DINEs represent physical systems as differential-algebraic equations with Dirac-structure algebraic constraints, jointly identifying component interconnections and learning component characteristics so subsystems can be isolated or composed without retraining
The work proposes Dirac-interconnected neural elements (DINEs), a neural network model that represents a physical system as a differential-algebraic equation (DAE) whose algebraic constraints are given by a Dirac structure in kernel representation, thereby simultaneously identifying from data the interconnection among components as a Dirac structure and learning the characteristics of the components as neural networks, which keeps learned subsystems in unreduced form so they can be isolated or composed into a new system without retraining, and which can handle partially observable systems.