BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Calibrated Physics-Informed Uncertainty Quantification - Vignesh G
 opakumar (University College London)
DTSTART:20250606T083000Z
DTEND:20250606T085000Z
UID:TALK230827@talks.cam.ac.uk
DESCRIPTION:Neural PDEs offer efficient alternatives to computationally ex
 pensive numerical PDE solvers for simulating complex physical systems. How
 ever\, their lack of robust uncertainty quantification (UQ) limits deploym
 ent in critical applications. We introduce a model-agnostic\, physics-info
 rmed conformal prediction (CP) framework that provides guaranteed uncertai
 nty estimates without requiring labelled data. By utilising a physics-base
 d approach\, we are able to quantify and calibrate the model's inconsisten
 cies with the PDE rather than the uncertainty arising from the data. Our a
 pproach uses convolutional layers as finite-difference stencils and levera
 ges physics residual errors as nonconformity scores\, enabling data-free U
 Q with marginal and joint coverage guarantees across prediction domains fo
 r a range of complex PDEs. We further validate the efficacy of our method 
 on neural PDE models for plasma modelling and shot design in fusion reacto
 rs.
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR