Parameter Selection Guide

Date:

2026-03-13

1 Overview

This document provides recommended parameter values for NEB Orchestrator based on system type and reaction class. These are starting points, not universal rules. Visual inspection of results is always required.

3 Parameter Descriptions

3.1 IRA kmax (Å⁻¹)

Controls permutation matching strictness in Iterative Rotational Alignment.

  • Low (1.0-1.5): Strict matching, works for well-behaved small molecules

  • Medium (1.5-2.5): Default range, balances strictness with flexibility

  • High (2.5-5.0): Loose matching for difficult atom mapping, may introduce errors

When to adjust:

  • If IRA fails to find alignment: increase kmax

  • If alignment looks chemically wrong: decrease kmax

  • For symmetric molecules: try multiple values (1.5, 2.0, 2.5)

3.2 Number of NEB Images

Total images including endpoints (reactant + product + intermediates).

  • 12-16 images: Simple reactions (single bond break/form)

  • 18-24 images: Moderate complexity (multiple bond changes)

  • 24-30 images: Complex rearrangements (pericyclic, sigmatropic)

Trade-offs:

  • More images → better path resolution, higher computational cost

  • Fewer images → faster, may miss intermediate features

3.3 Force Convergence Threshold (eV/Å)

  • 0.051 eV/Å (10⁻³ Ha/Bohr): Standard, sufficient for most applications

  • 0.026 eV/Å (0.5×10⁻³ Ha/Bohr): High precision, kinetics studies

  • 0.102 eV/Å (2×10⁻³ Ha/Bohr): Quick screening, not for publication

3.4 Cell Size (Å)

Artificial box size for gas-phase calculations. Must be large enough to avoid periodic boundary artifacts.

  • 20-25 Å: Small molecules (<20 atoms)

  • 25-30 Å: Medium molecules (20-50 atoms)

  • 30-40 Å: Large molecules (>50 atoms)

Rule of thumb: At least 10 Å vacuum between molecule and box edge in all directions.

3.5 SIDPP Growth α

Step size for adding new images in Sequential IDPP.

  • Default (0.33): Works for most systems

  • Lower (0.2-0.3): More conservative growth, better for complex reactions

  • Higher (0.4-0.5): Faster path generation, may introduce collisions

When to adjust: If IDPP path has atomic collisions, reduce α.

4 Decision Framework

Start with defaults for your system size
        ↓
Run workflow with default parameters
        ↓
Visualize IDPP path (ase gui results/idpp/*/path/*.xyz)
        ↓
    ┌───┴───┐
    │       │
  Good    Bad
    │       │
    │       └─→ Try different kmax (1.5, 2.0, 2.5)
    │           Re-run alignment + IDPP
    │           Re-inspect
    │
    ↓
Run full NEB optimization
    ↓
Validate saddle (dimer search, DFT single-point)

5 Why Not Bayesian Optimization?

5.1 The Temptation**

It’s natural to ask: “Why not use Optuna or similar tools to automatically find optimal parameters?”

5.2 Why It Doesn’t Work Here**

  1. No automated objective function

    • What metric determines “good” parameters?

    • Final energy? → Doesn’t indicate correct mechanism

    • Convergence speed? → Fast convergence to wrong saddle is worse

    • Barrier height? → No reference value for novel reactions

    • Collision count? → Binary (pass/fail), not continuous for optimization

  2. High computational cost

    • Bayesian optimization needs 20-50 trials to beat grid search

    • Each trial = full NEB calculation (hours to days)

    • Total time: weeks of compute vs. hours for 3-value grid search

  3. Low-dimensional search space

    • Only 3-4 meaningful parameters

    • Most have weak sensitivity (SIDPP α)

    • Grid search with 3 values each is sufficient

  4. Human validation still required

    • No automated metric correlates with “correct chemistry”

    • Visual inspection always needed

    • Automation provides false sense of security

5.3 When Bayesian Optimization Would Help**

  • High-throughput screening (1000+ reactions)

  • Automated quality metric (comparison to known barriers)

  • Cheap objective evaluation (MLIP, not DFT)

  • Many parameters (10+ hyperparameters)

None of these apply to typical NEB workflow usage.

5.4 Current Best Approach**

Simple grid search with visual inspection:

# Try 3 kmax values
for kmax in 1.5 2.0 2.5; do
    snakemake --config alignment.kmax=$kmax \
              --wildcards system=YourSystem
done

# Inspect all 3 paths
ase gui results/idpp/YourSystem_kmax*/path/*.xyz

# Pick best, run full NEB

Time: 3 × NEB runtime + 10 min inspection

6 Examples by Reaction Class

6.1 Proton Transfer (H-transfer)

systems:
  h_transfer:
    reactant: path/to/reactant.con
    product: path/to/product.con
    use_ira: true

alignment:
  kmax: 1.5  # Strict matching for light atoms

neb:
  optimization:
    images: 12
    converged_force: 0.0514221

Notes: H atoms have soft modes. Try kmax = 1.5, 1.8, 2.0 and compare.

6.2 Pericyclic Reactions (Claisen, Cope, Diels-Alder)

alignment:
  kmax: 2.0  # Medium strictness

neb:
  optimization:
    images: 18-24  # More images for complex path

Notes:

  • Backbone permutation critical

  • Visual inspection mandatory

  • Consider manual permutation based on mechanism

6.3 Isomerization (Simple)

alignment:
  kmax: 1.8  # Default works fine

neb:
  optimization:
    images: 12-16

Notes: Usually straightforward, default parameters sufficient.

6.4 Surface Reactions

# Requires modified workflow for PBC
alignment:
  kmax: 2.5
  # cell_size: not applicable (use slab model)

neb:
  optimization:
    images: 18-24
    converged_force: 0.0514221

Notes:

  • Current workflow targets gas-phase only

  • Surface reactions need PBC modifications

  • Larger cell in surface-normal direction

7 Troubleshooting

7.1 Problem: IRA alignment fails**

Symptoms: Error about permutation matching, very high RMSD

Solutions:

  1. Increase kmax (try 2.5, 3.0, 4.0)

  2. Check for missing atoms or different formulas

  3. Manually verify atom ordering in input files

7.2 Problem: IDPP path has collisions**

Symptoms: Atoms unrealistically close (< 0.5 Å)

Solutions:

  1. Reduce SIDPP α (try 0.25, 0.2)

  2. Increase number of images

  3. Check endpoint alignment (may be wrong permutation)

7.3 Problem: NEB doesn’t converge**

Symptoms: Max iterations reached, forces still high

Solutions:

  1. Increase maxiterations(2000, 3000)

  2. Add more images (better path resolution)

  3. Check initial path quality (may be too far from MEP)

  4. Reduce maxmove(0.05 instead of 0.1)

7.4 Problem: Wrong saddle point**

Symptoms: Saddle geometry doesn’t match expected mechanism

Solutions:

  1. Suspect permutation error

  2. Try different kmax values

  3. Manually permute based on chemical intuition

  4. Launch dimer from multiple IDPP images

  5. Compare with literature/reference calculations

8 References

  • IRA algorithm: Gunde et al., J. Chem. Inf. Model. 2021, 61, 5446-5457

  • SIDPP: Schmerwitz et al., J. Chem. Theory Comput. 2024, 20, 1, 155-163

  • NEB: Jonsson et al., 1998; Henkelman et al., J. Chem. Phys. 2000, 113, 9901

  • L-BFGS: Liu & Nocedal, 1989

9 See Also