Jastrow optimisation, the default in 2.12

[Philip_Hoggan]

Am applying CASINO to a 750 electron chunk of Cu(111)): a diatomic is adsorbed and the bond stretched. Correlation varies from init geom to stretched geom.

I optimised a Jastrow that is clearly not adequate, but the default option is much better for a small test system (CO). Now, I tried it as is for the whole chunk and there are over 4000 variable parameters. I'd like to take advantage of the approach and rationally select the parameters to optimise. How safe is it to truncate the polynomials sooner? Are there some parameters to exclude, at least initially from the vmc_opt step? Can I try keeping the same Jastrow data for copper and initialising the process from that ?

Another pb could be the copper pseudo, but its given reasonable results for copper surfaces in smaller systems.

[Mike Towler]

Hi Philip,

First of all, can you attach your Jastrow file for us to look at..?

Thanks,
Mike

[Philip_Hoggan]

Have emailed the original Jastrow Mike and would appreciate your advice.

Best,

Philip

MDT Edit: I've attached the file Philip sent me to this post.

[Mike Towler]

Hi Philip,

Your question seems to be based on a false premise; there are something like 45 optimizable parameters - not 4000 - in the correlation.data file that you sent to me. You are grouping the atoms into sets - and so the parameters for the atom-centred terms will be the same for all atoms in a given set. Having 45 parameters is not unreasonable. Are you thinking that different Cu atoms in the surface should actually belong to different sets, depending on their position relative to the adsorbate? If so, I wouldn't bother.

If you're having problems with the optimization, make sure (1) you're using enough configs and (2) you've got the atom labels in the different sets correct (I note that if you're absorbing a single diatomic onto a surface the labels for the adsorbate atoms don't look right to me.. but you've given me no information about the geometry. Is it a hetero- or homonuclear diatomic? How is it oriented?)

You don't tell me how you're trying to do the optimization either, but there are various other recent threads (and various sections of the manual) which give advice about optimization strategies. Note in particular that if you are just optimising a Jastrow factor and you have fixed the cutoff lengths, you can use opt_method=varmin_linjas - the optimised variance minimisation algorithm for the linear parameters in the Jastrow factor which is much faster than the other optimisation methods (I'm sure you know this..).

If you really want to reduce the number of parameters, you might consider not using f terms - at least initially. These are expensive to evaluate (possibly even by far the rate-limiting step in large systems like this). Also you could just fix the cutoff lengths at their defaults.

Keeping the same Cu Jastrow terms you obtained for the bare copper slab would probably be a useful starting point, yes.

Mike

[Philip_Hoggan]

Dear Mike,

Firstly, regarding the geometry, I copy co-ordinates below which I think show the atoms are numbered properly in the Jastrow of the above correlation.data file. The diatomic is H2. The Cell is an 'unfolded' 2 2 1 k-point grid, and I was warned to be careful about numbering in such a case by Dario Alfe, so hopefully there are no mistakes (it is written in cyclic order of co-ordinates, 16 Cu 2H atoms repeated four times)

Now to the remark about 4000 parameters. If you scrap the correlation.data then CASINO 2.12.1 provides a default Jastrow in parameter.casl. I tried using this in a small molecule (CO) and the vmc results were much more negative (than with the correlation.data I had optimised there). I decided to try that for my copper + H2 and THERE got a parameter.casl with 4400 variables.

So, if you check my suspect numbering for the correlation.data I think I'll just stick with that and try to get it optimised better. If the numbering has gone wrong, then I urgently need to learn how to rectify it in this and similar cases.

