Pith. sign in

REVIEW 3 major objections 3 minor 65 references

Local atomic geometries can be reconstructed from invariant descriptors with only a few tens of components, even when the representation is formally incomplete.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 10:05 UTC pith:D2QB7OPW

load-bearing objection Solid inversion study with a clean method, but read the reported success rates as descriptor-matching rates rather than geometric recovery; claims need qualification and code needs to ship. the 3 major comments →

arxiv 2607.20338 v1 pith:D2QB7OPW submitted 2026-07-22 physics.comp-ph physics.chem-ph

Reconstructing local environments from concise atomistic representations

Jigyasa Nigam , Tuong Phung , Ameya Daigavane , Aria Mansouri Tehrani , Tess Smidt This is my paper
classification physics.comp-ph physics.chem-ph
keywords atomistic descriptorsinverse problempower spectrumbispectrumstructural degeneracycoordinate reconstructiongradient descentlocal atomic environments
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks whether the inverse of an atomistic descriptor problem is tractable: given a rotation-invariant descriptor like the power spectrum or bispectrum, can you recover the local atomic geometry that produced it? The authors show that it can—using descriptors of only 9 (power spectrum) and 35 (bispectrum) components, built with a single radial basis, gradient descent on atomic coordinates reliably reproduces the target descriptor across molecular and materials datasets. This matters because descriptor inversion underpins inverse design, lets researchers expose descriptor degeneracies (distinct geometries that a representation cannot distinguish), and provides a concrete link between changes in descriptor space and physical distortions. They further use the framework to find new near-degenerate environment pairs and to show that linear interpolation in descriptor space maps to non-smooth, sometimes discontinuous geometric trajectories.

Core claim

Even formally incomplete, locally ill-conditioned invariant descriptors can be inverted to accurate atomic geometries. A 9-dimensional power spectrum and a 35-dimensional bispectrum (including five pseudoscalars) reconstruct local environments across molecular and materials datasets, with descriptor errors below 1e-4 in 67% and 90% of cases. Inversion treats neighbor coordinates as free parameters and minimizes descriptor mismatch by gradient descent; degenerate geometries sharing a descriptor are recovered as different initialization basins. Pseudoscalars are necessary to distinguish mirror-image environments. The paper produces new near-degenerate pairs separated by >2 Å and shows that lin

What carries the argument

The engine is a differentiable forward map from an atomic environment to a rotation-invariant feature vector: the local density is expanded in spherical harmonics with a single radial basis (R(r)=r), and correlation orders ν=2 (power spectrum), ν=3 (bispectrum), and ν=4 (trispectrum) are formed by symmetrized tensor products. Inversion is performed by gradient descent on the atomic coordinates themselves, so every intermediate trial is a discrete point cloud. A finite-difference version of the Jacobian sensitivity, Δ_degen, identifies directions along which the descriptor is nearly unchanged, which the authors turn into an optimization to find near-degenerate structure pairs.

Load-bearing premise

The reconstruction is judged successful by descriptor error below 1e-4 or 1e-5, and the paper assumes that this small descriptor error means the recovered geometry is close to the reference; its own Fig. 4 shows this can fail, with bispectra matching to ~1e-7 while geometries differ by more than an angstrom.

What would settle it

Take 100 environments from a standard molecular dataset, invert their bispectrum descriptors from 100 random initializations each, and count how many reconstructions with descriptor RMSE below 1e-5 nonetheless have distance-matrix deviation d_ref above 1 Å. If that fraction is non-negligible (say above 5%), the central claim that such descriptors invert to accurate geometries holds only in descriptor space, not real space; if the fraction is zero, the geometric-accuracy claim is confirmed. The vertical bands in the paper's Fig. 4 suggest the fraction will not be zero.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Compact invariant descriptors (9–35 components) can serve as invertible structural fingerprints rather than just ML input features, enabling descriptor-to-structure reconstruction for inverse design.
  • The same optimization provides a practical test for descriptor completeness: running many random restarts on one descriptor reveals the set of geometrically distinct environments consistent with it.
  • Including pseudoscalar components in the bispectrum is required to resolve mirror-image (chiral) ambiguity; scalar-only descriptors conflate enantiomers.
  • Linear paths in descriptor space do not correspond to smooth geometric transformations—neighbor counts can change discontinuously, so generative models that interpolate descriptors should expect non-physical intermediate structures.
  • Increasing the radial basis resolution partially lifts near-degeneracies, so some of the apparent incompleteness of these descriptors is an artifact of the aggressively truncated radial basis.

