arxiv: 2603.14737 · v1 · submitted 2026-03-16 · ⚛️ physics.chem-ph

Reproducible Orchestration of Best Practices for Reaction Path Optimization with the Nudged Elastic Band

Rohit Goswami (1) ((1) Institute IMX , Lab-COSMO , \'Ecole polytechnique f\'ed\'erale de Lausanne (EPFL) , Station 12 , CH-1015 Lausanne , Switzerland) This is my paper

Pith reviewed 2026-05-15 10:55 UTC · model grok-4.3

classification ⚛️ physics.chem-ph

keywords nudged elastic bandSnakemakemachine learning potentialsreproducibilityreaction path optimizationisomerizationworkflow automationgas-phase molecules

0 comments

The pith

A Snakemake workflow automates all steps of nudged elastic band calculations for gas-phase molecules using machine learning potentials and recovers known energy profiles without manual intervention.

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

The paper establishes a fully automated open-source pipeline that removes ad-hoc scripts and manual steps from NEB calculations by encoding every stage as a Snakemake dependency graph. The workflow retrieves PET-MAD machine learning potentials, performs endpoint minimization and alignment, generates an initial path, and optimizes the band inside the eOn package while pinning all software versions through conda-forge. Validation on the HCN to HNC isomerization shows the pipeline reproduces the accepted single-barrier profile and product energy. A sympathetic reader cares because these pre-processing steps have long been sources of irreproducibility and hidden errors in computational chemistry.

Core claim

The central claim is that an explicit Snakemake dependency graph, coupling PET-MAD potentials to eOn, can execute the complete NEB lifecycle for small gas-phase molecules from model download through converged minimum-energy path, producing identical results across platforms and matching the known HCN-HNC barrier without any user intervention at intermediate stages.

What carries the argument

The Snakemake workflow defined as an explicit dependency graph that orchestrates model retrieval, endpoint preparation, path initialization, and band optimization.

If this is right

All software dependencies are resolved identically on any platform that supports conda-forge, eliminating platform-specific installation errors.
Endpoint minimization, structural alignment, and initial-path generation occur automatically, removing sources of user-induced variation.
The same workflow can be applied to other small gas-phase molecules without rewriting scripts for each new reaction.
Machine learning potentials are directly usable inside established saddle-point search software without custom glue code.

Where Pith is reading between the lines

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

The approach lowers the barrier for non-specialists to run reliable NEB calculations by hiding the usual sequence of manual commands.
Because every step is version-pinned and logged, the workflow supplies a natural audit trail for publication-level reproducibility.
Extending the same dependency-graph pattern to larger systems or periodic boundary conditions would be a direct next test of generality.

Load-bearing premise

That the PET-MAD machine learning potentials are accurate enough for the molecules of interest and that the workflow rules cover every practical edge case that arises in gas-phase NEB runs.

What would settle it

Running the identical workflow on a second, independently verified isomerization reaction and finding that the automated path deviates from the manually confirmed minimum-energy path by more than the expected numerical tolerance.

Figures

Figures reproduced from arXiv: 2603.14737 by CH-1015 Lausanne, \'Ecole polytechnique f\'ed\'erale de Lausanne (EPFL), Lab-COSMO, Rohit Goswami (1) ((1) Institute IMX, Station 12, Switzerland).

**Figure 1.** Figure 1: Workflow pipeline showing the Snakemake dependency graph. Raw endpoints undergo minimization and IRA alignment before SIDPP path generation. The hybrid CI-NEB-MMF optimization then refines the path to the MEP and transition state. snakemake --configfile examples/hcn_isom/config.yaml -c4 This generates plots and intermediates in the results folder. To reproduce the exact figures: # Figures 1-4: TikZ schemat… view at source ↗

**Figure 2.** Figure 2: Hybrid CI-NEB-MMF optimization strategy. The CI-NEB phase performs global path optimization with a climbing image. When the climbing image force falls below threshold (≈ 0.5 eV/Å), the method switches to MMF for local saddle refinement via the lowest curvature mode. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Figure 3: Two-stage IRA alignment process. Raw endpoints are centered and aligned before minimization to establish consistent atom ordering. After geometry relaxation, alignment is reapplied to correct any atom mapping drift introduced during optimization. Both stages solve the joint rotation-permutation problem via the IRA algorithm. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

