Abstract:

Dislocation-mediated plasticity governs the mechanical behavior of structural alloys under deformation and extreme conditions. Accurate predictive modeling of strength, hardening, and failure that is needed to optimize and develop these alloys necessitates a multiscale framework that captures dislocation behavior from atomistic to continuum scales. This work presents a machine learning analysis of experimentally measured dislocation microstructure, then study the evolution and onset of dislocation microstructures through a physics-based, hierarchical approach that by extracting and utilizing physically meaningful descriptors of plastic deformation to inform mesoscale constitutive models. 

We first employ high-resolution differential-aperture X-ray microscopy (HR-DAXM) to measure lattice rotation and deviatoric elastic strain in deformed 304 stainless steel at small strain. Within single grains, the lattice rotation exhibits a multimodal distribution. To analyze this, we apply an unsupervised Cauchy mixture model, which resolves distinct rotational domains corresponding to subgrain formation. Mapping the data back into physical space reveals spatially contiguous regions of crystal rotation, validated by overlaying the dislocation density tensor to detect geometrically necessary boundaries (GND). 

To bridge the gap between experimentally observed static snapshot of patterns and their dynamic evolution, we utilize vector-density continuum dislocation dynamics (V-CDD). This curl-based transport model captures the collective interactions of dislocation ensembles, which is data-driven from lower-scale simulations. We present a computational method using streamline analysis of the dislocation and velocity vector fields to yield metrics such as mobile segment lengths and dislocation free paths. These V-CDD derived metrics and statistics provide a mechanistic basis for the parameters and statistics used in mesoscale constitutive models, a significant improvement over phenomenological approaches where parameters are fitted to the very properties the model is intended to predict. 

Ongoing efforts focus on strengthening the connection between experimental observations and modeling results by expanding the suite of comparable microstructural metrics and simulating the experimental process itself through digital twins. This approach will enhance the validation of continuum models and provide a predictive platform for guiding both alloy development and experimental technique optimization.