Leveraging advanced data analytics, machine learning, and metrology models to enable critical dimension metrology solutions for advanced integrated circuit nodes

Abstract

Integrated circuit (IC) technology is changing in multiple ways: 193i to extreme ultraviolet exposure, planar to nonplanar device architecture, from single exposure lithography to multiple exposure and directed self-assembly (DSA) patterning, and so on. Critical dimension (CD) control requirement is becoming stringent and more exhaustive: CD and process windows are shrinking, three-sigma CD control of <2 nm is required in complex geometries, and a metrology uncertainty of <0.2 nm is required to achieve the target CD control for advanced IC nodes (e.g., 14, 10, and 7 nm nodes). There are fundamental capability and accuracy limits in all the metrology techniques that are detrimental to the success of advanced IC nodes. Reference or physical CD metrology is provided by atomic force microscopy (CD-AFM) and TEM while workhorse metrology is provided by CD-SEM, scatterometry, and model-based infrared reflectrometry (MBIR). Precision alone is not sufficient for moving forward. No single technique is sufficient to ensure the required accuracy of patterning. The accuracy of CD-AFM is ∼1 nm and the precision in TEM is poor due to limited statistics. CD scanning electron microscopy (CD-SEM), scatterometry, and MBIR need to be calibrated by reference measurements for ensuring the accuracy of patterned CDs and patterning models. There is a dire need for a measurement with <0.5 nm accuracy and the industry currently does not have that capability with inline measurements. Being aware of the capability gaps for various metrology techniques, we have employed data processing techniques and predictive data analytics, along with patterning simulation and metrology models and data integration techniques to selected applications demonstrating the potential solution and practicality of such an approach to enhance CD metrology accuracy. Data from multiple metrology techniques have been analyzed in multiple ways to extract information with associated uncertainties and integrated to extract the useful and more accurate CD and profile information of the structures. This paper presents the optimization of scatterometry and MBIR model calibration and the feasibility to extrapolate not only in design and process space but also from one process step to a previous process step. A well-calibrated scatterometry model or patterning simulation model can be used to accurately extrapolate and interpolate in the design and process space for lithography patterning where AFM is not capable of accurately measuring sub-40 nm trenches. The uncertainty associated with extrapolation can be large and needs to be minimized. We have made use of measurements from CD-SEM and CD-AFM, along with the patterning and scatterometry simulation models to estimate the uncertainty associated with extrapolation and the methods to reduce it. For the first time, we have reported the application of machine learning (artificial neural networks) to the resist shrinkage systematic phenomenon to accurately predict the preshrink CD based on supervised learning using the CD-AFM data. The study lays out various basic concepts, approaches, and protocols of multiple source data processing and integration for a hybrid metrology approach. Impacts of this study include more accurate metrology, patterning models, and better process controls for advanced IC nodes.