**Figure 4.** Figure 4: Sequential IDPP (SIDPP) path growth. Unlike standard IDPP which interpolates all images simultaneously, SIDPP adds one image at a time, alternating between reactant and product sides. After each addition, all intermediate images are re-optimized. The step size parameter α controls placement of each new image relative to the frontier. This sequential growth avoids local minima that trap simultaneous interpo… view at source ↗

**Figure 5.** Figure 5: One-dimensional energy profile for the HCN → HNC isomerization. The single barrier of 2.46 eV separates the HCN reactant from the HNC product (0.57 eV above HCN). Colored traces show the optimization history from initial SIDPP guess (outer traces) to the converged path (black). Inset structures show the reactant, saddle point, and product geometries. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

**Figure 6.** Figure 6: Two-dimensional RMSD landscape for the HCN → HNC isomerization. Axes represent permutation-invariant RMSD from reactant (reaction progress s) and orthogonal deviation (d). The color scale indicates interpolated energy. The reactant (R), saddle point (SP), and product (P) basins are clearly resolved. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

read the original abstract

The nudged elastic band (NEB) method is the standard approach for finding minimum energy paths and transition states on potential energy surfaces. Practical NEB calculations require several pre-processing steps: endpoint minimization, structural alignment, and initial path generation. These steps are typically handled by ad-hoc scripts or manual intervention, introducing errors and hindering reproducibility. We present a fully automated, open-source Snakemake workflow for small gas phase molecules that couples modern machine learning potentials (PET-MAD) to the eOn saddle point search software. Each step of the calculation lifecycle is encoded as an explicit dependency graph, from model retrieval and endpoint preparation through path initialization and band optimization. The workflow resolves all software dependencies from conda-forge, ensuring identical execution across platforms. Validation on the HCN to HNC isomerization demonstrates that the automated pipeline recovers the known single-barrier energy profile and product energy without manual intervention.

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 packages a Snakemake workflow that automates NEB steps with PET-MAD and eOn, shown to work on one simple isomerization but not tested on harder cases.

read the letter

The paper gives a Snakemake workflow that fully automates NEB calculations for small gas-phase molecules by linking PET-MAD potentials with the eOn code. Every step from retrieving the model and minimizing endpoints to aligning structures, generating the initial path, and running the band optimization is turned into an explicit dependency graph. Dependencies are resolved through conda-forge so the same environment runs everywhere. This is new as a packaged end-to-end pipeline for these specific tools. It does well by cutting out the manual or script-based pre-processing that usually introduces variability. The validation on the HCN to HNC isomerization recovers the known energy profile and product energy, showing the automation works for that straightforward case without intervention. The main limitation is the single test case. HCN/HNC is a simple single-barrier isomerization, so the workflow has not been checked on multi-barrier paths, systems with symmetry issues, or cases where alignment fails. There are no error bars, no multiple molecules, and no direct comparison to a manual workflow, which leaves the claim that the rules cover all practical edge cases unproven. The evidence supports the result for the reported example but does not establish broad coverage. This paper is for computational chemists running NEB on small molecules who want a reproducible, shareable setup instead of custom scripts. Someone looking for a practical tool to standardize their calculations will get value from the open workflow and the clear description of the rules. The methods follow standard practice, and the citations to PET-MAD and eOn are on point. I would send it for peer review. The core is the workflow implementation, and referees could usefully ask for more test cases to strengthen the reproducibility argument.

Referee Report

1 major / 1 minor

Summary. The manuscript presents a Snakemake workflow that fully automates the NEB calculation lifecycle for small gas-phase molecules by coupling PET-MAD machine learning potentials to the eOn software. The workflow encodes endpoint minimization, structural alignment, initial path generation, and band optimization as an explicit dependency graph, with all dependencies resolved from conda-forge to guarantee cross-platform reproducibility. Validation is shown on the HCN to HNC isomerization, where the automated pipeline recovers the known single-barrier energy profile and product energy without manual intervention.

