Updates on high-throughput DFPT · LiMgAs P2Ir Si Li3Sb K2O Ga3Os Be3P2 CsAu ZnSe MgO AgCl SiC YWN3 SrO PbF2 RuS2 MgSiP2 SiO2 GaP Be2C SnHgF6 MgMoN2 ZnO VCu3S4 ZrSiO4 Ba(MgP)2 Ba(MgAs)2

Updates on high-throughput DFPT

Guido Petretto, Matteo Giantomassi, Henrique P. C. Miranda,Michiel J. van Setten, David Waroquiers, Shyam Dwaraknath,

Donald Winston, Xavier Gonze, Kristin A. Persson, Geoffroy Hautier,Gian-Marco Rignanese

MODL, Institute of Condensed Matter and Nanosciences, Université catholique de Louvain,1348 Louvain-la-neuve, Belgium

22/05/2019

G. Petretto (MODL, UCL) High-throughput DFPT 22/05/2019 1

Outline

1 Introduction

2 High-throughput DFPT

3 Phonons database

4 Abinit for HT

5 Further developments


Outline

1 Introduction


3 Phonons database

4 Abinit for HT



Aim of the project

Diffusion of large databases based on DFT calculations

⇓High-throughput workflows for Abinit

⇓DFPT phonon band structures at Materials Project


Where were we?

ABIDEV 2017:

Infrastructure to run high-throughput calculations with Abinit

dependencies on different python frameworks

high-throughput framework: Abiflows

Preliminary results:

Convergence study

Workflows at NERSC

ABIDEV 2019:

Results

Materials project

Problems encountered

Next steps


Outline

1 Introduction


3 Phonons database

4 Abinit for HT



High-throughput framework

DFT + DFPT code

Abinit python analysis framework Workflow manager

Abinit workflows framework

Generic material analysis framework

Available on Github:https://github.com/abinit/abiflows


DFPT workflow

Extended workflow to cover all possible DFPT calculation available


Convergence study

Find optimal k-points and q-points sampling for high-throughputPetretto, Gonze, Hautier, Rignanese, Comp. Mat. Sci., 144, 331 (2018)

Set of 48 semiconductors

Various sizes, crystal symmetries, gaps

Several K and Q grids

Statistic on error with respect to dense grids:

relative and absolute errormean and maximum error

NaLi2Sb Ca(CdP)2 CdS SrLiP InS GaN RbYO2 Si4P4RuSiO2 BP AlSb LiZnP MgCO3 ScF3 ZnGeN2 CsBrLiMgAs P2Ir Si Li3Sb K2O Ga3Os Be3P2 CsAuZnSe MgO AgCl SiC YWN3 SrO PbF2 RuS2MgSiP2 SiO2 GaP Be2C SnHgF6 MgMoN2 ZnO VCu3S4ZrSiO4 Ba(MgP)2 Ba(MgAs)2 Ca(MgAs)2 C RbI FeS2 ZnTe


Convergence study: grids density

absolute and relative errors on ω, Eat, Z∗ and �

1500 points per reciprocal atom ⇒ Nkpt*Natoms w 1500⇒ All materials converged with 0.5 cm−1 MAE and 0.6% MARE

Better using a Q-grid commensurable with K-grid

⇒ Smoother close to Γ

UL LK KW WXΓ Γ0

25

50

75

100

125

150

175

Fre

quen

cy(c

m−

1)

AgCl - qpt 6× 6× 6AgCl - qpt 12× 12× 12

0.3 0.6 0.9 1.2 1.5

600800

100012001400160018002000

kpra

ω

0.3 0.6 0.9 1.2 1.5εMARE (%)

600800

100012001400160018002000

qpra

ω

512 1000 1728 2744 4096kpra

−0.5+0.5

−0.5+0.5

−0.5+0.5

−0.5+0.5

−0.5+0.5

−0.5+0.5

Err

oron

ω(%

)38 cm−1

46 cm−1

62 cm−1

103 cm−1

147 cm−1

156 cm−1AgCl - qpt: U


Convergence study: symmetry of the grid

Convergence rate of phonon frequencies ω for symmetric versusnon-symmetric grids

500 1000 1500 2000 2500 3000 3500kpra

10−2

10−1

100

101

MxA

RE

(%)

