Characterizing Atomistic Transitions Using Cross-scale Graph-pooled Chebyshev Signatures
Rostyslav Hnatyshyn, Danny Perez · 2026
Read Characterizing Atomistic Transitions Using Cross-scale Graph-pooled Chebyshev Signatures by Rostyslav Hnatyshyn, Danny Perez (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-27381.txt.
Introduction
Hnatyshyn and Perez characterize atomistic transitions with cross-scale graph-pooled Chebyshev signatures. Long simulations bury the key changes in sheer data volume. Their signatures are permutation invariants of an operator. It maps a Coulomb matrix of the starting state to the final state's matrix. A distance between them then sorts them into families. Tests used one long run of a small metallic nanoparticle. These groups showed patterns and hierarchies that older tools miss. Clustering them this way eases analysis of huge runs.
Catalog record
| CREATOR | Rostyslav Hnatyshyn, Danny Perez |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55518 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.27381
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Characterizing Atomistic Transitions Using Cross-scale Graph-pooled Chebyshev Signatures Rostyslav Hnatyshyn, Danny Perez Abstract Large-scale atomistic simulations can produce extreme volumes of information in the form of long trajectories. Reliably and automatically extracting key information from such datasets remains a formidable challenge, especially as it pertains to the analysis of the structural transitions affecting the system. We present a novel approach to characterize and compare atomistic transitions using cross-scale graph-pooled Chebyshev signatures. These signatures are permutation invariants of an operator that transform a Coulomb matrix representation of the initial state of the system into that corresponding to the final state. Using a long-time trajectory of a small metallic nanoparticle, we show that these signatures can be used to define a natural distance metric between transitions that allows for classification and clustering into physically meaningful families. This approach is shown to capture complex patterns and hierarchies of transition types that are inaccessible to traditional techniques, dramatically facilitating the analysis of large-scale simulations. 1 Introduction Atomistic simulations track the evolution of materials and molecules as they undergo rich sequences of structural transformations. Individual simulations can contain millions of such structural transitions, especially when generated with long-time methods such as Accelerated Molecular Dynamics [ 26 ] or kinetic Monte Carlo [ 36 ] . The interpretation of these simulations presents a considerable challenge, as it involves the identification and isolation of key structural events and their classification into meaningful families. This analysis currently overwhelmingly relies on manual visualization and analysis by domain experts. Automating simulation post-processing therefore has the potential to significantly increase the amount of information that can be extracted from computationally intensive simulations and to consequently enable the development of powerful machine learning approaches. Conventional approaches to trajectory analysis cluster individual frames of a simulation in an appropriate structural space (e.g., Cartesian or internal coordinates, RMSD matrices, or more sophisticated descriptors) and interpret clusters as metastable conformations or structural groups. In materials simulations, order parameters such as local bond environments or other structural fingerprints (Steinhardt order parameters [ 33 ] , common neighbor analysis [ 27 ] ) are commonly used to differentiate crystalline phases, amorphous regions, or defect types. These descriptors can be used to track the structural evolution of the material along the simulation trajectory and hence to robustly identify different regimes or trends. These conventional state-based approaches focus on the sequence of configurations that are visited, but not on the process by which individual states transform into each other. Kinetic frameworks instead construct coarse-grained models that can be used to cluster states based on dynamical consideration. For example, Markov state models [ 23 , 4 ] and related transfer-operator or Koopman approaches [ 38 ] estimate a global propagator over a configuration or feature space. Spectral analysis of these operators, through eigenvalues and eigenfunctions or through methods such as Perron cluster cluster analysis [ 6 ] , can then be used to identify metastable macrostates and slow dynamical modes that capture key kinetic events. However, these methods do not intrinsically enable the comparison of the different transition pathways observed in different stages of a simulations. A newer generation of methods pushes further toward a transition-centric view by directly clustering trajectory segments or paths. CATBOSS [ 5 ] , for example, combines change-point detection with density-peak clustering to classify short trajectory segments into metastable pieces and brief transition intervals, thereby explicitly labeling rare events in time. Trajectory map visualizations [ 17 ] use both structural similarity and temporal ordering to highlight interstate transitions and their relative rarity. In biomolecular applications, Path Similarity Analysis [ 31 ] quantifies distances between entire high-dimensional transition paths and clusters them to reveal distinct mechanistic routes between the same end states. Together, these developments underscore a growing recognition that complex systems evolve through ensembles of transitions pathways whose diversity and organization are scientifically important. Despite this progress, there remains a noticeable gap for systems — particularly hard materials — in which simulations generate very large, heterogeneous populations of relatively local events: vacancy or interstitial hops in different environments, slips on various planes, local bond-rearrangement motifs during amorphization or recrystallization, or shape fluctuations in metallic nanoparticles. Indeed, even apparently simple systems are known to evolve through large sets of collective multi-atoms reaction pathways [ 9 ] that can occur in rapid succession. In such cases, the objects of interest are the individual elementary transitions that represent the system before and after structural changes localized in space and time. Existing tools provide limited support for automatically classifying and organizing such events at scale . Frame clustering tells us which states are visited, but only through manual inspection can one identify recurring the local rearrangements mediating those transitions. Global kinetic models reveal which state-to-state jumps are slow or frequent, but they do not, by themselves, compare or classify the mechanisms that connect different sets of metastable states. Segment- and path-based approaches are powerful when one has a modest number of well-resolved trajectories, but they become cumbersome when confronted with millions of short, heterogeneous events distributed across large systems. There is therefore a need for methods that treat transitions in metastable systems as first-class objects. In this work, we take steps in that direction by showing how transitions can be labeled and compared in a way that (i) captures how the underlying topology changes between initial and final configurations, (ii) is agnostic to where in the system the event occurs, and (iii) can be readily analyzed using unsupervised learning approaches to discover families of recurrent transition motifs. Our approach enables the automatic construction of “transition taxonomies” for complex MD datasets: organizing local events into mechanistic families, quantifying their relative frequencies, and relating them to macroscopic observables. The paper is organized as follows. In Section 2 , we formally define transition operators between representations of the physical system before and after a transition and propose permutation-invariant signatures that can be used to characterize and compare transitions, which we refer to as cross-scale Chebyshev signatures. In Section 3 , we then illustrate how these signatures can be used to automate the analysis of very large simulations of metallic nanoparticles that evolve through an extremely rich set of complex collective transitions. We finally discuss the properties of this method in comparison with traditional approaches in Section 4 before concluding. 2 Methods In the following, we construct a transition operator that mathematically describe the changes between the representations of a complex atomistic system before and after a structural transition. We then show how invariants of this operator can be used as signatures in a way that enables meaningful labeling and pairwise comparisons. 2.1 Problem setting We consider pairs of atomic configurations \mathbf{R}_{0},\mathbf{R}_{1}\in\mathbb{R}^{N\times 3} containing the same number of atoms N . Each element of a pair corresponds to the initial and final states of a structural transition \mathbf{R}_{0}\rightarrow\mathbf{R}_{1} . In the following, we consider representations of these atomic configurations in terms of matrices \mathbf{C}_{0}(\mathbf{R}_{0}),\mathbf{C}_{1}(\mathbf{R}_{1})\in\mathbb{R}^{N\times N} . The choice of the mapping between \mathbf{R} and \mathbf{C} is not unique; we here directly draw from the concept of Coulomb matrices [ 29 ] to define each element i,j of \mathbf{C} in terms of functions of the Cartesian distance between atoms i and j . Throughout,