Significance. The provision of a fully specified, open-source Snakemake workflow with conda-forge dependency resolution is a clear strength that directly supports reproducibility in computational chemistry. If the encoded rules prove robust, the approach could standardize best practices for reaction path optimization and reduce errors from ad-hoc scripting. The integration of modern ML potentials further enhances accessibility for small-molecule studies. The current single-case validation, however, leaves the generality of the automation claim only partially demonstrated.

major comments (1)

[Validation] Validation section: The demonstration is limited to the HCN to HNC isomerization, a simple single-barrier case that recovers the expected profile. This does not test handling of multi-barrier paths, symmetry-related alignments, multiple minima, or failed alignments, which the introduction identifies as practical edge cases. The claim that the dependency graph covers all practical edge cases in gas-phase NEB runs therefore rests on narrower evidence than the title and abstract suggest.

minor comments (1)

[Abstract] Abstract: The statement that the pipeline recovers the known profile would be strengthened by including at least one quantitative metric, such as the computed barrier height or its deviation from the literature value.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback and positive overall assessment of the manuscript. We address the major comment on validation below.

read point-by-point responses

Referee: [Validation] Validation section: The demonstration is limited to the HCN to HNC isomerization, a simple single-barrier case that recovers the expected profile. This does not test handling of multi-barrier paths, symmetry-related alignments, multiple minima, or failed alignments, which the introduction identifies as practical edge cases. The claim that the dependency graph covers all practical edge cases in gas-phase NEB runs therefore rests on narrower evidence than the title and abstract suggest.

Authors: We agree that the validation is limited to a single benchmark system. The HCN to HNC isomerization is a canonical test case that fully exercises the workflow's core steps (endpoint minimization, structural alignment, initial path generation, and band optimization) without manual intervention. However, the manuscript does not claim that the dependency graph covers every practical edge case; the title and abstract emphasize reproducible orchestration of standard best practices for small gas-phase molecules. The rules are intentionally modular to allow extension (e.g., chaining for multi-barrier paths or custom alignment rules), but we acknowledge that additional test cases would better demonstrate generality. We will therefore revise the abstract to clarify the validation scope and add a short discussion paragraph on extensibility and limitations. This is a partial revision. revision: partial

Circularity Check

0 steps flagged

No circularity: workflow description and external validation are self-contained

full rationale

The paper describes a Snakemake workflow that automates NEB steps using external components (PET-MAD potentials, eOn software, conda-forge dependencies) and validates the pipeline on the HCN/HNC isomerization case, recovering a known single-barrier profile. No equations, derivations, fitted parameters, or self-citations appear in the load-bearing claims. The central result is an empirical demonstration on an independent benchmark molecule rather than any reduction of outputs to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The workflow depends on the accuracy of an external ML potential and the completeness of its rule set for NEB preprocessing.

axioms (1)

domain assumption PET-MAD potentials accurately represent the potential energy surface for the studied molecules
Invoked when the workflow uses the model without additional fitting or benchmarking beyond the single validation case.

pith-pipeline@v0.9.0 · 5494 in / 1117 out tokens · 44517 ms · 2026-05-15T10:55:28.616311+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.

Validation on the HCN to HNC isomerization demonstrates that the automated pipeline recovers the known single-barrier energy profile

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.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

[1]

Jacobsen

Hannes Jónsson, Gregory Mills, and Karsten W. Jacobsen. Nudged elastic band method for finding minimum energy paths of transitions. InClassical and Quantum Dynamics in Condensed Phase Simulations, pages 385–404. World Scientific, 1998. doi:10.1142/9789812839664_0016

work page doi:10.1142/9789812839664_0016 1998
[2]

Uberuaga, and Hannes Jónsson

Graeme Henkelman, Blas P. Uberuaga, and Hannes Jónsson. A climbing image nudged elastic band method for finding saddle points and minimum energy paths.Journal of Chemical Physics, 113(22): 9901–9904, November 2000. ISSN 0021-9606. doi:10.1063/1.1329672

work page doi:10.1063/1.1329672 2000
[3]

Vilhjálmur Ásgeirsson, Benedikt Orri Birgisson, Ragnar Bjornsson, Ute Becker, Frank Neese, Christoph Riplinger, and Hannes Jónsson. Nudged Elastic Band Method for Molecular Reactions Using Energy- Weighted Springs Combined with Eigenvector Following.Journal of Chemical Theory and Computation, 17(8):4929–4945, August 2021. ISSN 1549-9618. doi:10.1021/acs.j...