Si - brkSi - sym

BP - brkBP - sym

ZnO - brkZnO - sym

⇓Grids should preserve the symmetries of the system


Results validation

Validation versus experimental data

Vibrational entropy at 300K

0 20 40 60 80 100 120 140

Sexp (J/(K mol))

0

20

40

60

80

100

120

140

Ssi

m(J

/(K

mol

))

InS

MgCO3

SiO2

K2O

AlSb

ZnSe

ZrSiO4

BP

MgOSiO2

C

GaN

CdS

ScF3

SrO

Si

RbI

GaP

PbF2

SiC

Be2C

ZnO

CsBr

ZnTe

RuS2FeS2

AgCl

CB

e 2C

SiC S

iB

PS

iO2

MgO

GaN

SiO

2

Zn

OG

aPF

eS2

Ru

S2

−15−10−5

05

101520

Err

or(%

)

SrO

Cd

SA

lSb

MgC

O3

InS

Zn

Se

Zn

Te

ZrS

iO4

AgC

lS

cF3

K2O

CsB

rP

bF

2

Rb

I

−15−10−5

05

101520

Err

or(%

)

Γ phonon frequencies

0 200 400 600 800 1000 1200

ω (cm−1)

−40

−30

−20

−10

0

10

20

Err

or(%

)


Outline

1 Introduction


3 Phonons database

4 Abinit for HT



Materials Project phonons database

Open access database: 1521 semiconductor materials (and growing. . . )1508 of those materials with less than 13 atomic sites

∼ 5M CPU-hours

Interatomic forceconstants (DDB files)

Phonon dispersion

Born effective charges

Dielectric tensor

Thermodynamic prop.:

∆F , ∆Eph, Cv , S

Γ X YΣ Γ Z Σ N P Y Z X P0

500

1000

Fre

quen

cy(c

m−

1) β-cristobalite

Γ X M Γ Z R A Z X R M A

stishovite

Γ M K Γ A L H A L M K H0

500

1000

Fre

quen

cy(c

m−

1) α-quartz

Petretto, Dwaraknath, Miranda, Winston, Giantomassi, Van Setten, Gonze, Persson, Hautier, Rignanese, Scientific Data (2018)


Materials Project phonons database

All the data available on the websiteand through REST service.

Interactive visualization of the phonondispersion using the phononwebsite

http://henriquemiranda.github.io/

phononwebsite/


http://henriquemiranda.github.io/phononwebsite/http://henriquemiranda.github.io/phononwebsite/

Outline

1 Introduction


3 Phonons database

4 Abinit for HT



Reliability: workflows

For the 1521 phonon band structures:

Relax workflow

Phonon workflow (+ anaddb)

Out of all the submitted workflows only ∼ 30 did not completesuccessfully

Main reasons:

Too slow relaxation/relaxation did not converge

Too small gap (switched to metallic)

Presence of La

Poor choice of materials


Reliability: results

Warnings available in the database:

Negative ω close to Γ: 24 materials

ASR break >30 cm−1: 72 materials

CNSR break >0.2e: 92 materials

Wave Vector

0

10

20

30

40

50

60

Freq

uenc

y (c

m1 )

0 20 40 60 80asr value

0

50

100

150

200

250

300

350

400

Num

ber o

f mat

eria

ls

0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75chneut value

0

200

400

600

800

1000

Num

ber o

f mat

eria

ls


Problems encountered

Current MP cluster: KNL nodes

not optimizedreserve full noderelatively poor performancesdifficult to fine tune parallelization at high-throughput level

Relax - ionmov 22: seems faster but may fail at small tolmxf (1e−6)

⇒ switch ionmov at python level.Autoparal for DFPT

always gives the maximum number of preocesses allowedoften parallelizing over just the k-points is advantegeous

⇒ could be improved?Memory

moving to larger materials already caused a few jobs to fail due tomemory issues

⇒ might be needed to rely on estimation of total memory


Outline

1 Introduction


3 Phonons database

4 Abinit for HT



More data on the MP database

Calculations have proceed: almost 500 more materials

More physical quantities will be extracted from the phonon data:

Sound velocity as slope of acoustic modes

Low-frequency dielectric permittivity tensor �αβ(ω)

Thermal displacement ellipsoids (Debye-Waller)

0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35Wave Vector

0

20

40

60

80

100

120

140

160

Freq

uenc

y (c

m1 )

0.7071, 0.0000, 0.7071 - X

0 200 400 600 800 1000Frequency (cm 1)

20

0

20

40

60

()

Re{ 00}Re{ 11}Re{ 22}Im{ 00}Im{ 11}Im{ 22}


Phonons for metals

Extend the calculation to metals as well. Materials project:

24356 metals8242 with less than 6 atoms

Potential issues:

Denser k-point grids

Q-point grids?

Smearing

Kohn anomalies

Fourier interpolation 0

20

40

60

80

100

Freq

uenc

y (c

m )-1

Γ X

Transverse

Longitudinal

non-uniform gridspline interpolationuniform grid (20 q-points)uniform grid (25 q-points)

ε=0.09

He, Liu, Li, Rignanese, Zhou (2019)

⇓New convergence study required


Volume: Grüneisen parameters

Phonons at different volumes (e.g. ±2%) ⇒ GrüneisenTools already available in Abinit (netcdf) and Abipy

Preferable to have separate workflows

g = GrunsNcFile.from_ddb_list(["-2_DDB", "+0_DDB", "+2_DDB"])

g.plot_phbands_with_gruns(with_doses=None)

g.plot_gruns_scatter()

phbst.plot_phbands(units="cm-1")

M K A L H A L M K HWave Vector

0

50

100

150

200

250

300

Freq

uenc

y (c

m1 )

0 50 100 150 200 250 300Frequency (cm 1)

4

2

0

2

4

Grun

eise

n


Volume: Quasi-Harmonic Approximation

Tools for QHA implemented in Abipy:

Standard QHA object

generated from GSR and PHDOS netcdf files

Fittings

Thermal expansion coefficient

Interface to Phonopy for further functionalities

Cheaper QHA: QHA-3P (Nath et al. arXiv:1807.04669)

several electronic energies at different V and 3 phonon calculations

extrapolate phonon contribution at other V

Satisfactory results

⇒ Interesting for high-throughput


Volume: Quasi-Harmonic Approximation

Example QHA-3P for Si

qha = QHA.from_files(gsr_paths, dos_paths)

qha3p = QHA3P.from_files(gsr_paths, gruns_path, ind_doses=[1,2,3])

fig = qha.plot_thermal_expansion_coeff()

qha3p.plot_thermal_expansion_coeff(ax=fig.axes[0])

0 200 400 600 800 1000 1200 1400 1600

T (K)

0.0

0.5

1.0

1.5

α(K−

1)

×10−5

qha

qha3p


Thank you for your attention


High-throughput framework

What do we need for high-throughput with ?

Python interface to DFT codes

Inputs

pseudopotentials and cutoffs

automatic generation

Workflow management

Database interface

Workflows

Error handling and data analysis


Materials Project phonons database: rester

Fetching DDB files from MP and analyze results with Abipy

ddb = abilab.DdbFile.from_mpid("mp-1265") # MgO

phbst, phdos = ddb.anaget_phbst_and_phdos_files(ndivsm=20, nqsmall=20,

lo_to_splitting=True)

phbst.plot_phbands(units="cm-1")


q-points convergence

Use subgrids of the q-point grid to check the convergence w.r.t. qpt

smoothed average

average for each q density

smoothed max

Material dependent

Suggest good convergence level at 1500 qppa

reducing by a factor 2 may lead to sizeable errors on average


IntroductionHigh-throughput DFPTPhonons databaseAbinit for HTFurther developments

Documents

Updates on high-throughput DFPT · LiMgAs P2Ir Si Li3Sb K2O Ga3Os Be3P2 CsAu ZnSe MgO AgCl SiC YWN3 SrO PbF2 RuS2 MgSiP2 SiO2 GaP Be2C SnHgF6 MgMoN2 ZnO VCu3S4 ZrSiO4 Ba(MgP)2 Ba(MgAs)2