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:20260421T090514Z
LOCATION:Plenary Room (Bldg. 6 - 001)
DTSTART;TZID=Europe/Stockholm:20260629T193700
DTEND;TZID=Europe/Stockholm:20260629T193800
UID:submissions.pasc-conference.org_PASC26_sess124_pos111@linklings.com
SUMMARY:Generalization of Long-Range Machine Learning Potentials in Comple
 x Chemical Spaces
DESCRIPTION:Michał Sanocki (Technical University of Munich)\n\nThe vastnes
 s of chemical space makes generalization a fundamental challenge for machi
 ne learning interatomic potentials (MLIPs). Although MLIPs enable near–qua
 ntum-accuracy atomistic simulations at greatly reduced computational cost,
  their practical reliability is often limited by poor transferability to o
 ut-of-distribution systems.  Here, we systematically assess how explicit l
 ong-range modeling influences both accuracy and transferability of MLIPs a
 cross diverse chemical spaces, using metal–organic frameworks as a stringe
 nt test case. We benchmark Allegro, MACE, and DimeNet++ architectures comb
 ined with Euclidean Fast Attention (EFA), Charge Equilibration Layer for L
 ong-range Interactions (CELLI), and Latent Ewald Summation (LES). To rigor
 ously probe generalization, we introduce biased train–test splitting strat
 egies that enforce structural and chemical dissimilarity between training 
 and test sets.\n\nWe find that long-range corrections are essential for ro
 bust transferability, with physics-based models showing the most consisten
 t performance across out-of-distribution regimes. In contrast, neither CEL
 LI nor LES can reliably infer meaningful partial charges from energies and
  forces alone in complex systems without reference data. These results dem
 onstrate that out-of-distribution transferability, is a prerequisite for t
 rust in MLIPs and provide a general framework for diagnosing systematic fa
 ilures across chemical space.\n\n
END:VEVENT
END:VCALENDAR