| About this Abstract |
| Meeting |
TMS Specialty Congress 2025
|
| Symposium
|
The 7th International Congress on 3D Materials Science (3DMS 2025)
|
| Presentation Title |
Unlocking 3D Nanoparticle Shapes From 2D HRTEM images: Deep Learning for Classification and Denoising at Atomic Resolution |
| Author(s) |
Romain Moreau, Hakim Amara, Maxime Moreaud, Jaysen Nelayah, Adrien Moncomble, Christian Ricolleau, Damien Alloyeau, Riccardo Gatti |
| On-Site Speaker (Planned) |
Romain Moreau |
| Abstract Scope |
The study focuses on leveraging Deep Learning (DL) to enhance the analysis of nanoparticles (NPs) using High Resolution Transmission Electron Microscopy (HRTEM), down to atomic resolution. A Convolutional Neural Network (CNN) model was developed to automate the identification of 3D NP shapes from 2D images, overcoming challenges related to manual post-processing and noise. The model was trained on a carefully curated dataset of simulated HRTEM images, capturing various orientations, defocuses and NP sizes. Additionally, a UNet-type model was created for denoising and enhancing contrast between NPs and substrates, addressing limitations in image clarity due to microscope aberrations and environment. Even though this denoising model gives accurate predictions while processing simulated images, we have noticed that the inference on experimental images could sometimes fail. Therefore, we have added a segmentation model to improve the robustness of predictions on experimental micrographs. |
| Proceedings Inclusion? |
Undecided |