Atom Atno Type Position (fractional) Position (Cartesian au)
------------------------------------------------------------------------------
1 29 1 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
2 29 1 0.5000 0.2887 0.0000 4.8238 0.0000 0.0000
3 29 1 -0.2500 0.3557 0.0000 -2.4119 4.1776 0.0000
4 29 1 0.2500 0.6443 0.0000 2.4119 4.1776 0.0000
5 29 1 0.2500 0.3110 -0.1073 2.4119 1.3925 -3.8985
6 29 1 0.7500 0.5997 -0.1073 7.2357 1.3925 -3.8985
7 29 1 0.0000 0.6667 -0.1073 0.0000 5.5701 -3.8985
8 29 1 0.5000 0.9553 -0.1073 4.8238 5.5701 -3.8985
9 29 1 0.0000 0.3333 -0.2155 0.0000 2.7850 -7.8291
10 29 1 0.5000 0.6220 -0.2155 4.8238 2.7850 -7.8291
11 29 1 -0.2500 0.6890 -0.2155 -2.4119 6.9626 -7.8291
12 29 1 0.2500 0.9777 -0.2155 2.4119 6.9626 -7.8291
13 29 1 0.0000 0.0000 -0.3239 0.0000 0.0000 -11.7673
14 29 1 0.5000 0.2887 -0.3239 4.8238 0.0000 -11.7673
15 29 1 -0.2500 0.3557 -0.3239 -2.4119 4.1776 -11.7673
16 29 1 0.2500 0.6443 -0.3239 2.4119 4.1776 -11.7673
17 1 2 0.2500 0.2610 0.0605 2.4119 0.9751 2.1996
18 1 2 0.2500 0.0276 0.0605 2.4119 -0.9751 2.1996
19 29 1 -0.5000 0.7113 0.0000 -4.8238 8.3551 0.0000
20 29 1 -0.0000 1.0000 0.0000 -0.0000 8.3551 0.0000
21 29 1 -0.7500 1.0670 0.0000 -7.2357 12.5327 0.0000
22 29 1 -0.2500 1.3557 0.0000 -2.4119 12.5327 0.0000
23 29 1 -0.2500 1.0223 -0.1073 -2.4119 9.7476 -3.8985
24 29 1 0.2500 1.3110 -0.1073 2.4119 9.7476 -3.8985
25 29 1 -0.5000 1.3780 -0.1073 -4.8238 13.9252 -3.8985
26 29 1 -0.0000 1.6667 -0.1073 -0.0000 13.9252 -3.8985
27 29 1 -0.5000 1.0447 -0.2155 -4.8238 11.1402 -7.8291
28 29 1 -0.0000 1.3333 -0.2155 -0.0000 11.1402 -7.8291
29 29 1 -0.7500 1.4003 -0.2155 -7.2357 15.3177 -7.8291
30 29 1 -0.2500 1.6890 -0.2155 -2.4119 15.3177 -7.8291
31 29 1 -0.5000 0.7113 -0.3239 -4.8238 8.3551 -11.7673
32 29 1 -0.0000 1.0000 -0.3239 -0.0000 8.3551 -11.7673
33 29 1 -0.7500 1.0670 -0.3239 -7.2357 12.5327 -11.7673
34 29 1 -0.2500 1.3557 -0.3239 -2.4119 12.5327 -11.7673
35 1 2 -0.2500 0.9724 0.0605 -2.4119 9.3302 2.1996
36 1 2 -0.2500 0.7390 0.0605 -2.4119 7.3800 2.1996
37 29 1 1.0000 0.5774 0.0000 9.6475 0.0000 0.0000
38 29 1 1.5000 0.8660 0.0000 14.4714 0.0000 0.0000
39 29 1 0.7500 0.9330 0.0000 7.2356 4.1776 0.0000
40 29 1 1.2500 1.2217 0.0000 12.0594 4.1776 0.0000
41 29 1 1.2500 0.8884 -0.1073 12.0594 1.3925 -3.8985
42 29 1 1.7500 1.1770 -0.1073 16.8833 1.3925 -3.8985
43 29 1 1.0000 1.2440 -0.1073 9.6475 5.5701 -3.8985
44 29 1 1.5000 1.5327 -0.1073 14.4714 5.5701 -3.8985
45 29 1 1.0000 0.9107 -0.2155 9.6475 2.7850 -7.8291
46 29 1 1.5000 1.1994 -0.2155 14.4714 2.7850 -7.8291
47 29 1 0.7500 1.2663 -0.2155 7.2356 6.9626 -7.8291
48 29 1 1.2500 1.5550 -0.2155 12.0594 6.9626 -7.8291
49 29 1 1.0000 0.5774 -0.3239 9.6475 0.0000 -11.7673
50 29 1 1.5000 0.8660 -0.3239 14.4714 0.0000 -11.7673
51 29 1 0.7500 0.9330 -0.3239 7.2356 4.1776 -11.7673
52 29 1 1.2500 1.2217 -0.3239 12.0594 4.1776 -11.7673
53 1 2 1.2500 0.8384 0.0605 12.0594 0.9751 2.1996
54 1 2 1.2500 0.6050 0.0605 12.0594 -0.9751 2.1996
55 29 1 0.5000 1.2887 0.0000 4.8237 8.3551 0.0000
56 29 1 1.0000 1.5773 0.0000 9.6475 8.3551 0.0000
57 29 1 0.2500 1.6443 0.0000 2.4118 12.5327 0.0000
58 29 1 0.7500 1.9330 0.0000 7.2356 12.5327 0.0000
59 29 1 0.7500 1.5997 -0.1073 7.2356 9.7476 -3.8985
60 29 1 1.2500 1.8884 -0.1073 12.0594 9.7476 -3.8985
61 29 1 0.5000 1.9553 -0.1073 4.8237 13.9252 -3.8985
62 29 1 1.0000 2.2440 -0.1073 9.6475 13.9252 -3.8985
63 29 1 0.5000 1.6220 -0.2155 4.8237 11.1402 -7.8291
64 29 1 1.0000 1.9107 -0.2155 9.6475 11.1402 -7.8291
65 29 1 0.2500 1.9777 -0.2155 2.4118 15.3177 -7.8291
66 29 1 0.7500 2.2663 -0.2155 7.2356 15.3177 -7.8291
67 29 1 0.5000 1.2887 -0.3239 4.8237 8.3551 -11.7673
68 29 1 1.0000 1.5773 -0.3239 9.6475 8.3551 -11.7673
69 29 1 0.2500 1.6443 -0.3239 2.4118 12.5327 -11.7673
70 29 1 0.7500 1.9330 -0.3239 7.2356 12.5327 -11.7673
71 1 2 0.7500 1.5497 0.0605 7.2356 9.3302 2.1996
72 1 2 0.7500 1.3163 0.0605 7.2356 7.3800 2.1996

