BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//modelica.simtek.cc//amfc2026//speaker//QB7AC7
BEGIN:VTIMEZONE
TZID:CST
BEGIN:STANDARD
DTSTART:20000101T000000
RRULE:FREQ=YEARLY;BYMONTH=1
TZNAME:CST
TZOFFSETFROM:+0800
TZOFFSETTO:+0800
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-amfc2026-C9HZ8L@modelica.simtek.cc
DTSTART;TZID=CST:20260921T114000
DTEND;TZID=CST:20260921T120500
DESCRIPTION:Physics-Informed Neural Nets for Control (PINCs) extend traditi
 onal Physics-Informed Neural Networks (PINNs) by incorporating control inp
 uts and variable initial conditions\, therefore enabling autoregressive lo
 ng-horizon simulation by chaining short-horizon predictions\, and making t
 hem suitable for control-oriented applications. However\, the existing lit
 erature commonly adopts PINC training procedures where the governing dynam
 ics are embedded in an explicit analytical form\, whereas in many engineer
 ing workflows\, the system models are often available only as Functional M
 ock-up Units (FMUs). This paper proposes an FMU-based training framework f
 or PINC that uses an FMI 3.0 Model Exchange (ME) FMU directly as the descr
 iption of system dynamics. The physics-informed residual is evaluated in t
 he forward pass by querying the FMU for state derivatives\, while gradient
  backpropagation is enabled through FMI 3.0 adjoint derivatives integrated
  into PyTorch via a custom autograd.Function. Case studies on a Van der Po
 l oscillator and on a mechanical system show that the proposed framework c
 an train PINCs without manually implementing the analytical governing equa
 tions and can achieve long-horizon simulation performance comparable to co
 nventional ODE-based training when the required FMI 3.0 features are avail
 able.
DTSTAMP:20261004T070612Z
LOCATION:FMI & MBSE (R2001)
SUMMARY:FMU-based Training of Physics-Informed Neural Nets for Control - Yi
 ding Wang\, Luca Bascetta\, Gianni Ferretti
URL:https://modelica.simtek.cc/amfc2026/talk/C9HZ8L/
END:VEVENT
END:VCALENDAR