dfocc: canonical uhf df/cd ccsd/(t)/(at) E/G #2739

loriab · 2022-10-05T22:53:54Z

Description

Finally, the first of @bozkaya's new methods. These are the canonical (non-orbital-optimized, non-FNO) CC methods with UHF reference by density-fitting and Cholesky decomposition. Gradients available for the most popular methods.

This is PR No. 5 in the mega-dfocc-remp series.

User API & Changelog headlines

new methods!
- uhf df ccd E & G
- uhf df ccsd E & G
- uhf df ccsd(t) E & G (present but experimental (require set qc_module occ) until further optimized)
- uhf df a-ccsd(t) E (present but experimental (require set qc_module occ) until further optimized)
- uhf cd-ccd E
- uhf cd-ccsd E
- uhf cd-ccsd(t) E (present but experimental (require set qc_module occ) until further optimized)
- uhf cd a-ccsd(t) E (present but experimental (require set qc_module occ) until further optimized)

Dev notes & details

pick over canonical methods and run some basic timings tests
TODO: kill new pass
TODO: there might be UHF CCD in there, too.
TODO: merge capabilities maintenance #2731 first. it has a lot of stdsuite changes that I don't want to rebase through. nevertheless, stdsuite on the new methods is working fine locally. the azure errors are that uhf cc methods are running rather than throwing NYI as it expects.
TODO: df ccsd(t) gradients need to be marked experimental and semi-hidden until optimization

REVISED 25 Oct

Module comparison timings [s] for CCSD & CCSD(T) energy & gradient
* Benzene, C6H6
* Either aug-cc-pVDZ (192 nbf) or cc-pVDZ (119 nbf); all-electron
* Same singlet system run as RHF & UHF
* CC converged to 1e-7
* 20 GiB, 8 threads

                              GRADIENT                                 ENERGY
                            symm     c1                             symm     c1
                            ----    ----                            ----    ----
<<<  CCSD/aug-cc-pVDZ  >>>

cfour (vcc) rhf   conv       132    1650                              52    1202 
ccenergy    rhf   conv       266    2037                              92    1099
dfocc       rhf   df         n/a     420                             n/a     230
fnocc       rhf   df                                                 n/a     149

cfour       uhf   conv       225    3473                              90    2413 
ccenergy    uhf   conv       463    5196                             222    2576
dfocc       uhf   df         n/a    1545                             n/a     826

cfour       u/r ratio        1.7     2.1                             1.7     2.0                                        
ccenergy    u/r ratio        1.7     2.5                             2.4     2.3
dfocc       u/r ratio        n/a     3.7                             n/a     3.6

Checklist

Tests added for any new features
All or relevant fraction of full tests run

Status

Ready for review
Ready for merge

JonathonMisiewicz

Very initial comments for today:

JonathonMisiewicz · 2022-11-03T20:25:47Z

doc/sphinxman/source/occ.rst

@@ -374,7 +374,24 @@ Starting in v1.4, MP2.5 and MP3 default to the density-fit algorithm. Set |globa

 Publications resulting from the use of the non-OO CC codes should cite the following publications:

-* **REMP** [Behnle:2019:REMP]_, [Behnle:2022:OREMP]
+* **MP2** [Bozkaya:2011:omp2]_, [Bozkaya:2013:omp2grad]_, and [Bozkaya:2014:dfomp2grad]_


I would much rather we separate out the DF and the non-DF implementations. A paper using conventional OMP2 has no business citing the DF-OMP2 paper. That'll just confuse people following citations.

JonathonMisiewicz · 2022-11-03T20:27:10Z

psi4/driver/procrouting/proc.py

@@ -1043,6 +1055,12 @@ def select_ccsd_at_(name, **kwargs):
 if mtd_type == 'CONV':
 if module in ['', 'MRCC'] and which("dmrcc", return_bool=True):
 func = run_mrcc
+ elif mtd_type == "DF":
+ if module in ["OCC"]: # SOON "",


As in, soon this will be ["OCC", ""]? Why?

This is the "semi-disabling". By not having ["", "OCC"], the naive input set reference uhf; set cc_type df; energy("a-ccsd(t)") won't run and you won't be maddened by worse-than-conventional-CC scaling. However, if you're determined to try the code, it is accessible by adding a set qc_module occ.

psi4/src/psi4/dfocc/ccd_3index_intr.cc

JonathonMisiewicz · 2022-11-03T20:32:32Z

psi4/src/psi4/dfocc/ccd_F_intr.cc

+ SharedTensor2d K, T, U, Tau;
+
+ // OO block
+ // F_mi = (1-\delta_{mi}) f_mi


Why is the definition of the F intermediate changing? What other parts of the code have to change, even in the restricted case, in order for the restricted code to give the same results as before?

I bet it's preparation for frozen virtual and fno. @bozkaya ?

