Computational particle beam microscopy
OA Version
Citation
Abstract
Particle beam microscopes have enabled researchers to understand material properties by imaging a specimen at near-atomic resolution. In a particle beam microscope, a beam of particles, such as helium ions or electrons, is incident on a sample, producing secondary electrons that reflect sample topography and chemical composition information. The generated secondary electrons are then collected by a detector to produce images of the sample surface. One limiting factor of these microscopes is the sample damage induced by the particle beam, especially for delicate biological samples. Standard techniques for sample damage reduction generally involve source dose reduction, in the form of reduced dwell time or reduced beam current. As a consequence, the image quality tends to be reduced as well.
In this thesis, we develop several techniques to improve image quality without changing the dose or to reduce the average dose without loss in image quality. We first propose a detailed probabilistic model for particle beam microscope measurements at the level of secondary electron counts. We introduce time resolution to this form of microscopy by dividing any given dose into repeated low-dose measurements. We then design different discrete-time time-resolved estimators that operate on these low-dose measurements. We demonstrate the improvement of the proposed methods through Monte Carlo simulation and real experimental data.
Next, we establish a rigorous framework for understanding the potential for the time-resolved approach. We first introduce idealized continuous-time abstractions of particle beam microscopy with direct electron detection. We show that the low-dose measurements are more informative per incident particle than high-dose measurements through Fisher information analysis. The analysis indicates that our method is able to mitigate the randomness of incident particles, i.e. source shot noise, by comparing the Fisher information of continuous-time time-resolved measurements with that of the unattainable oracle case with a deterministic incident particle beam. Novel estimators for use with continuous-time measurements are introduced and analyzed, and estimators for use with discrete-time measurements are analyzed and shown to approach their continuous-time counterparts as time resolution is increased. We show that our method is able to achieve performance improvement by a factor that is approximately equal to the secondary electron yield and that the improvement is uniform over the dose.
The previously proposed time-resolved estimation techniques focus on pixel-wise estimation methods that do not consider image-domain priors. We then explore methods that exploit inter-pixel correlation. We adopt the plug-and-play method to combine our understanding of the image acquisition model with sophisticated image priors, including a deep neural network denoiser as an implicit prior. We also seek deep learning-based denoising methods for particle beam micrographs using synthetic training data generated from the accurate generative model. Both methods surpass previous pixel-wise time-resolved methods by a large margin.
Lastly, we demonstrate that the time-resolved method is also valuable in feature detection, which is of great use in assessing semiconductor manufacturing accuracy. We explore how time-resolved methods can reduce the probability of error in distinguishing background pixels from feature pixels via error exponent analysis.