Speaker: Ivo F. Sbalzarini (TU Dresden)
Group Leaderwww.mpi-cbg.de
Host: Smita Krishnaswamy
“Data-driven inference and physics-informed machine learning of active material models”
Abstract:
Living materials, like cells and tissues, are mechanically active. They are able to move and deform by themselves, generating internal mechanical stresses by consuming a chemical fuel. Active material models describe the spatiotemporal dynamics of such non-equilibrium materials. They are key to understanding self-organization, e.g., in biological morphogenesis. Different descriptions of active materials exist, from interacting stochastic particles to mean-field partial differential equations. It is often unclear which description is sufficient to explain a given phenomenon, and many are hard to numerically solve or simulate. In this talk, I will show how stability-guided model selection enables noise-robust inference of minimal partial differential equation models of active materials from data with guaranteed physical consistency. I will also exploit the connection between numerical analysis and machine learning to remedy training pathologies in physics-informed neural networks for multi-scale active phenomena, such as active turbulence. These developments present exciting opportunities in applications from spatial biology, which are dominated by nonlinear processes in space and time with often unknown physics. This is showcased in our work on understanding biological tissue morphogenesis as a self-organized mechano-chemical process.