Thoughts on the instability of HSE06+ADMM

[Nicholas Winner]

I'm running GaAs calculation using HSE06 and ADMM basis sets, and running into some issues. 

I'm trying to initialize with a non hybrid calculation and use SCREEN_ON_INITIAL_P. This is because at 128 atoms (minimum size required for the HSE06 pot.), the memory requirements get to be pretty large even compared to other systems.

Problem is that while initializing with PBE may sometimes lead to a tighter screening threshold, initializing with PBE in this case invariable leads to a destabilized SCF, with no trend to convergence. I thought that, since PBE leads to very low band gap, it might be too poor of an initial guess. So I tried initializing it with (a) r2scan, (b) LDA+OPTX, and (c) LDA+TB09. These each give better estimates of the band gap and so I would expect them to provide better initial density matrices, but all of them failed to create a stable calculation. 

Are there any guesses for what could be done and why this is happening? Does GaAs require no initial guess? That would require some pretty large memory nodes, but is theoretically doable I guess.

Calculation details:

&QS

                        METHOD GPW

                        EPS_DEFAULT 1e-12

                        EXTRAPOLATION ASPC 

                        EPS_PGF_ORB 1e-16 

                &END QS

                ! Parameters needed to perform an SCF run.

                &SCF

                        MAX_SCF 20 ! Max number of steps for an inner SCF loop

                        EPS_SCF 1e-06 ! Convergence threshold for SCF

                        SCF_GUESS RESTART ! How to initialize the density matrix

                        MAX_ITER_LUMO 400 ! Iterations for solving for unoccupied levels when running OT

                        &OT

                                MINIMIZER diis

                                PRECONDITIONER FULL_SINGLE_INVERSE

                                ENERGY_GAP -1

                                ALGORITHM IRAC

                                LINESEARCH 2PNT

                                ROTATION True

                                OCCUPATION_PRECONDITIONER False

                        &END OT

                        &OUTER_SCF

                                MAX_SCF 20

                                EPS_SCF 1e-06

                        &END OUTER_SCF

                &END SCF

                &MGRID

                        CUTOFF 350.0 

                        REL_CUTOFF 50 

                        NGRIDS 5 

                        PROGRESSION_FACTOR 3

&XC

                        &XC_FUNCTIONAL

                                &PBE

                                        PARAMETRIZATION ORIG

                                        SCALE_C 1

                                        SCALE_X 0

                                &END PBE

                                &XWPBE

                                        omega .11

                                        scale_x -.25

                                        scale_x0 1

                                &end

                        &END XC_FUNCTIONAL

                        &HF

                                FRACTION 0.25

                                &SCREENING

                                        EPS_SCHWARZ 1e-07

                                        EPS_SCHWARZ_FORCES 1e-07

                                        SCREEN_ON_INITIAL_P T

                                        SCREEN_P_FORCES False

                                &END SCREENING

                                &INTERACTION_POTENTIAL

                                        CUTOFF_RADIUS 6.63

                                        T_C_G_DATA t_c_g.dat

                                        omega .11

                                        POTENTIAL_TYPE shortrange

                                &END INTERACTION_POTENTIAL

                                &LOAD_BALANCE

                                        RANDOMIZE True

                                &END LOAD_BALANCE

                                &MEMORY

                                        EPS_STORAGE_SCALING 0.1

                                        MAX_MEMORY 2000

                                &END MEMORY

                        &END HF

                &END XC

[Krack Matthias (PSI)]

Did you try already OT MINIMIZER CG (and possibly no ALGORITHM IRAC but the default STRICT)?

--
You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/62f09ab5-01d6-497e-8c27-f11b2693756dn%40googlegroups.com.

[Nicholas Winner]

I was not able to get convergence using algo strict. In fact, I always thought that IRAC was a little more flexible for optimization than strict orthogonality. Maybe I'm wrong.

CG including 3PNT (where I turned off rotation) also does not work.

[Matt Watkins]

Hi, 

seemed to work OK for me running a 3x3x3 of the conventional cell. See attached starting from PBE guess.
Still HSE complains about box being too small ... but convergence is smooth.

Matt

[Nicholas Winner]

Hi Matt. It looks like the reason yours converged and mine didn't is the auxiliary basis. I was running comparing the admm-{dz, dzp, tzp} basis sets from the new BASIS_ADMM_UZH file. For that case I was eventually able to get dz quality converged only if I removed the initial density screening, while the others diverged like I said. These basis sets, however, seem to have a worse condition number and more diffuse functions than the FIT type. Using fit9 and cfit10 like you did, 360Gb of integrals are reduced to 120Gb of integrals, and the convergence is good again. I'll have to be a little more careful with which auxiliary basis I choose, I suppose.

HSE always complains about the box being too small. I think its a bug? So long as the minimum distance between images is 10A it should be fine since the decay length is 5A.

[Matt Watkins]

I will have a look at the HSE error message - indeed, I suspect it just checks whether a truncated potential is used then complains ...

JGH may have comment on the BASIS_ADMM_UZH sets. They are definitely more diffuse and probably 'more accurate' but I am not sure if they are practical for routine condensed phase DFT calculations.

Matt