Where Pith is reading between the lines

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

  • The paper's success metric (descriptor error below 1e-4 or 1e-5) may overstate geometric reconstruction quality: its own Fig. 4 shows bispectrum mismatches around 1e-7 coexisting with geometric deviations over 1 Å. A structural-quality metric (d_ref or RMSD) should accompany descriptor-error reporting in future work.
  • Because near-degenerate directions persist along extended coordinate ranges, any property model built on a descriptor that cannot resolve these directions will be unreliable for properties sensitive to them; this suggests a practical diagnostic: train a model on a descriptor and inspect the reconstruction landscape for flat directions before trusting extrapolations.
  • The inversion framework could be repurposed as an interpretability tool for learned representations: project a model's latent space onto descriptor space, invert selected points, and inspect the resulting geometries to see what features the model encodes.
  • If radial basis resolution lifts near-degeneracies, then the right way to think about descriptor completeness is as a joint property of correlation order and radial basis, not of the descriptor family alone; benchmark suites for completeness should therefore specify both.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper addresses the inverse problem of recovering a local atomic environment from its invariant atom-centered density-correlation descriptors. Using a single radial basis (n_max=1), the authors consider power-spectrum (ν=2, d=9), bispectrum (ν=3, d=35), and trispectrum (ν=4, d=56) descriptors, and reconstruct point clouds by gradient descent on atomic coordinates with multiple random restarts, with a neighbor-merging step to fix coordination. Experiments cover environments from QM9, the Materials Project, transition-metal complexes, high-symmetry polyhedra, known degenerate B8 pairs, and newly discovered near-degenerate ν=3 and ν=4 pairs. The paper also studies linear interpolation in descriptor space for the Bailar twist and for coordination-changing paths. The central claims are that even formally incomplete or locally ill-conditioned compact descriptors can be inverted to accurate geometric reconstructions, and that the inversion framework provides a numerical tool for probing descriptor degeneracies and sensitivity.

Significance. If substantiated, the result would be of clear value to the atomistic machine-learning community: it turns the forward descriptor map into a practical inverse-design tool, enables systematic identification of descriptor degeneracies, and offers a way to interpret how paths in descriptor space translate into structural distortions. The paper's strengths are its clean differentiable formulation (Eq. (12)), the use of an independent higher-order descriptor (ν=4 trispectrum) to confirm recovery of degenerate B8 geometries, and the constructive search for near-degenerate environments with a higher-order check. The main weakness is that the dataset-wide success criterion is defined purely by descriptor error, which, as the paper itself shows in Fig. 4, does not by itself imply geometric nearness.

