Image-Dependent Pair Potential (IDPP)

• IDPP is a framework that defines surrogate objectives via interpolated pairwise distances to generate realistic atomic pathways.
• It uses a weighted cost function and NEB-based relaxation to avoid unphysical atom overlaps and bond distortions during interpolation.
• The method significantly reduces computational iterations and errors in minimum energy path searches across diverse chemical systems.

The Image-Dependent Pair Potential (IDPP) is a methodological framework for constructing physically realistic initial pathways between two known atomic configurations, thereby facilitating efficient and robust minimum energy path (MEP) searches in high-dimensional chemical and materials systems. The central idea is to define a surrogate objective function based on deviations from target pairwise distances, which are linearly interpolated between reactant and product states, and then relax an initial path on this surface using a nudged elastic band (NEB) approach. The IDPP method avoids unphysical atom-atom overlaps and bond stretches present in traditional linear Cartesian interpolation, significantly reducing computational overhead and convergence failures in subsequent electronic structure calculations (Smidstrup et al., 2014, Schmerwitz et al., 2023).

1. Mathematical Formulation of the IDPP Objective

The IDPP objective is anchored in the pairwise distances between atoms. For a system of $N$ atoms with Cartesian endpoints $r^\alpha$ (reactant) and $r^\beta$ (product), let $d_{ij}^\alpha$ and $d_{ij}^\beta$ denote the pairwise distances between atoms $i$ and $j$ at the endpoints. For each intermediate image $k$ (in a discretized path with $p+1$ total images), the target pairwise distance is interpolated:

$$d_{ij}^k = d_{ij}^\alpha + \frac{k}{p} (d_{ij}^\beta - d_{ij}^\alpha)$$

The cost for a trial image $r^\alpha$0 is

$r^\alpha$1

with weight $r^\alpha$2, penalizing close contacts disproportionately and thereby maintaining physical plausibility by disfavoring atomic overlap. The total path objective is the sum over non-fixed images:

$r^\alpha$3

The optimized path is obtained by minimizing $r^\alpha$4 with reactant and product endpoints held fixed (Smidstrup et al., 2014, Schmerwitz et al., 2023).

2. Rationale for Pairwise Distance Interpolation

Direct linear interpolation of atomic Cartesian coordinates frequently yields intermediate configurations exhibiting nonphysical features—atoms passing through each other or bonds unreasonably stretched or compressed. By instead interpolating in the space of pairwise distances ($r^\alpha$5 for $r^\alpha$6 atoms), the IDPP approach distributes bond-length changes more uniformly along the path.

Linear interpolation in distance space is computationally efficient and robust, as higher-order schemes (e.g., cubic interpolation) offer little added benefit for most applications. The framework generalizes seamlessly to systems with periodic boundary conditions or large numbers of atoms (Smidstrup et al., 2014, Schmerwitz et al., 2023).

3. Path Optimization via NEB on the IDPP Surface

The NEB algorithm is employed to minimize the IDPP objective across the set of intermediate images. For each image, the "true" force from the IDPP surrogate is

$r^\alpha$7

A uniform spring force is defined between adjacent images, and the NEB projection procedure partitions force components to preserve the path's integrity: the perpendicular component of the IDPP force and the parallel component of the spring force are retained. An optimizer (e.g., velocity projection or quasi-Newton) updates images iteratively until a convergence criterion is met (typical tolerances: max force $r^\alpha$8). The resulting IDPP path is a physically plausible and smooth sequence of atomic configurations, suitable for initializing subsequent high-accuracy MEP searches on realistic potential energy surfaces (Smidstrup et al., 2014).

4. Sequential IDPP (S-IDPP) and Limitations of Linear Initialization

Linear interpolation-based IDPP initialization (LI-IDPP) may still yield unphysical artifacts—especially in systems involving rotations of large fragments or where atoms traverse a geometric partition between endpoints. Such artifacts induce strong repulsive gradients in the IDPP surface, resulting in fragmentation or incorrect dissociation during NEB relaxation. The S-IDPP variant addresses this by iteratively extending the path from both endpoints:

• Images are added one-by-one at idealized geometric intervals, with local spring constants dynamically scaled so that relaxation brings new images to the correct spacing.
• After each insertion, the partial path is re-optimized using IDPP-NEB.
• Once the target number of images is reached, a final NEB step evens out spacing across the path.

This procedure prevents error propagation from endpoint-induced artifacts and produces stable initial paths even for complex reorientations or large systems (Schmerwitz et al., 2023).

Illustrative Failures of LI-IDPP and Successes of S-IDPP

System	LI-IDPP Artifacts	S-IDPP Outcome
Ethylene + cyclopentadiene (Diels–Alder)	Ring bond broken midpath	Five-membered ring preserved
TMBPI isomerization	Ligand bond dissociation, H atom ejected	Rigid ligand rotation, correct isomerization path
[3+2] azidoethyl cycloaddition	Side-chain fragment detachment	Side-chain rotates cleanly to correct saddle
9,9′-Bianthracene rotation	Peripheral aromatic C, H lost	Aromatic integrity intact, correct transition state

All examples demonstrate S-IDPP's efficacy in maintaining chemical integrity and producing near-MEP initial guesses (Schmerwitz et al., 2023).

5. Computational Efficiency and Parameterization

IDPP evaluations involve only the calculation of interatomic distances and application of the prescribed weighting function; no quantum chemical or empirical energies are required, making the construction of IDPP paths orders of magnitude cheaper than even a single ab initio force evaluation. Pathway generation is typically completed in milliseconds to seconds, even for systems with hundreds of atoms.

Recommended parameters:

• Weighting: $r^\alpha$9
• Spring constants: uniform for LI-IDPP; segment-specific for S-IDPP ($r^\beta$0)
• Image count: sufficient to resolve critical rearrangements; the method tolerates large numbers of images due to the path's smoothness
• Convergence tolerances: for IDPP NEB, typically max force $r^\beta$1 to $r^\beta$2, or tighter ($r^\beta$3 max force, $r^\beta$4 RMS for S-IDPP) (Smidstrup et al., 2014, Schmerwitz et al., 2023)

6. Performance Gains and Practical Impact

Empirical data demonstrate that the IDPP method yields initial paths significantly closer to true minimum energy paths than Cartesian linear interpolation. In representative systems:

• Ethane methyl rotation: 3× reduction in SCF and atomic-step iterations
• Al$r^\beta$5Ni cluster atom exchange: nearly 10× fewer SCF iterations
• Amorphous Si atom exchange: ∼1.5× fewer SCF iterations (with greater wall-time savings due to load balancing)
• Systems with challenging rotational or large-amplitude motion: S-IDPP resolves failures present in LI-IDPP

The overall impact is a substantial reduction in computational expense and improved stability of subsequent ab initio NEB calculations, especially evident in parallel architectures where load balancing is sensitive to path smoothness (Smidstrup et al., 2014, Schmerwitz et al., 2023).

7. Implementation Considerations and Outlook

The IDPP method is compatible with Cartesian coordinates and periodic boundary conditions, requiring no special internal coordinates or system-specific adjustments. Both LI-IDPP and S-IDPP integrate seamlessly into NEB-based workflow pipelines. S-IDPP is recommended in scenarios involving large molecular rotations, significant side chain motion, or when initial Cartesian interpolation is known to produce unphysical intermediate structures.

The adoption of IDPP, particularly the sequential insertion (S-IDPP) algorithm, is expected to enable robust and scalable MEP optimization for increasingly complex systems and reactions, minimizing resource expenditure in high-accuracy quantum chemical simulations (Smidstrup et al., 2014, Schmerwitz et al., 2023).