Invited Lecture
Microstructural origin of work hardening in FCC Cu single crystals

The efficiency of discrete dislocation dynamics (DDD) simulations has been improved by several orders of magnitude through the development of the subcycling algorithm and its recent GPU implementation on the ParaDiS program. This enabled DDD simulations of single crystal Cu to exhibit a clear strain hardening rate using a single GPU for 2-3 weeks. Such simulations generated a large amount of data from which machine learning tools can be applied to extract physical understandings of the work hardening phenomena. For example, we found that the dislocation network formed under uniaxial [001] loading consists of links whose lengths satisfy an exponential distribution. DDD simulation data for ~100 different loading orientations show that the dislocation link length on each slip system satisfies the exponential distribution (characterized by density ρ_i and a dimensionless parameter ϕ_i). These data enabled us to train neural networks to predict the plastic strain rate and multiplicate rate on each slip system, for all loading orientations. A generalized Taylor relation is obtained that accurately predicts the flow stress given ρ_i and ϕ_i of every slip system.

We also report the first direct comparison between DDD simulations and experimental stress-strain measurements of bulk single crystals under identical loading conditions. Such a comparison is essential in establishing the fundamental premise of dislocation-based theory of crystal plasticity. The experiments are performed using the desktop Kolsky bar apparatus on single crystal Cu samples, with a nominal strain rate of ~ 10⁴ s^-1. The actual strain rate is measured as a function of time and applied to the DDD simulation cell. We will present the insights gained from the comparisons between DDD predictions and these experimental observations.