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:20260624T171344Z
LOCATION:Bldg. 6 - 003
DTSTART;TZID=Europe/Stockholm:20260701T150000
DTEND;TZID=Europe/Stockholm:20260701T153000
UID:submissions.pasc-conference.org_PASC26_sess117_msaSC103@linklings.com
SUMMARY:Beyond Benchmark Performance: Trustworthy AI, Generalization, and 
 Translational Impact in Scientific Computing
DESCRIPTION:Sally Ellingson (University of Kentucky)\n\nArtificial intelli
 gence (AI) and machine learning are increasingly used within high-performa
 nce computing workflows to accelerate scientific discovery, particularly i
 n data-intensive domains such as drug discovery. However, strong benchmark
  performance does not guarantee scientific or translational value. Models 
 that fail to generalize beyond narrowly defined datasets cannot generate n
 ew hypotheses, support real-world decision-making, or meaningfully advance
  science. In discovery-driven settings, such limitations undermine trust i
 n AI systems and limit their practical impact.\n\nThis work examines how d
 ataset bias, overfitting, and validation design can produce misleading per
 formance metrics that reward memorization rather than transferable learnin
 g. When evaluation protocols do not reflect real-world novelty, AI models 
 may appear effective while remaining incapable of extrapolation to new tar
 gets, conditions, or chemical space. Such models lack translational merit,
  regardless of computational scale or architectural complexity.\n\nWe argu
 e that trustworthy AI in scientific computing requires explicit quantifica
 tion of dataset bias, transparent evaluation practices, and validation str
 ategies aligned with discovery goals. By prioritizing generalization, repr
 oducibility, and honest performance assessment, the scientific computing c
 ommunity can develop AI systems that earn trust and contribute to genuine 
 scientific progress rather than reinforcing existing knowledge.\n\nDomain:
  Climate, Weather, and Earth Sciences, Applied Social Sciences and Humanit
 ies, Engineering, Computational Methods and Applied Mathematics\n\nSession
  Chair: Tobias Hodel (University of Bern, Switzerland)\n\n
END:VEVENT
END:VCALENDAR