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

This paper introduces uncertainty-aware Nudged Elastic Band and Dimer methods that incorporate anisotropic force covariance directly into the optimizer's geometric constraints to preserve saddle-search equations, demonstrating significantly improved convergence and accuracy in locating transition states compared to standard stochastic approaches.

The Gist

Imagine you are trying to find the lowest point in a vast, foggy mountain range to cross from one valley to another. In the world of atoms and molecules, this "lowest point" is called a saddle point, and finding it is crucial for predicting how materials change, react, or break.

Scientists use two main tools for this job: the Nudged Elastic Band (NEB) and the Dimer method.

• NEB is like stretching a rubber band between two valleys. You pull the band tight, and it naturally settles into the path of least resistance (the "Minimum Energy Path").
• Dimer is like a tightrope walker balancing on a pole. They wiggle the pole to find the direction of the steepest slope and walk that way to find the peak of the hill (the saddle point).

The Problem: The Foggy Map

Usually, these tools rely on a perfect map of the terrain. But in modern science, we often use "learned" maps (AI models) that aren't perfect. These maps have uncertainty.

The paper points out a tricky problem: This uncertainty isn't the same everywhere.

• Sometimes the map is blurry in one direction (like a thick fog to your left) but clear in another (a sunny path to your right).
• Sometimes the fog moves around as you walk.
• Standard tools treat all directions the same. If the map is blurry to the left, they might just take a smaller step everywhere, or they might get confused and walk off the path entirely because they don't know which way is "safe."

The Solution: The "Smart Compass"

The authors, Yifan Yu and Yangshuai Wang, invented new versions of these tools called UA-NEB and UA-Dimer (Uncertainty-Aware).

Instead of just taking a smaller step when the map is blurry, their new tools act like a smart compass that knows exactly which directions are foggy and which are clear.

Here is how they work, using simple analogies:

1. The Rubber Band (NEB) with a Flexible Guide

Imagine your rubber band is being pulled by a guide who knows the terrain.

• Old Way: If the guide is unsure about the terrain to the left, they might just tell the whole band to move slower. This is inefficient.
• New Way (UA-NEB): The guide says, "The terrain to the left is foggy, so don't push the band that way. But the terrain to the right is clear, so push hard there!"
• The Magic: They do this without changing the destination. The band still aims for the exact same lowest path as before; it just gets there more efficiently by ignoring the foggy directions and trusting the clear ones. They call this "preserving the geometry."

2. The Tightrope Walker (Dimer) with a Weighted Pole

Imagine the tightrope walker holding a pole.

• Old Way: If the wind (uncertainty) is blowing hard from the side, the walker might just stop or spin wildly.
• New Way (UA-Dimer): The walker feels the wind. If the wind is strong from the left, they tilt the pole to compensate, using the clear air on the right to stabilize their movement. They adjust their balance based on where the uncertainty is, not just how much there is.

Why Does This Matter?

The paper tested these new tools in two ways:

1. A Mathematical Test: They created a fake mountain with a known path and added "fog" (noise) in specific directions.

   • Result: The new tools found the path with 21% less error than the old tools.
   • Key Insight: Simply knowing how much fog there is (a single number) wasn't enough. You needed to know which direction the fog was in (a map of the fog).

2. A Real-World Test (Tungsten Vacancy): They simulated a hole (vacancy) in a block of Tungsten metal, a common problem in nuclear materials.

   • Result: The new tools reduced the error in predicting the energy barrier by 56% compared to the old standard method, and by 23% compared to a method that only looked at simple, one-dimensional uncertainty.
   • Why it helps: In this metal, the uncertainty was "anisotropic" (different in different directions). The old tools got confused by the complex fog, but the new tools navigated right through it.

The Big Takeaway

The paper argues that when you have a map with uneven fog, you shouldn't just slow down your whole journey. Instead, you should change how you walk.

• Don't change the destination: The goal (the saddle point) stays the same.
• Change the steps: Use the "fog map" to decide which steps to take boldly and which to take carefully.

By embedding this "fog awareness" directly into the walking rules (the math of the algorithm) rather than just using it as a warning sign, the new methods find the correct path much faster and more accurately, especially in complex, real-world materials.

Technical Summary

Technical Summary: Geometry-Preserving Nudged Elastic Band and Dimer Methods Under Anisotropic Force Uncertainty

Problem Statement
Atomistic transition rates depend exponentially on index-one saddle barriers; small errors in barrier height (e.g., a few millielectronvolts) can lead to significant errors in predicted reaction rates. The Nudged Elastic Band (NEB) and Dimer methods are standard tools for locating minimum-energy paths (MEPs) and saddle points. However, in large-scale searches, these optimizers increasingly rely on stochastic, ensemble, or machine-learned interatomic potentials (MLIPs) rather than exact deterministic forces. These models provide mean forces accompanied by anisotropic, spatially varying covariance estimates, which are often largest near defects and transition states.