Looking at this again, this term is zero for canonical orbitals. Presumably, this term is in the full equations but was neglected previously.

JonathonMisiewicz

The chunk for today. Includes a new question for the original code authors.

JonathonMisiewicz · 2022-11-04T19:58:24Z

pytest.ini

@@ -47,6 +47,7 @@ markers =
 dfccsd-grad
 dfccsd-t-grad
 dfccsd
+ dfccsdt-grad


This change exposes an inconsistency in pytest.ini:

Just two lines above, dfccsd-t-grad is defined, but dfccsdt is also defined. Let's be consistent about which we pick. I prefer dfccsd-t. For good measure, let's have dfccsdat be synchronized as to hyphen-or-not, as well.

The markers are definitely a mess, mostly because they're a mirror of CTest markers that are also haphazard. A difference is that ctest -L matches partial markers but pytest -m requires full match. If we do try to reform markers, there's also https://github.com/psi4/psi4/blob/master/psi4/driver/p4util/exceptions.py#L490-L496 as standard transformation. The test names in stdsuite use this.

JonathonMisiewicz · 2022-11-04T20:03:25Z

psi4/src/psi4/dfocc/ccd_W_intr.cc

+ W.reset();
+ Tau.reset();
+ T2new->write_anti_symm(psio_, PSIF_DFOCC_AMPS);
+ T2new.reset();


Suggested change

T2new.reset();

Shared pointers are cleared immediately upon leaving scope: this serves no purpose.

psi4/src/psi4/dfocc/ccd_3index_intr.cc

psi4/src/psi4/dfocc/ccd_F_intr.cc

JonathonMisiewicz · 2022-11-04T20:06:47Z

psi4/src/psi4/dfocc/ccd_F_intr.cc

+ // F(m,i) += (1-kronDelta(m,i)) f(m,i)
+ for(int m=0; m<naoccB; ++m) {
+ for(int i=0; i<naoccB; ++i) {
+ int mi = (m * naoccB) + i;


Suggested change

int mi = (m * naoccB) + i;

JonathonMisiewicz · 2022-11-04T20:06:57Z

psi4/src/psi4/dfocc/ccd_F_intr.cc

+ // F(A,E) = (1-kronDelta(A,E)) f(A,E)
+ for(int a=0; a<navirA; ++a) {
+ for(int e=0; e<navirA; ++e) {
+ int ae = (a * navirA) + e;


Suggested change

int ae = (a * navirA) + e;

JonathonMisiewicz · 2022-11-04T20:08:19Z

psi4/src/psi4/dfocc/ccd_F_intr.cc

+ SharedTensor2d Tau, T;
+
+ // F(M,I) OO Alpha Block
+ // F(M,I) = (1-kronDelta(M,I)) f(M,I) + \sum_(E) 0.5*t(I,E)*f(M,E) + \sum_(Q) tQ*b(Q,MI) + \sum_(Q,E) Tau"(Q,IE)*b(Q,ME)


Make this file consistent on delta vs kronDelta, and on whether the overall F is defined before each piece is. This is inconsistent between the restricted and unrestricted parts, at present.

JonathonMisiewicz · 2022-11-04T20:08:33Z

psi4/src/psi4/dfocc/ccd_F_intr.cc

+ // F(a,e) = (1-kronDelta(a,e)) f(a,e)
+ for(int a=0; a<navirB; ++a) {
+ for(int e=0; e<navirB; ++e) {
+ int ae = (a * navirB) + e;


Suggested change

int ae = (a * navirB) + e;

JonathonMisiewicz · 2022-11-04T20:10:54Z

psi4/src/psi4/dfocc/ccd_W_intr.cc

+ SharedTensor2d T2, Tau, T2new;
+
+ // W_MNIJ Alpha Block
+ // W_MNIJ = <MN||IJ> + \sum_(EF) T(IJ,EF) * <MN|EF>


Once again, the new code includes a comment defining the quantity in full, but the older restricted code includes a comment describing each line. (In other words, where is the restricted code's counterpart to l. 165? I can find its counterpart to 166.)

@bozkaya, how much of this is deliberate? Is this a new code style that you're introducing, but not wanting to change all the code right away?

JonathonMisiewicz

The major sticking point here is that I need to know the source for all these references.

JonathonMisiewicz · 2022-11-10T16:02:55Z

psi4/src/psi4/dfocc/ccd_F_intr.cc

+ SharedTensor2d K, T, U, Tau;
+
+ // OO block
+ // F_mi = (1-\delta_{mi}) f_mi


Looking at this again, this term is zero for canonical orbitals. Presumably, this term is in the full equations but was neglected previously.

psi4/src/psi4/dfocc/ccd_W_intr.cc

psi4/src/psi4/dfocc/uccsdl_W_intr.cc

psi4/src/psi4/dfocc/uccsdl_l2_amps.cc

JonathonMisiewicz · 2022-11-10T17:41:22Z

psi4/src/psi4/dfocc/uccsdl_l2_amps.cc

+// DIIS
+ SharedTensor2d RAA, LAA, RBB, LBB, RAB, LAB, LA, LB;
+
+ // RAA


Remove the extra indent.

