Reliability of Deep Learning Models for Scanning Electron Microscopy Analysis

Chuen-Wun Pai, HUNG-WEI HSUEH, Shu-han Hsu

Submission Track: Paper Track (Tiny Paper)

Submission Category: Automated Material Characterization

Keywords: Scanning electron microscopy, deep learning, reliability, bit flip, fault injection

TL;DR: We examine how single-event upsets in weight memory degrade SEM analysis accuracy, highlighting critical vulnerabilities in deep learning models and the need for robust fault-tolerant solutions.

Abstract: Scanning electron microscopy (SEM) provides high-resolution nanoscale imaging of materials, critical for advanced materials characterization. Deep learning has accelerated SEM analysis by identifying nanoscale features with high accuracy. However, the reliability of these models remains insufficiently explored, particularly under soft errors, which may be triggered by radiation or transient disturbances. Such faults can lead to bit flips in both weight and neuron memory, leading to inference inaccuracy. We investigate a popular deep learning architecture, ResNet-50, by systematically injecting faults into their weight parameters for SEM characterization. Our analysis not only demonstrates how performance degrades under varying fault scenarios but also uncovers which layers and bit positions exhibit heightened vulnerability. These findings provide insights for developing robust fault-tolerant systems, including protective hardware measures and resilient model-training pipelines, thereby paving the way for more reliable deployment of SEM-based deep learning in industrial and research environments.

AI4Mat Journal Track: Yes

Submission Number: 6