BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Through the eyes of a descriptor: Constructing complete\, invertib
 le\, descriptions of atomic environments - Martin Uhrin\, EPFL
DTSTART:20210614T153000Z
DTEND:20210614T160000Z
UID:TALK159703@talks.cam.ac.uk
CONTACT:Bingqing Cheng
DESCRIPTION:In this work we apply methods for describing 3D images to the 
 problem of\nencoding atomic environments in a way that is invariant to rot
 ations\,\ntranslations\, and permutations of the atoms and\, crucially\, c
 an be\ndecoded back into the original environment modulo global orientatio
 n\nwithout the need for training a model. From the point of view of\ndecod
 ing\, the descriptor is optimally complete and can be extended to\narbitra
 ry order\, allowing for a systematic convergence of the fidelity\nof the d
 escription. In experiments on molecules ranging from 3 to 29\natoms in siz
 e\, we demonstrate that positions can be decoded with a 96%\nsuccess rate 
 and positions plus species with a 60% rate of success\,\nrising to 95% if 
 a second fingerprint is used. In all cases\, consistent\nrecovery is obser
 ved for molecules with 14 or fewer atoms. Additionally\,\nwe evaluate the 
 descriptor's performance in predicting the energies and\nforces of bulk ir
 on by means of a neural network model trained on DFT\ndata\, achieving roo
 t-mean-square deviations of 3.7 meV/atom and 0.19\neV/Å for energies and 
 forces respectively. The combined ability to both\ndecode and make propert
 y predictions from a representation that does not\nneed to be learned lays
  the foundations for a novel way of building\ngenerative models that are t
 asked with solving the inverse problem of\npredicting atomic arrangements 
 that are statistically likely to have\ncertain desired properties.
LOCATION:virtual ZOOM meeting ID: 263 591 6003\, Passcode: 000042\, https:
 //us02web.zoom.us/j/2635916003?pwd=ZlBEQnRENGwxNmJGMENGMWxjak5nUT09
END:VEVENT
END:VCALENDAR