Nudged elastic band

The Nudged Elastic Band method is a technique for finding transition paths (and corresponding energy barriers) between given initial and final states. The method involves constructing a “chain” of “replicas” or “images” of the system and relaxing them in a certain way.

Relevant literature References:

  1. H. Jonsson, G. Mills, and K. W. Jacobsen, in ‘Classical and Quantum Dynamics in Condensed Phase Systems’, edited by B. J. Berne, G. Cicotti, and D. F. Coker, World Scientific, 1998 [standard formulation]
  2. ‘Improved Tangent Estimate in the NEB method for Finding Minimum Energy Paths and Saddle Points’, G. Henkelman and H. Jonsson, J. Chem. Phys. 113, 9978 (2000) [improved tangent estimates]
  3. ‘A Climbing-Image NEB Method for Finding Saddle Points and Minimum Energy Paths’, G. Henkelman, B. P. Uberuaga and H. Jonsson, J. Chem. Phys. 113, 9901 (2000)
  4. ‘Improved initial guess for minimum energy path calculations.’, S. Smidstrup, A. Pedersen, K. Stokbro and H. Jonsson, J. Chem. Phys. 140, 214106 (2014)

The NEB class

This module defines one class:

class ase.neb.NEB(images, k=0.1, fmax=0.05, climb=False, parallel=False, remove_rotation_and_translation=False, world=None, method='aseneb', dynamic_relaxation=False)[source]

Nudged elastic band.

Paper I:

  1. Henkelman and H. Jonsson, Chem. Phys, 113, 9978 (2000).

Paper II:

G. Henkelman, B. P. Uberuaga, and H. Jonsson, Chem. Phys, 113, 9901 (2000).

Paper III:

E. L. Kolsbjerg, M. N. Groves, and B. Hammer, J. Chem. Phys, submitted (2016)
images: list of Atoms objects
Images defining path from initial to final state.
k: float or list of floats
Spring constant(s) in eV/Ang. One number or one for each spring.
climb: bool
Use a climbing image (default is no climbing image).
parallel: bool
Distribute images over processors.
remove_rotation_and_translation: bool
TRUE actives NEB-TR for removing translation and rotation during NEB. By default applied non-periodic systems
dynamic_relaxation: bool
TRUE calculates the norm of the forces acting on each image in the band. An image is optimized only if its norm is above the convergence criterion. The list fmax_images is updated every force call; if a previously converged image goes out of tolerance (due to spring adjustments between the image and its neighbors), it will be optimized again. This routine can speed up calculations if convergence is non-uniform. Convergence criterion should be the same as that given to the optimizer. Not efficient when parallelizing over images.
method: string of method

Choice betweeen three method:

  • aseneb: standard ase NEB implementation
  • improvedtangent: Paper I NEB implementation
  • eb: Paper III full spring force implementation

Example of use, between initial and final state which have been previously saved in A.traj and B.traj:

from ase import io
from ase.neb import NEB
from ase.optimize import MDMin
# Read initial and final states:
initial ='A.traj')
final ='B.traj')
# Make a band consisting of 5 images:
images = [initial]
images += [initial.copy() for i in range(3)]
images += [final]
neb = NEB(images)
# Interpolate linearly the potisions of the three middle images:
# Set calculators:
for image in images[1:4]:
# Optimize:
optimizer = MDMin(neb, trajectory='A2B.traj')

Be sure to use the copy method (or similar) to create new instances of atoms within the list of images fed to the NEB. Do not use something like [initial for i in range(3)], as it will only create references to the original atoms object.

Notice the use of the interpolate() method to obtain an initial guess for the path from A to B.



Interpolate path linearly from initial to final state.


From a linear interpolation, create an improved path from initial to final state using the IDPP approach [4].


Generate an idpp pathway from a set of images. This differs from above in that an initial guess for the IDPP, other than linear interpolation can be provided.

Only the internal images (not the endpoints) need have calculators attached.

See also

Information about energy minimization (optimization). Note that you cannot use the default optimizer, BFGSLineSearch, with NEBs. (This is the optimizer imported when you import QuasiNewton.) If you would like a quasi-newton optimizer, use BFGS instead.
How to use calculators.



If there are \(M\) images and each image has \(N\) atoms, then the NEB object behaves like one big Atoms object with \(MN\) atoms, so its get_positions() method will return a \(MN \times 3\) array.


The code:

from ase.optimize import BFGS
opt = BFGS(neb, trajectory='A2B.traj')

will write all images to one file. The Trajectory object knows about NEB calculations, so it will write \(M\) images with \(N\) atoms at every iteration and not one big configuration containing \(MN\) atoms.

The result of the latest iteration can now be analysed with this command: ase gui A2B.traj@-5:.

For the example above, you can write the images to individual trajectory files like this:

for i in range(1, 4):
    opt.attach(io.Trajectory('A2B-%d.traj' % i, 'w', images[i]))

The result of the latest iteration can be analysed like this:

$ ase gui A.traj A2B-?.traj B.traj -n -1


Restart the calculation like this:

images ='A2B.traj@-5:')

Climbing image

The “climbing image” variation involves designating a specific image to behave differently to the rest of the chain: it feels no spring forces, and the component of the potential force parallel to the chain is reversed, such that it moves towards the saddle point. This depends on the adjacent images providing a reasonably good approximation of the correct tangent at the location of the climbing image; thus in general the climbing image is not turned on until some iterations have been run without it (generally 20% to 50% of the total number of iterations).

To use the climbing image NEB method, instantiate the NEB object like this:

neb = NEB(images, climb=True)


Quasi-Newton methods, such as BFGS, are not well suited for climbing image NEB calculations. FIRE have been known to give good results, although convergence is slow.

Parallelization over images

Some calculators can parallelize over the images of a NEB calculation. The script will have to be run with an MPI-enabled Python interpreter like GPAW’s gpaw-python. All images exist on all processors, but only some of them have a calculator attached:

from ase.parallel import rank, size
from ase.calculators.emt import EMT
# Number of internal images:
n = len(images) - 2
j = rank * n // size
for i, image in enumerate(images[1:-1]):
    if i == j:

Create the NEB object with NEB(images, parallel=True). For a complete example using GPAW, see here.

Analysis of output

A class exists to help in automating the analysis of NEB jobs. See the Diffusion Tutorial for some examples of its use.

class ase.neb.NEBTools(images)[source]

Class to make many of the common tools for NEB analysis available to the user. Useful for scripting the output of many jobs. Initialize with list of images which make up a single band.

get_barrier(fit=True, raw=False)[source]

Returns the barrier estimate from the NEB, along with the Delta E of the elementary reaction. If fit=True, the barrier is estimated based on the interpolated fit to the images; if fit=False, the barrier is taken as the maximum-energy image without interpolation. Set raw=True to get the raw energy of the transition state instead of the forward barrier.


Returns the parameters for fitting images to band.


Returns fmax, as used by optimizers with NEB.


Plots the NEB band on matplotlib axes object ‘ax’. If ax=None returns a new figure object.