how to use preconditioner Kerker for convergence?

[kaypu]

Dear QuantumWise staff

my two-probe system is quasi-one-dimensional armchair-nanoribbon, and it is hard to converge for 0.4V,  the parameter: T=1000k, meshcutoff=75.0*Hartree, k-point=1*1*100, damping factor=0.01 history step=10
*************************************************
........
307 E = -365.631 dE =  1.123687e-03 dH =  6.392827e-03
308 E = -365.632 dE =  6.391485e-04 dH =  2.665628e-03
309 E = -365.632 dE =  5.746915e-06 dH =  2.754449e-03
310 E = -365.632 dE =  1.763446e-04 dH =  2.789024e-03
311 E = -365.632 dE =  2.103748e-04 dH =  1.979505e-03
312 E = -365.632 dE =  5.505047e-05 dH =  2.247296e-03
*************************************************
dE meets the needs of condition of convergence but dH, i don't know whether the parameter of preconditioner Kerker can help me. if this could be, can you tell me how to set the parameter of preconditioner Kerker.

[kstokbro]

If the problem only happens at finite bias it might be that you need more bias contour points.
Check if there is any warnings in the logfile.
Make sure to use DoubleContourIntegralParameters and lower the real_axis_point_density relative to the default

[kaypu]

thank you so much professor kstokbro

there is no warnings in the logfile, i'll try to lower the real_axis_point_density to 0.005eV(double contour integral).

another question: Do lower the real_axis_point_density  change the results? i mean the transmisson spectrum

Regards

[Anders Blom]

You can also try more history steps.

[kaypu]

Thank you professor Anders, i also increase the history step to 15, but for my system it did't work well under bias.

Actually, my system indeed converge  more than 350 steps when the bias higher than 0.4V, it cost lots of time. could you give me some advices to speed up the convergence?

another question: Do lower the real_axis_point_density  change the results? i mean the transmisson spectrum

[kstokbro]

Lowering the real axis point density improves the accuracy of the real axis integral. If the parameter is already low enough it will not change the results to make it smaler.

[kaypu]

thank you professor kstokbro

 i use real_axis_point_density to 0.005eV, it still hard to converge at bias 0.9V(dH is hard to converge), should i increase bias contour points to 50-70?

or i should increase the scattering region, my scattering region is the same length as cell of electrode. should i double the scattering region?

[Anders Blom]

Yes, most definitely. If your central region is only just the combined size of the two electrodes, it's too small.

[kaypu]

thank you professor Anders, i increase the central region (quadruple the electrode), under the low bias it converge well, but when the bias is 1.4V, it is still hard to converge, should i use electrode constraint? maybe the electrode constraint (DM) can lead to converge?

[Anders Blom]

Which ATK version are you using?

[kaypu]

atk 11.8.2-linux
 Is that the convergence of 12.2 better than 11.8?

[kstokbro]

The newer the  better,
12.8 will be out soon, it should be the best,

[Anders Blom]

Try increasing history steps as much as possible (it costs a bit in memory). We have seen cases where 10 works really well, and others where 30-40 is needed.

[kaypu]

i increase the history steps to 30(my Cluster can't sustain more than 35) and mesh cutoff to 200Ry, but it doesn't work. what should i do? Can i use electrode constraint? is it helpful?

[kaypu]

or using LDA instead of GGA, is that LDA converged better than GGA?