We discuss several mechanistic approaches and experimental data for improving post-CMP cleaning of W plugs with TiN as barrier liner, and dielectric substrates SiO2 and Si3N4 for use at the 10 nm technology node (metal pitch of 40 nm). Particle charge in the low pH, W CMP slurries are usually positive, and the W surface is always negatively charged at pH >3. Therefore, a strong electrostatic attraction is expected to occur between the W surface and the residual particles during post-CMP cleaning. Two main approaches were chosen to break down the strong particles-W surface post-CMP electrostatic interactions, as well as particles dispersion and prevention of redeposition: (1) using cleaning additives able to adsorb at the W surface and reverse the W surface charge; (2) using organic additives to reverse the particle charge. The latter approach results in two strongly negative charged surfaces, which are able to repulse each other, and leads to the best cleaning.

Chemical Mechanical Planarization (CMP) is a key process for IC manufacturers. Tungsten (W) is an important material for connecting logic elements and for connecting memory elements, thanks to its excellent planarization, filling, mechanical and electromigration properties. W slurries are developed to remove high amounts of W via an abrasive, in conjunction with an oxidizer. After the polishing process, the planarized surface is contaminated with abrasive particles, organic residue, pad debris and metal cations through covalent or hydrogen-bonding, electrostatic and Van der Waals attractions. Post-CMP cleaning is required to remove all these contaminants while exhibiting low galvanic and chemical corrosion. Formulated cleans are needed to meet all these requirements. The performance of formulated W/TiN post-CMP cleaners for N10 and N7 has been evaluated. The newly developed formulations show a factor 4 reduction in metal surface contamination (from ~2 x 1012atoms/cm2to ~ 5 x 1011atoms/cm2), which is important to prevent dielectric breakdown. Very low particulate and organic residue defectivity was additionally confirmed by different surface characterization techniques: XPS, FTIR, contact angle/surface energy.

Image thresholding is one of the simplest but still effective image segmentation methods [1] that can find applications in many areas, especially in microscopy. Most microscopy images are grey scale images, from which thresholding produces binary images so that additional information such as phase volume, number of particles amongst others can be calculated. While there are several commercial software and freeware available for image segmentation, auto-thresholding and auto-analysis is preferred due to the advantages of faster throughput of analysis and consistency. This is especially useful in manufacturing environments where higher numbers of images are generated.While there are several auto-thresholding methods that are currently available, namely, histogram shapebased method [2], clustering-based method [3], entropy-based method [4], and local method [5], the work discussed utilizes a combination of un-supervised and supervised machine learning method to auto-threshold SEM images of chemical-mechanical-planarization (CMP) silica slurry particles on a substrate. The method works by calculating the area fraction of slurry particles from the SEM images generated. This measured parameter is then used as a baseline to evaluate the effectiveness of the post-CMP solutions in removing residue slurry particles. Figure 1 shows the auto-thresholding results using K-means clustering algorithm [6]. Herein, two clusters are selected: one corresponding to the slurry particles and the other corresponding to the background of the SEM images. Figures 1(a) and 1(b), indicate the effectiveness of the algorithm for SEM images with particle area fraction larger than 0.1%. However, in figures 1(c) and 1(d), the approach does not seem to work well due to the smaller area fractions < 0.1%, figure 1(e) and 1(f). Similar problems occur with some other auto-thresholding method as well.To circumvent the challenges encountered, a novel approach is proposed which uniquely combines the un-supervised and supervised learning method. Since the un-supervised K-means clustering algorithm works well with particles area fraction > 0.1%, these images along with the K-means clustering thresholding results can be used as training data for supervised learning algorithms and the trained model can be applied to auto-threshold images with much less slurry particles. This method is implemented in Python with Sklearn [6] and Skimage [7] libraries. The results are shown in Figure 2. Figure 2(a) shows good linear relationship between area fractions measured from manual and automatic thresholding. Figure 2(b) and 2(c) are example original and segmented binary images with particle area fraction < 0.1% using this method. This method can be applied in atomic force microscopy (AFM) images as well where the surface particles need to be subtracted from images for surface roughness calculations and the results will be discussed as well.