[CP2K-user] Problem with geometry optimization using LS_SCF

[Fabian Ducry]

Dear Torstein,

On Test 2:
I am no expert on LS_SCF but my understanding ist that PURIFICATION_METHOD 
SIGN should only be used if the chemical potential is know and set using 
MU. This is useful for AIMD where MU is known and does not fluctuate too 
much. Switching PURIFICATION_METHOD to TRS4 and enabling DYNAMIC_THRESHOLD 
will improve convergence if you do not know MU beforehand. I assume that 
the reason for the long walltime/step for LS_SCF is because of poor 
convergence originating in the purification.

On Test 3:
As OT and LS_SCF are numerical methods they will not result in identical 
densies, similar but not the same. So you can expect the forces to be 
slightely different too. If I run your input with tighter settings EPS_SCF 
1.0e-8 and EPS_FILTER 1.0e-8 the max. force at step 126 (starting val 123) 
is already below the convergence threshold:

 --------  Informations at step =   126 ------------
  Optimization Method        =                 BFGS
  Total Energy               =     -3498.3787056930
  Real energy change         =        -0.0000126132
  Predicted change in energy =        -0.0000107671
  Scaling factor             =         0.0000000000
  Step size                  =         0.0038980052
  Trust radius               =         0.4724315332
  Decrease in energy         =                  YES
  Used time                  =               49.665

  Convergence check :
  Max. step size             =         0.0038980052
  Conv. limit for step size  =         0.0030000000
  Convergence in step size   =                   NO
  RMS step size              =         0.0004691139
  Conv. limit for RMS step   =         0.0015000000
  Convergence in RMS step    =                  YES
  Max. gradient              =         0.0004187196
  Conv. limit for gradients  =         0.0004500000
  Conv. in gradients         =                  YES
  RMS gradient               =         0.0000483787
  Conv. limit for RMS grad.  =         0.0003000000
  Conv. in RMS gradients     =                  YES
 ---------------------------------------------------

Best,
Fabian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20200525/f2f28265/attachment.htm>