Inquiry Regarding High Energy Barrier in CI-NEB Calculation

[esad223]

Dear All,

I am performing CI-NEB calculations to investigate Li diffusion in amorphous Mo₃S₁₃. I inserted a single lithium atom into the system and generated the initial and final structures, followed by using nebmake.pl to generate intermediate images. Both the initial and final structures were fully optimized to the required accuracy before starting the NEB calculation. However, my NEB results indicate an unexpectedly high diffusion barrier, which seems questionable despite the absence of direct experimental data for comparison. Given these results, I suspect there may be an issue with my setup or calculation procedure.

Here are the key energy values from my NEB calculation:

Image|Reaction Coordinate|Energy (eV)|Relative Energy (eV)
0 0.009413 −1482.7993 0.0000
1 0.325824 −1008.2604 474.5389
2 0.372460 −1245.6743 237.1250
3 0.357279 −1394.2229 88.5764
4 0.261281 −1406.7407 76.0586
5 0.008624 −1482.7350 0.0643
Given the large energy difference between images, I am concerned that the computed barrier may not be physically accurate. Although the calculation nearly converged, it eventually stopped and could not proceed further. I have attached a plot for reference.

After further research, I attempted the calculation again using IDPP interpolation to generate intermediate images, as suggested in online sources and forum discussions. While the energy profile appears somewhat different during the run, the barrier remains high. Since this is my first time performing a CI-NEB calculation, I am unsure whether my approach is correct. I have been troubleshooting this issue for about a month and would greatly appreciate any guidance.

I have attached my input files for your reference.

Thank you for your time and assistance.

Best regards,
Emmanuel

[graeme]

You need to have the same settings for your endpoints and your band. I see several differences including the precision, lmixmax, VOSKOWN, ISPIN, NBANDS, ... - I mean it's a little crazy in that it seems like the band parameters are completely different from the endpoints.

[esad223]

Thank you for your response. Just to clarify, should I use the same INCAR parameters from my geometry optimization for the CI-NEB calculation, with only the necessary CI-NEB parameters added? It seems like my band calculation had different settings from the endpoints, so I want to make sure I correct that properly.

[graeme]

yes, that's right. The parameters related to the electronic structure need to be the same.

[esad223]

Dr. Graeme,

I appreciate your guidance on the CI-NEB calculation. I have the calculation running now, though it has not yet converged. Here are the current results:
Step Energy Change Total Energy Force
0 0.009413 -1482.799300 0.000000
1 0.090774 -1482.761300 0.038000
2 0.095373 -1482.675500 0.123800
3 0.260201 -1482.583700 0.215600
4 0.064114 -1482.550500 0.248800
5 0.150448 -1482.558300 0.241000
6 0.401064 -1482.570300 0.229000
7 0.271121 -1482.624700 0.174600
8 0.074422 -1482.706700 0.092600
9 0.008624 -1482.735000 0.064300

I will continue monitoring the convergence and adjust parameters as needed.

Best regards,
Emmanuel

[graeme]

that looks much better

[esad223]

Dear Dr. Graeme,

I hope you are doing well. I wanted to share the results of my CI-NEB calculation and sincerely appreciate your insights on my approach.

For my study, I inserted a single lithium atom into my AIMD-quenched (300K) amorphous Mo₃S₁₃ structure, selecting both the initial Li placement and final positions based on an educated guess. Given the amorphous nature of my system, which contains several voids, I am wondering if this is the best approach for studying Li-ion diffusion.

Would you recommend a more systematic way to determine the initial lithium placement and final positions, considering the lack of experimental data to guide my choices? Your expertise would be invaluable in refining my methodology.

I look forward to your feedback. Thank you for your time and guidance.

Best regards,
Emmanuel

[graeme]

These looks much better and physically reasonable.

You have, however, identified an important challenge. This is a disordered material and so it will take more than a single barrier to quantify diffusion in the material.

For that, one simple (but risky) approach is to run MD at high temperature in order to extract diffusivity values, which can be extrapolated to room temperature. The problem is that you will have to go to a very high temperature to see Li diffusion on a timescale that you can afford to run, and then it is possible for the material structure to change or for Li diffusion mechanisms to change.

An alternative approach that we an others have worked on is called adaptive Kinetic Monte Carlo (aKMC). In that kind of calculation, such as can be done with the EON code, Li barriers can be found using forces and energies from vasp. Each possible hopping mechanisms is added to an event table and one is selected using the standard KMC algorithm. In that way, a state-to-state trajectory is calculated on the time scale of Li hopping, rather than vibrations.

If you want to try a calculation like this, we can help with setting it up. We can also collaborate or work with someone in your group. This does look like a system of scientific interest and one for which aKMC is pretty ideal.

One more point: there are now fairly accurate machine learning (ML) potentials which can accelerate these calculations. EON can also link to ML potentials for diffusion calculations. Resulting diffusion pathways and barriers can be verified or refined with DFT.

[esad223]

Dear Dr. Graeme,

Thank you for your detailed response and insights. I appreciate your suggestions regarding alternative approaches to quantify diffusion in this disordered material.

I understand the challenges of running MD at high temperatures and the potential structural changes that could arise. The adaptive Kinetic Monte Carlo (aKMC) approach sounds very promising, especially with its ability to calculate Li hopping events over longer timescales. I’d definitely be interested in exploring this further.

I spoke with my PI, and he is open to setting up a meeting with our group if you’re interested. We can coordinate a time that works for you, and I would also be happy to collaborate with you on this. Additionally, I have been exploring machine learning potentials for diffusion studies, but I am still determining the best approach. I believe your expertise would be invaluable in refining this aspect of the study.

Looking forward to your thoughts.

Best,
Emmanuel

[graeme]

Sounds great; let's discuss by email.