Interatomic  Poten-als  in  LAMMPS  

 Aidan  P.  Thompson  Sandia  Na1onal  Laboratories  

2nd  LAMMPS  Users  Workshop  August  2011,  Albuquerque,  New  Mexico  

Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company,for the United States Department of Energy’s National Nuclear Security Administration under contract DE-AC04-94AL85000."

Recently  Added  Poten-als  •  COMB  poten-al  (Genera-on  2)  

Tzu-­‐Ray  Shan  (U  Florida),  talk  on  Tues  PM  metal  and  semiconductors  and  their  oxides  

•  Embedded  ion  method  poten-als  (EIM)  Xiaowang  Zhou  (Sandia),  ionic  compounds  combo  of  Li,  Na,  K,  Rb,  Cs,  F,  Cl,  Br,  and  I  

•  C++  version  of  ReaxFF  Me1n  Aktulga  (LBNL),  talk  on  Wed  PM  can  be  faster  than  Fortran  version  by  2-­‐3x  

•  Electron  force  field  (eFF)  Andres  Jaramillo  (Caltech),  explicit  electron  dynamics  in  extreme  condi1ons  

Recently  Added  Poten-als  •  AIREBO  poten-al  bug  fixes  

Marcel  Fallet  &  Steve  Stuart  (Clemson),  one  more  bug-­‐fix,  upgrade  is  imminent  

•  Mishin  ADP  poten-al  Chris  Weinberger  (Sandia)  &  Chandra  Veer  Singh  (Cornell)  angular-­‐dependent  EAM  for  metals  and  alloys  

•  Dreiding  poten-al  Tod  Pascal  (Caltech)  hydrogen  bonding  for  solvated  bio-­‐molecules  

•  New  Peridynamics  poten-als  Mike  Parks  &  Stuart  Silling  (Sandia),  talk  on  Wed  PM  fracture  at  the  meso  and  con1nuum  scales  

Upcoming  Poten-als  •  Core/shell  poten-al  

Mike  Chandross  (Sandia),  zero-­‐order  model  for  polariza1on,  uranium  and  other  nuclear  fuel  materials  

•  COMB  poten-al  (Genera-on  3)  Tzu-­‐Ray  Shan  (U  Florida),  more  materials  with  polariza1on  effects  

•  MGPT  poten-al  from  John  Moriarty  &  Jaime  Marian  (LLNL),  tantalum  and  other  transi1on  metals  

•  BOP  poten-al  Xiaowang  Zhou  and  Don  Ward  (Sandia)  High-­‐accuracy  poten1al  for  semiconductors  

 

LAMMPS  Poten-al  benchmarks  

Compute  Cost  of  Interatomic  Poten-als    Growing  Exponen-ally  

 Drivers  •   Cycles  are  cheap  •   Availability  of  quantum  calcula1ons  (N  <  100)  •   Qualita1ve  accuracy  no  longer  enough  

h`p://lammps.sandia.gov/bench.html#poten1als  

Compute  cost  of  LAMMPS  poten1als  versus  publica1on  date      

CHARMm

EAMStillinger-Weber

Tersoff

AIREBO

MEAM

ReaxFF

ReaxFF/C

eFF

COMB

EIM

GAP(Estim.)

1980 1990 2000 2010Year Published

10-6

10-5

10-4

10-3

10-2

Cost

[cor

e-se

c/at

om-ti

mes

tep]

Challenges  with  Complex  Poten-als  

•  How  to  Implement  in  LAMMPS?  – Rewrite  code  from  scratch  (REBO)  –  Integrate  exis1ng  serial  code  (ReaxFF)  – Access  via  general  API  (KIM)  

•  How  to  Validate  LAMMPS  Version?  •  How  to  Handle  New  Versions?  

How  to  Fit  Poten-als  to  New  Materials?  

•  Automated  fieng  procedures  exist  for  certain  classes  of  materials  and  poten1als  (EAM)  

•  More  commonly,  good  fits  can  be  obtained  only  by  gurus  (Baskes,  van  Duin)  

•  Combina1on  of  nonlinear  op1miza1on  and  physical  intui1on  •  Increasing  interest  in  automated  machine-­‐learning  

approaches  –  Splines  –  Gene1c  programs  – Mul1-­‐objec1ve  op1miza1on  –  Neural  networks  –  Series  expansions  

GAP  Approach  for  Interatomic  Poten-als  

Diamond:  Force  errors  for  GAP  fi`ed  to  DFT.    Adding  higher-­‐order  GAP  coefficients  systema1cally  increases  accuracy    Bartok  et  al.,  PRL  104  136403  (2010)  

GAP:  A  systema-c,  informa-cs  approach  •  Based  on  QM  and  mathema1cs  rather  than  

empiricism.  •  Local  density  around  each  atom  expanded  in  

4D  hyperspherical  harmonics,  analogous  to  Fourier  series  

•  Atomic  configura1ons  described  by  bispectrum  of  lowest-­‐order  coefficients  in  series  

•  Preserves  universal  physical  symmetries:  invariance  w.r.t.  rota1on,  transla1on,  permuta1on  

•  Gaussian  process  (GP)  regression  used  to  interpolate  energy  of  QM  configura1ons  

•  100-­‐1000x  more  expensive  than  MEAM    •  Far  cheaper  than  QM,  linear  scaling  •  Can  trade  performance  and  accuracy    

r ! r( ) = c j !m ,mYj!m ,m ",#,$( )

!m ,m=" j

j

#j=0

$

#rcut

0 50 100 150 200 250 300Pzz

[GPa]

0

1000

2000

3000

4000

5000

6000

Tem

per

ature

[K

]

Dr09 a-axisBa09i a-axisDFTDFT Melt LineEOS Benedict et al.

GAP  Poten-al  for  Beryllium  

Beryllium  

Working  with  GAP  developers,  fieng  directly  to  forces  and  energies  from  high-­‐temperature  DFT  MD  simula1ons  of  small  systems  (from  Mike  Desjarlais,  1640)    

!!" !#$ !#" !$ " $ #" #$ !"!#$

!#"

!$

"

$

#"

#$

!"

GAP  Forces  versus  DFT  Forces  for    Liquid  Be  at  5,000  K,  250  atoms  

0 50 100 150 200 250 300Pxx

[GPa]

0

1000

2000

3000

4000

5000

6000

Tem

per

atu

re [

K]

Dr09 a-axisBa09i a-axisDFTDFT Melt LineEOS Benedict et al.

•  Ini1al  fit  to:  −  ambient  HCP  −  high-­‐pressure  BCC  and  liquid  

•  Accurately  reproduced  that  data  •  Problems  with:  

−  ambient  elas1c  constants  −  high-­‐pressure  HCP  

•  Need  to  refit  with  more  data