Geometry-Preserving Nudged Elastic Band and Dimer Methods under Anisotropic Force Uncertainty

Yangshuai Wang; Yifan Yu

arxiv: 2605.24401 · v1 · pith:UQPMHG3Tnew · submitted 2026-05-23 · 🧮 math.NA · cs.NA· physics.comp-ph

Geometry-Preserving Nudged Elastic Band and Dimer Methods under Anisotropic Force Uncertainty

Yifan Yu , Yangshuai Wang This is my paper

Pith reviewed 2026-06-30 13:34 UTC · model grok-4.3

classification 🧮 math.NA cs.NAphysics.comp-ph

keywords nudged elastic banddimer methoduncertainty quantificationsaddle point searchatomistic simulationsstochastic optimizationminimum energy pathcovariance weighting

0 comments

The pith

UA-NEB and UA-Dimer embed covariance directly into NEB projection and Dimer rotation to cut barrier errors while preserving mean-potential equations.

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

The paper develops uncertainty-aware variants of the nudged elastic band and dimer methods for finding minimum-energy paths and saddle points when force evaluations come from surrogate models with anisotropic, spatially varying errors. It modifies the normal projection step in NEB to an oblique version and the rotation-translation steps in Dimer to covariance-weighted versions, treating the covariance matrix as an optimizer-level reliability weight. These changes keep the underlying equations identical to those of the mean-potential problem. Benchmarks on an analytic test case and a 127-atom tungsten-vacancy system show measurable error reductions, and the stochastic recursions are shown to converge almost surely inside a local stability region under a Lyapunov hypothesis.

Core claim

The central claim is that anisotropic force uncertainty is most effectively used when folded into the constrained geometry of the optimizer—an oblique normal projection for UA-NEB and covariance-weighted rotation and translation for UA-Dimer—rather than collapsed into a scalar weight, and that this geometry-preserving construction yields lower mean barrier errors (21 % versus stochastic NEB on the analytic case, 56 % on the tungsten vacancy) while the mean-potential saddle equations remain unchanged.

What carries the argument

Oblique normal projection (NEB) and covariance-weighted rotation and translation (Dimer), which treat the covariance matrix as a reliability metric inside the constrained update while leaving the mean-potential equations intact.

If this is right

UA-NEB reduces mean barrier error by 21% relative to stochastic NEB on the analytic benchmark.
Full UA-NEB reduces mean barrier error by 56% versus stochastic NEB and by 23% versus diagonal covariance weighting in the 127-atom tungsten-vacancy system.
UA-Dimer reduces reflected-gradient residual by 22% on the analytic benchmark.
Both methods fit Robbins–Monro recursions and converge almost surely inside the local stability neighborhood under the stated Lyapunov hypothesis.

Where Pith is reading between the lines

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

The same covariance-weighted projection idea could be tested on other path-finding algorithms that rely on normal constraints, such as string methods.
If surrogate models supply position-dependent covariances near defects, the geometry-preserving approach may improve reliability of transition-state searches in larger defect-containing crystals.
The local Lyapunov hypothesis could be checked numerically on additional material systems to map the size of the guaranteed convergence basin.

Load-bearing premise

The provided covariance matrix accurately describes the anisotropic force errors and the stochastic iterates remain inside a locally stable neighborhood where the Lyapunov stability hypothesis holds.

What would settle it

A numerical test in which the covariance matrix matches the true anisotropic error distribution yet the reported barrier-error reductions fail to appear or the iterates diverge from the claimed local neighborhood.

read the original abstract

The nudged elastic band (NEB) and Dimer methods are standard tools for computing minimum-energy paths and index-one saddle points in atomistic transition problems. They are increasingly driven by surrogate or learned force models, whose force errors are often anisotropic and spatially varying near transition states and defect cores, where saddle-search iterations are most sensitive. We introduce uncertainty-aware NEB and Dimer methods (UA-NEB, UA-Dimer) that use covariance as an optimizer-level reliability metric while preserving the mean-potential saddle-search equations: an oblique normal projection for NEB and covariance-weighted rotation and translation for Dimer. Both algorithms fit Robbins--Monro recursions; under a local Lyapunov stability hypothesis, verified explicitly for a canonical UA-NEB setting and stated as a hypothesis for UA-Dimer, the stochastic iterations converge almost surely within the corresponding local stability neighborhood. In the analytic benchmark, UA-NEB reduces mean barrier error by $21\%$ relative to stochastic NEB and UA-Dimer reduces the reflected-gradient residual by $22\%$; in the 127-atom tungsten-vacancy benchmark, full UA-NEB reduces mean barrier error by $56\%$ relative to stochastic NEB and by $23\%$ relative to diagonal covariance weighting. These results show that anisotropic uncertainty is most useful when embedded in the constrained geometry of the optimizer rather than collapsed into a scalar acquisition or trust criterion.