JonathonMisiewicz · 2022-11-10T17:43:36Z

tests/dfccsd-grad2/input.dat

@@ -0,0 +1,31 @@
+#! DF-CCSD cc-pVDZ gradient for the NH molecule.
+
+ref = psi4.Matrix.from_list([ #TEST


I need sources for all these references.

…) E, cd-ccsd E, cd-ccsd(t) E, cd-a-ccsd(t) E

… update docs tables.

Co-authored-by: Jonathon Misiewicz <[email protected]>

loriab mentioned this pull request Oct 6, 2022

[Feature addition] Add the REMP and OO-REMP hybrid perturbation theories to OCC/DFOCC #2354

Closed

8 tasks

loriab added testing cc For all issues involving the CC module, ground-state energies to response properties. dfocc For issues in the DFOCC module. labels Oct 6, 2022

loriab added this to the Psi4 1.7 milestone Oct 6, 2022

loriab force-pushed the dfocc_dfcc branch from 6f85123 to c77ca62 Compare October 25, 2022 07:15

JonathonMisiewicz reviewed Nov 3, 2022

View reviewed changes

JonathonMisiewicz reviewed Nov 4, 2022

View reviewed changes

JonathonMisiewicz reviewed Nov 10, 2022

View reviewed changes

JonathonMisiewicz approved these changes Nov 14, 2022

View reviewed changes

loriab force-pushed the dfocc_dfcc branch from 4966490 to 2dc0bd2 Compare November 14, 2022 20:22

bozkaya and others added 8 commits December 4, 2022 01:18

dfocc: canonical uhf cc: df-ccsd E & G, df-ccsd(t) E & G, df-a-ccsd(t…

8881fe9

…) E, cd-ccsd E, cd-ccsd(t) E, cd-a-ccsd(t) E

dfocc: canonical uhf cc: de-new-ing pass

d9142b6

dfocc: add uhf ccd df/cc. fix up stdsuite

1e8941a

dfocc: disable occ df/cd ccsd(t)/a-ccsd(t) E/G without qc_module=occ.…

83335f1

… update docs tables.

dfocc: fix CI and add doc refs

761f17a

review comments. tighten conv check

44cfb9e

Apply suggestions from code review

5e8a104

Co-authored-by: Jonathon Misiewicz <[email protected]>

bump qcengine 0.25.0

9af8c87

loriab force-pushed the dfocc_dfcc branch from 2dc0bd2 to 9af8c87 Compare December 4, 2022 09:01

maxscheurer approved these changes Dec 4, 2022

View reviewed changes

loriab merged commit 38cca14 into psi4:master Dec 4, 2022

loriab deleted the dfocc_dfcc branch December 4, 2022 21:58

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

dfocc: canonical uhf df/cd ccsd/(t)/(at) E/G #2739

dfocc: canonical uhf df/cd ccsd/(t)/(at) E/G #2739

loriab commented Oct 5, 2022 •

edited

Loading

JonathonMisiewicz left a comment

JonathonMisiewicz Nov 3, 2022

JonathonMisiewicz Nov 3, 2022

loriab Nov 3, 2022

JonathonMisiewicz Nov 3, 2022

loriab Nov 3, 2022

JonathonMisiewicz Nov 10, 2022

JonathonMisiewicz left a comment

JonathonMisiewicz Nov 4, 2022

loriab Nov 4, 2022

JonathonMisiewicz Nov 4, 2022

JonathonMisiewicz Nov 4, 2022

JonathonMisiewicz Nov 4, 2022

JonathonMisiewicz Nov 4, 2022

JonathonMisiewicz Nov 4, 2022

JonathonMisiewicz Nov 4, 2022

JonathonMisiewicz left a comment

JonathonMisiewicz Nov 10, 2022

JonathonMisiewicz Nov 10, 2022

JonathonMisiewicz Nov 10, 2022

		@@ -0,0 +1,31 @@
		#! DF-CCSD cc-pVDZ gradient for the NH molecule.

		ref = psi4.Matrix.from_list([ #TEST

dfocc: canonical uhf df/cd ccsd/(t)/(at) E/G #2739

dfocc: canonical uhf df/cd ccsd/(t)/(at) E/G #2739

Conversation

loriab commented Oct 5, 2022 • edited Loading

Description

User API & Changelog headlines

Dev notes & details

Checklist

Status

JonathonMisiewicz left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

JonathonMisiewicz left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

JonathonMisiewicz left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

loriab commented Oct 5, 2022 •

edited

Loading