Speaker
Description
Field-Programmable Gate Arrays (FPGAs) are increasingly becoming pivotal in the advancement of artificial intelligence (AI) and deep learning applications. Their unique architecture allows for customizable hardware acceleration, which is instrumental in handling the intensive computational demands of modern AI algorithms.
Transmission Electron Microscopy (TEM) provides exceptional high-resolution imaging capabilities essential for exploring materials at the atomic scale. However, real-time analysis of TEM data poses significant computational challenges due to the sheer volume and complexity of the generated images. We present a deep learning approach for object detection that accurately locates electron hits within high-resolution TEM images. Our model is trained on a curated dataset collected from TEM experiments, focusing on images containing up to three electron hits. Despite the initial limitation in electron count per image, the model demonstrates robust performance in accurately identifying and localizing electron events.
To bridge the gap between high computational demands and real-time processing requirements, we deploy the trained TensorFlow model onto Field-Programmable Gate Arrays (FPGAs). This deployment leverages the parallel processing capabilities of FPGAs, significantly accelerating inference times and enabling on-the-fly data analysis during TEM operations. The integration of our deep learning model with FPGA hardware showcases a scalable solution for real-time electron hit detection, potentially extending to images with higher electron counts in future work.
Our approach not only enhances the efficiency of TEM data analysis but also opens avenues for dynamic experimentation where immediate feedback is crucial. This fusion of high-resolution data acquisition with accelerated deep learning inference sets a new precedent for real-time computational microscopy.