major comments (3)
  1. [§III.A, Eqs. (13)–(14), Fig. 4] The operational definition of 'success' is a descriptor-error threshold: L<10^{-4} in §III.A and L<10^{-5} in §III.B. Since the forward map is non-injective, small descriptor error does not guarantee geometric proximity. Fig. 4 explicitly shows bispectrum mismatches of order 10^{-7} accompanying d_ref values above 1 Å. Therefore the reported success rates (67% for ν=2, 90% for ν=3, 30% for ν=4) are rates of descriptor matching, not demonstrated rates of geometric recovery. The RMSD values in Fig. 2b are computed only for trials that passed the descriptor-error threshold, and no geometric success threshold is reported. Because the abstract's central claim is 'accurate geometric reconstructions', the paper should report a geometric success rate (e.g., the fraction of trials with RMSD or d_ref below a physically meaningful threshold) and/or show the joint distribution of Δξ and RMSD over al
  2. [§III.B, Fig. 4, and §IV] The manuscript itself states in §III.B that 'a low descriptor reconstruction error need not necessarily imply the recovered geometry is close to either reference.' This is precisely the issue that affects the dataset-wide interpretation in §III.A. For the B8 degeneracy experiment, the use of the independent ν=4 trispectrum to check that recovered structures cluster near A+ or A- is a legitimate external confirmation. The same logic should be applied to the dataset experiments: report how many of the 'successful' reconstructions actually lie near the reference geometry rather than in the vertical bands illustrated in Fig. 4. Without this, the paper's abstract and conclusion overstate the strength of the dataset-level evidence.
  3. [§III.C, Eq. (17)] The near-degeneracy search in §III.C minimizes Δdegen while including an additional penalty on d_ref to avoid trivial degeneracies. This is reasonable, and the ν=5 check in Fig. 5b provides independent evidence that the optimized pair is genuinely nearly degenerate. However, for the ν=4 pair in Fig. 5a, the SI only provides a random-direction comparison for the ν=3 pair (Fig. S3). Adding the analogous random-baseline comparison for the ν=4 pair would make the claim that the found direction is specifically unconstrained by the descriptor more robust.
minor comments (3)
  1. [Eq. (13) and §III.A] The text calls L a mean absolute error, while Eq. (13) defines Δξ as a root mean square error over squared deviations. Please harmonize the terminology for the optimization loss and the reported descriptor error.
  2. [§II.B and SI] Several optimization hyperparameters (learning rate schedule, number of gradient steps, number of random initializations, θ_thresh for merging, and the rule for reintroducing atoms) are described qualitatively. For a methods paper, these should be specified precisely, or the code should be made available at the time of review.
  3. [Throughout] The notation for the distance-based metric alternates among d_ref, dref, and d_{ref}; please use a single notation. Also, the caption of Fig. 2b refers to the mean absolute error while the text refers to RMSE.

Circularity Check

0 steps flagged

No significant circularity: inversion minimizes descriptor loss by design, and geometric recovery is checked against independent metrics (RMSD, d_ref, trispectrum).

full rationale

The paper's forward map (Eqs. 2-6) is a fixed, parameter-free transformation, and the inversion in Eq. 12 is a coordinate-space optimization of descriptor mismatch. Defining success by a threshold on that same loss (L<1e-4 or L<1e-5) is a convergence criterion, not a hidden parameter fit being relabeled as prediction. The geometric claims are separately evaluated: RMSD (Eq. 14), the permutation-invariant distance d_ref (Eq. 15), and, for the B8 degeneracy study, a trispectrum metric that was not used during optimization (Sec. III.B and Fig. 3). That independent higher-order check is legitimate external confirmation. The near-degenerate pair construction in Sec. III.C does minimize the very quantity reported (Δdegen, Eq. 17), but the paper explicitly calls this a 'constructive approach'; it is an existence/synthesis result, not an empirical prediction derived from the descriptor. The added d_ref penalty also ensures the pairs are not trivially related by symmetry. Self-citations (e.g., Refs. 18, 52, 59) supply descriptor implementations and degenerate-pair data, but the core inversion framework does not rest on an unverified self-cited uniqueness theorem; the known degeneracies originate from independent analytical constructions. The manuscript itself flags the central interpretive limitation in Sec. III.B/Fig. 4: 'a low descriptor reconstruction error need not necessarily imply the recovered geometry is close to either reference.' This undercuts the strength of the abstract's 'accurate geometric reconstructions' phrasing and means the reported success rates are descriptor-matching rates, but this is an evidential/correctness concern, not circularity. Overall, the derivation chain is self-contained: no fitted parameter is renamed as a prediction, and no load-bearing conclusion reduces to its own input by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests mainly on hand-chosen descriptor truncations and optimization thresholds, plus the assumption that descriptor-space proximity implies configuration-space proximity (contradicted in the paper's Fig. 4) and that gradient descent explores the non-convex landscape adequately. No new physical entities are introduced.

