Feature: Different max-force criterion of ABACUS from ASE&VASP

[QuantumMisaka]

Details

ABACUS have the different max-force criterion from ASE and VASP.

In ABACUS, the max-force, namely LARGEST GRAD in print-out and Largest gradient in running*.log, which seems to be calculated by source/module_relax/relax_new/relax.cpp:

//=========================================
//set gradient for ions degrees of freedom
//=========================================

    grad_ion.zero_out();
    ModuleBase::matrix force_eva = force * (ModuleBase::Ry_to_eV / ModuleBase::BOHR_TO_A); //convert to eV/A

	int iat=0;
	for(int it = 0;it < GlobalC::ucell.ntype;it++)
	{
		Atom* atom = &GlobalC::ucell.atoms[it];
		for(int ia =0;ia< GlobalC::ucell.atoms[it].na;ia++)
		{	
            double force2 = 0.0;
			if(atom->mbl[ia].x == 1)
			{
				grad_ion(iat, 0) = force_eva(iat, 0);
                if( std::abs(force_eva(iat,0)) > max_grad) max_grad = std::abs(force_eva(iat,0));
			}
			if(atom->mbl[ia].y == 1)
			{
				grad_ion(iat, 1) = force_eva(iat, 1);
                if( std::abs(force_eva(iat,1)) > max_grad) max_grad = std::abs(force_eva(iat,1));
			}
			if(atom->mbl[ia].z == 1)
			{
				grad_ion(iat, 2) = force_eva(iat, 2);
                if( std::abs(force_eva(iat,2)) > max_grad) max_grad = std::abs(force_eva(iat,2));
			}
			++iat;
		}
	}
    assert(iat==nat);

    if(max_grad > force_thr_eva) force_converged = false;
	if(GlobalV::OUT_LEVEL=="ie")
	{
        std::cout << " ETOT DIFF (eV)       : " << etot - etot_p << std::endl;
		std::cout << " LARGEST GRAD (eV/A)  : " << max_grad << std::endl;
        etot_p = etot;
	}

The max force is directly calculated by the maximum of TOTAL-FORCE along x, y, and z.
An example is :

><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><

    TOTAL-FORCE (eV/Angstrom)

 ><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><

     atom              x              y              z
       H1        +0.012106811       -0.0042683316       +0.0037308624
       H2        -0.024080247        -0.018872146        +0.047958005
........
 Ion relaxation is converged!

 Energy difference (Ry) = +0.000000

 Largest gradient is (eV/A) = +0.047958
 end of geometry optimization

But this method is different from ASE and VASP, which use the number of real force calculated by the vector sum of force along x, y, z, direction, respectively.

In ASE, the max-force calculation code is

if forces is None:
        forces = self.atoms.get_forces()
        fmax = sqrt((forces ** 2).sum(axis=1).max())

using the same example above, but by ASE-ABACUS, the fmax will be 0.057059 eV/Ang.
in which the force is read from ABACUS calculation result.

in VASP, an example is :

POSITION                                       TOTAL-FORCE (eV/Angst)
 -----------------------------------------------------------------------------------
      9.58597      2.23321      7.72570         0.010849      0.007507     -0.021642
      4.66193      2.20102      9.10501        -0.008403      0.036297     -0.019736
........
  FORCES: max atom, RMS     0.042162    0.016426
  FORCE total and by dimension    0.179943    0.040538

The max-force is calculated by the same method of ASE.

I consider that the max-force criterion of ASE and VASP is more reasonable, and the largest gradient in ABACUS my cause confusion to user and may raise problem during relaxation ?

By the way, in QE, the relaxation criterion is by total-force, which is the sum of all number of real-force.
@kirk0830 Thanks for checking

Task list for Issue attackers (only for developers)

• Reproduce the performance issue on a similar system or environment.
• Identify the specific section of the code causing the performance issue.
• Investigate the issue and determine the root cause.
• Research best practices and potential solutions for the identified performance issue.
• Implement the chosen solution to address the performance issue.
• Test the implemented solution to ensure it improves performance without introducing new issues.
• Optimize the solution if necessary, considering trade-offs between performance and other factors (e.g., code complexity, readability, maintainability).
• Review and incorporate any relevant feedback from users or developers.
• Merge the improved solution into the main codebase and notify the issue reporter.

[YuLiu98]

I think all these criteria are reasonable if you describe them in detail to let users know what they are doing. If modified, the new version of ABACUS would be inconsistent with the old version, which is also confusing.

[QuantumMisaka]

I think all these criteria are reasonable if you describe them in detail to let users know what they are doing. If modified, the new version of ABACUS would be inconsistent with the old version, which is also confusing.

I do think max-force or Largest Gradient is describing the things ASE & VASP is done, and the so-called inconsistent will only affect the ion step convergence criterion, but not the energy & force & stress results themselves.

User may be more confusing about the max-force and relax convergence criterion from ABACUS and ASE-ABACUS are different.

[QuantumMisaka]

Any other discussion ?