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 embeds full anisotropic covariance into the projection and rotation steps of NEB and Dimer while keeping the mean-potential equations, and the benchmarks show error drops, but UA-Dimer convergence rests on an unverified local Lyapunov hypothesis.

read the letter

The core contribution is a pair of modified NEB and Dimer schemes that insert the full covariance matrix from a surrogate force model directly into the geometry constraints—an oblique normal projection for UA-NEB and covariance-weighted rotation plus translation for UA-Dimer—without altering the underlying mean saddle-search equations. Both are cast as Robbins-Monro recursions.

What stands out is the explicit construction that preserves the original deterministic updates while treating the covariance as a reliability weight at the projection level rather than as a scalar trust region or acquisition function. The analytic test and the 127-atom tungsten-vacancy example report concrete gains: 21 % lower mean barrier error for UA-NEB versus plain stochastic NEB, and 56 % versus stochastic NEB (23 % versus diagonal weighting) in the larger system. UA-Dimer shows a 22 % drop in reflected-gradient residual on the analytic case.

The main limitation is the convergence argument. The abstract states that almost-sure convergence holds inside a local stability neighborhood under a local Lyapunov hypothesis, and that this hypothesis is checked explicitly only for one canonical UA-NEB instance; for UA-Dimer it remains a hypothesis. Because the reported error reductions presuppose that the stochastic iterations have in fact converged, the practical reliability claim for UA-Dimer is weaker than the NEB side. No other load-bearing gaps appear in the abstract-level description.

The work is aimed at people who run minimum-energy-path or saddle searches with learned or surrogate potentials whose force errors are anisotropic near defects or transition states. A reader already using NEB or Dimer inside a materials workflow would find the projection modifications and the benchmark numbers directly usable.

It is worth sending to peer review. The idea is practical, the quantitative comparisons are present, and the open theoretical point on UA-Dimer is narrow enough that referees can address it without rejecting the whole contribution.

Referee Report

1 major / 1 minor

Summary. The paper introduces uncertainty-aware NEB and Dimer methods (UA-NEB, UA-Dimer) that embed anisotropic force covariance into the optimizer geometry via oblique normal projection (NEB) and covariance-weighted rotation/translation (Dimer) while exactly preserving the mean-potential saddle-search equations. Both are formulated as Robbins-Monro recursions; under a local Lyapunov stability hypothesis (explicitly verified for one canonical UA-NEB case, hypothesized for UA-Dimer), the iterations converge almost surely. Analytic benchmarks report 21% mean barrier error reduction for UA-NEB and 22% reflected-gradient residual reduction for UA-Dimer versus stochastic baselines; the 127-atom tungsten-vacancy benchmark shows 56% and 23% gains for full UA-NEB versus stochastic NEB and diagonal weighting, respectively.

Significance. If the convergence statements hold, the work shows that anisotropic uncertainty is most effective when retained inside the constrained geometry of the saddle-search optimizer rather than collapsed to a scalar weight. The preservation of the exact mean-potential equations, the parameter-free construction of the uncertainty embedding, and the quantitative error reductions (21-56%) provide a concrete route to more reliable minimum-energy path and index-1 saddle computations when forces come from surrogate or learned models. The approach is internally consistent with the Robbins-Monro framework and avoids ad-hoc parameters.

major comments (1)

[Abstract / Convergence analysis] Abstract (convergence paragraph) and the section stating the local Lyapunov hypothesis: the almost-sure convergence guarantee for UA-Dimer is asserted only as a hypothesis, whereas the same hypothesis is explicitly verified for a canonical UA-NEB setting. Because the reported performance improvements (22% residual reduction analytically; 56% barrier error reduction in the 127-atom benchmark) presuppose that the stochastic recursions have converged inside the local stability neighborhood, the unverified hypothesis for UA-Dimer is load-bearing for the central claim that anisotropic uncertainty can be reliably embedded in the Dimer geometry.

minor comments (1)

[Benchmarks] Table or figure captions for the tungsten-vacancy benchmark should explicitly state the number of independent stochastic runs used to compute the reported mean errors and whether the full UA-NEB or a diagonal-covariance variant was employed in each comparison.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and for highlighting this point on the convergence analysis. We respond to the major comment below.

read point-by-point responses

Referee: [Abstract / Convergence analysis] Abstract (convergence paragraph) and the section stating the local Lyapunov hypothesis: the almost-sure convergence guarantee for UA-Dimer is asserted only as a hypothesis, whereas the same hypothesis is explicitly verified for a canonical UA-NEB setting. Because the reported performance improvements (22% residual reduction analytically; 56% barrier error reduction in the 127-atom benchmark) presuppose that the stochastic recursions have converged inside the local stability neighborhood, the unverified hypothesis for UA-Dimer is load-bearing for the central claim that anisotropic uncertainty can be reliably embedded in the Dimer geometry.

