Pregledni rad

https://doi.org/10.21857/mnlqgcpq4y

On physics informed neural networks - imposing constraints in infinitesimal strain continuum mechanics

Marko Čanađija orcid.org/0000-0001-6550-0258 ; Faculty of Engineering University of Rijeka
Martin Zlatić ; Faculty of Engineering University of Rijeka

Sažetak

The paper presents methods for implementing physical constraints from continuum mechanics in neural networks that serve as constitutive models. The focus is on infinitesimal strains in nonlinear elasticity. Basic constraints are presented with a brief overview of continuum mechanics, and then implemented in feed-forward neural networks. Among others, the topics of convexity, objectivity, thermodynamic consistency and normalization are addressed. Where appropriate, several possible techniques have been presented. Performance is demonstrated by developing a constitutive model for a single-walled carbon nanotube using physics informed neural networks

Ključne riječi

physics informed neural networks; physical constraints; continuum mechanics; convexity

Hrčak ID:

343849

URI

https://hrcak.srce.hr/343849

Datum izdavanja:

27.1.2026.

Podaci na drugim jezicima: hrvatski

Posjeta: 257 *