Bestimmung der Physikalischen Eigenschaften eines Physikalischen Systems durch ein Neuronales Graphnetzwerk zur Durchsetzung von Konservierungen Paarweiser Linearer und Winkelmomente

Details

Veröffentlichungs-Nr.

EP4764895

Anmeldetag

18. Dezember 2024

Veröffentlichung

24. Juni 2026

Rechtsraum

EP

Abstract

The invention is notably directed to a computer-implemented method of determining one or more physical properties of a physical system (10i, 10f, 10) represented by a graph neural network (GNN). The physical system is a multi-body dynamical system of components interacting through pairwise interactions, such as a 3D granular system. The GNN includes nodes and bidirectional edges that are mapped onto the components and the pairwise interactions, respectively. The method comprises: (i) accessing an initial state obtained (S121) in accordance with initial dynamics of the physical system; (ii) iteratively executing (S125) the GNN based on the initial state accessed to iteratively update (S1256) a virtual state of the physical system as represented by the GNN, whereby multiple iterations are performed, which respectively correspond to multiple time steps; and (iii) determining (S126) one or more physical properties of the physical system at one or more of the multiple time steps based on one or more virtual states of the physical system as updated (S1256) at one or more of the corresponding iterations, respectively. At each iteration, the virtual state is updated (S1256) through message passing steps, which iteratively update (S1252, S1256) representations of the nodes and the bidirectional edges, so as to enforce (S1255), for each bidirectional edge of the bidirectional edges, a conservation of both a linear momentum and an angular momentum of the pairwise interaction corresponding to said each bidirectional edge. The angular momentum is measured about a point of action of internal force of the corresponding pairwise interaction. The invention is further directed to characterization systems and computer program products.

Anmelder

Firma

ECOLE POLYTECHNIQUE FEDERALE DE LAUSANNE (EPFL)

Land

🇨🇭 Schweiz