Machine learning techniques applied for the detection of nanoparticles on surfaces using coherent Fourier scatterometry

Abstract

We present an efficient machine learning framework for detection and classification of nanoparticles on surfaces that are detected in the far-field with coherent Fourier scatterometry (CFS). We study silicon wafers contaminated with spherical polystyrene (PSL) nanoparticles (with diameters down to λ/8). Starting from the raw data, the proposed framework does the pre-processing and particle search. Further, the unsupervised clustering algorithms, such as K-means and DBSCAN, are customized to be used to define the groups of signals that are attributed to a single scatterer. Finally, the particle count versus particle size histogram is generated. The challenging cases of the high density of scatterers, noise and drift in the dataset are treated. We take advantage of the prior information on the size of the scatterers to minimize the false-detections and as a consequence, provide higher discrimination ability and more accurate particle counting. Numerical and real experiments are conducted to demonstrate the performance of the proposed search and cluster-assessment techniques. Our results illustrate that the proposed algorithm can detect surface contaminants correctly and effectively.

Convolutional neural network applied for nanoparticle classification using coherent scatterometry data

“…Based on a total of 1302 experimental images rather than synthetic data, with a simple CNN with two convolutional layers and batch normalization, we demonstrated 95% accuracy on the test data. The proposed approach outperforms the existing algorithm for analysis of scattered maps, which is based on thresholding and search [29]. For relatively small particles, with diameters (classes) of 40 and 50 nm, the accuracy has been improved by a factor of two.…”

“…We compared the performance of our CNN classifier, pretrained on the five classes (Section 3.A) with a method that has been recently implemented by some of the authors of this paper [29]. We did this on new test sets of separately 40, 50, 60, and 80 nm particles, with roughly 40 cut-out images per class.…”

“…By providing the samples that were unseen during training, our results for the first time highlight the importance of the novelty detection to capture the confusing inputs in a contamination detection problem. The results show that the proposed method has superior capabilities compared to classification with the traditional search algorithm [29]. A vital issue for future research is to merge the proposed classification CNN with the network for automatic object detection.…”

“…Recently, 2D scattered maps generated by the CFS technique have been studied with line-by-line search algorithms resulting in histograms that rely on the features of characteristic electronic signals that are generated when a particle is detected. The cumbersome search routines are associated with the hyperparameters defined by the user for each specific input dataset, e.g., an expected amplitude and width of the characteristic signal, density, and number of points in a cluster of isolated particles, and the zeroing parameters for different iterations of the search [29]. There is still a need for discussion on minimization of user-algorithm interaction [30,31].…”

Identification of Model Particle Mixtures Using Machine-Learning-Assisted Laser Diffraction

“…The technique hereby proposed, in which relevant sample information is reliably extracted from a single and static far-field diffraction pattern, offers two main advantages over typical LD devices. Firstly, it allows for efficient particle identification by detecting the signal within a small angle (∼0.26 • ) with respect to the light propagation axis, thus effectively reducing the number of detectors needed for its implementation, as is also the case in recent micro-and nano-particle identification proposals that make use of Machine Learning (ML) algorithms [31,[39][40][41][42][43].…”