optimize fail at update atomic coordinates

[pchong90]

I am running MPQC with caspt2 from Molcas.

previous iteration:

gradient in internal coordinates
1 0.0000005000
gradient in Cartesian coordinates
1 0.0000000000
2 -0.0000000000
3 -0.0000003536
4 0.0000000000
5 0.0000000000
6 0.0000003536
.......................................
Starting "backtrack" line optimization.
RASSCF energy [au]: -100.083249970000
CASPT2 energy [au]: -0.223376470000
Total CASPT2 energy [au]: -100.306626440000
Sufficient decrease cannot be achieved for the requested value accuracy. Skipping line search. MolcasPT2R12: changing atomic coordinates: Molecular formula: HF molecule: ( symmetry = c1 unit = "angstrom" { n atoms geometry }={ 1 H [ -0.0000000000 0.0000000000 -0.8283798730] 2 F [ 0.0000000000 -0.0000000000 0.0924798730] } ) Atomic Masses: 1.00783 18.99840 Reference origin = [ 0.0000000000 0.0000000000 0.0000000000] Finished "backtrack" line optimization. ....................................... Max Gradient : 0.0000003536 0.0000001000 no Max Displacement : 0.0000002899 0.0000001000 no Gradient*Displace: 0.0000000000 0.0000001000 yes taking step of size 0.000000 Optimization iteration 4 complete //////////////////////////////////////////////////////////////////////

next iteration: gradient in internal coordinates 1 0.0000005000 gradient in Cartesian coordinates 1 0.0000000000 2 -0.0000000000 3 -0.0000003536 4 0.0000000000 5 -0.0000000000 6 0.0000003536 NOTICE: maxabs_gradient increased from 3.5355e-07 to 3.5355e-07 ....................................... Starting "backtrack" line optimization. RASSCF energy [au]: -100.083249980000 CASPT2 energy [au]: -0.223376460000 Total CASPT2 energy [au]: -100.306626440000 MolcasPT2R12: changing atomic coordinates: Molecular formula: HF molecule: ( symmetry = c1 unit = "angstrom" { n atoms geometry }={ 1 H [ 0.0000002902 -0.0000000228 -2048799803.9995157719] 2 F [ 0.0000000020 -0.0000009555 2048799803.4597973824] } ) Atomic Masses: 1.00783 18.99840 Reference origin = [ 0.0000000000 0.0000000000 0.0000000000]

I will take a look on the QNewton algorithm about updating stepsize

[pchong90]

update:

the problem come from the inverse BFGS hessian 1 2
1 1.0000000000
2 -0.0000000000 1.0000000000
3 0.9399626144 2.1093753639
4 0.0000000000 0.0000000000
5 0.0000000000 0.0000000000
6 -0.9399626139 -2.1093753628
3 4
3 4032902427836434.0000000000
4 0.0034431655 1.0000000000
5 -2.9931933212 -0.0000000000
6 -4032902425739279.5000000000 -0.0034431655
5 6
5 1.0000000000
6 2.9931933197 4032902423642127.5000000000

[evaleev]

Weird, it seems it's singular? It should be using a pseudoinverse anyway On Mar 20, 2015 5:55 PM, "Chong Peng" notifications@github.com wrote:

update:

the problem come from the inverse BFGS hessian 1 2

1 1.0000000000

2 -0.0000000000 1.0000000000

3 0.9399626144 2.1093753639

4 0.0000000000 0.0000000000

5 0.0000000000 0.0000000000

6 -0.9399626139 -2.1093753628

                    3                             4

3 4032902427836434.0000000000

4 0.0034431655 1.0000000000

5 -2.9931933212 -0.0000000000

6 -4032902425739279.5000000000 -0.0034431655

                    5                             6

5 1.0000000000

6 2.9931933197 4032902423642127.5000000000

— Reply to this email directly or view it on GitHub https://github.com/ValeevGroup/mpqc/issues/36#issuecomment-84164556.

[pchong90]

The problem come from the gradient. it happens that g(n-1) = g(n) at x(n-1) and x(n). when updating BFGS, g(n-1)-g(n) will just be zero, and it will appear in the denominator.

when x(n-1) and x(n) is very close. Molcas round it to 6 decimal digit. so x(n-1) = x(n) after the round up.

I will look into how to make Molcas more accurate.

But it is still worth to add a check that g(n-1) != g(n). I will do it later.

[pchong90]

Problem is the energy MPQC get from Molcas is not accurate enough. Temporary solution is printing more digits for energy in Molcas.