Speaker
Description
Authors: William Lavery, Sara Hamis
We adapt the classic physics-informed neural network architecture to uncover the spatiotemporal dynamics of cellular data. This approach trains a supervised learning framework while enforcing a generalised reaction-diffusion partial differential equation (PDE) in two dimensions. By representing the diffusion and reaction terms as multilayer perceptrons, the method learns their forms with minimal prior assumptions while simultaneously solving the governing PDE. We have evaluated the method with respect to two scenarios: (1) numerically simulated cell density data obtained by forward-solving the original PDE, and (2) agent-based simulations that converge to the continuous cell density case in the limit of infinite cells and infinitesimal time steps. The next step is to validate the framework using experimental in vitro cell data. The overall approach can be extended to other governing PDEs (e.g., incorporating additional terms on the right-hand side of the reaction-diffusion equation) and broader particle interaction models.
