Reducing Rework In CMP: An Enhanced Machine Learning-Based Hybrid Metrology Approach

By Vamsi Velidandla, John Hauck, Zhuo Chen, Joshua Frederick, and Zhihui Jiao

The semiconductor industry is constantly marching toward thinner films and complex geometries with smaller dimensions, as well as newer materials. The number of chemical mechanical planarization (CMP) steps has increased and, with it, a greater need for within-wafer uniformity and wafer-to-wafer control of the thin film layers.

Process engineers have typically adopted over-polishing and re-working as part of the standard operating procedure to reach the desired end point and required film uniformity on the wafers. This is because the current generation of integrated metrology toolsets are based on relatively simple optical techniques, such as reflectometry, and do not have the inherent high resolution offered by off-line techniques, such as ellipsometry, which are technically complicated and cost prohibitive to implement as integrated solutions on process tools.

When it comes to thin film residuals, the current steps in the CMP process — with both over-polish and rework steps playing prominent roles — are inefficient and result in lower yields.

The next generation of CMP tools from leading suppliers are targeting a 100% increase over current throughput, going from 80 to 100 wafers per hour to more than 200 wafers per hour. In order to achieve the expected increase in throughput, the time currently being spent on offline feedback and rework is simply not feasible as a part of a process control strategy.

If the true potential of these next-gen CMP tools is to be reached, these CMP tools must be installed with integrated metrology capable of measuring extremely thin films and accurately reporting the end point, thereby eliminating the need for offline metrology. With this requirement, integrated metrology modules will need additional input and data processing capability to measure sub-50Å residual films in a CMP environment.

A recent internal study between Onto Innovation and Micron indicates that a hybrid metrology approach can be effective in improving the measurement accuracy of thinner films. This approach combines measurements from different steps in the process and then uses that information to enhance the data analysis of the integrated metrology tool via machine learning. Such an approach provides accurate film thickness discrimination and enables the proper end point in CMP. This reduces the need for over-polishing and significantly reduces the rework rate.

In this study, an oxide CMP step (to stop on nitride) toward the end of the 3D NAND manufacturing process was selected (figure 1). The CMP experiment was designed with nominal residual oxide thickness targets of 100Å, 50Å, 25Å and 0Å.

Accurate ellipsometry-based measurements were carried out on the wafers prior to the deposition of the oxide layer. These measurements were used in a machine-learning model on the integrated metrology tool. This step enhanced the physical model that was used and improved the spectral analysis.

Fig. 1: 3D NAND Test structure selected for oxide residue study.

Post-CMP wafers were measured using advanced normal incidence reflectometry. The data was then analyzed, with and without machine learning. In both cases the pre-deposition information was feed-forward from the ellipsometer.

Data from the test indicated that a simple feed-forward hybrid metrology approach enables the discrimination of the four different film thickness on the post-CMP wafers, 100Å, 50Å, 25Å and 0Å (figure 2). By applying machine learning to enhance the data analysis, a hybrid metrology approach enabled discrimination and provided an even more accurate measurement of film thickness (figure 3).

Fig. 2: Simple feed-forward from ellipsometer to reflectometer enables discrimination of thin film residual on integrated metrology tools in oxide CMP.

Fig. 3: A comprehensive hybrid metrology with a machine learning approach results in better discrimination and accuracy of thin film residuals in oxide CMP.

Based on this study, machine learning-based hybrid metrology can be implemented in a fab using existing toolsets. In addition, it can be accomplished with minimal additional capital investment (figure 4) by combining pre-CMP information from high-resolution standalone metrology with post-CMP information from integrated metrology. Furthermore, this unique approach improved not just the data analysis, but the data acquisition recipes, as well.

Fig. 4: Machine learning based Hybrid Metrology approach to enhance the sensitivity of integrated thin film metrology on a CMP tool.

With an enhanced machine learning-based hybrid metrology approach, we can measure and differentiate sub-10Å top oxide from the rest of the film deck. This enables inline CMP process control with fast throughput and higher productivity and eventually helps improve the CMP process to reduce rework rate.

Vamsi Velidandla is the senior director of product marketing at Onto Innovation.

John Hauck is the senior metrology engineer at Micron Technology.

Zhuo Chen is a technical manager at Onto Innovation.

Joshua Frederick is a metrology engineer at Micron Technology.

Zhihui Jiao is an application scientist at Onto Innovation.