Accelerating Electron Diffraction Analysis with Machine Learning Inference on FPGAs

Abstract

Reflection High-Energy Electron Diffraction (RHEED) is used to study the crystal structure and growth of thin films in material science. The RHEED systems available today take several minutes to analyze the crystal’s structure from an image captured during the crystal's growth. This thesis captures the design and deployment of a high-speed camera pipeline with a LeNet5 neural network using the HLS4ML library on an FPGA. After deployment, a latency of 450-750 us was obtained depending on the input size to the neural network. It establishes the possibility of an RHEED system that can process images from a camera up to a rate of 1000 Hz. This system would enable users to gain more insights about their experiments in real-time and allow them to modify or stop the deposition process during the crystal growth to achieve desired characteristics.