free parameters (4)
  • Descriptor truncation (n_max=1, l_max per ν) = n_max=1; l_max=8 (ν=2), 6 (ν=3), 3 (ν=4)
    The compactness of the descriptors is the central object of study; these hand-chosen truncations determine all results. Single radial basis Rn(r)=r creates additional degeneracies (Eq. 7–8).
  • Optimization hyperparameters = not fully specified (learning rate η, initial point count, merge threshold θ_thresh, steps T)
    Control the success rate; the paper does not report sweeps or sensitivity for η/θ_thresh; SI gives some details.
  • Success thresholds = MAE <1e-4 (dataset reconstructions); RMSE <1e-5 (B8 degeneracy)
    These thresholds define 'successful' reconstructions and therefore the headline success rates; they are not derived from physical criteria.
  • Degeneracy-optimization penalty on d_ref = not reported numerically
    Added 'to favor accidental degeneracies over trivial structural degeneracies' — its weight shapes the optimized pair but the value is not given.
axioms (4)
  • standard math Spherical harmonics form a complete orthonormal basis on S² and Clebsch-Gordan coefficients are exact; NICE/ACDC invariants equal RTP/Gaunt up to scaling.
    Used throughout §II to define density coefficients and invariants; the implementation-equivalence claim is stated in §II A.
  • domain assumption Locality: a local atom-centered environment (within cutoff) contains enough information to reconstruct the environment; neighbors outside cutoff are ignored.
    Defines the environment A_i in Eq. (1) and the entire reconstruction target; standard for atomistic descriptors.
  • ad hoc to paper Gradient descent on atomic coordinates explores the highly non-convex reconstruction landscape adequately across random seeds.
    The whole method relies on this; no guarantee of global convergence is given, and the observed success rates below 100% indicate the assumption does not always hold.
  • ad hoc to paper CrystalNN/Voronoi cutoff retains the relevant nearest-neighbor environment and suppresses collinear-atom degeneracies of the single-radial-basis descriptor.
    Used to extract environments in §III A; the single radial basis Rn(r)=r makes collinear atoms degenerate (Eq. 8), so the preprocessing is load-bearing for the dataset results.

pith-pipeline@v1.3.0-alltime-deepseek · 23896 in / 11753 out tokens · 92312 ms · 2026-08-01T10:05:58.333228+00:00 · methodology

0 comments
read the original abstract

Symmetry-based representations of local atomic structure, such as the power spectrum or bispectrum, are routinely used to characterize the structural diversity of datasets and as input features for atomistic machine learning. Although these descriptors systematically incorporate increasingly complex geometric correlations, it remains unclear if a given feature can be mapped back to a discrete point cloud, whether such a reconstruction is unique, and how changes in the descriptor are reflected in the underlying atomic geometry. The choice and discretization of the radial and angular bases, as well as the high dimensionality of the resulting feature vectors -- which may contain hundreds or thousands of components -- make this interpretation even more challenging. In this work, we investigate the inverse problem of recovering atomic structures from invariant descriptors. We show that accurate reconstructions can be obtained from remarkably compact descriptors of different correlation orders, each comprising only a few tens of features. Even representations that are formally incomplete or locally ill-conditioned can be inverted to accurate geometric reconstructions of atomic environments across molecular and material datasets. Our reconstruction framework provides a general algorithmic means of identifying approximate degeneracies of invariant descriptors and recovering distinct atomic environments that cannot be distinguished by a given representation. Finally, by reconstructing atomic configurations from descriptors, we examine how perturbations in invariant descriptors of different correlation orders translate into structural distortions.

Figures

Figures reproduced from arXiv: 2607.20338 by Ameya Daigavane, Aria Mansouri Tehrani, Jigyasa Nigam, Tess Smidt, Tuong Phung.

