Visualization of Deep Convolutional Neural Networks to Investigate Porous Nanocomposites for Electromagnetic Interference Shielding

Constructing porous structures in electromagnetic interference (EMI) shielding materials is a common strategy to decrease the secondary pollution caused by the reflection of electromagnetic waves (EMWs). However, the lack of direct analysis methods makes it difficult to fully understand the effect of porous structures on EMI, hindering EMI composites' development. Furthermore, while deep learning techniques, such as deep convolutional neural networks (DCNNs), have significantly impacted material science, their lack of interpretability limits their applications to property predictions and defect detection tasks. Until recently, advanced visualization techniques provided an approach to reveal the relevant information behind DCNNs' decisions. Inspired by it, a visual approach for porous EMI nanocomposite mechanism studies is proposed. This work combines DCNN visualization with experiments to investigate EMI porous nanocomposites. First, a rapid and straightforward salt-leaked cold-pressing powder sintering method is employed to prepare high-EMI CNTs/PVDF composites with various porosities and filler loadings. Notably, the solid sample with 30 wt % loading maintains an ultrahigh shielding effectiveness of 105 dB. The influence of porosity on the shielding mechanism is discussed macroscopically based on the prepared samples. To determine the shielding mechanism, a modified deep residual network (ResNet) is trained on a dataset of scanning electron microscopy (SEM) images of the samples. The Eigen-CAM visualization of the modified ResNet intuitively shows that the amount and depth of the pores impact the shielding mechanisms and that shallow pore structures contribute less to EMW absorption. This work is instructive for material mechanism studies. Besides, the visualization has the potential as a porous-like structure marking tool.