Crystal does not appear to have inversion symmetry

PS I'll also email the out that is not in a format you can upload.

Best,

Philip

[Philip_Hoggan]

BY the way:

1-H2 is parallel to the surface.

2-the varmin_linjas seems to only take real wavefunctions and mine is complex.

Best,

Philip

[Mike Towler]

Hi Philip,

OK - the atoms are indeed numbered and grouped into sets correctly; you have four horizontal H2 molecules per cell in an asymmetric slab configuration. Fine.

OK - the 'default Jastrow' thing:

First of all, all the documentation and examples and summer schools refer to the 'standard Jastrow' used in CASINO for many years which is stored in the correlation.data file - this is currently considered the default for general use.

The parameters.casl file refers to the new and quasi-undocumented 'general Jastrow factor'. Pablo has been developing this for years and shows no inclination for making it the default for general use (as far as I understand it because it's slower to evaluate in VMC/DMC and doesn't have full functionality). Making it the default should include documenting it in the manual and examples and so on, so it's probably a bit naughty of him to have it appear as a default if you specify use_jastrow = T in input and don't supply a Jastrow of your own - I'll have him punished later. Though of course if you're happy with it and understand what it is then go ahead (I put some disconnected documentation in doc/gjastrow_guide.pdf a while back).

Creating a 'default Jastrow file' in a large system like this using an automatic procedure can be problematic; the process of grouping the parameters of atom-centred functions into sets presumably produces large numbers of different sets and vast numbers of parameters because the H atoms are at different distances from all the Cu atoms and there is bugger all symmetry. It's far more efficient to design your own default with the atom-centred parameters grouped into sensible sets.

You are correct that varmin_linjas doesn't work for complex wave functions yet. It's definitely on the TODO list.. as this is his baby perhaps Neil could comment on the likelihood of this being done in the near future?

Mike

PS: you can attach anything to a forum post providing you gzip it first. The phpbb3 forum software seems to be allergic to text files for security reasons, and there doesn't appear to be an option for turning this off.

[Neil Drummond]

Dear Philip / Mike,

Varmin-linjas for complex wave functions is on my to-do list, although it's not likely to get done for a while.

Note that optimisation for complex wave functions is rarely needed. When twist averaging, the same Jastrow factor is normally used at all twists. Hence one can perform the optimisation at a simulation-cell Bloch vector that permits a real wave function, and then use that Jastrow factor at all twists.

Best wishes,

Neil.

[Philip_Hoggan]

Dear Mike and Neil,

Many thanks for the helpful information.

Now (maybe particularly to Neil) I've been using varmin_linjas today on the real wave-function obtained by defining a single k-point at the gamma point.

I'd hoped you could then use the results in the correlation.data file for the 2 2 1 grid. Is this likely to be the case or do I need to do exactly what Neil describes?

Needless to say, there's a big finite size effect problem with these copper surfaces and I estimate that the 2 2 1 grid has 1/25 the error due to this of the single k-point. Basically, I need to do as big a grid as will fit. The 2 2 1 here gets you a complex wave-function (no inversion).

Best,

Philip

[Neil Drummond]

Dear Philip,

I suggest using a 2x2x1 grid with inversion symmetry, which will allow a real wave function, for the optimisation of the Jastrow factor. You can then use that Jastrow factor for all offsets to your 2x2x1 grid of k vectors.

Best wishes,

Neil.