Figure 1
Figure 1. Figure 1: FIG. 1: Mapping a reference local environment [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: a) PCA projection of the bispectrum ( [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: a) Distribution of successful reconstructions [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: For each recovered geometry, we plot its [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: a) A pair of distinct environments containing 7 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: Comparison between the bispectrum ( [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7: Descriptor norms (top) and the structural [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8: Linear interpolation of descriptors ( [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

65 extracted references · 1 canonical work pages

  1. [1]

    O. T. Unke, S. Chmiela, H. E. Sauceda, M. Gastegger, I. Poltavsky, K. T. Schutt, A. Tkatchenko, and K.-R. Mueller, Machine learning force fields, Chemical reviews 121, 10142 (2021)

    2021

  2. [2]

    Westermayr, M

    J. Westermayr, M. Gastegger, K. T. Sch¨ utt, and R. J. Maurer, Perspective on integrating machine learning into computational chemistry and materials science, The Journal of Chemical Physics154(2021)

    2021

  3. [3]

    J. Xia, Y. Zhang, and B. Jiang, The evolution of machine learning potentials for molecules, reactions and materials, Chemical Society Reviews54, 4790 (2025)

    2025

  4. [4]

    Musil, A

    F. Musil, A. Grisafi, A. P. Bart´ ok, C. Ortner, G. Cs´ anyi, and M. Ceriotti, Physics-Inspired Structural Representa- tions for Molecules and Materials, Chem. Rev.121, 9759 (2021)

    2021

  5. [5]

    A. E. Allen, E. Shinkle, R. Bujack, and N. Lubbers, Op- timal invariant sets for atomistic machine learning, npj Computational Materials (2026)

    2026

  6. [6]

    Kohn, Density Functional and Density Matrix Method Scaling Linearly with the Number of Atoms, Phys

    W. Kohn, Density Functional and Density Matrix Method Scaling Linearly with the Number of Atoms, Phys. Rev. Lett.76, 3168 (1996)

    1996

  7. [7]

    Prodan and W

    E. Prodan and W. Kohn, Nearsightedness of electronic matter, Proc. Natl. Acad. Sci.102, 11635 (2005)

    2005

  8. [8]

    Yang, Direct calculation of electron density in density-functional theory, Physical review letters66, 1438 (1991)

    W. Yang, Direct calculation of electron density in density-functional theory, Physical review letters66, 1438 (1991)

    1991

  9. [9]

    Yang and T.-S

    W. Yang and T.-S. Lee, A density-matrix divide-and- conquer approach for electronic structure calculations of large molecules, The Journal of chemical physics103, 5674 (1995)

    1995

  10. [10]

    Jensen,Introduction to computational chemistry (John wiley & sons, 2017)

    F. Jensen,Introduction to computational chemistry (John wiley & sons, 2017)

    2017

  11. [11]

    Behler and M

    J. Behler and M. Parrinello, Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces, Phys. Rev. Lett.98, 146401 (2007)

    2007

  12. [12]

    A. P. Bart´ ok, R. Kondor, and G. Cs´ anyi, On representing chemical environments, Phys. Rev. B87, 184115 (2013)

    2013

  13. [13]

    A. V. Shapeev, Moment Tensor Potentials: A Class of Systematically Improvable Interatomic Potentials, Mul- tiscale Model. Simul.14, 1153 (2016)

    2016

  14. [14]

    B. J. Braams and J. M. Bowman, Permutationally in- variant potential energy surfaces in high dimensionality, Int. Rev. Phys. Chem.28, 577 (2009)

    2009

  15. [15]

    Xie and J

    Z. Xie and J. M. Bowman, Permutationally Invariant Polynomial Basis for Molecular Energy Surface Fitting via Monomial Symmetrization, J. Chem. Theory Com- put.6, 26 (2010)

    2010

  16. [16]

    Drautz, Atomic Cluster Expansion for Accurate and 14 Transferable Interatomic Potentials, Phys

    R. Drautz, Atomic Cluster Expansion for Accurate and 14 Transferable Interatomic Potentials, Phys. Rev. B99, 014104 (2019)

    2019

  17. [17]

    Van de Walle, A complete representation of structure– property relationships in crystals, Nature materials7, 455 (2008)

    A. Van de Walle, A complete representation of structure– property relationships in crystals, Nature materials7, 455 (2008)

    2008

  18. [18]

    Nigam, S

    J. Nigam, S. Pozdnyakov, and M. Ceriotti, Recursive evaluation and iterative contraction ofN-body equiv- ariant features, J. Chem. Phys.153, 121101 (2020)

    2020

  19. [19]

