BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Machine learning an interatomic potential without (much) human eff
 ort - Noam Bernstein\, U. S. Naval Research Laboratory
DTSTART:20200817T153000Z
DTEND:20200817T160000Z
UID:TALK150262@talks.cam.ac.uk
CONTACT:Bingqing Cheng
DESCRIPTION:While many of the applications of machine learning (ML) method
 s in materials have been to the\ndirect prediction of experimentally obser
 vable properties\, the same methods for regression in\nhigh dimensions hav
 e also been very successful in approximating high accuracy\, e.g. density\
 nfunctional theory (DFT)\, potential energy surfaces. Such an approximatio
 n constitutes an\ninteratomic potential: an explicit formula for the bondi
 ng energy as a function of the atomic\npositions. Since the input space is
  large and the effective functional form of ML methods is\nextremely flexi
 ble\, it is easy to inadvertently produce a fit that is accurate near the 
 input data\,\nbut extremely inaccurate for reasonable configurations that 
 are just a bit different. Since such a\nfit leads to artifacts in any real
  simulation\, it has proven essential to develop an extensive fitting\ndat
 abase that includes not only all the relevant configurations but also the 
 boundary of the\nimportant\, i.e. relatively low energy\, region. Doing th
 is ensures that the resulting potential has\nan increasing energy as it le
 aves the low energy configuration subspace. Our experience in\ndeveloping 
 several such fitting databases has shown that doing this manually is a tim
 e\nconsuming procedure that would greatly benefit from automation and broa
 dly applicable\nheuristics. We have developed such a process\, which itera
 tes over a sequence of random-\nstructure searches and potential fitting\,
  with only a minimal number of reference (DFT)\nevaluations. The resulting
  potentials are accurate and robust for the configurations that occur\ndur
 ing the random structure search\, from the high-energy regions all the way
  to local minima\;\nthey also give at least qualitatively reasonable value
 s for defects such as vacancies and\nsurfaces. We apply the process to a r
 ange of materials with different chemical nature and\ncoordination environ
 ments\, including insulating\, semiconducting\, and metallic bonding.
LOCATION:virtual ZOOM meeting ID: 263 591 6003\, https://zoom.us/j/2635916
 003
END:VEVENT
END:VCALENDAR