work page doi:10.1021/acs.jctc.1c00462 2021
[4]

Efficient exploration of chemical kinetics, October 2025

Rohit Goswami. Efficient exploration of chemical kinetics, October 2025

work page 2025
[5]

Enhanced climbing image nudged elastic band method with hessian eigenmode alignment, January 2026

Rohit Goswami, Miha Gunde, and Hannes Jónsson. Enhanced climbing image nudged elastic band method with hessian eigenmode alignment, January 2026

work page 2026
[6]

Rohit Goswami, Mikhail Masterov, Sreyan Kamath, Alberto Peña-Torres, and Hannes Jónsson. Efficient implementation of Gaussian Process Regression Accelerated Saddle Point Searches with application to molecular reactions.Journal of Chemical Theory and Computation, 21(16):7935–7943, 2025. ISSN 1549-9618. doi:10.1021/acs.jctc.5c00866

work page doi:10.1021/acs.jctc.5c00866 2025
[7]

Adaptive pruning for increased robustness and reduced computa- tional overhead in Gaussian process accelerated saddle point searches.ChemPhysChem, page e202500730,

Rohit Goswami and Hannes Jónsson. Adaptive pruning for increased robustness and reduced computa- tional overhead in Gaussian process accelerated saddle point searches.ChemPhysChem, page e202500730,

work page
[8]

Adaptive Pruning for Increased Robustness and Reduced Computational Overhead in Gaussian Process Accelerated Saddle Point Searches.ChemPhysChem, 27 (4):e202500730, February 2026

ISSN 1439-7641. doi:10.1002/cphc.202500730

work page doi:10.1002/cphc.202500730
[9]

Peterson

Andrew A. Peterson. Acceleration of saddle-point searches with machine learning.The Journal of Chemical Physics, 145(7):074106, August 2016. doi:10.1063/1.4960708

work page doi:10.1063/1.4960708 2016
[10]

Hall, Christopher H

Felix Mölder, Kim Philipp Jablonski, Brice Letcher, Michael B. Hall, Christopher H. Tomkins-Tinch, Vanessa Sochat, Jan Forster, Soohyun Lee, Sven O. Twardziok, Alexander Kanitz, Andreas Wilm, Manuel Holtgrewe, Sven Rahmann, and Johannes Köster. Sustainable data analysis with Snakemake. F1000Research, 10:33, April 2021. doi:10.12688/f1000research.29032.2

work page doi:10.12688/f1000research.29032.2 2021
[11]

PET-MAD as a lightweight universal interatomic potential for advanced materials modeling.Nature Communications, 16(1):10653, November 2025

Arslan Mazitov, Filippo Bigi, Matthias Kellner, Paolo Pegolo, Davide Tisi, Guillaume Fraux, Sergey Pozdnyakov, Philip Loche, and Michele Ceriotti. PET-MAD as a lightweight universal interatomic potential for advanced materials modeling.Nature Communications, 16(1):10653, 2025. ISSN 2041-1723. doi:10.1038/s41467-025-65662-7

work page doi:10.1038/s41467-025-65662-7 2025
[12]

Abbott, Philip Loche, Arslan Mazitov, Davide Tisi, Marcel F

Filippo Bigi, Joseph W. Abbott, Philip Loche, Arslan Mazitov, Davide Tisi, Marcel F. Langer, Alexander Goscinski, Paolo Pegolo, Sanggyu Chong, Rohit Goswami, Sofiia Chorna, Matthias Kellner, Michele 7 A Preprint Ceriotti, and Guillaume Fraux. Metatensor and metatomic: Foundational libraries for interoperable atomistic machine learning, August 2025

work page 2025
[13]

Nocedal ,\ 10.1090/S0025-5718-1980-0572855-7 journal journal Mathematics of Computation \ volume 35 ,\ pages 773 ( year 1980 ) NoStop