    A. P. Bart´ ok, M. C. Payne, R. Kondor, and G. Cs´ anyi, Gaussian Approximation Potentials: The Accuracy of Quantum Mechanics, without the Electrons, Phys. Rev. Lett.104, 136403 (2010)

    2010

  20. [20]

    V. L. Deringer, A. P. Bart´ ok, N. Bernstein, D. M. Wilkins, M. Ceriotti, and G. Cs´ anyi, Gaussian process regression for materials and molecules, Chemical reviews 121, 10073 (2021)

    2021

  21. [21]

    Bochkarev, Y

    A. Bochkarev, Y. Lysogorskiy, and R. Drautz, Graph atomic cluster expansion for semilocal interactions be- yond equivariant message passing, Physical Review X14, 021036 (2024)

    2024

  22. [22]

    C. B. Mahmoud, A. Anelli, G. Cs´ anyi, and M. Ceriotti, Learning the electronic density of states in condensed matter, arXiv preprint arXiv:2006.11803 (2020)

    Pith/arXiv arXiv 2006

  23. [23]

    V. Fung, P. Ganesh, and B. G. Sumpter, Physically in- formed machine learning prediction of electronic density of states, Chemistry of Materials34, 4848 (2022)

    2022

  24. [24]

    Grisafi, A

    A. Grisafi, A. Fabrizio, B. Meyer, D. M. Wilkins, C. Corminboeuf, and M. Ceriotti, Transferable machine- learning model of the electron density, ACS Central Sci- ence5, 57 (2018)

    2018

  25. [25]

    Z. Lou, A. M. Lewis, and M. Rossi, Long-range machine learning of electron density for twisted bilayer moir\’e materials, arXiv preprint arXiv:2602.09938 (2026)

    Pith/arXiv arXiv 2026

  26. [26]

    Hegde and R

    G. Hegde and R. C. Bowen, Machine-learned Approxima- tions to Density Functional Theory Hamiltonians, Scien- tific reports7, 1 (2017)

    2017

  27. [27]

    Zhang, B

    L. Zhang, B. Onat, G. Dusson, A. McSloy, G. Anand, R. J. Maurer, C. Ortner, and J. R. Kermode, Equivariant Analytical Mapping of First Principles Hamiltonians to Accurate and Transferable Materials Models, npj Com- putational Materials8, 158 (2022)

    2022

  28. [28]

    Nigam, M

    J. Nigam, M. J. Willatt, and M. Ceriotti, Equivariant Representations for Molecular Hamiltonians andN- Center Atomic-Scale Properties, J. Chem. Phys.156, 014 (2022)

    2022

  29. [29]

    S. De, A. P. Bart´ ok, G. Cs´ anyi, and M. Ceriotti, Compar- ing molecules and solids across structural and alchemical space, Physical Chemistry Chemical Physics18, 13754 (2016)

    2016

  30. [30]

    C. W. Rosenbrock, E. R. Homer, G. Cs´ anyi, and G. L. Hart, Discovering the building blocks of atomic systems using machine learning: application to grain boundaries, NPJ Computational Materials3, 29 (2017)

    2017

  31. [31]

    R. K. Cersonsky, B. A. Helfrecht, E. A. Engel, S. Kli- avinek, and M. Ceriotti, Improving sample and feature selection with principal covariates regression, Machine Learning: Science and Technology2, 035038 (2021)

    2021

  32. [32]

    Glielmo, B

    A. Glielmo, B. E. Husic, A. Rodriguez, C. Clementi, F. No´ e, and A. Laio, Unsupervised learning methods for molecular simulation data, Chemical Reviews121, 9722 (2021)

    2021

  33. [33]

    T. C. Nicholas, E. V. Alexandrov, V. A. Blatov, A. P. Shevchenko, D. M. Proserpio, A. L. Goodwin, and V. L. Deringer, Visualization and quantification of geometric diversity in metal–organic frameworks, Chemistry of Ma- terials33, 8289 (2021)

    2021

  34. [34]

    E. D. Donkor, A. Laio, and A. Hassanali, Do machine- learning atomic descriptors and order parameters tell the same story? the case of liquid water, Journal of Chemical Theory and Computation19, 4596 (2023)

    2023

  35. [35]

    C. Zeni, R. Pinsler, D. Z¨ ugner, A. Fowler, M. Horton, X. Fu, Z. Wang, A. Shysheya, J. Crabb´ e, S. Ueda,et al., A generative model for inorganic materials design, Na- ture639, 624 (2025)

    2025

  36. [36]

    Merchant, S

    A. Merchant, S. Batzner, S. S. Schoenholz, M. Aykol, G. Cheon, and E. D. Cubuk, Scaling deep learning for materials discovery, Nature624, 80 (2023)

    2023

  37. [37]

    Zunger, S.-H

    A. Zunger, S.-H. Wei, L. G. Ferreira, and J. E. Bernard, Special quasirandom structures, Physical review letters 65, 353 (1990)

    1990

  38. [38]

    J.-C. Lian, L. Li, G.-F. Huang, W. Hu, and W.-Q. Huang, Highly efficient search strategy for special quasirandom structures, Physical Review B111, 224207 (2025)

    2025

  39. [39]

    van de Walle, P

    A. van de Walle, P. Tiwary, M. de Jong, D. L. Olm- sted, M. Asta, A. Dick, D. Shin, Y. Wang, L.-Q. Chen, and Z.-K. Liu, Efficient stochastic generation of special quasirandom structures, Calphad42, 13 (2013)

    2013

  40. [40]

    S. N. Pozdnyakov, M. J. Willatt, A. P. Bart´ ok, C. Or- tner, G. Cs´ anyi, and M. Ceriotti, On the complete- ness of atomic structure representations, arXiv preprint arXiv:2001.11696 (2020)

    Pith/arXiv arXiv 2001

  41. [41]

    S. N. Pozdnyakov, L. Zhang, C. Ortner, G. Cs´ anyi, and M. Ceriotti, Local invertibility and sensitivity of atomic structure-feature mappings, Open Research Europe1, 126 (2021)

    2021

  42. [42]

    Maennel, O

    H. Maennel, O. T. Unke, and K.-R. Mueller, Complete and efficient covariants for three-dimensional point con- figurations with application to learning molecular quan- tum properties, The Journal of Physical Chemistry Let- ters15, 12513 (2024)

    2024

  43. [43]

    Uhrin, Through the eyes of a descriptor: Constructing complete, invertible descriptions of atomic environments, Phys

    M. Uhrin, Through the eyes of a descriptor: Constructing complete, invertible descriptions of atomic environments, Phys. Rev. B104, 144110 (2021)

    2021

  44. [44]

    V. Fung, S. Jia, J. Zhang, S. Bi, J. Yin, and P. Ganesh, Atomic structure generation from reconstructing struc- tural fingerprints, Machine Learning: Science and Tech- nology3, 045018 (2022)

    2022

  45. [45]

    Cobelli, P

    M. Cobelli, P. Cahalane, and S. Sanvito, Local inver- sion of the chemical environment representations, Physi- cal Review B106, 035402 (2022)

    2022

  46. [46]

    Elijoˇ sius, F

    R. Elijoˇ sius, F. Zills, I. Batatia, S. W. Norwood, D. P. Kov´ acs, C. Holm, and G. Cs´ anyi, Zero-shot molec- ular generation via similarity kernels, Nature Com- munications16, 10.1038/s41467-025-60963-3 (2025), arXiv:2402.08708

    Pith/arXiv arXiv 2025

  47. [47]

    M. J. Willatt, F. Musil, and M. Ceriotti, Atom-density representations for machine learning, J. Chem. Phys. 150, 154110 (2019)

    2019

  48. [48]

    M. A. Caro, Optimizing many-body atomic descrip- tors for enhanced computational performance of machine learning based interatomic potentials, Physical Review B 100, 024112 (2019)

    2019

  49. [51]

    S. Luo, T. Chen, and A. Krishnapriyan, Enabling effi- cient equivariant operations in the fourier basis via gaunt tensor products, inInternational Conference on Learning Representations, Vol. 2024 (2024) pp. 24742–24777