Authors: We acknowledge the distinction drawn in the manuscript: the local Lyapunov stability hypothesis is explicitly verified for a canonical UA-NEB case but stated as a hypothesis for UA-Dimer. The empirical benchmarks (analytic and 127-atom tungsten-vacancy) demonstrate the reported error reductions, providing practical evidence that the recursions reach the claimed performance within the stability neighborhood. The central claim concerns the benefit of embedding anisotropic covariance inside the optimizer geometry rather than collapsing it; this is supported by the quantitative comparisons even under the hypothesized convergence. To strengthen clarity, we will revise the abstract and convergence section to explicitly note that the UA-Dimer results rest on the stated hypothesis and briefly motivate its plausibility via the structural analogy to the verified UA-NEB case. We view this as a clarification rather than a fundamental change to the claims. revision: partial

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper defines UA-NEB and UA-Dimer via explicit geometric constructions (oblique normal projection for NEB; covariance-weighted rotation/translation for Dimer) that are stated to preserve the mean-potential saddle-search equations. Convergence is asserted under an external local Lyapunov stability hypothesis (verified for UA-NEB, hypothesized for UA-Dimer) rather than any parameter fitted inside the paper or self-referential definition. Reported error reductions are empirical benchmark outcomes. No quoted step reduces a claimed prediction or uniqueness result to its own inputs by construction, and no self-citation chain is load-bearing for the central claims.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The convergence claim rests on one domain-assumption hypothesis; no free parameters or invented entities are introduced in the abstract.

axioms (1)

domain assumption local Lyapunov stability hypothesis for stochastic iterations
Invoked to guarantee almost-sure convergence of the Robbins-Monro recursions inside a local neighborhood.

pith-pipeline@v0.9.1-grok · 5787 in / 1251 out tokens · 39124 ms · 2026-06-30T13:34:34.054472+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

4 extracted references · 3 canonical work pages

[1]

Braun, C

J. Braun, C. Ortner, Y. Wang, and L. Zhang , Higher-order far-field boundary conditions for crystalline defects , SIAM Journal on Numerical Analysis, 63 (2025), pp. 520--541, https://doi.org/10.1137/24M165836X

work page doi:10.1137/24m165836x 2025
[2]

H. J. Kushner and G. G. Yin , Stochastic Approximation and Recursive Algorithms and Applications , Springer, 2 ed., 2003, https://doi.org/10.1007/b97441

work page doi:10.1007/b97441 2003
[3]

D. R. Trinkle , Lattice green function for extended defect calculations: Computation and error estimation with long-range forces , Physical Review B, 78 (2008), p. 014110, https://doi.org/10.1103/PhysRevB.78.014110

work page doi:10.1103/physrevb.78.014110 2008
[4]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTION or pop #1 'skip if FUNCTION new.block.checka empty 'skip 'new.block if FUNCTION field.or.null duplicate empty pop "" 'skip ...

[1] [1]

Braun, C

J. Braun, C. Ortner, Y. Wang, and L. Zhang , Higher-order far-field boundary conditions for crystalline defects , SIAM Journal on Numerical Analysis, 63 (2025), pp. 520--541, https://doi.org/10.1137/24M165836X

work page doi:10.1137/24m165836x 2025

[2] [2]

H. J. Kushner and G. G. Yin , Stochastic Approximation and Recursive Algorithms and Applications , Springer, 2 ed., 2003, https://doi.org/10.1007/b97441

work page doi:10.1007/b97441 2003

[3] [3]

D. R. Trinkle , Lattice green function for extended defect calculations: Computation and error estimation with long-range forces , Physical Review B, 78 (2008), p. 014110, https://doi.org/10.1103/PhysRevB.78.014110

work page doi:10.1103/physrevb.78.014110 2008

[4] [4]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTION or pop #1 'skip if FUNCTION new.block.checka empty 'skip 'new.block if FUNCTION field.or.null duplicate empty pop "" 'skip ...