Mining structure-property linkage in nanoporous materials using an interpretative deep learning approach

Abstract

In this study, a 3D convolutional neural network (CNN) is designed and trained with simulated NP materials to investigate their structure-property relationship. It is demonstrated that our approach is able to predict the effective stiffness of NP structure with a wide range of microstructures while exhibiting high accuracy and low computational cost. It is well-accepted that CNNs are purely predictive ‘black-box’ models that are not easily interpretable and therefore are not practical in identifying the microstructure parameters that largely impact the mechanical properties. To tackle this limitation, we develop a unique interpretation method for providing meaningful insights into learning the structure and property linkage of nanoporous material based on our well-trained CNN model. Using this method, it is revealed that the CNN identifies relative density and the saddle-shaped regions of the ligaments as the two most important features that strongly impact stiffness. While the effect of relative density is already known from previous scaling relationships, verifying that the training process as well as the predicted values of stiffness from our CNN model are trustworthy, our curvature analysis candidates a new microstructural parameter with negative contribution on the effective stiffness of NP metals knowing as: saddle-shaped regions of the ligaments that could be included in future scaling relationships to improve their accuracy.