Recent Posts21
General Questions and Answers / Re: Elastic Constants at higher Temperature
« Last post by Anders Blom on November 21, 2024, 23:48 »The concern I have with that is that whichever MD image you pick, it will have a lot of randomness to it. That is, picking 2 or 3 different images (spaced out decently in time, but in the steady-state phase) you might get quite different results. So it is probably necessary to sample several images, which not only gives you an average but also a variation, i.e. error bars.
The idea with the thermal displacement method is that it provides a single structure which is a statistical ensemble for a given temperature and thus eliminates the need for the averaging (that is, you can do just a single DFT calculation for the elastic constants). The "price" you pay is a more expensive calculation of the dynamical matrix, but with a forcefield it's not an issue anyway.
Whether or not these two methods give the same result would of course be interesting to compare! Formally they are supposed to...
Finally, keep in mind that the accuracy of the forcefield should perhaps be checked for the specific material. I doubt ZrC was specifically included when they derived the ReaxFF potenial... (In fact, the original paper is for SiC, and I think all those metal elements are just included as possible impurities.) This might be an example where no forcefield exists and you have to resort to DFT. Or, why not try the MACE or M3GNet neural network potential (included in W-2024.09) - that might actually work really well here! You can also fit a machine-learned forcefield (MTP) using QuantumATK - it should actually be relatively easy for ZrC if you "only" need phonons and not melted or amorphous structures.
Also, when using the displacement method you need to use a larger unit cell (and you also need that for the MD method). For anything involving finite temperature, you cannot operate on small unit cells; a simple way to understand that is that the phonon wavelengths are typically much larger than lattice constants (at least a few nanometers, typically).
The idea with the thermal displacement method is that it provides a single structure which is a statistical ensemble for a given temperature and thus eliminates the need for the averaging (that is, you can do just a single DFT calculation for the elastic constants). The "price" you pay is a more expensive calculation of the dynamical matrix, but with a forcefield it's not an issue anyway.
Whether or not these two methods give the same result would of course be interesting to compare! Formally they are supposed to...
Finally, keep in mind that the accuracy of the forcefield should perhaps be checked for the specific material. I doubt ZrC was specifically included when they derived the ReaxFF potenial... (In fact, the original paper is for SiC, and I think all those metal elements are just included as possible impurities.) This might be an example where no forcefield exists and you have to resort to DFT. Or, why not try the MACE or M3GNet neural network potential (included in W-2024.09) - that might actually work really well here! You can also fit a machine-learned forcefield (MTP) using QuantumATK - it should actually be relatively easy for ZrC if you "only" need phonons and not melted or amorphous structures.
Also, when using the displacement method you need to use a larger unit cell (and you also need that for the MD method). For anything involving finite temperature, you cannot operate on small unit cells; a simple way to understand that is that the phonon wavelengths are typically much larger than lattice constants (at least a few nanometers, typically).