Abstract
Variations in the chemical composition or thermomechanical processing of metallic materials result in a vast landscape of possible microstructural morphologies. While this creates ample opportunities for alloys with improved mechanical performance, the design process is challenging due to the morphological complexity and the range of deformation micro-mechanisms involved. Empirically based statistical approaches are well suited to address some of these challenges. Previously, a model based on n-point statistics and principal component analysis was successfully used for predicting damage nucleation based on the microstructural morphology of dual-phase steels. Here, we give an in-depth exploration and analysis of such algorithms as applied to experimental data. First, we investigate model architecture by implementing and testing over 1000 model variants. This leads to improved predictive ability and several alternate architectures including one with a Fourier transformation instead of a n-point statistics transformation. Second, we analyze the noise, resolution and data quantity impact to give guidelines on the necessary data required to train a predictive model. Third, we investigate which morphological features are utilized by the model to make predictions by inputting artificiallyconstructed microstructures, inverting the model, examination of the basis image, and variation of model hyperparameters. It is found that grain boundary fluctuations less than 1 mu m are correlated to damage nucleation. This is consistent with observations in the literature on the effect of grain size and interconnected martensite regions on damage nucleation. Furthermore, it may give insights into the superior mechanical properties of alloys with bimodal grain size distributions. This demonstrates a unique approach of elucidating morphological effects from a single alloy by exploiting microstructural heterogeneity. It may be applied to other microstructures as well.
NSTL
Elsevier