Learning non-Markovian physics from data - Archive ouverte HAL Access content directly
Journal Articles Journal of Computational Physics Year : 2020

Learning non-Markovian physics from data


We present a method for the data-driven learning of physical phenomena whose evolution in time depends on history terms. It is well known that a Mori-Zwanzig-type projection produces a description of the physical phenomena that depends on history, and also incorporates noise. If the data stream is sampled from the projected Mori-Zwanzig manifold, the description of the phenomenon will always depend on one or more unresolved variables, a priori unknown, and will also incorporate noise. The present work introduces a novel technique able to unveil the presence of such internal variables—although without giving it a precise physical meaning—and to minimize the inherent noise. The method is based upon a refinement of the scale at which the phenomenon is described by means of kernel-PCA techniques. By learning the metriplectic form of the evolution of the physics, the resulting approximation satisfies basic thermodynamic principles such as energy conservation and positive entropy production. Examples are provided that show the potential of the method in both discrete and continuum mechanics.
Fichier principal
Vignette du fichier
PIMM_JCP_2020_CHINESTA2.pdf (1.56 Mo) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

hal-03019897 , version 1 (23-11-2020)



David González, Francisco Chinesta, Elías Cueto. Learning non-Markovian physics from data. Journal of Computational Physics, 2020, pp.1-14. ⟨10.1016/j.jcp.2020.109982⟩. ⟨hal-03019897⟩
41 View
67 Download



Gmail Facebook X LinkedIn More