Release notes

Release notes#

[v2.9.0] - 2026-01-27#

This release introduces significant advancements in Nudged Elastic Band (NEB) path initialization and saddle point refinement. The update features the novel Off-path Climbing Image NEB (OCI-NEB) method, a comprehensive suite of Image Dependent Pair Potential (IDPP) initializers, and enhanced support for multi-headed Metatomic potentials.

✨ Major Features#

Off-path Climbing Image NEB (OCI-NEB)#

Standard Climbing Image NEB constrains the highest energy image to move along the elastic band tangent. This release implements OCI-NEB (previously referred to as RONEB), a hybrid method that integrates Min-Mode Following (MMF) directly into the climbing image phase.

  • Mechanism: When the climbing image force drops below a specified threshold (trigger_force or relative trigger_factor), the system switches to the dimer search. This allows the image to break orthogonality with the band and follow the true lowest eigenmode toward the saddle, even if the elastic band curvature deviates from the Minimum Energy Path (MEP).

  • Stability: The implementation includes robust fallback strategies. If the dimer mode and the path tangent diverge significantly (controlled by angle_tol), the system penalizes the trust radius or reverts to standard CI-NEB.

  • Curvature Recovery: The dimer method now caches and restores the configuration with the most negative curvature if the rotation fails to converge, ensuring partial results contribute to the optimization.

The modalities of this method are described in the accompanying publication.

Advanced Path Initialization (IDPP & S-IDPP)#

Linear interpolation often results in high-energy atomic overlaps. This version provides a suite of interpolation strategies based on the Image Dependent Pair Potential (IDPP).

  • Collective IDPP: Solves the IDPP objective function for all images simultaneously using the global optimizer (e.g., L-BFGS).

  • Sequential IDPP (S-IDPP): Grows the path sequentially from the reactant and product inward. This method proves superior for complex reaction coordinates by resolving clashes at the frontiers before interpolating the center.

  • ZBL Repulsion: An optional Ziegler-Biersack-Littmark (ZBL) repulsive potential can now wrap the IDPP objective (sidpp_zbl). This prevents atomic fusion during the initialization of dense paths.

  • Oversampling: The system can now generate an initial path with a higher density of images (e.g., 3x), relax them via IDPP, and decimate the path back to the target image count using cubic Hermite splines.

🚀 Enhancements#

  • Metatomic Variants: Added support for multi-headed machine learning potentials. Users may now specify variant_base, variant_energy, or variant_energy_uncertainty in the configuration to target specific output heads (e.g., energy/pbe0 or energy_uncertainty/ensemble).

  • Onsager-Machlup Action: Implemented spring dynamics based on the Onsager-Machlup action. This allows for adaptive spring constants (om_optimize_k) that scale with the local path curvature and force, improving resolution in curved regions.

  • Peak Analysis: The NEB job now automatically detects local maxima along the spline. It writes these configurations (peakXX_pos.con) and estimates their reaction modes (peakXX_mode.dat), facilitating subsequent dimer searches.

  • Timing Reports: The client now reports breakdown of Real, User, and System time in results.dat and standard logs.

  • Relative triggers: The NEB now uses relative thresholds along with absolute ones for starting the climbing image.

🔧 Configuration Changes#

New parameters are available in the [Nudged Elastic Band] block:

  • initializer: Options include linear, idpp, idpp_collective, sidpp, and sidpp_zbl.

  • ci_mmf: Boolean to enable OCI-NEB.

  • onsager_machlup: Boolean to enable OM-based spring dynamics.

  • setup_mmf_peaks: Boolean to toggle the writing of peak estimates.

Along with updates for the [Metatomic] variants.

Contents