BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.com//16thmodelicafmiconference//talk//YTGZHW
BEGIN:VTIMEZONE
TZID:Europe/Zurich
BEGIN:DAYLIGHT
DTSTART:20240909T000000
TZNAME:CEST
TZOFFSETFROM:+0200
TZOFFSETTO:+0200
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20241027T030000
RDATE:20251026T030000
TZNAME:CET
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20250330T030000
RDATE:20260329T030000
TZNAME:CEST
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
SUMMARY:Efficient Training of Physics-enhanced Neural ODEs via Direct Coll
 ocation and Nonlinear Programming - Linus Langenkamp\, Bernhard Bachmann\,
  Philip Hannebohm
DTSTART;TZID=Europe/Zurich:20250909T104500
DTEND;TZID=Europe/Zurich:20250909T111000
DTSTAMP:20260816T003001Z
UID:pretalx-16thmodelicafmiconference-YTGZHW@pretalx.com
DESCRIPTION:We propose a novel approach for training Physics-enhanced Neur
 al ODEs (PeN-ODEs) by expressing the training process as a dynamic optimiz
 ation problem. The full model\, including neural components\, is discretiz
 ed using a high-order implicit Runge-Kutta method with flipped Legendre-Ga
 uss-Radau points\, resulting in a large-scale nonlinear program (NLP) effi
 ciently solved by state-of-the-art NLP solvers such as Ipopt. This formula
 tion enables simultaneous optimization of network parameters and state tra
 jectories\, addressing key limitations of ODE solver-based training in ter
 ms of stability\, runtime\, and accuracy. Extending on a recent direct col
 location-based method for Neural ODEs\, we generalize to PeN-ODEs\, incorp
 orate physical constraints\, and present a custom\, parallelized\, open-so
 urce implementation. Benchmarks on a Quarter Vehicle Model and a Van-der-P
 ol oscillator demonstrate superior accuracy\, speed\, generalization with 
 smaller networks compared to other training techniques. We also outline a 
 planned integration into OpenModelica to enable accessible training of Neu
 ral DAEs.
LOCATION:Audi-Midi
URL:https://pretalx.com/16thmodelicafmiconference/talk/YTGZHW/
END:VEVENT
END:VCALENDAR