Quercus Hernández - Learning Physics with Graphs
Quercus Hernándes · Jornadas de jóvenes investigadores del I3A · 2021
We take advantage of two inductive biases [1]: Geometric structure: We are able to exploit the geometric constraints of the system by performing computations over graphs based on the nodal connectivities.This enables the algorithm to learn more complex interactions, even in non-Euclidean manifolds [2]. Metriplectic structure:The time prediction is achieved via a thermodynamically consistent integrator based on the GENERIC formalism [3].It divides the system into conservative dynamics, related to Hamiltonian mechanics, and dissipative dynamics. METHODSComputational modelling has become a standard tool in a wide variety of scientific fields, in order to simulate reality phenomena and predict its future behaviour.The aim of this work is to learn physical simulators with the correct mathematical structure using graph-based deep learning.