BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260624T171345Z
LOCATION:Bldg. 8 - B 102
DTSTART;TZID=Europe/Stockholm:20260629T133000
DTEND;TZID=Europe/Stockholm:20260629T140000
UID:submissions.pasc-conference.org_PASC26_sess111_msa166@linklings.com
SUMMARY:From Dataset-Induced Biases to Uncertainty-Driven Learning of Inte
 ratomic Potentials
DESCRIPTION:Viktor Zaverkin (Saarland University, Leibniz Institute for Ne
 w Materials)\n\nMachine-learned interatomic potentials have significantly 
 advanced atomistic modeling by combining near first-principles accuracy wi
 th the efficiency of classical potentials. However, their reliability in s
 imulations strongly depends on the datasets used for training, both for sp
 ecialized and foundation models. While large and diverse datasets can enab
 le generalization across broader regions of configurational and chemical s
 pace, their coverage is rarely balanced. As a result, dataset-induced bias
 es can persist, leading to dataset-dependent observables in atomistic simu
 lations.\n\nAddressing these challenges requires more systematic strategie
 s for constructing datasets, for which active learning provides a natural 
 framework. Its effective application, however, relies on the ability to ef
 ficiently identify uncertain configurations from large candidate pools and
  select diverse batches that reduce the need for repeated model retraining
 . It further requires sampling strategies that promote efficient explorati
 on of relevant regions of configurational space.\n\nThis contribution pres
 ents uncertainty quantification approaches based on gradient features deri
 ved from the linearization of nonlinear models, batch active learning stra
 tegies for selecting diverse configurations, and uncertainty-biased molecu
 lar dynamics to guide exploration of configurational space. Combined, thes
 e approaches enable the construction of machine-learned interatomic potent
 ials with uniform accuracy across complex configurational landscapes while
  significantly reducing the number of required reference calculations.\n\n
 Domain: Chemistry and Materials, Physics, Computational Methods and Applie
 d Mathematics\n\nSession Chairs: Michał Sanocki (Technical University of M
 unich); Ian Störmer (Technical University of Munich); and Philip Loche (Te
 chnical University of Munich, EPFL)\n\n
END:VEVENT
END:VCALENDAR