SYSTEMS FOR PARSING MATERIAL PROPERTIES FROM WITHIN SHG SIGNALS
Semiconductor metrology systems based on directing radiation on a wafer, detecting second harmonic generated (SHG) radiation from the wafer and correlating the second harmonic generated (SHG) signal to one or more electrical properties of the wafer are disclosed. The disclosure also includes parsing the SHG signal to remove contribution to the SHG signal from one or more material properties of the sample such as thickness. Systems and methods described herein include machine learning methodologies to automatically classify obtained SHG signal
1 .- 31 . (Canceled)
32 . An automated method of characterizing electrical properties of a sample, the method comprising:
directing a beam of electro-magnetic radiation from an optical source to a wafer comprising said sample;
receiving a Second Harmonic Generation (SHG) signal from the sample;
detecting the SHG signal using an optical detector to generate a detected SHG signal; and
under the control of a hardware computing device, processing the detected signal to extract features from the detected SHG signal to estimate a desired characteristic of the sample from the processed detected SHG signal using machine learning.
33 . The automated method of claim 32 , wherein processing the received signal comprises processing the received signal using a physically derived machine learning model.
34 . The automated method of claim 33 , wherein the physically derived machine learning model is trained to remove from the SHG signal the effect of a characteristic of the sample different from the desired characteristic.
35 . The automated method of claim 34 , wherein the physically derived machine learning model comprises layer thickness, artifact detection, artifact identification and artifact quantification.
36 . The automated method of claim 32 , wherein extracting the features comprises extracting features using a machine learning kernel.
37 . The automated method of claim 32 , wherein estimating the desired characteristic of the sample comprises decoding the extracted features using a linear decoder trained based on an ensemble of SHG signals together with training labels using a model.
38 . The automated method of claim 37 , wherein the model comprises Logistic regression, L1-regularized logistic regression, Support vector machine (SVM), sparse support vector machine, Neural network and/or Deep learning.