UT theoretical chemistry code forum

Hi, NEB users,

I am wondering the following questions related to NEB calculations and your comments are welcome:

1. To my understanding, for NEB calculation, once the image POSCARs are created, starting vasp to simultaneously optimize all the images. I am wondering if we start normal vasp relaxation with each of the image POSCAR and compare the energies of the ending and images points. For barriers found in these two ways, is there any difference in the physical meaning and interpretations?

2. The second one is related to the number of images and nodes. Say I have resources of 8 nodes and 2 cpu per nodes computing power, that is 16 cpus. Since 16 must be the number of images created. I can only have image number to be 4(4 cpu per image) , 8(2 cpu) , 16(1cpu) . Since I am afraid that 1 or 2 cpu is enough to finish job within wall time, i start choosing 4(4 cpu per image). Then I am asking myself :
2.1 If I am interesting to know the situation in the middle point between staring and ending points, can I manually insert one image directory ?
2.2 Would missing odd number of images be a problem?
2.3 Also Can I restart a new NEB calculation with two intermediate image structures as new starting and ending points to get denser images ?
2.4 If the job ended abnormally(out of wall time or get killed unexpectedly), is there any way to restart the calculation. In normal vasp calcualtion, I have a tool to monitor and restart with the CONTCAR as the new POSCAR for the continuing calculations. What is the situation in NEB? Will the method of copying the CONTCAT to POSCAR in each image directory and resubmit the job be valid here too?
2.5 I have some experiences with TS search in Dmol from Material studio. The manual suggests that we need to verify a transition state: you must perform a vibrational analysis. A true transition state will have one imaginary vibrational frequency whose normal mode corresponds to the reaction coordinate; all other eigenvalues will be real. How to verify the TS after the NEB calculation ?

I asked a lot and thanks for your time. It is snowing outside and i am nebing inside :)

1. If you start a normal relaxation at each of the images of an neb the result will be that all your images will minimize to either the initial state or the final state. Thus you can not calculate the TS this way.

2. In my experience 16 images far exceeds the number of images needed for a typical path. I would regularly use 3-5 images to resolve a single transition state.

I good rule of thumb for neb runs is 1 node per image. This way you take advantage of the fast intanode connection while the neb inherently need little commutation between images. So I would run say a 5 image neb on 5 nodes (10cpus). Or if your system is large run a 3 image neb on 6 nodes (12 cpus)
If you are interested in finding a high resolution minimum energy path. I would suggest that you use a course climbing image neb to find the saddle. Once the saddle is found you can just follow the steepest decent path in both directions to map out the MEP.
We have a dynamic matrix script that you can use to determine the normal modes at both the TS and the minimum we typically use it to find the harmonic pre-factor to the Arrehnius rate
http://theory.cm.utexas.edu/vtsttools/scripts/#DYM

Here are a few additional comments:

2.1. You can run an NEB with a single image. For simple pathways, the climbing image technique will move this image to the saddle point. For more complex pathways additional images are needed to resolve the path. You can set up the initial configurations any way that you want. Just be aware that the images should be fairly equally spaced. If they are not, use a conservative optimizer which can deal with high forces such as IBRION=3 and a small POTIM of <0.1.

2.3. This can work, people do use this approach, but I don't recommend it. The danger is that the intermediate images are not along the true reaction pathway. If they were, and did define the reaction path, then a climbing image would converge to the saddle. Putting it another way, as long as the intermediate images define the reaction coordinate so that the climbing image can converge to the saddle, there is no need for a higher density of images. Using a min-mode following method (e.g. the dimer in vasp) is a better single-image approach than a 1-image NEB between fixed intermediate images.

2.4. Restarting an NEB calculation works in the same way as a regular calculation. If you use our vtstscripts, the 'vfin.pl dir' script will compress the calculation into the directory dir and leave the original directory ready for continuation by replacing the POSCAR files with the CONTCAR files.

2.5. If the climbing image converges, you can be fairly sure (but not 100%) that you have a first-order saddle point. Then, as Dan mentioned, you can then use finite-difference to calculate the normal modes and verify that there is one negative mode.