DPG Phi
Verhandlungen
Verhandlungen
DPG

Dresden 2026 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

BP: Fachverband Biologische Physik

BP 14: Poster Session II

BP 14.61: Poster

Dienstag, 10. März 2026, 18:00–21:00, P2

PINNs based inference in reaction-diffusion systems — •Lukas Pöschl1,2 and Vasily Zaburdaev1,21Friedrich-Alexander Universität Erlangen-Nürnberg — 2Max Planck Zentrum für Physik und Medizin

Physics-informed neural networks (PINNs) promise to serve as universal function approximators that embed governing equations and biophysical constraints to operate under noisy, low data conditions. However, PINNs exhibit various training pathologies that limit broad practical application beyond simple benchmarking problems. We identify situations in biophysical models where PINNs fail, including stiff kinetics, sharp gradients and chaotic dynamics and map these to corresponding failure modes in the standard PINN formulation, together with the respective mitigations. We assess this improved formulation on four problems with synthetically generated experimental data. These tasks include the classical forward and inverse problems for parameter estimation and inference of experimentally inacessible species. Furthermore, we address the issue of discovering chaotic and pattern forming dynamics and thus the optimization of experimental parameters to explore these biophysically relevant regimes. Across selected problems, the modified PINNs suggest promising performance in handling the tested systems and the ability to operate with sparse measurements and noisy data.

Keywords: PINNs; Reaction-Diffusion; Neural Networks; Machine Learning; Inference

100% | Mobil-Ansicht | English Version | Kontakt/Impressum/Datenschutz
DPG-Physik > DPG-Verhandlungen > 2026 > Dresden