    2024

  50. [52]

    Nigam, S

    J. Nigam, S. N. Pozdnyakov, K. K. Huguenin-Dumittan, and M. Ceriotti, Completeness of atomic structure rep- resentations, APL Machine Learning2(2024)

    2024

  51. [53]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al., Pytorch: An imperative style, high-performance deep learning library, inAdvances in neural information processing systems(2019) pp. 8026–8037

    2019

  52. [54]

    A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. A. Persson, The Materials Project: A materials genome approach to accelerating materials innovation, APL Materials1, 011002 (2013)

    2013

  53. [55]

    Ramakrishnan, P

    R. Ramakrishnan, P. O. Dral, M. Rupp, and O. A. von Lilienfeld, Quantum chemistry structures and properties of 134 kilo molecules, Scientific Data1(2014)

    2014

  54. [56]

    Balcells and B

    D. Balcells and B. B. Skjelstad, tmqm dataset—quantum geometries and properties of 86k transition metal com- plexes, Journal of chemical information and modeling60, 6135 (2020)

    2020

  55. [57]

    Daigavane and S

    A. Daigavane and S. Kim, atomic datasets (2025)

    2025

  56. [58]

    S. P. Ong, W. D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia, D. Gunter, V. L. Chevrier, K. A. Persson, and G. Ceder, Python materials genomics (py- matgen): A robust, open-source python library for mate- rials analysis, Computational Materials Science68, 314 (2013)

    2013

  57. [59]

    Nigam and M

    J. Nigam and M. Ceriotti, Bispectrum degenerate B8 data,https://doi.org/10.5281/zenodo.8003294 (2023). Reconstructing local environments from concise atomistic representations Supplementary Information Jigyasa Nigam,1,∗ Tuong Phung,2 Ameya Daigavane,2 Aria Mansouri Tehrani, 1 and Tess Smidt 1, 2,† 1Research Laboratory of Electronics, Massachusetts Institu...

    work page doi:10.5281/zenodo.8003294 2023

  58. [60]

    Here, we repeat the reconstruction while constraining every optimized environment to comprise six neighbors

    Optimizing to environments with six neighbors In the main text, we observed that the linear interpolation in descriptor space frequently reconstructs to structures with fewer than six neighbors. Here, we repeat the reconstruction while constraining every optimized environment to comprise six neighbors. As shown in Fig. S5, this causes the descriptor error...

  59. [61]

    Geiger and T

    M. Geiger and T. Smidt, e3nn: Euclidean neural networks, arXiv preprint arXiv:2207.09453 (2022)

    Pith/arXiv arXiv 2022

  60. [62]

    Nigam, S

    J. Nigam, S. Pozdnyakov, and M. Ceriotti, Recursive evaluation and iterative contraction of N -body equivariant features, J. Chem. Phys. 153, 121101 (2020)

    2020

  61. [63]

    S. Luo, T. Chen, and A. Krishnapriyan, Enabling efficient equivariant operations in the fourier basis via gaunt tensor products, in International Conference on Learning Representations , Vol. 2024 (2024) pp. 24742–24777

    2024

  62. [64]

    Y. Xie, A. Daigavane, M. Kotak, and T. Smidt, The price of freedom: Exploring expressivity and runtime tradeoffs in equivariant tensor products, arXiv preprint arXiv:2506.13523 (2025)

    Pith/arXiv arXiv 2025

  63. [65]

    Pietrucci and W

    F. Pietrucci and W. Andreoni, Graph Theory Meets Ab Initio Molecular Dynamics: Atomic Structures and Transformations at the Nanoscale, Phys. Rev. Lett. 107, 085504 (2011)

    2011

  64. [66]

    J. Hu, W. Yang, and E. M. Dilanga Siriwardane, Distance matrix-based crystal structure prediction using evolutionary algorithms, The Journal of Physical Chemistry A 124, 10909 (2020)

    2020

  65. [67]

    A. P. Bart´ ok, R. Kondor, and G. Cs´ anyi, On representing chemical environments, Phys. Rev. B87, 184115 (2013). 26

    2013