Towards a data-driven paradigm for characterizing plastic anisotropy using principal components analysis and manifold learning

It is understood that anisotropic plastic behavior is strongly affected by the initial crystallographic texture and subsequent evolutions in the same during any kind of mechanical deformation. This means that macroscopic phenomenological modeling is inadequate since it leaves out any and all data related to texture, such as pole figures. While this shortcoming has long provided an argument for necessarily including micromechanical models (using crystal plasticity theory), the reality is that such simulations are time consuming and computationally expensive. In this paper, we present some developments that could be used in a novel model-free data-driven method for linking mechanical behavior directly with crystallographic texture, using manifold learning. Illustrated using numerically build data, our paradigm detects the intrinsic dimensionality of crystallographic texture data and represents it in feature space, as a function of micro-mechanical parameters. This would eventually enable local learning and interpolation to facilitate data-driven simulations.