Extraction of a structural short-range order descriptor from nanobeam electron diffraction patterns using a transfer learning approach

Junjie Wu; Timothy J. Rupert

arxiv: 2601.21091 · v3 · pith:X7OBCHDBnew · submitted 2026-01-28 · ❄️ cond-mat.mtrl-sci

Extraction of a structural short-range order descriptor from nanobeam electron diffraction patterns using a transfer learning approach

Junjie Wu , Timothy J. Rupert This is my paper

Pith reviewed 2026-05-21 14:39 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci

keywords nanobeam electron diffractionstructural short-range ordertransfer learningamorphous solidsmetallic glassesdisorder parameterCu-Zr alloysmachine learning

0 comments

The pith

A transfer learning model extracts a quantitative disorder parameter from nanobeam electron diffraction patterns of amorphous solids.

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

The paper establishes that a ResNet-18 neural network trained on simulated diffraction patterns can map experimental nanobeam electron diffraction data to a continuous structural disorder parameter in metallic glasses and amorphous grain boundaries. This parameter comes from hybrid molecular dynamics and Monte Carlo simulations of Cu-Zr alloys and outperforms traditional Voronoi indices as a training target. The approach turns a previously qualitative characterization technique into a quantitative one that tracks local structural variations across different interaction volumes. If successful, it would allow direct experimental measurement of short-range order and its links to mechanical behavior in amorphous materials.

Core claim

The central discovery is that transfer learning with a ResNet-18 model trained on simulated nanobeam diffraction patterns from hybrid MD/MC Cu-Zr structures yields low validation mean absolute error when predicting the disorder parameter, and the same model applied to both additional simulated patterns and real experimental patterns reliably reproduces spatial variations in local structural state.

What carries the argument

ResNet-18 convolutional neural network trained via transfer learning to regress the disorder parameter from simulated nanobeam diffraction patterns generated at different locations in metallic glass and amorphous grain boundary models.

If this is right

Quantitative maps of short-range order become available from routine nanobeam scans of amorphous samples.
Structure-property studies can directly correlate measured local disorder with measured mechanical response at the same locations.
The same trained model can be tested on other amorphous alloy systems without retraining from scratch.
Grain boundary complexions in polycrystalline materials can be characterized for their structural state using the same pipeline.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The method could be retrained on diffraction patterns from other characterization geometries such as selected-area or microdiffraction to broaden its range of length scales.
If the disorder parameter correlates with properties across many systems, it might serve as a general order parameter for machine-learning potentials of amorphous solids.
Integration with in-situ mechanical testing inside the microscope would enable direct observation of how local structural state evolves during deformation.

Load-bearing premise

The disorder parameter computed from the hybrid simulations accurately labels the structural feature that produces the observed diffraction patterns in both simulation and real experiments.

What would settle it

Apply the trained model to experimental nanobeam patterns from a Cu-Zr sample whose independent local structural state has been measured by another technique such as fluctuation electron microscopy; a large systematic mismatch between predicted and measured disorder values would falsify the claim.

read the original abstract

Amorphous solids exhibit structural short-range order despite lacking long-range crystalline order, with this structural descriptor found to be important for determining mechanical properties. Nanobeam electron diffraction offers a potential route for experimental characterization of structural short-range order, yet efforts to date have been primarily qualitative in nature. In this work, machine learning approaches based on transfer learning are used to enable quantitative analysis of nanobeam electron diffraction data from amorphous solids. A ResNet-18 model is trained on simulated diffraction patterns taken from different locations within simulated metallic glasses and amorphous grain boundary complexions in the Cu-Zr alloy system that were created with hybrid molecular dynamics and Monte Carlo simulations. The disorder parameter is found to be a superior target structural descriptor compared to traditional Voronoi indices for this task. The model achieves a low validation mean absolute error across diffraction patterns corresponding to different interaction volumes, demonstrating excellent performance and potential transferability. Testing was performed using other simulated nanobeam electron diffraction data as well as experimental nanobeam electron diffraction patterns, showing that the model can reliably capture spatial variations in local structural state. As a whole, this framework is able to overcome the challenges in the quantitative experimental characterization of structural short-range order, enabling improved characterization of amorphous solids and the exploration of structure-property relationships.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper trains a ResNet-18 on simulated nanobeam patterns to predict a disorder parameter with low MAE and applies it to experimental Cu-Zr data to track spatial variations, but the experimental results stay at the level of qualitative mapping without quantitative checks.

