Stefanos Papanikolaou1,Kamran Karimi1
NOMATEN CoE1
Stefanos Papanikolaou1,Kamran Karimi1
NOMATEN CoE1
There has been a long-standing notion that alloys’ micro/nano hardness is strongly tied to the underlying microstructure. Polycrystals, for instance, consist of multitudes of disoriented grains within a complex polycrystalline network that dictate the mechanical response (i.e. hardness) across nano and micro scales. Nevertheless, the nature of such inherent microstructure-property correlations remains elusive and debated to this date. Conventional physics-based frameworks such as the Hall–Petch relationship empirically describe grain boundary strengthening effects by a single (mean) grain size parameter ignoring inherent grain scale hierarchies and intricate topology of the grain boundary network at micro/nano-structural levels. Here we use a data-driven approach based on the state-of-the-art machine learning (ML) and Graph Neural Net (GNN) model to infer grain-scale hardness from the (initial) grain boundary microstructural information. We trained our GNN model using an Electron backscatter diffraction (EBSD) map containing local lattice orientation information which was supplemented by a nano-mechanical dataset corresponding to a nanoindented polycrystalline steel. The trained ML model was able to make robust predictions of the load-depth curves over a broad range of grain scales. We further investigated that the model performance strongly depends on some certain set of grain-level (topological) attributes such as individual grain size, number of (nearest) neighbors, and grain-grain misorientation angles. On top of mechanical properties (such as hardness), our model can accurately forecast intermittent displacement bursts (i.e. pop-ins) and associated size and statistical distributions solely based on microstructural metrics.