Molcas Forum

emiranda

Member

Registered: 2024-05-28

Posts: 5

Discrepancies in CMS-PDFT Gradients Compared to CASPT2

Hello everyone,

I am currently using OpenMolcas for non-adiabatic dynamics simulations and need to compute CMS-PDFT energies and analytical gradients at each time step. While the calculations work well with CASPT2  method, I’m encountering unexpected discrepancies with the CMS-PDFT gradients, which are impacting the simulation’s energy conservation.

The CMS-PDFT gradients I obtained differ significantly from those computed with CASPT2 for the same molecular system. These differences lead to substantial fluctuations in the kinetic energy, which disrupt energy conservation in my dynamics simulations.

For instance, using CMS-PDFT, I employed this input:

&GATEWAY
Coord = openmolcas.xyz
Basis = 6-31g*
Group = NoSymm
RICD

&SEWARD
grid input
grid=ultrafine
end of grid input
DoAnalytical

&RASSCF
FileOrb = openmolcas.RasOrb
Spin = 1
Nactel = 6 0 0
Inactive = 18
Ras2 = 6
CiRoot = 2 2 1
CMSI

&MCPDFT
KSDFT = T:PBE
GRAD
MSPDFT

&ALASKA
Root = 2
PNEW

and got the following gradients:

 C    -5.360996598404      2.248281820449     -0.045630534881
 C    -4.777863040592     -1.183342044244      0.042323028164
 C     2.277917405371     -4.549946125983      0.771513050590
 C     4.723364116583     -2.124149470602     -1.039029597186
 C     1.139537104008      2.554480195694      0.225818211927
 C     1.672803128558      2.777365985260      0.233389900519
 H    -0.214232043288      0.090894804122     -0.127834071912
 H    -0.178296822667     -0.280042068749      0.060763776436
 H     0.018136248168     -0.329189002966     -0.066332731831
 H     0.419565696224      0.078578750294      0.002865069232
 H    -0.008260248008      0.414281345096     -0.013869311510
 H     0.288325054048      0.302785811630     -0.043976789548

CASPT2 input was:

&GATEWAY
Coord = openmolcas.xyz
Basis = 6-31g*
Group = NoSymm
RICD

&SEWARD
DoAnalytical

&RASSCF
FileOrb = openmolcas.RasOrb
Spin = 1
Nactel = 6 0 0
Inactive = 18
Ras2 = 6
CiRoot = 2 2 1

&CASPT2
IPEA=0.00
IMAGINARY=0.10
XMultistate=all
GRDT

&ALASKA
Root = 2
PNEW

However, the gradients obtained with CASPT2 were:

C     -0.059769051039      0.080207264599     -0.017108060487
C     -0.133317682759      0.012239121495     -0.002695568138
C      0.125540503687      0.004513165271      0.013931201043
C      0.008135953952     -0.117312999764     -0.030976272416
C     -0.019161593820      0.107947967693      0.021097519899
C     -0.036931659620     -0.156068354750     -0.012719491373
H      0.059692344889     -0.058489057418      0.009062409808
H      0.051667884212      0.023682757985      0.002430704547
H     -0.011426275654      0.028922979525     -0.003807101560
H      0.013030442421      0.015863689528      0.010121752807
H      0.009326957676      0.030193728898      0.006418973477
H     -0.006787823944      0.028299736938      0.004243932392

The large discrepancy between these values is creating substantial variations in kinetic energy, which makes it challenging to propagate the dynamics effectively using the FSSH (Fewest Switches Surface Hopping) method.

Has anyone encountered similar issues with CMS-PDFT gradients in OpenMolcas, or are there any recommended steps to ensure CMS-PDFT gradient accuracy? I’d appreciate any advice or suggestions to troubleshoot or potentially mitigate these discrepancies.

The geometry of the molecule is:

  C     6.0  -10.75299170    4.17802233   -0.10858163   12.00000000
  C     6.0  -11.04331990    1.46528756    0.08540350   12.00000000
  C     6.0   -8.35987364    0.39336198    0.12145113   12.00000000
  C     6.0   -6.64232461    2.29602093   -0.17320844   12.00000000
  C     6.0   -8.17653621    4.72490846    0.03659356   12.00000000
  C     6.0   -7.00765634    6.96607855    0.11959307   12.00000000
  H     1.0  -12.23237530    5.30010800   -0.41131631    1.00782504
  H     1.0  -12.47071740    0.17138599    0.05993853    1.00782504
  H     1.0   -8.22152064   -1.52693829   -0.37003826    1.00782504
  H     1.0   -4.59151904    2.54532677    0.18260388    1.00782504
  H     1.0   -7.92178086    8.87601890    0.45764793    1.00782504
  H     1.0   -5.03753737    7.54128582    0.31918608    1.00782504

Thank you for your help!

emiranda

Member

Registered: 2024-05-28

Posts: 5

Re: Discrepancies in CMS-PDFT Gradients Compared to CASPT2

I’ve encountered additional issues in frequency calculations that might be related to the gradient discrepancies.

