Looking at your MC, at least with the CE you trained, you have reached the lowest energy structure for the chosen temperature.
It would be fair to conclude, according to the trained CE, that the predicted structure is the lowest-energy configuration.
However, since you have DFT at hand, it’s always expected to check your claim.
1 Like
Yes I think so and I agree with @bsreeharsha , and note if you’ve done enumeration in this supercell then you have generated all unique occupations in this supercell and hence the lowest energy structure among those have to be the ground-state (for this cluster-expansion and this supercell).
But as seen above, there could still be larger supercells that allow for lower energy structures at the same composition.
Hello Erik,
Thanks for the reply. I really appreciate your help. However, I have another question about the effect of entropy in MC runs.
The CE model is trained with all the energies calculated at 0 K, but when we do the MC annelling to verify the ground state structures (e.g., from 2000 K to 300 K), we need to take configurational entropy into account. I wonder how mchammer address the entropy here.
When searching for the information online, I found that it indicates
MC does not directly output a number like “E−TS”. It samples arrangements so that ensemble averages you compute from the trajectory correspond to finite-T thermodynamics. Entropy shows up through how many arrangements are visited and with what frequency, not as an extra term added to each energy.
I don’t really understand it. Can you explain how the entropy effect is addressed here and how it does to ensure the accuracy of the result in annealing when considering configurational entropy?
Thank you so much.
The CE is trained on energies, the MC sampling (if converged) will sample structures according to the correct thermodynamic probability distribution.
So yes, you dont get a number for the entropy S as output from a simple MC simulation, but it is “included” correctly in the sampling.
This question you’re asking is very general and not really related to icet, or mchammer.
I suggest do some reading or watching some lectures on statistical mechanics and how it connects to thermodynamics, and maybe specifically look up Monte Carlo simulations for configurational sampling.
1 Like