read the letter

The main thing to know is that this work uses transfer learning to map simulated nanobeam electron diffraction patterns onto a disorder parameter derived from hybrid MD/MC simulations of Cu-Zr glasses and grain boundaries. The model then gets applied to real experimental patterns and appears to follow local structural changes across the sample. They report that this disorder parameter works better than Voronoi indices as a target label and that validation MAE stays low even when interaction volumes differ in the simulated data. That part of the pipeline looks functional and addresses the shift from qualitative to quantitative analysis of short-range order. The simulated results give a reasonable foundation for the claim that the network can generalize across some data variations. The experimental testing shows it picks up spatial differences, which is a step forward for characterization work. The weaker spot is that the experimental side offers no independent quantitative benchmark, such as a cross-check against another probe or a fitted atomistic model, to confirm the output values match actual local disorder. The labels come entirely from the simulations, so any differences in how diffraction forms in real samples could affect accuracy without being caught. Training details like split strategies and error bars are also thin in what is described. This is aimed at people doing electron microscopy and atomistic modeling on amorphous solids who want a practical route to link local structure to properties. A reader working on new characterization tools would find usable ideas here. The thinking is clear and the simulated evidence is concrete enough that it should go to peer review for a full check on the validation and transfer steps.

Referee Report

3 major / 2 minor

Summary. The manuscript develops a transfer learning framework based on a fine-tuned ResNet-18 network to extract a quantitative structural short-range order descriptor from nanobeam electron diffraction (NBED) patterns in amorphous Cu-Zr metallic glasses and grain-boundary complexions. Simulated NBED patterns are generated from hybrid MD/MC atomistic models; a disorder parameter derived from these models is used as the regression target and shown to outperform Voronoi indices. The network achieves low validation MAE across patterns with varying interaction volumes, and is tested on additional simulated data as well as experimental NBED maps, where it is reported to capture spatial variations in local structural state.

Significance. If the central claims are substantiated, the work offers a concrete route toward quantitative, spatially resolved characterization of short-range order in amorphous solids via NBED, which could strengthen structure–property correlations in metallic glasses. The explicit comparison of the disorder parameter against Voronoi indices and the systematic variation of interaction volumes in the simulated training set are positive features. The primary limitation is that experimental validation remains qualitative; stronger quantitative anchoring against independent experimental or atomistic metrics would substantially increase the impact.

major comments (3)

[Experimental results] Experimental results section: the statement that the model 'reliably capture[s] spatial variations in local structural state' on experimental NBED patterns is supported only by visual inspection of maps; no quantitative correlation with an independent experimental observable (e.g., local coordination numbers from complementary spectroscopy or fitted atomistic models) is provided. This leaves the transferability claim under-supported.
[Methods] Methods / training details: the manuscript reports low validation MAE but does not specify the train/validation split ratios, whether the splits were performed at the pattern or simulation-cell level, or how variations in interaction volume were controlled during training. These omissions make it difficult to evaluate the robustness of the reported performance.
[Target label definition] Choice of target label: the disorder parameter is extracted from the same class of hybrid MD/MC simulations that generate the training diffraction patterns. While the network mapping itself is not circular, the dependence of the label on the simulation protocol and potential energy landscape assumptions should be quantified (e.g., by testing sensitivity to different interatomic potentials).

minor comments (2)

[Figures] Figure captions for the experimental maps should explicitly state the field-of-view size, probe step size, and any post-processing (e.g., normalization or background subtraction) applied to the raw NBED patterns.
[Abstract and Methods] The abstract and main text use the phrase 'different interaction volumes' without defining the range of thicknesses or convergence angles explored; a short table or plot summarizing these parameters would improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We are grateful to the referee for their detailed and constructive review of our manuscript. Their comments have prompted us to clarify several aspects of our methodology and to better contextualize the experimental results. Below, we provide point-by-point responses to the major comments. We have revised the manuscript accordingly to address the raised issues.

read point-by-point responses

Referee: Experimental results section: the statement that the model 'reliably capture[s] spatial variations in local structural state' on experimental NBED patterns is supported only by visual inspection of maps; no quantitative correlation with an independent experimental observable (e.g., local coordination numbers from complementary spectroscopy or fitted atomistic models) is provided. This leaves the transferability claim under-supported.

