Deep learning-based microstructure reconstruction for characterizing the electromechanical property of nanomaterial networks

Various microscopic imaging techniques, such as optical microscopy, scanning electron microscopy, and X-ray microscopy, are essential for characterizing multi-scale material behavior and understanding structure-property relationships. However, establishing a large microstructural image library of materials through experimental imaging can be expensive, time-consuming, and inaccessible. There is an emerging trend in using computational techniques (e.g., traditional statistical approaches and machine learning) to more efficiently generate microstructural images to be utilized in material studies. This study focuses on developing deep learning approaches to reconstruct and artificially generate microstructures of strain-sensing nanomaterial networks based on microscope imagery. Here, the nanomaterial networks were fabricated using corona-enabled electrostatic printing, which leverages high voltage discharged in the air to attract feedstock material particles – specifically carbon nanotubes – onto substrates. The fabricated carbon nanotube networks were previously found to exhibit strain-sensing performance with high linearity. To better understand this electromechanical behavior, optical microscope images of carbon nanotube networks under different tensile strains were experimentally obtained and used for developing deep learning models. In this study, we employ a Generative Adversarial Network (GAN), incorporating skip-layer excitation and differentiable augmentation. The GAN model is conditioned on specific microstructure metrics, such as porosity, and uses the Learned Perceptual Image Patch Similarity as part of its loss function to ensure realistic image generation. The model is trained on 2048×2048 images, allowing it to accurately replicate complex microstructural patterns. Overall, this study establishes a technique based on deep learning for generating a large database for characterizing the electrical properties of nanomaterial networks at micro-scale in a low-cost and efficient manner.