RONEB: Adaptive Transition-State Search

• The paper introduces RONEB, which dynamically couples NEB’s global chain-of-states with MMF’s rapid local convergence to overcome stagnation on potential energy surfaces.
• It employs adaptive force triggers and alignment-based penalties to decouple climbing images for targeted MMF refinement, enhancing efficiency.
• Performance benchmarks reveal up to a 46% reduction in gradient calls while maintaining chemical accuracy and robust convergence.

Reorienting Off-path Nudged Elastic Bands (RONEB) is an adaptive hybrid algorithm for transition-state search and path refinement on potential energy surfaces (PES). It dynamically couples the double-ended chain-of-states stability of the Nudged Elastic Band (NEB) method with the rapid local convergence of single-ended minimum-mode–following (MMF) techniques, notably the Dimer method, to overcome stagnation on flat or rough energy landscapes, accelerate saddle-point identification, and decouple climbing images from ill-defined path tangents when warranted (Goswami et al., 19 Jan 2026).

1. Conceptual Foundation

RONEB was developed in response to limitations of existing NEB approaches. Standard NEB ensures pathway connectivity between reactant and product states via a discretized chain of images, with a "climbing image" driven toward the transition state by inverting the component of the force tangent to the reaction path. However, on flat or noisy regions of the PES, the path tangent may become ill-defined, causing force projections to stall and convergence to degrade. In contrast, MMF methods, such as the Dimer algorithm, offer robust local refinement near saddles by inverting forces along the lowest-curvature (minimum mode) direction but lack global path constraints.

RONEB introduces a hybrid strategy: it tracks the optimization history, applies force-based triggers, and uses alignment-based penalties to determine when to temporarily decouple the climbing image from the band and refine it with MMF steps. This preserves NEB's global stability but allows rapid local convergence and systematically reorients images off the original path only when favorable, enhancing efficiency and avoiding the entrapment associated with noise or ambiguous tangents (Goswami et al., 19 Jan 2026).

2. Mathematical Structure and Algorithms

RONEB is formalized by integrating NEB's energy decomposition, MMF saddle search routines, and adaptive decision rules.

NEB Energy and Forces

Given a chain of $P+2$ images $\{R_0, ..., R_{P+1}\}$ with fixed endpoints and spring constant $k$, the NEB energy is: $$E_{NEB} = \sum_{i=0}^{P+1} V(R_i) + \frac{k}{2} \sum_{i=0}^{P} (\|R_{i+1} - R_i\| - \Delta)^2$$ For movable images $i \in [1,P]$, forces are split: $$F_i^{NEB} = F_i^\perp + F_i^{\parallel, \text{spring}}$$ where

$$F_i^\perp = F_i^{true} - (F_i^{true} \cdot \tau_i)\tau_i$$

$$F_i^{\parallel, \text{spring}} = k \left(\|R_{i+1} - R_i\| - \|R_i - R_{i-1}\|\right) \tau_i$$

$\tau_i$ is the improved path tangent (Henkelman–Uberuaga scheme).

For the climbing image $k$: $\{R_0, ..., R_{P+1}\}$0

MMF (Dimer) Integration

A MMF step operates at $\{R_0, ..., R_{P+1}\}$1, constructing two copies displaced along unit vector $\{R_0, ..., R_{P+1}\}$2: $\{R_0, ..., R_{P+1}\}$3 Curvature is minimized: $\{R_0, ..., R_{P+1}\}$4 The force used for translation is: $\{R_0, ..., R_{P+1}\}$5

Trigger and Back-off Criteria

RONEB hand-off to MMF is conditionally executed based on:

• Relative force trigger:

$\{R_0, ..., R_{P+1}\}$6

when $\{R_0, ..., R_{P+1}\}$7, MMF is initiated.

• Mode-tangent alignment:

$\{R_0, ..., R_{P+1}\}$8

MMF step is aborted if $\{R_0, ..., R_{P+1}\}$9 or positive curvature occurs.

• Adaptive penalization:

$k$0

$k$1

Algorithmic workflow is encapsulated in a loop where climbing-image index stability, force thresholds, and Dimer success criteria determine alternation between global NEB relaxation and MMF refinement (see pseudocode in (Goswami et al., 19 Jan 2026)).

3. Off-path Reorientation and Symmetry Adaptation

RONEB can systematically reorient images off the initial NEB path using group-theoretic and quaternionic methods.

Distortion Symmetry Group Analysis

Distortion symmetry, as implemented in DiSPy (Munro et al., 2018), identifies symmetry elements acting on the chain of NEB images. The distortion symmetry group $k$2 combines conventional image symmetries and reversal operations ($k$3: $k$4).

By decomposing the distortion space $k$5 into irreducible representations, symmetry-adapted projectors $k$6 extract components allowing controlled off-path perturbations: $k$7 Symmetry reduction via group-theoretic projections is a rigorous mechanism for reorienting NEB images in directions that potentially unlock lower-energy pathways.

Quaternion-based Reorientation in Periodic Systems

For molecular crystals, interpolation of periodic NEB paths is further enabled by quaternion-based SLERP for fragment orientations (Goncharova et al., 2024). Each intermediate image uses: $k$8 where

$k$9

Off-path reorientation in RONEB proceeds by identifying images with large perpendicular NEB forces, then applying corrective local rotations, possibly followed by smoothing adjustments via SLERP on neighboring images.