Jorge Nocedal. Updating quasi-Newton matrices with limited storage.Mathematics of Computation, 35 (151):773–782, 1980. doi:10.1090/S0025-5718-1980-0572855-7

work page doi:10.1090/s0025-5718-1980-0572855-7 1980
[14]

IRA: A shape matching approach for recognition and comparison of generic atomic patterns.Journal of Chemical Information and Modeling, 61(11):5446–5457, November 2021

Miha Gunde, Nicolas Salles, Arnaud Hémeryck, and Livia Martin-Samos. IRA: A Shape Matching Approach for Recognition and Comparison of Generic Atomic Patterns.Journal of Chemical Information and Modeling, 61(11):5446–5457, November 2021. ISSN 1549-9596. doi:10.1021/acs.jcim.1c00567

work page doi:10.1021/acs.jcim.1c00567 2021
[15]

Yorick L. A. Schmerwitz, Vilhjálmur Ásgeirsson, and Hannes Jónsson. Improved Initialization of Optimal Path Calculations Using Sequential Traversal over the Image-Dependent Pair Potential Sur- face.Journal of Chemical Theory and Computation, 20(1):155–163, January 2024. ISSN 1549-9618. doi:10.1021/acs.jctc.3c01111

work page doi:10.1021/acs.jctc.3c01111 2024
[16]

Improved initial guess for minimum energy path calculations.The Journal of Chemical Physics, 140(21):214106, June 2014

Søren Smidstrup, Anders Pedersen, Kurt Stokbro, and Hannes Jónsson. Improved initial guess for minimum energy path calculations.Journal of Chemical Physics, 140(21):214106, 2014. ISSN 0021-9606. doi:10.1063/1.4878664

work page doi:10.1063/1.4878664 2014
[17]

Incorporation of Internal Coordinates Interpolation into the Freezing String Method.Journal of Chemical Theory and Computation, November 2025

Jonah Marks and Joseph Gomes. Incorporation of Internal Coordinates Interpolation into the Freezing String Method.Journal of Chemical Theory and Computation, November 2025. doi:10.1021/acs.jctc.5c01492

work page doi:10.1021/acs.jctc.5c01492 2025
[18]

Harris, Oleg L

Tanja van Mourik, Gregory J. Harris, Oleg L. Polyansky, Jonathan Tennyson, Attila G. Császár, and Peter J. Knowles. Ab initio global potential, dipole, adiabatic, and relativistic correction surfaces for the HCN–HNC system.The Journal of Chemical Physics, 115(8):3706–3718, 2001. doi:10.1063/1.1383586

work page doi:10.1063/1.1383586 2001
[19]

Two-dimensional RMSD projections for reaction path visualization and validation

Rohit Goswami. Two-dimensional RMSD projections for reaction path visualization and validation. MethodsX, page 103851, March 2026. ISSN 2215-0161. doi:10.1016/j.mex.2026.103851

work page doi:10.1016/j.mex.2026.103851 2026
[20]

From A to B in free energy space

Davide Branduardi, Francesco Luigi Gervasio, and Michele Parrinello. From A to B in free energy space. Journal of Chemical Physics, 126(5):54103, February 2007. ISSN 0021-9606. doi:10.1063/1.2432340

work page doi:10.1063/1.2432340 2007
[21]

doi:10.1038/s41597-022-01870-w

Mathias Schreiner, Arghya Bhowmik, Tejs Vegge, Jonas Busk, and Ole Winther. Transition1x - a dataset for building generalizable reactive machine learning potentials.Scientific Data, 9(1):779, December 2022. doi:10.1038/s41597-022-01870-w

work page doi:10.1038/s41597-022-01870-w 2022
[22]

Huguenin-Dumittan, Melika Honarmand, Qianjun Xu, Egor Rumiantsev, Wei Bin How, Marcel F

Philip Loche, Kevin K. Huguenin-Dumittan, Melika Honarmand, Qianjun Xu, Egor Rumiantsev, Wei Bin How, Marcel F. Langer, and Michele Ceriotti. Fast and flexible long-range models for atomistic machine learning.Journal of Chemical Physics, 162(14):142501, April 2025. doi:10.1063/5.0251713. 8 A Preprint Figure 2.Hybrid CI-NEB-MMF optimization strategy. The C...

work page doi:10.1063/5.0251713 2025