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:20260624T171341Z
LOCATION:Bldg. 8 - B 102
DTSTART;TZID=Europe/Stockholm:20260629T143000
DTEND;TZID=Europe/Stockholm:20260629T150000
UID:submissions.pasc-conference.org_PASC26_sess111_msa254@linklings.com
SUMMARY:End-to-End Uncertainty Quantification for Atomistic Machine Learni
 ng
DESCRIPTION:Matthias Kellner (EPFL)\n\nMachine-learning interatomic potent
 ials (MLIPs) have found widespread adoption in atomistic simulation workfl
 ows. Their use, however, introduces statistical uncertainties from finite 
 training data and approximation errors, as well as systematic discrepancie
 s inherited from the underlying electronic-structure reference. These unce
 rtainties demand rigorous quantification to signal to the modeller when si
 mulations are not trustworthy.\n\nIn this talk, I will present recent deve
 lopments toward end-to-end uncertainty quantification for atomistic machin
 e learning. I will first discuss shallow ensemble approaches for obtaining
  and propagating calibrated uncertainty estimates with minimal computation
 al overhead. [1]\nI will then show how explicit probabilistic force-traini
 ng objectives can improve the calibration of force uncertainty estimates, 
 and how fine-tuning protocols can recover high-quality force uncertainty e
 stimates at substantially reduced training cost. [2]\n\nFinally, I will il
 lustrate how these methods can be used beyond single-point predictions by 
 propagating model uncertainty to thermodynamic observables such as melting
  points through statistical reweighting. [3] I will also introduce PET-UAF
 D an ensemble of universal MLIPs trained on multiple electronic-structure 
 references and calibrated against experimental data. [4] These approaches 
 provide uncertainty estimates not only relative to the reference theory, b
 ut also with respect to experimental observables.\n\n[1] doi.org/10.1088/2
 632-2153/ad594a\n[2] doi.org/10.1021/acs.jctc.6c00310\n[3] doi.org/10.1038
 /s41467-025-65662-7\n[4] doi.org/10.48550/arXiv.2604.24607\n\nDomain: Chem
 istry and Materials, Physics, Computational Methods and Applied Mathematic
 s\n\nSession Chairs: Michał Sanocki (Technical University of Munich); Ian 
 Störmer (Technical University of Munich); and Philip Loche (Technical Univ
 ersity of Munich, EPFL)\n\n
END:VEVENT
END:VCALENDAR