About the energy window of transmission spectrum calculation

[yongjunwinwin]

The current is obtained by integrating the transmission spectrum at a given bias voltage(e.g. -1 V) and the range of integration is the chemical potential of the right and left electrode. So is it reasonable to define the energy window for transmission spectrum by the bias( e.g. [-0.5, 0.5] and zero energy is aligned to average Fermi level) thus reduce calculation time.

I notice one tutorial "Graphene Junction Device" (page 11)mention that both valence and conduction band edges of the central region should within the energy range of the transmission spectrum to obtain accurate conductance, so I am a little confused that whether the strategy I mention before is improper or not?

[Anders Blom]

Your assumption of the integration window is not entirely correct. Actually you integrate from -infinity to +infinity, but the transmission spectrum is weighted by the composite Fermi distributions of the left and right electrodes (see http://quantumwise.com/documents/manuals/latest/ReferenceManual/index.html/ref.transmissionspectrum.html). So depending on the electron temperature, the active integration range can extend quite far beyond the bias window; normally you should at least have 5kT extra to be on the safe side.

You can see the "bias window coverage" in the report in the log file from the current calculation. The number is an error estimate for the current; if it's below 100% you have a too narrow integration window.

You do have a point however that using a transmission from -5 to 5 eV is overkill for a bias of up to say 2 V. So if the transmission calculation is the primary bottleneck of the calculation, you can reduce time by making the interval smaller. Just look at that report to make sure it's not too small. Often it's also interesting to see what the spectrum looks like, also outside the bias window - it can give you an idea of how large bias you need to have to see NDR, or something like that.

Also note that the transmission spectrum is almost linearly parallelized, up to hundreds of nodes. So you can easily gain a factor of 10 by running the transmission calculation in parallel MPI on 16 nodes or so.

[esp]

what about the number of steps ... would a calculation of -5 to 5 with 100 steps run quicker or slower than one with 50? I would imaging 100 is more accurate, but i am assumming 50 is also faster, which may not be the case .. can you comment on this?

[Anders Blom]

50 is surely faster (scaling is pretty much linear here) but 100 more accurate (unless it's wasted accuracy on a spectrum which is very smooth). The energy points are parallel in MPI, so if you only have 1x1 k-points, this is the primary source of parallel scaling.