Mimicking complex dislocation dynamics by interaction networks

TL;DR

This paper explores using Interaction Networks, an AI approach, to replicate the complex dislocation dynamics observed in two-dimensional models, focusing on individual and collective velocities during creep.

Contribution

It demonstrates that Interaction Networks can effectively learn and reproduce the complex dynamics of dislocation systems from creep data, including interaction kernels.

Findings

Interaction Networks accurately reproduce dislocation velocities

The method captures both individual and collective dynamics

Discussion on the quality of interaction kernel learning

Abstract

Two-dimensional discrete dislocation models exhibit complex dynamics in relaxation and under external loading. This is manifested both in the time-dependent velocities of individual dislocations and in the ensemble response, the strain rate. Here we study how well this complexity may be reproduced using so-called Interaction Networks, an Artificial Intelligence method for learning the dynamics of complex interacting systems. We test how to learn such networks using creep data, and show results on reproducing individual and collective dislocation velocities. The quality of reproducing the interaction kernel is discussed.