I calculated the normal modes with both XMS-CASPT2 and CMS-PDFT. Here are the setups and results I obtained:

1) For XMS-CASPT2

&GATEWAY
Coord=geom.xyz
Basis=6-31g*
Group=NoSymm
RICD

&SEWARD
DoAnalytical
grid input
grid=ultrafine
end of grid input

&RASSCF
FileOrb = openmolcas.RasOrb
Spin = 1
Nactel = 6 0 0
Inactive = 18
Ras2 = 6
CiRoot = 2 2 1
CMSI

&CASPT2
IPEA=0.00
IMAGINARY=0.10
XMultistate=all

&MCKINLEY

 [Molden Format]
 [N_FREQ]
                   30
 [FREQ]
   195.28135698871139
   408.04662361635735
   453.03802795226710
   595.51497044017640
   651.79557556099110
   677.75126497949611
   725.63344503930591
   748.60099572283923
   774.73436589201776
   837.65863374773471
   872.04677119308064
   873.27875880809825
   920.59873188851429
   926.63329627515873
   1021.7969356468493
   1148.6269715336214
   1151.8356604607800
   1194.5405483945271
   1337.8580158721572
   1413.9117615727625
   1505.9915993781947
   1575.6085087673232
   1685.6696784134949
   1732.2828325034745
   3309.5607597710523
   3370.6681864199895
   3380.1970695186465
   3394.3315096870469
   3399.8492823386478
   3410.4489785475362

2_ For PDFT

&GATEWAY
Coord=geom.xyz
Basis=6-31g*
Group=NoSymm
RICD

&SEWARD
DoAnalytical
grid input
grid=ultrafine
end of grid input

&RASSCF
FileOrb = openmolcas.RasOrb
Spin = 1
Nactel = 6 0 0
Inactive = 18
Ras2 = 6
CiRoot = 2 2 1
CMSI

&MCPDFT
KSDFT = T:PBE
GRAD
MSPDFT

&MCKINLEY

 [Molden Format]
 [N_FREQ]
                   30
 [FREQ]
  -5835.6595037510406
  -4691.5573160648037
  -4357.9524494125544
  -3898.2840308437048
  -3629.5727076625853
  -3107.2841063466922
  -2393.0424327055921
  -2269.1323438159129
  -2157.7374368702272
  -1590.8645945798628
  -735.03823132495359
   956.16725982878484
   1077.1896435606845
   1124.4489962781879
   1306.3804313262042
   1726.9542826680479
   2217.5581423985095
   2394.4856032086036
   2762.7960823795042
   2763.8359070527172
   2885.4251278428305
   3204.7646634088833
   3558.2967639586900
   4019.3245171495519
   4382.5260203891812
   4436.7963111380805
   4698.7536172249347
   4715.3297729956530
   4831.7538469359542
   4851.0827857667646

The significant number of negative frequencies in CMS-PDFT suggests issues with the gradients, which aligns with the fluctuations in kinetic energy I've observed during the dynamics simulations. The large discrepancies in gradient values between CMS-PDFT and CASPT2/RASSCF methods might contribute to this problem.

Ignacio

Administrator

From: Uppsala

Registered: 2015-11-03

Posts: 1,085

Re: Discrepancies in CMS-PDFT Gradients Compared to CASPT2

I guess this is relevant:

 ###############################################################################
 ###############################################################################
 ###                                                                         ###
 ###                                                                         ###
 ###    ERROR: MS-PDFT gradients with density fitting not implemented        ###
 ###                                                                         ###
 ###                                                                         ###
 ###############################################################################
 ###############################################################################

matthew-hennefarth

Member

From: University of Chicago

Registered: 2024-11-10

Posts: 5

Re: Discrepancies in CMS-PDFT Gradients Compared to CASPT2

Yes, CMS-PDFT gradients are not implemented with density fitting. Current master branch (after PR 747) will crash if one tries to compute MS-PDFT gradients with density fitting.