Stack Exchange
log in

Pro

 The mathematics and science behind Ma…

mattermodeling.stackexchange.com
Pro
Dec 26, 2024 19:44
What you have said is right. But there are right answers for wrong(?) reasons from dft. Don't you think?
Pro
Dec 24, 2024 13:40
@user1271772 FYI, There have been attempts to use CCSD for FF developments. About the hydrophobic pockets, No, they were also not done from CCSD. But from MBD framework on top of DFT. I have suggested that reference since classical FFs do fail in such cases and one need QM corrections there.
Pro
Dec 24, 2024 12:25
@user1271772 thanks, I have joined the discord server there. Well, i couldn't understand your first question? Now for machine learning, in the same paper of Dral, they have used delta learning to learn the CCSD energies from dft ( afaik). They were pretty much accurate. But even to generate the ML training data one needs huge CCSD calculation for example
Pro
Dec 24, 2024 12:23
@user1271772
Pro
Dec 24, 2024 12:15
@user1271772 I have added some equations that may be correct in the edited question.
Pro
Dec 24, 2024 12:10
@user1271772 well, people use that to machine learn the CCSD energies. See recent (2015 or 2019 year) papers by Dral et al in JCTC. Also, see Papers from Bowman group.
Pro
Dec 24, 2024 12:09
@user1271772 No, I have never used MRCC before. Can you suggest a suitable working manual? Are these manuals available widely?
Pro
Dec 24, 2024 12:08
@user1271772 In MP2, only the 2-e terms are correlated, unlike CCSD/CCSD(T) where even 1-e terms are also correlated. This gives CC methods an edge over MP2.
Pro
Dec 24, 2024 12:07
@user1271772 Well, that paper was long ago, I am working in the same group and have successfully used CCSD, CCSD(T) for smaller systems.
Pro
Dec 24, 2024 12:07
@user1271772 That is also true. Okay, I will try use MRCC for now and see if there is any progress with it.
Pro
Dec 24, 2024 12:00
@user1271772 So, if CCSD(T) DMs are not available, then I can't go further with it.
Pro
Dec 24, 2024 12:00
@user1271772 As Prof. Kallay wrote, this should be. But the target is CCSD(T). So I was looking at CFOUR program as well.
I want a standard program for most of the work. Which save time for normalization of DMs, setting up desired accuracies and benchmarking for a suite of jobs that we use.
Pro
Dec 24, 2024 11:57
@user1271772 Oh,, yes. I am looking at every corner to find the answer.
Pro
Dec 24, 2024 11:56
@user1271772
I am currently using CCSD, but plan to use CCSD(T).
Pro
Dec 24, 2024 11:56
@user1271772
I have never used casting. So, I will do a bit of search and try to implement that.
Pro
Dec 24, 2024 11:55
@user1271772 Sorry for that.
A lot to learn. I am not used to this chat box, but it is very handy.
Thank you for introducing this to me very nicely.
I will keep things in mind for a better chat experience.
Pro
Dec 24, 2024 11:53
@user1271772 Whenever you have some time. :D :D.
Pro
Dec 24, 2024 11:52
@user1271772 I agree with you. You can read that paper, it is interesting. There are other works as well, if you are interested.
That is the problem I am trying to tackle. Getting more accuracy by discarding any parametrization in conventional FFs, say discarding LJ parameters with pure correlation energy.
See, there was a paper by Tkatchenko et al in Science (2019) where the correlation interactions were prevalent even at 10A distance. Which was not picked up by conventional FFs, but with MBD method. That was responsible for a hydrophobic pocket. I think one can modify the conventional FFs…
Pro
Dec 24, 2024 11:44
@user1271772 Thank you very much for the tricks. Yes, indeed we can use sparse matrix formulation to discard values say below 10^(-7) or 10^(-10) by testing the accuracy. But even for that, I need to generate the 2RDM from any of the codes. That is becoming painful. Or even, calculate 2RDM block by blocks so that I can only pick a block to work with.
- No, I don't know precisely that how many of the matrix elements will be zero.
Pro
Dec 24, 2024 11:35
@user1271772 Yes, 600-800
Pro
Dec 24, 2024 11:35
@us
Pro
Dec 24, 2024 11:33
@user1271772 Thank you very much for your response. I am looking into 6-800 basis functions (aug-cc-pVTZ basis). So the 2RDM will be huge, ofcourse, N_basis ^4. That is the target.
- If you look at this paper, https://doi.org/10.1007/s11224-020-01495-y Which briefly says it all. It is for calculation of true AIM charge, Multipole moments leading formally for analysis for forces between atoms. Which will ultimately lead to Force field development.
Pro
Dec 24, 2024 10:58
@user1271772
-I think PySCF requires more memory than CFOUR or G09. I can quickly put up a test and check about that. I will let you know about the results.
- About the ccsd_2rdm function, I think that rewriting that part in Fortran will effectively reduce the memory usage. Also, Instead of calculating it fully, one may calculate it block-wise, say OOOO part, VVVV part etc and combine them later altogether. PySCF does that as well, but I have checked that calculation of full 2RDM takes a huge amount of memory.
Pro
Dec 24, 2024 10:51
@user1271772 Sorry about the pinging, I didn't know that but definitely I will use the same from now on.
Pro
Dec 22, 2024 15:17
- Yes you are true there. I just don't want to complicate the stuff there. Essentially, from the 2RDM, all the properties can also be calculated, including 2-e correlations which are absent in 1PDMs.
- I can't understand what you are saying about "not very rigorous". Can you please explain?
Thanks
Pro
Dec 22, 2024 15:14
- I have compared the memory usages of Gaussian, PySCF and CFOUR. PySCF can not solve the CCSD lambda equations which CFOUR and Gaussian can at the certain memory threshold. The ccsd_rdm.py function in PySCF also takes a large memory requirement. This is why I intend to find the amplitudes and calculate the 2PDM myself, also, exploit the symmetry there. Needless to say, I have to use fortran and OMP there.
Pro
Dec 22, 2024 15:10
I agree with you,
 

