The numerical grid point for DFT in GAMESS



You may suffer from SCF convergence problems often. This is mostly related to the Numerical grid of DFT. The default grid of GAMESS is NRAD = 96, NLEB = 302. NRAD is the number of radial spheres, while NLEB-s the angular (Lebedev) grid points on each sphere. These numbers are a bit too high. In fact, the particular numbers should also depend on the types of functionals you use. For example, (50,194) is recommended for hybrid or GGA functionals, which you mostly use for the calculations.



I took this table from https://manual.q-chem.com/5.2/Ch5.S5.SS2.html.

GAMESS has an option of SG1 in $DFT group. You can set it as SG1=.t. GAMESS recommends its use for the early steps of geometry optimization since it is extremely FAST. But there is confusion. The SG1=.t. in GAMESS is NOT identical to what the Peter Gill recommended above. The SG1 in GAMESS sets (24, 194), while the above table indicates SG-1 as (50,194). I found a bug in the GAMESS manual, which indicates (24, 194) but what the GAMESS actually utilizes is (50, 194). So the SG1 in GAMESS is identical to the above table.

So this is my recommendation. Use SG1=.t. for initial geometry optimization. Once it is done, use NRAD=50 NLEB=194 in your $DFT group for all types of calculations.
However, there is a difference between SG1=.t. and directly setting (NRAD=50, NLEB=194). The former utilizes a pruning process, which removes unnecessary grid points, while the latter doesn't. Therefore, your final results will depend on the choices.

There is one more complication though. GAMESS strangely utilizes a procedure of HF —> coarse grid DFT —> fine grid DFT for its SCF procedure. This could be Ok, if you don't have any previous converged orbitals. But if you do, this procedure creates a problem. If you provide nearly converged $VEC group in your input, GAMESS turns off automatically HF step and directly goes to the next step, which is coarse grid DFT. This is a problem since your $VEC was optimized with find grid in the previous run. But GAMESS uses it with coarse grid option, making the already good $VEC worse. There is no way to turn off the coarse grid part of SCF. But you can make the coarse grid equal to fine grid of NRAD and NLEB. You can do this by setting NRAD0 and NLEB0 coarse grid options equal to NRAD and NLEB.

Meta-GGA functionals require a tighter grid to achieve the same accuracy. For this reason a tighter default grid of NRAD=99 and NLEB=590 is chosen by default with all meta-GGA functionals.

Comments

Post a Comment

Popular posts from this blog

The ROHF of GAMESS for TDDFT calculation

Unfortunately, there are many definitions of ROHF wavefunction. The Fock matrix in the MO basis has the form                                               closed       open        virtual         closed      F2      |     Fb     | (Fa+Fb)/2                  -----------------------------------         open        Fb      |     F1     |    Fa                  -----------------------------------         virtual   (Fa+Fb)/2 |     Fa     |    F0 where Fa and Fb are the usual alpha and beta Fock matrices any UHF program produces.  All ROHF methods agree on these, as they are the variational conditions that separate the doubly occupied, alpha occupied, and empty orbital spaces. The diagonal blocks can be written                F2 = Acc*Fa + Bcc*Fb                F1 = Aoo*Fa + Boo*Fb                F0 = Avv*Fa + Bvv*Fb Some choices for the canonicalization coefficients to define the diagonal blocks are                           Acc  Bcc  Aoo Boo  Avv  Bvv  Guest and Saunders
Read more

Improving SCF convergency of DFT

As compared to HF, DFT density typically converges slowly or often fails to converge. This post describes a general strategy dealing with these situations. GAMESS has two popular SCF convergers of DIIS and SOSCF. The DIIS works well when the density matrix is a bit far from its minimum, while SOSCF is a preferred choice once the density matrix is near to it. In addition, it is generally true that HF converges much better than DFT. Therefore, it is better to start with HF and DIIS. The switch to SOSCF optimizer needs to be activated, once the density is near to its minimum. This can be realized by  $dft DFTTYP=B3LYP SWOFF=1e-6 sg1=.t. $end  $scf couple=.true. alpha(1)=0.5,0.5,0.5 beta(1)=0.5,0.5,0.5    dirscf=.t. diis=.t. soscf=.f. damp=.t. shift=.t.    swdiis=1e-4 $end The important keywords are SWOFF = 1e-6 DIIS=.t. SWDIIS = 1e-4 The SWOFF = 1e-6 ensures that HF is utilized in the SCF cycle until the density change becomes below 1x10^-6. The DIIS=.t. sets the initial op
Read more

Fortran compilation on Linux

1. gfortran Type fortran yourfile.f90 -o yourfile.x Then, it will create yourfile.x in the same directory. Simply type yourfile.x will execute it. See more examples at  https://gcc.gnu.org/wiki/GFortranGettingStarted 2. Intel Fortran Students as well as instructors can receive free copy of intel compilers at  https://software.intel.com/en-us/qualify-for-free-software/student Once you installed it, you should put the following lines in your .cshrc file. for c shell. If you use b shell, compilervars.csh should be replaced with compilervars.sh. <install_dir>/bin/compilervars.csh intel64 where <install_dir> is the directory structure containing the compiler /bin directory The default path for <install_dir> is /opt/intel/ setenv LD_LIBRARY_PATH "/usr/lib64" set path=( $path /opt/intel/bin ) After this, you should run .cshrc by typing    source .cshrc Then you can compile your file as   ifort yourfile.f90 -o yourfile.x
Read more