SKM 2023 – wissenschaftliches Programm

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

CPP: Fachverband Chemische Physik und Polymerphysik

CPP 45: Emerging Topics in Chemical and Polymer Physics, New Instruments and Methods

CPP 45.7: Vortrag

Donnerstag, 30. März 2023, 11:00–11:15, ZEU 255

Solving inverse transport problems across irregular interfaces of sorptive porous media via physics-informed neural networks — •Alexandra Serebrennikova and Karin Zojer — Institute of Solid State Physics, TU Graz, Petersgasse 16, 8010, Graz, Austria

We show how state of the art extended physics-informed neural networks serve us to solve inverse transport problems with jump conditions across irregular interfaces. This approach reveals the material constants which govern reactive diffusion of organic volatiles migrating across an interface between porous sorptive packaging and food media if we provide experimental data and a transport model.

In such scenarios, associated differential equations (PDE) imply jumps not only in the solution, but also in the solution gradient across interfaces. The idea is to use multiple NN to construct the solution; each NN approximates the solution function of the PDE associated to a domain within the defined interfaces. The networks are coupled across interfaces such that the boundary conditions are satisfied.

As NNs are required to fit underlying physics by minimizing PDE residuals, they are inherently suited to solve inverse problems for the parameters involved in the equations. As further benefit, the discretized experimental data can be represented with a continuous function which offers a meshfree and compact surrogate model for the solution function.