Converge problem for Cu-CNT two probe device

[perfetti]

Hello everyone.

I am now running a device's transmission calculation with about 150 Cu atoms in each electrode, and carbon nanotube in the central region. There are about 450 atoms in total.

The calculation is expensive,and I am now facing two phenomena:
1) It doesn't give the transmission calculation after the Device DFT calculation in the log file. But in my another test it gives the transmission spectrum calculation after the Device DFT calculation finished.
2) The procedure seems like: Left Electrode-->Right Electrode-->Equivalent Bulk. Now it goes to the second round, and in the Equivalent Bulk Part,which has 450 atoms, it seems to be dangling ,and refuse to converge. Or converge very slowly.
    The data is as following:

  0 E = -724.807 dE =  1.000000e+00 dH =  5.504534e+01    
  1 E = -792.305 dE =  6.749808e+01 dH =  2.657772e+01                        
  2 E = -807.447 dE =  1.514179e+01 dH =  1.475684e+01                        
  3 E = -817.859 dE =  1.041207e+01 dH =  1.180922e+01                      
  4 E = -845.698 dE =  2.783917e+01 dH =  1.294242e+01                      
  5 E = -842.232 dE =  3.466064e+00 dH =  1.257266e+01                      
  6 E = -873.413 dE =  3.118168e+01 dH =  1.189997e+01                      
  7 E = -877.852 dE =  4.438391e+00 dH =  1.140631e+01                      
  8 E = -875.366 dE =  2.485674e+00 dH =  9.900852e+00                      
  9 E = -883.274 dE =  7.908225e+00 dH =  1.054715e+01                      
10 E = -884.964 dE =  1.689212e+00 dH =  1.036553e+01                      
11 E = -904.116 dE =  1.915216e+01 dH =  1.003756e+01                      
13 E = -910.438 dE =  4.249501e+00 dH =  8.955506e+00    
14 E = -915.376 dE =  4.937311e+00 dH =  8.254702e+00                      
15 E = -918.475 dE =  3.099397e+00 dH =  8.678539e+00                      
16 E = -922.951 dE =  4.476517e+00 dH =  8.288408e+00  
17 E = -944.823 dE =  2.187161e+01 dH =  8.209600e+00  
18 E = -927.055 dE =  1.776837e+01 dH =  7.335138e+00  
19 E = -943.449 dE =  1.639448e+01 dH =  7.367521e+00
20 E = -940.267 dE =  3.182627e+00 dH =  7.047239e+00
21 E =  -947.79  dE =  7.523456e+00 dH =  6.967868e+00
22 E = -957.677 dE =  9.886761e+00 dH =  6.402009e+00                        
23 E = -970.576 dE =  1.289936e+01 dH =  6.602399e+00  
24 E = -958.176 dE =  1.240045e+01 dH =  5.074202e+00  
25 E = -970.406 dE =  1.223067e+01 dH =  6.218405e+00  

I have three questions:

1) It that any chance that i didn't include the transmission spectrum module in the script? If so,could I make up later? The calculation is still running and I don't want to kill it.

2)I changed the history step to 3,cause i saw a post who said the lower history step helps the converge for Cu. It did help at the first, but now it seems to face the same problem. I am really a little worrying cause it already ran for 2 days......

3)From which place do we decide it's converging? The dE or dH? Or the total energy?


   Can anybody gave any advice? Especially about the second problem......Thanks a lot.

[Anders Blom]

Need to see script details to advise.

[perfetti]

Dear Dr. Blom,
       I've sent you the script file via email. Thank you.

[perfetti]

Dear everyone,
       I have a problem: I am calculating the transmission spectrum of Cu/carbon nanotube system. However, The system doesn't converge even after 400 iterations. I have set the electron temperature to 1500K, and density mesh cut-off as 40 Hartree,  and according to the advice from this forum, I set the history steps to 4 to make it converge faster. The LCAO basis set is double zeta polarized for both C and Cu .
      I am not sure where's the problem. I've already tried all the methods I found on this board, but I still cannot get a convergence. The script is attached and thank anyone who could give me a hand.
      

[Anders Blom]

I'm afraid you have moved the mesh cut-off the wrong way, it should be higher than default, not lower. You can try 150 H or even 200.

However, the primary trick to improving convergence is to increase the number of history steps. Try 15 or 20. Also, if you do reach some level of convergence like 1e-4, that's actually plenty enough, it can be hard to reach all the way to 4e-5, but by our (recent) experience all physical quantities are well converged at 1e-4 (possibly with the exception of complicated relaxation cases).

[perfetti]

