eResearch2013.pdf

Transcript

1. Shreyas Cholia! Lawrence Berkeley National Laboratory The Materials Project -­‐  1  -­‐   July  3,  2013  
2. An  open  science  ini3a3ve  that  makes   available  a  huge  database  of  computed   materials  proper3es  for  all  materials   researchers.       The Materials Project Wordcloud  showing   frequencies  of  elements  in   Materials  Project's  database     ..except  Oxygen,  which  appears  12,751   Emes  (3.5x  as  much  as  the  next  most   frequent,  Phosphorus)  
3. Big Data Science  is  increasingly   data-­‐driven       Computa3onal   cycles  are  cheap     Take  an  –omics   approach  to     science     Compute  all     interes/ng  things     first,  ask  ques/ons     later      
4. The –omics approach •  Instead  of  trying  to  derive  a  solu3on  and  compute   the  results,  just  compute  the  space  of  all   possibili3es  and  look  for  the  op3mal  result  in  there.  
5. 18  years  from   creaEon  to   commercial   manufacture!   Teflon Titanium Velcro Polycarbonate GaAs Diamond-like Thin Films Materials Data from: Eagar, T.; King, M. Technology Review (00401692) 1995, 98, 42. invented 1960 1970 1950 "Need for speed" in new materials Lithium  ion   S.  WhiLngham   Sony   1980 1990 2000
6. Infrastructure SubmiMed   Materials   Materials   Data   (MongoDB)   Materials  Proper-es   Supercomputers   •  Over  10  million  CPU  hours  of   calculaEons  in  <  6  months   •  Over  40,000  successful  VASP  runs   (30,000+  materials)   •  Generalizable  to  other  high-­‐ throughput  codes/analyses   CalculaEon   Workflows   Supercomputers   Codes  to  run   (in  sequence)   Atomic  posi-ons  
7. 0 10000 20000 30000 40000 50000 60000 Jul 2011 Oct 2011 Jan 2012 Apr 2012 Jul 2012 Oct 2012 Date Number of runs state Failed Successful Computation •  Run  VASP  on   NERSC   supercompu3ng   resources   •  Use  Fireworks  to   manage  large   groups  of  runs   •  Results  in  …  data   for  MP   Just sit back and enjoy the automation.. Total  
8. Application Stack •  Python  Django  Web  Portal   •  Pymatgen  and  other  scien3fic  Python  libraries   •  Fireworks  +  VASP    MongoDB  under  all  these  layers   –  Currently  at  Mongo  2.2   8  
9. Why we like MongoDB •  Flexibile  Schema   •  Developer-­‐friendliness   •  JSON   •  Great  for  read-­‐heavy  paZerns   9  
10. The Materials Project hMp://materialsproject.org/  
11. Materials API – Data As A Service hZps://www.materialsproject.org/rest/v1/materials/Fe2O3/vasp/energy   Preamble  URL   Unique  Iden3fier.  Eg.   a  formula  (Fe2O3),  id   (1234)  or  chemical   system  (Li-­‐Fe-­‐O)   Data  type   (vasp,  exp,   etc.)   Property   Materials  API  maps  URLs  to  data  objects  
12. Sample JSON output GET  hZps://www.materialsproject.org/rest/v1/materials/Fe2O3/vasp/energy   { "created_at": "2013-03-17T09:14:58.158081", "valid_response": true, "version": { "pymatgen": "2.5.4", "db": "2013.02.25", "rest": "1.0" }, "response": [{ "energy": -132.33005625, "material_id": 542309 }, { "energy": -66.62512425, "material_id": 24972 }], "copyright": "Copyright 2012, The Materials Project" } Just  the  energy  and  the   id  of  the  material  
13. Takeaways •  Bigdata  X-­‐omics:  Compute  lots  of  results  and  look   for  solu3ons  in  the  resultant  data  set   •  Schema  Last  Data:  Using  a  MongoDB  NoSQL   database  for  science  data  to  allow  for  evolving   schemas   •  Data  as  a  service:  Make  data  available  through  web   interfaces  and  REST  APIs  to  reduce  the  barrier  for   end  user  scien3sts   -­‐  13  -­‐  
14. Thanks   •  Materials  Project   –  hMp://materialsproject.org   •  Our  code   –  hMps://github.com/materialsproject/pymatgen   –  hMps://github.com/materialsproject/fireworks   •  Contact:   –  Shreyas  Cholia:  scholia@lbl.gov   •  Ques3ons?     -­‐  14  -­‐