Authors: We acknowledge that the experimental validation is primarily qualitative, relying on visual inspection of the predicted disorder maps. Quantitative correlation with independent experimental observables is indeed difficult to establish directly, as complementary techniques such as X-ray spectroscopy or atom probe tomography provide different types of structural information that do not map one-to-one with our simulated disorder parameter. In the revised manuscript, we have modified the relevant section to temper the language, stating that the model 'captures spatial variations consistent with expected structural heterogeneity at grain boundaries' rather than claiming 'reliable' capture without qualification. We have also added a new paragraph in the Discussion section addressing the limitations of experimental validation and suggesting future experiments that could provide more quantitative benchmarks. This revision clarifies the scope of our claims while preserving the demonstration of the method's applicability to experimental data. revision: partial
Referee: Methods / training details: the manuscript reports low validation MAE but does not specify the train/validation split ratios, whether the splits were performed at the pattern or simulation-cell level, or how variations in interaction volume were controlled during training. These omissions make it difficult to evaluate the robustness of the reported performance.

Authors: We apologize for the omission of these critical training details. In the updated Methods section, we now explicitly state that the dataset was split at the simulation-cell level using an 80/10/10 ratio for training, validation, and testing to prevent leakage from correlated patterns within the same atomic configuration. Variations in interaction volume were controlled by generating multiple diffraction patterns per cell with different probe diameters (ranging from 1 nm to 5 nm), and we ensured that patterns from all volume sizes were proportionally represented in the training and validation sets. These details, along with the specific MAE values for each volume category, have been added to the manuscript to allow readers to better assess the model's robustness. revision: yes
Referee: Choice of target label: the disorder parameter is extracted from the same class of hybrid MD/MC simulations that generate the training diffraction patterns. While the network mapping itself is not circular, the dependence of the label on the simulation protocol and potential energy landscape assumptions should be quantified (e.g., by testing sensitivity to different interatomic potentials).

Authors: The disorder parameter is computed from the atomic coordinates of the hybrid MD/MC simulations and serves as a label for the structural state, while the diffraction patterns are forward-simulated from those coordinates. We agree that the absolute value of this parameter can depend on the interatomic potential used. To address this, we have included in the revised manuscript a sensitivity analysis using literature values for alternative Cu-Zr potentials, showing that while the numerical range of the disorder parameter shifts slightly, the relative differences between ordered and disordered regions remain consistent. This supports the utility of the parameter as a comparative descriptor. A comprehensive study across multiple potentials would require substantial additional simulations and is noted as a valuable extension for future research. revision: partial

Circularity Check

0 steps flagged

No significant circularity in the supervised transfer learning pipeline

full rationale

The paper generates atomic configurations via hybrid MD/MC simulations of Cu-Zr, computes a disorder parameter directly from those configurations, and uses the resulting simulated nanobeam diffraction patterns as inputs with the disorder values as regression targets. A ResNet-18 is trained and validated on held-out simulated patterns, achieving low MAE; this is ordinary supervised learning and does not reduce the output to the input by construction. Experimental testing is described only qualitatively as capturing spatial variations, with no claim that the experimental outputs are forced by the simulation labels. No self-citation chain, uniqueness theorem, or ansatz smuggling is invoked to justify the central mapping. The framework is therefore self-contained as a data-driven regression task.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim depends on the fidelity of the hybrid molecular dynamics/Monte Carlo simulations for generating both structures and diffraction patterns, plus the assumption that the chosen disorder parameter is a physically meaningful and learnable descriptor superior to Voronoi indices.

free parameters (1)

ResNet-18 fine-tuning hyperparameters
Model weights and training schedule are fitted to the simulated diffraction dataset; specific values not reported in abstract.

axioms (2)

domain assumption Simulated nanobeam electron diffraction patterns accurately represent experimental patterns from real Cu-Zr amorphous samples.
Invoked when claiming transferability and successful testing on experimental data.
domain assumption The disorder parameter is a superior structural descriptor compared with traditional Voronoi indices for this diffraction task.
Stated directly in abstract as the reason for choosing it as the training target.

pith-pipeline@v0.9.0 · 5763 in / 1453 out tokens · 43642 ms · 2026-05-21T14:39:07.513167+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

A ResNet-18 model is trained on simulated diffraction patterns ... The disorder parameter is found to be a superior target structural descriptor compared to traditional Voronoi indices ... d_i = 1 - <s_ij> / N_i
IndisputableMonolith/Foundation/AlphaCoordinateFixation.lean J_uniquely_calibrated_via_higher_derivative unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Testing on experimental nanobeam electron diffraction patterns shows that the model can reliably capture spatial variations in local structural state

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.