As an update, I changed the mesh-cut off to 150H, and the history steps to 15, and also changed the tolerance to 1e-4,however, the calculation for the bulk part still dangles and refuse to converge like this:
0 E = -649.231 dE =  1.000000e+00 dH =  3.061515e+02
 1 E =  -462.26 dE =  1.869710e+02 dH =  1.541593e+01
2 E =  -565.39 dE =  1.031294e+02 dH =  2.096157e+02
3 E = -763.144 dE =  1.977541e+02 dH =  2.376388e+02    
4 E = -468.965 dE =  2.941786e+02 dH =  5.509043e+01      
5 E = -767.877 dE =  2.989117e+02 dH =  5.360779e+02
 6 E = -462.653 dE =  3.052235e+02 dH =  4.909141e+01      
 7 E = -452.875 dE =  9.778729e+00 dH =  2.286163e+01
8 E = -461.033 dE =  8.157908e+00 dH =  2.771539e+01
 9 E = -468.244 dE =  7.211182e+00 dH =  4.494197e+01                  
 10 E = -454.093 dE =  1.415036e+01 dH =  1.936802e+01  
 11 E = -452.705 dE =  1.388138e+00 dH =  1.389084e+01  
 12 E = -456.179 dE =  3.473517e+00 dH =  1.353404e+01
 13 E = -457.176 dE =  9.966808e-01 dH =  1.282225e+01  
 14 E = -459.926 dE =  2.750394e+00 dH =  1.584601e+01  
15 E = -474.547 dE =  1.462107e+01 dH =  2.645947e+01  
 16 E = -552.829 dE =  7.828156e+01 dH =  1.898858e+02  
17 E =  -574.61 dE =  2.178147e+01 dH =  2.323332e+02  
 18 E = -576.849 dE =  2.238520e+00 dH =  2.338102e+02    
19 E = -590.474 dE =  1.362579e+01 dH =  2.599107e+02
20 E = -587.911 dE =  2.563215e+00 dH =  2.581052e+02  
21 E = -571.423 dE =  1.648838e+01 dH =  2.287264e+02    
  22 E =  -457.57 dE =  1.138525e+02 dH =  1.391865e+01  
23 E = -452.489 dE =  5.081609e+00 dH =  1.332160e+01                                      
 24 E = -451.717 dE =  7.718476e-01 dH =  1.351076e+01          
  25 E = -574.382 dE =  1.226650e+02 dH =  2.151903e+02
  25 E = -574.382 dE =  1.226650e+02 dH =  2.151903e+02
27 E = -454.387 dE =  1.374729e+02 dH =  1.729166e+01    
28 E = -461.523 dE =  7.135213e+00 dH =  3.135542e+01    
29 E = -460.242 dE =  1.280163e+00 dH =  1.827711e+01
 30 E = -583.144 dE =  1.229017e+02 dH =  2.450235e+02
31 E = -462.571 dE =  1.205735e+02 dH =  3.156188e+01  
32 E = -455.328 dE =  7.242995e+00 dH =  1.917901e+01    
33 E = -453.494 dE =  1.833932e+00 dH =  1.673408e+01    
  34 E = -568.153 dE =  1.146591e+02 dH =  2.064697e+02    
 35 E = -566.517 dE =  1.635918e+00 dH =  1.973332e+02    
36 E = -680.774 dE =  1.142575e+02 dH =  2.127642e+02
37 E = -665.206 dE =  1.556826e+01 dH =  2.106167e+02    
......

Now it has gone on for 165 iterations, but it still gives a result of
165 E = -491.197 dE =  1.210922e+01 dH =  9.094218e+01 .

I've attached my script , and thanks for any advice beforehand.    

I'm afraid you have moved the mesh cut-off the wrong way, it should be higher than default, not lower. You can try 150 H or even 200.

However, the primary trick to improving convergence is to increase the number of history steps. Try 15 or 20. Also, if you do reach some level of convergence like 1e-4, that's actually plenty enough, it can be hard to reach all the way to 4e-5, but by our (recent) experience all physical quantities are well converged at 1e-4 (possibly with the exception of complicated relaxation cases).

[Anders Blom]

Your system seems reasonably set up, except for the device Poisson solver. Using FFT is not the right choice, you should always use FFT2D (or Multigrid, if there are gates) for transport calculations.

But you say it's the bulk part which doesn't converge...?

Ah! I think I see the problem. For some reason your central region is not periodic - this can cause all kinds of problems!

I suggest you change the device algorithm method from EquivalentBulk to NeutralAtoms (and fix the Poisson solver above, and the temperature), and run again. Skipping the EquivalentBulk can mean convergence is slower for the device part, but on the other hand you save time by not having to go through the EquivalentBulk, so often this evens out. Even better - the device part parallelizes well, which the EquivalentBulk does not (esp. in your case with 1x1 k-points).

Alternatively you can try to make the central region periodic, but if the above works then all is fine anyway.

[perfetti]

Dr. Blom,
        I am sorry I made a mistake here:It's not the equivalent bulk here not converge, but the Density Matrix Report after the equivalent calculation. All the left electrode, the right electrode ,and the equivalent bulk have converged.
        Do you have any other advice? And what do you mean by mentioning the temperature?
        On the other hand, how to decide whether the central region is periodic or not?
        Thanks a lot.

                    

Your system seems reasonably set up, except for the device Poisson solver. Using FFT is not the right choice, you should always use FFT2D (or Multigrid, if there are gates) for transport calculations.

But you say it's the bulk part which doesn't converge...?

Ah! I think I see the problem. For some reason your central region is not periodic - this can cause all kinds of problems!

I suggest you change the device algorithm method from EquivalentBulk to NeutralAtoms (and fix the Poisson solver above, and the temperature), and run again. Skipping the EquivalentBulk can mean convergence is slower for the device part, but on the other hand you save time by not having to go through the EquivalentBulk, so often this evens out. Even better - the device part parallelizes well, which the EquivalentBulk does not (esp. in your case with 1x1 k-points).

Alternatively you can try to make the central region periodic, but if the above works then all is fine anyway.

[Anders Blom]

The Poisson method is still a critical issue. It's always better to run with as low temperature as possible for good accuracy, higher temperature is only need when convergence is poor, but for this system I'm rather sure the problem can be solved without changing temperature.