[gpaw-users] bfgs line search

[Jess Wellendorff]

On 05/17/2012 05:09 PM, Jess Wellendorff wrote:
> Hi Sampsa,
>
> yes, the BFGSLineSearch algorithm tries to avoid taking "bad" steps
> during structure relaxation, so it looks around in the potential-energy
> landscape before deciding on the next step to take.
> Therefore, each relaxation step will often be preceded by several
> electronic structure calculations.
>
> NB: Since the BFGSLineSearch algorithm relies on an accurate "picture"
> of the surrounding potential-energy landscape, it is important to employ
> well-converged electronic structures.
> Otherwise the algorithm may get the wrong impression of which step to take.
> This appears to be the case in the example you have provided. The
> algorithm is supposed to only take steps that will lead to a smaller
"smaller" --> more negative
>
> total energy than that of the preceding step. This has clearly failed.
>
> Best
> - Jess Wellendorff
>
> On 05/17/2012 04:38 PM, Sampsa Riikonen wrote:
>> Dear List,
>>
>> This is a weird detail of ASE behaviour..
>>
>> For the optimization of a small molecule, aserun.log looks like this:
>> ---------
>> cat aserun.log
>>
>> BFGSLineSearch:   0  15:32:19       -6.000802       0.0392
>> BFGSLineSearch:   1  15:33:29       -6.000746       0.0266
>> BFGSLineSearch:   2  15:34:32       -6.000747       0.0337
>> BFGSLineSearch:   3  15:35:22       -6.000710       0.0375
>> BFGSLineSearch:   4  15:36:04       -6.000737       0.0188
>>
>> ----------
>>
>> aserun.out has the following:
>> ----------
>> grep -i "converged" aserun.out
>>
>> Converged After 32 Iterations.
>> Converged After 17 Iterations.
>> Converged After 17 Iterations.
>> Converged After 20 Iterations.
>> Converged After 4 Iterations.
>> Converged After 3 Iterations.
>> Converged After 19 Iterations.
>> Converged After 13 Iterations.
>> Converged After 3 Iterations.
>> Converged After 9 Iterations.
>> Converged After 8 Iterations.
>> Converged After 3 Iterations.
>> Converged After 3 Iterations.
>> Converged After 14 Iterations.
>> Converged After 13 Iterations.
>> Converged After 6 Iterations.
>> Converged After 4 Iterations.
>> Converged After 3 Iterations.
>> Converged After 3 Iterations.
>> Converged After 3 Iterations.
>> Converged After 13 Iterations.
>> Converged After 14 Iterations.
>> Converged After 10 Iterations
>> ----------
>>
>> So, gpaw has calculated the electronic structure for quite many
>> geometries, while bfgs line search only
>> reports 5 steps.
>>
>> Does it report only successive "line minizations" and nothing else?
>>
>> Regards,
>>
>> Sampsa
>>
>>
>> _______________________________________________
>> gpaw-users mailing list
>> gpaw-users at listserv.fysik.dtu.dk
>> https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users
> _______________________________________________
> gpaw-users mailing list
> gpaw-users at listserv.fysik.dtu.dk
> https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users