4. Implementation Workflow

Implementation of RONEB proceeds via:

1. Initialization with S-IDPP-generated NEB path, spring constants, and force thresholds.
2. Monitoring climbing-image index and stability counters.
3. Conditional triggering of MMF phase when force and stability prerequisites are met, including dimer construction, rotation toward minimum mode, translation, and update rules for force and alignment.
4. Adaptive adjustment of trigger thresholds and reset of optimizer histories if MMF leads to significant geometric change.
5. Iterative alternation between NEB steps and MMF refinements until the global force tolerance is achieved.

Key pseudocode structure: $i \in [1,P]$3 (Goswami et al., 19 Jan 2026)

5. Performance Benchmarks and Metrics

RONEB achieves substantial reductions in computational effort while retaining transition-state accuracy:

• On the Baker–Chan gas-phase test set (PET-MAD v1.1.0 potential, S-IDPP initialization, 8 images, $$E_{NEB} = \sum_{i=0}^{P+1} V(R_i) + \frac{k}{2} \sum_{i=0}^{P} (\|R_{i+1} - R_i\| - \Delta)^2$$0 eV/Ų), Bayesian negative-binomial regression yields a median reduction of 46.3% in gradient calls [$$E_{NEB} = \sum_{i=0}^{P+1} V(R_i) + \frac{k}{2} \sum_{i=0}^{P} (\|R_{i+1} - R_i\| - \Delta)^2$$1 CrI: $$E_{NEB} = \sum_{i=0}^{P+1} V(R_i) + \frac{k}{2} \sum_{i=0}^{P} (\|R_{i+1} - R_i\| - \Delta)^2$$2].
• On the OptBench Pt(111) heptamer diffusion set (Morse potential, 5 images, $$E_{NEB} = \sum_{i=0}^{P+1} V(R_i) + \frac{k}{2} \sum_{i=0}^{P} (\|R_{i+1} - R_i\| - \Delta)^2$$3 eV/Ų), call reduction averages 28% (397→286), with median 20%.
• Accuracy: saddle configurations converge to within $$E_{NEB} = \sum_{i=0}^{P+1} V(R_i) + \frac{k}{2} \sum_{i=0}^{P} (\|R_{i+1} - R_i\| - \Delta)^2$$4–$$E_{NEB} = \sum_{i=0}^{P+1} V(R_i) + \frac{k}{2} \sum_{i=0}^{P} (\|R_{i+1} - R_i\| - \Delta)^2$$5 Å RMSD, wall-times scale proportionally.
• Bayesian modeling incorporates splines on initial RMSD, group intercepts, and negative-binomial overdispersion; speedup factor posterior median is 0.5 [$$E_{NEB} = \sum_{i=0}^{P+1} V(R_i) + \frac{k}{2} \sum_{i=0}^{P} (\|R_{i+1} - R_i\| - \Delta)^2$$6].

Parameter defaults that yield robust performance include $$E_{NEB} = \sum_{i=0}^{P+1} V(R_i) + \frac{k}{2} \sum_{i=0}^{P} (\|R_{i+1} - R_i\| - \Delta)^2$$7, $$E_{NEB} = \sum_{i=0}^{P+1} V(R_i) + \frac{k}{2} \sum_{i=0}^{P} (\|R_{i+1} - R_i\| - \Delta)^2$$8 eV/Å, stability latch $$E_{NEB} = \sum_{i=0}^{P+1} V(R_i) + \frac{k}{2} \sum_{i=0}^{P} (\|R_{i+1} - R_i\| - \Delta)^2$$9, and alignment tolerance $i \in [1,P]$0 (Goswami et al., 19 Jan 2026).

6. Theoretical Extensions and Practical Considerations

RONEB generalizes beyond standard NEB by direct symmetry adaptation, enabling systematic pathway reorientation for both atomic and molecular crystals:

• Representation-theory methods (as in DiSPy) allow for systematic symmetry reduction, facilitating exploration of otherwise inaccessible paths. Iterative application of symmetry-adapted projectors enables "reorient-and-relax" cycles to bias saddle searches (Munro et al., 2018).
• In periodic systems, hybrid cell+SLERP interpolation and off-path quaternionic corrections further improve pathway initialization and real-time reorientation, mitigating atom overlaps and unrealistic geometries (Goncharova et al., 2024).

Computational guidelines for practical deployment include choosing spring constants ($i \in [1,P]$1) appropriate for system stiffness, adjusting MMF rotation limits ($i \in [1,P]$2), and constraining image displacement magnitudes to prevent optimizer instability. ML force fields (SO3krates architecture) accelerate the NEB cycle while maintaining near-DFT accuracy, supporting automated high-throughput transition-state screening (Goncharova et al., 2024).

7. Context, Impact, and Implications

RONEB represents a rigorous, versatile framework for saddle-point refinement, seamlessly integrating double-ended global optimization with local single-ended acceleration. It provides a robust solution to stagnation and pathway ambiguity in complex PES landscapes, enhances convergence rates, and is compatible with symmetry adaptation and periodic reorientation protocols. As demonstrated across benchmark transition-state datasets and complex surface diffusion mechanisms, RONEB nearly halves computational effort without sacrificing chemical accuracy, establishing itself as a foundational method for automated chemical discovery, materials design, and mechanistic exploration (Goswami et al., 19 Jan 2026, Munro et al., 2018, Goncharova et al., 2024).

A plausible implication is that future developments in transition-state search algorithms may increasingly leverage the dynamic hand-off philosophy, integrating symmetry analysis and machine-learned potentials for fully automated, high-throughput pathway exploration in diverse chemical and materials systems.