the same 3D angle relative to the reference; this is true for ideal, undeformed grains but not guaranteed for real polycrystals.
ad hoc to paper Each experimental reflection can be paired with a unique 'indexing-twin' reflection in the simulated reference grain.
Section 3.2 lists this pairing as a required step but provides no algorithm. The central result depends on this correspondence existing and being computable.
domain assumption The simulated reference grain contains all possible reflections relevant to the experimental setup.
Section 3.1 acknowledges that covering the full solid angle is difficult in practice and proposes simulation instead; if the reference list is incomplete, pairing will fail for some spots.
domain assumption Grain orientations are uniformly distributed on a sphere, all variables have equal variance, and the prior probability of finding k grains is uniform.
Section 4 lists these assumptions to justify the choice of clustering algorithms, but they are not justified by EDLD physics and are not checked against experimental data.

invented entities (1)

AT-map (orientation stamp epsilon)
purpose: A 3D feature representation that maps each Laue spot to an orientation angle relative to a simulated reference grain, enabling clustering of spots by grain.
The AT-map is a constructed feature with no external observable handle. Its validity depends entirely on the assumed pairing to reference twins, so it cannot independently corroborate the clustering result.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Application of machine learning in grain-related clustering of Laue spots in a polycrystalline energy dispersive Laue pattern." pith.science (2026). https://pith.science/paper/EFZYRMFS

@misc{pith2026241212224,
  author       = {Pith},
  title        = {Pith review of: Application of machine learning in grain-related clustering of Laue spots in a polycrystalline energy dispersive Laue pattern},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EFZYRMFS}},
  note         = {Machine review of arXiv:2412.12224}
}

read the original abstract

We address the identification of grain-corresponding Laue reflections in energy dispersive Laue diffraction (EDLD) experiments by formulating it as a clustering problem solvable through unsupervised machine learning (ML). To achieve reliable and efficient identification of grains in a Laue pattern, we employ a combination of clustering algorithms, namely hierarchical clustering (HC) and K-means. These algorithms allow us to group together similar Laue reflections, revealing the underlying grain structure in the diffraction pattern. Additionally, we utilise the elbow method to determine the optimal number of clusters, ensuring accurate results. To evaluate the performance of our proposed method, we conducted experiments using both simulated and experimental datasets obtained from nickel wires. The simulated datasets were generated to mimic the characteristics of real-world EDLD experiments, while the experimental datasets were obtained from actual measurements.

Figures

Figures reproduced from arXiv: 2412.12224 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

and Pietsch, U

Abboud, A., Kirchlechner, C., Keckes, J., Nurdan, T.C., Send, S., Micha, J-S., Ulrich, O., Hartmann, R., Struder, L. and Pietsch, U. (2017) ‘Single-shot full strain tensor determination with microbeam X-ray Laue diffraction and a two-dimensional energy-dispersive detector’, Journal of Applied Crystallography, Vol. 50, No. 3, pp.901–908. Brinkmann, U. (200...

work page 2017
[524]

(2020) ‘Introduction to the k-means clustering algorithm based on the elbow method’, Accounting, Auditing and Finance, Vol

Cui, M. (2020) ‘Introduction to the k-means clustering algorithm based on the elbow method’, Accounting, Auditing and Finance, Vol. 1, No. 1, pp.5–8. Deng, D. (2020) ‘DBSCAN clustering algorithm based on density’, in 2020 7th International Forum on Electrical Engineering and Automation (IFEEA), IEEE, September, pp.949–953. Dingel, K., Liehr, A., Vogel, M....

work page 2020

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

and Pietsch, U

Abboud, A., Kirchlechner, C., Keckes, J., Nurdan, T.C., Send, S., Micha, J-S., Ulrich, O., Hartmann, R., Struder, L. and Pietsch, U. (2017) ‘Single-shot full strain tensor determination with microbeam X-ray Laue diffraction and a two-dimensional energy-dispersive detector’, Journal of Applied Crystallography, Vol. 50, No. 3, pp.901–908. Brinkmann, U. (200...

work page 2017

[2] [524]

(2020) ‘Introduction to the k-means clustering algorithm based on the elbow method’, Accounting, Auditing and Finance, Vol

Cui, M. (2020) ‘Introduction to the k-means clustering algorithm based on the elbow method’, Accounting, Auditing and Finance, Vol. 1, No. 1, pp.5–8. Deng, D. (2020) ‘DBSCAN clustering algorithm based on density’, in 2020 7th International Forum on Electrical Engineering and Automation (IFEEA), IEEE, September, pp.949–953. Dingel, K., Liehr, A., Vogel, M....

work page 2020