In Situ Self-Healing of Beam Damaged MXenes under TEM

With the significant improvements and developments in transmission electron microscopes (TEM) techniques, accurate observation and identification is now more commonly adopted in angstrom-level precision, which reveals as sequential images about the real-space information. However, it still has difficulties to analyze the spatial defect distribution, phase transformation and dynamic phenomena since the TEM images can only record the appearances of the surface atoms while the out-of-plane information is severely missing. In this work, the healing process, characterized by the regeneration of crystalline structures within the e-beam burned holes, has been experimentally observed using TEM images. Modeling based on density functional theory and molecular dynamics was conducted to study the energy path as well as the stabilized phases of all stages before and after the healing of MXene. Furthermore, a first principle based deep learning algorithm is proposed and adopted to reconstruct the 3D transition states of a MXene healing process experimentally observed. Starting from the TEM image. Around hole atoms are assumed to have crystalline structure and set as the input along with the in-plane information of atoms that can be observed directly within the hole. Out-of-plane information of the latter group of atoms are then treated as the output of the algorithm along with the spatial distribution of all underlying atoms. The input and output information are reformed into graphical representations and treated as a graph reference problem. The trunk of this algorithm is set as a deep convolutional neural network incorporated with attention-based and non-attention-based components. To conform to the distance prediction, smooth potentials are constructed for Ti-Ti, Ti-C and C-C interactions based on first principle calculation. The minimum potential energy accumulation of all atoms within the hole is set as one of the convergence conditions of the algorithm along with the fitting of the training data sets. Our models are trained on defected MXene structures optimized with DFT, where different sizes and number of atoms are considered. The predictions of this algorithm fit well with the TEM images and help with the understanding of the healing processes observed in MXene materials.