Existing uncertainty quantification (UQ) approaches typically use covariance to select configurations for high-fidelity labeling or to construct probabilistic path surrogates (e.g., GP-NEB). These methods often act outside the constrained geometry of the optimizer. A critical design challenge arises: a covariance preconditioner that is benign for unconstrained gradient descent can be detrimental to NEB and Dimer methods if it alters the stationarity equations defining the target MEP or saddle. Specifically, generic covariance preconditioning may not commute with the normal–tangential decomposition, potentially shifting the stationary set from the true MEP of the mean potential to a metric-distorted path.

Methodology
The authors propose Uncertainty-Aware NEB (UA-NEB) and Uncertainty-Aware Dimer (UA-Dimer), algorithms that integrate anisotropic covariance directly into the step geometry while strictly preserving the deterministic stationarity equations of the mean potential.

• Uncertainty-Aware NEB (UA-NEB):

  • Oblique Projection: Instead of a standard Euclidean normal projection, UA-NEB employs an oblique projection $Q_{\perp, G}$ defined by the inverse covariance metric $G = (\Sigma_F + \lambda I)^{-1}$. This projection maps the covariance-weighted gradient $G\nabla E$ onto the Euclidean normal subspace.
  • Geometry Preservation: The authors prove that the zero set of the oblique projected force, $Q_{\perp, G} G \nabla E = 0$, is equivalent to the classical MEP condition $\nabla E \parallel \tau$. This ensures that the algorithm seeks the saddle of the mean potential $E$, not a metric-shifted version.
  • Spring Force: The spring component redistributing images along the path is also weighted by the metric $G$ to maintain consistency in the image spacing.
  • Climbing Image: A climbing-image variant is derived that reverses the tangential force using the same oblique projection logic.

• Uncertainty-Aware Dimer (UA-Dimer):

  • Weighted Rotation: The rotation step, which aligns the dimer with the lowest-curvature eigenvector, utilizes a covariance-weighted residual based on the Hessian-vector product (HVP) uncertainty.
  • Weighted Translation: The translation step uses a covariance metric to precondition the reflected gradient. Unlike NEB, the Dimer translation uses a full-rank reflection; the metric dampens unreliable components of the reflected gradient without altering the critical point set.
  • Handoff: A signal-to-noise test is defined to transition from the NEB path search to the local Dimer refinement, ensuring the initial guess for the Dimer is within the stability neighborhood relative to its uncertainty.

• Theoretical Framework:

  • Both algorithms are cast as Robbins–Monro stochastic approximation recursions.
  • Under a local Lyapunov stability hypothesis (explicitly verified for a canonical UA-NEB setting), the authors prove that the stochastic iterations converge almost surely to the target stationary set within a local stability neighborhood.
  • A local $O(1/k)$ mean-square residual convergence rate is established under strengthened contraction conditions.

Key Contributions

1. Algorithmic Derivation: The derivation of covariance-weighted updates for NEB and Dimer that preserve the mean-potential saddle-search equations. This involves constructing oblique projections for NEB and metric-preconditioned reflections for Dimer.
2. Convergence Theory: Casting the methods as stochastic approximations and proving local almost-sure convergence. The paper provides an explicit Lyapunov verification for the canonical UA-NEB case and states the hypothesis for UA-Dimer.
3. Numerical Validation: Testing the methods on an analytic saddle-search problem and a 127-atom body-centered cubic (bcc) tungsten vacancy hop, using paired seeds and matched force-evaluation counts to isolate the effect of the covariance geometry.

Results

• Analytic Benchmark: In a controlled anisotropic noise setting, full UA-NEB reduced the mean barrier error by 21% relative to standard stochastic NEB. The metric-based updates (oblique projection) were identified as the primary driver of this improvement, outperforming scalar penalties or label-refresh strategies.
• Dimer Refinement: In the Dimer test, UA-Dimer reduced the reflected-gradient residual by 22% compared to the standard method, demonstrating improved local refinement of saddle candidates.
• Atomistic Benchmark (W-Vacancy): In the 127-atom tungsten vacancy benchmark, full UA-NEB reduced the mean barrier error by 56% relative to stochastic NEB and by 23% relative to a diagonal covariance weighting (which ignores off-diagonal correlations).
• Anisotropy Sensitivity: The results indicate that full tensor covariance information is most beneficial when the reliable and unreliable force directions are not aligned with the coordinate axes (non-Cartesian regimes), a condition common in defect cores.

Significance
The paper claims that anisotropic uncertainty is most effectively utilized when embedded directly into the constrained geometry of the optimizer (as a step metric) rather than collapsed into a scalar acquisition criterion or used solely for active learning triggers. By preserving the stationarity equations of the mean potential, UA-NEB and UA-Dimer allow the use of stochastic, learned, or ensemble force models without compromising the definition of the target saddle or MEP. The methods are shown to be scalable, relying on covariance-vector products rather than dense matrix inversions, making them compatible with local or low-rank covariance models common in large-scale atomistic simulations.