Abstract: Methods and systems for calibrating simulated measurement signals generated by a parametric measurement model are described herein. Regression on real measurement signals is performed using a parametric model. The residual fitting error between the real measurement signals and simulated measurement signals generated by the parametric model characterizes the error of the parametric model at each set of estimated values of the one or more floating parameters. Simulated measurement signals are generated by the parametric model at specified values of the floating parameters. A residual fitting error associated with the simulated measurement signals generated at the specified values of the floating parameters is derived from the residual fitting errors calculated by the regression on the real measurement signals. The simulated measurement signals are calibrated by adding the residual fitting error to the uncalibrated, simulated measurement signals.

Abstract: Methods and systems for selecting measurement locations on a wafer for subsequent detailed measurements employed to characterize the entire wafer are described herein. High throughput measurements are performed at a relatively large number of measurement sites on a wafer. The measurement signals are transformed to a new mathematical basis and reduced to a significantly smaller dimension in the new basis. A set of representative measurement sites is selected based on analyzing variation of the high throughput measurement signals. In some embodiments, the spectra are subdivided into a set of different groups. The spectra are grouped together to minimize variance within each group. Furthermore, a die location is selected that is representative of the variance exhibited by the die in each group. A spectrum of a measurement site and corresponding wafer location is selected to correspond most closely to the center point of each cluster.

Abstract: Machine learning techniques are used to predict values of fixed parameters when given reference values of critical parameters. For example, a neural network can be trained based on one or more critical parameters and a low-dimensional real-valued vector associated with a spectrum, such as a spectroscopic ellipsometry spectrum or a specular reflectance spectrum. Another neural network can map the low-dimensional real-valued vector. When using two neural networks, one neural network can be trained to map the spectra to the low-dimensional real-valued vector. Another neural network can be trained to predict the fixed parameter based on the critical parameters and the low-dimensional real-valued vector from the other neural network.

Abstract: Methods and systems for calibrating metrology tool offset values to match measurement results across a fleet of metrology tools are presented herein. The calibration of offset values is based on measurements of inline, production wafers and does not require the use of specially fabricated and characterized quality control (QC) wafers. In this manner, the entire process flow to calibrate metrology tool offset values is automated and fully integrated within a high volume semiconductor fabrication process flow. In a further aspect, the implementation of a new offset value is regulated by one or more predetermined control limit values. In another further aspect, the measured values of a parameter of interest are adjusted to compensate for the effects of measurement time on the wafer under measurement.

Abstract: Machine learning techniques are used to predict values of fixed parameters when given reference values of critical parameters. For example, a neural network can be trained based on one or more critical parameters and a low-dimensional real-valued vector associated with a spectrum, such as a spectroscopic ellipsometry spectrum or a specular reflectance spectrum. Another neural network can map the low-dimensional real-valued vector. When using two neural networks, one neural network can be trained to map the spectra to the low-dimensional real-valued vector. Another neural network can be trained to predict the fixed parameter based on the critical parameters and the low-dimensional real-valued vector from the other neural network.