[CP2K-user] [CP2K:11324] Weird "convergence" of L-BFGS CELL_OPT

[Thomas Kühne]

Dear Daniele, 

in the LBFGS method convergence used to be decided upon the projected 
gradient and the relative force error only, the convergence checks are just 
printed for information. You can use the BFGS method instead, which converges 
only if all checks are met. However, the implementation details are slightly 
different, which is why typically more BFGS steps are required. 

Best, 
Thomas

> Am 22.02.2019 um 14:39 schrieb Daniele Ongari <daniele... at gmail.com>:
> 
> Dear CP2K mailing list,
> 
> We experience a strange behavior of the L-BFGS optimization algorithm when dealing with cases where SCF convergence is problematic.
> 
> After 20 minimization steps the cp2k printed the following:
> 
> 
> --------  Informations at step =    21 ------------
>  Optimization Method        =                LBFGS
>  Total Energy               =     -2345.6430825804
>  Internal Pressure [bar]    =        23.9760390703
>  Real energy change         =         0.0028283710
>  Decrease in energy         =                   NO
>  Used time                  =             5864.105
> 
>  Convergence check :
>  Max. step size             =         0.0000000000
>  Conv. limit for step size  =         0.0030000000
>  Convergence in step size   =                  YES
>  RMS step size              =         0.0000000000
>  Conv. limit for RMS step   =         0.0015000000
>  Convergence in RMS step    =                  YES
>  Max. gradient              =         0.0082916384
>  Conv. limit for gradients  =         0.0004500000
>  Conv. for gradients        =                   NO
>  RMS gradient               =         0.0009846623
>  Conv. limit for RMS grad.  =         0.0003000000
>  Conv. for gradients        =                   NO
>  Pressure Deviation [bar]   =       -76.0239609297
>  Pressure Tolerance [bar]   =       100.0000000000
>  Conv. for  PRESSURE        =                  YES
> ---------------------------------------------------
> 
> --------------------------
> OPTIMIZATION STEP:     22
> --------------------------
> 
> ***********************************************
> * Specific L-BFGS convergence criteria
> * WANTED_PROJ_GRADIENT and WANTED_REL_F_ERROR
> * satisfied .... run CONVERGED!
> ***********************************************
> 
> 
> 
> The two things that look strange for us are:
> 
> 1) Max. step size and RMS step size were 'artificially' set to 0.0000000000
> 2) Not all checks are showing convergence, while L-BFGS algorithm has decided to stop.
> 
> 
> Therefore, we would like to know if this is an expected behavior of the L-BFGS algorithm and if yes, what is the reason behind working this way?
> 
> Please see input and output files attached.
> 
> With best regards,
> Daniele and Sasha
> 
> -- 
> You received this message because you are subscribed to the Google Groups "cp2k" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns... at googlegroups.com <mailto:cp2k+uns... at googlegroups.com>.
> To post to this group, send email to cp... at googlegroups.com <mailto:cp... at googlegroups.com>.
> Visit this group at https://groups.google.com/group/cp2k <https://groups.google.com/group/cp2k>.
> For more options, visit https://groups.google.com/d/optout <https://groups.google.com/d/optout>.
> <aiida.coords.xyz><aiida.inp><aiida.out><aiida-pos-1.xyz>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20190222/2e91e339/attachment.htm>