D3 scaling.

[rebeltroopers]

Hello,

I have been working with GPUMD to create a reactive potential for polyethylene. I have shown that NEP is able to accurately capture PE chemistry with the pairwise cutoff. One problem I am encountering is that the van der Waals forces are handled through the implementation of D3. However, for PE, D3 overbinds, leading to "stiff" polymers.

I was wondering where in the code I can multiply the D3 contribution by a constant less than 1 to penalize the D3 interaction. I believe this will resolve the issue of excessive stiffness, and in the future, I can help integrate this modification into the DFT-D3 add-on so that it can be scaled by a constant.

Just FYI the D3 addon seems to make the GPUMD run about a factor of 4 times slower. With the ase implementation of nep with D3 it is about a factor of 100 times slower.

Let me know if you need any additional information.

[brucefan1983]

Which functional did you use? You can scale the s6 and s8 parameters in the code. When you run GPUMD with dftd3, you should get outputs generated by the following code fragment:

  std::cout << "    Add DFT-D3:" << std::endl;
  std::cout << "        potential cutoff = " << rc_potential << " Angstrom" << std::endl;
  std::cout << "        coordination number cutoff = " << rc_coordination_number << " Angstrom"
            << std::endl;
  std::cout << "        functional = " << functional << std::endl;
  std::cout << "            s6 = " << dftd3_para.s6 << std::endl;
  std::cout << "            s8 = " << dftd3_para.s8 << std::endl;
  std::cout << "            a1 = " << dftd3_para.a1 << std::endl;
  std::cout << "            a2 = " << dftd3_para.a2 << " Angstrom" << std::endl;

It is from the file https://github.com/brucefan1983/GPUMD/blob/master/src/force/dftd3.cu

You can scale down dftd3_para.s6 and dftd3_para.s8.

[rebeltroopers]

Hello Bruce,

Thank you for pointing me to the location. I will adjust the parameters slightly and see if I can reduce the issue. I used the PBE functional from D3, specifically dftd3 pbe 12 6.

One more question, does GPUMD have a shell that can be accessed or an easy way to send information through? I am working on an application that needs to run GPUMD from a shell or via ase. It would need to use the GPU as the CPU will be to slow for this as the D3 in the CPU will take on the order of weeks to finish.

One last thing. Is there a way to get a single point calculations with GPUMD?

Best regards,

[brucefan1983]

Have you tried calorine?

[brucefan1983]

Have you tried calorine?

https://gitlab.com/materials-modeling/calorine

[rebeltroopers]

Hello Bruce,

I was looking through the code. I think that calorine could work with the GPUMD calculator it has. I will go write it up and see if this can work.

Thanks for your time.

[brucefan1983]

OK, I think this is better to be moved to discussion.