 Python based Matter Modeling software

phonopy.github.io/phonopy, pypi.org/project/pysmiles, github.c...
Pro
Jan 2, 2024 12:10
@user1271772 Thanks. Happy new year to you too..
Pro
Dec 27, 2023 11:48
@user1271772 I have tried with other methods, but with NAO, IBO I have had some problems with the implementation. But with ER method, I find that the CCSD energy is -1.255729963139254 which is correct to canonical, PM and FB method up to five decimal digits. So, it is okay now. Actually, the next phase of my research is targeted towards sparseness of 1PDM and 2PDM (all in AO basis). We are looking at the sparseness definition of H-Head_Gordon in this paper JPCA, 1998, 102, 12, 2215-2222
Pro
Dec 11, 2023 10:18
Also, I will do other localization and CCSD thereafter as you suggested. I think it will be a nice discussion on stack-exchange about why the correlation energies are increased than the canonical ones in some localization methods. What do you think?
Pro
Dec 11, 2023 10:18
@user1271772 Hi @Nike. Thank you very much for the detailed explanation and the answer. I think I have read about cons of Pipek-Mezey localization in LED-DLPNO CCSD(t) method. Due to that, Foster-Boys algorithm is preferred. However, I will let you know if I quickly find that paper.
Pro
Dec 9, 2023 20:14
From these outputs, it seems,
1. One can use localized MOs as post-HF calculation
2. Of course the correlation energy recovered is different in each localization method.
Pro
Dec 9, 2023 20:12
Hi Nike. Thank you again for your help. My quest was to find if CCSD followed by localization was possible in PYSCF or not. I asked it in this platform since experts like you can help.
As a test of a novice, I tried Boys localization. The code may have failed due to wrong inputs from my side or wrong arguments to the CCSD calculation.
After your advice, I ran the code again in a fresh notebook. Now it seems it runs fine. Let me share the input and output.

import numpy
from pyscf import gto,scf,cc,lo
Pro
Dec 8, 2023 23:52
1. If passing LMO's as input guess for the post-HF methods is allowed in PySCF
2. If the above is possible, how to do that in right way?
Pro
Dec 8, 2023 23:51
I have two doubts,
Pro
Dec 8, 2023 23:48
Python 3.10.11 | packaged by conda-forge | (main, May 10 2023, 18:58:44) [GCC 11.3.0]
Type 'copyright', 'credits' or 'license' for more information
IPython 8.15.0 -- An enhanced Interactive Python. Type '?' for help.

...: atom = [
...: [ 'H' , 0.000000 , 0.000000 , 0.100000],
...: [ 'H1' , 0.000000 , 0.000000 , 0.860000],
...: [ 'H' , 2.000000 , 0.000000 , 0.100000],
...: [ 'H1' , 2.000000 , 0.000000 , 0.860000]],
Pro
Dec 8, 2023 23:48
This is the first time I am using any chatroom. So please bear with me. Thank you very much Nike for your precious time. I think the confusion about the MP2 energies and NaNs of CCSD energies were produced by different tools. Again I tried the same code that I have shown in the question using ipython3 notebook. Which shows different answers. Let me share this.
Pro
Dec 8, 2023 23:38
:64831085