image: a, Working principle of the demonstrated method. The samples-under-test were multilayer semiconductor devices with alternating layers of oxide (SiO2) and nitride (Si3N4) on a silicon substrate. To obtain the spectroscopic data, commercial ellipsometers and reflectometers installed in the semiconductor production lines were used. For the machine learning model, measured spectral data and each layer thickness were used as input and output, respectively. b, Thickness prediction results for the 23 test samples. The predicted thickness (red circles) matches well with the actual thickness (blue triangles), regardless of the material or layer position, with an average prediction RMSE of approximately 1.6 Å. c, Outlier device detection results. Seventeen normal samples and two outlier samples were prepared for the test. All of the normal and outlier samples are successfully classified. view more
Credit: by Hyunsoo Kwak, Sungyoon Ryu, Suil Cho, Junmo Kim, Yusin Yang, and Jungwon Kim
With the recent explosive demand for data storage, ranging from data centers to various smart and connected devices, the need for higher-capacity and more compact memory devices is constantly increasing. As a result, semiconductor devices are now moving from 2D to 3D. The 3D-NAND flash memory is the most commercially successful 3D semiconductor device today, and its demand for supporting our data-driven world is now growing exponentially.
The scaling law for 3D devices is achieved by stacking more and more semiconductor layers, well above 100 layers, in a more reliable way. As each layer thickness corresponds to the effective channel length, accurate characterization and control of layer-by-layer thickness is critical. To date, unfortunately, non-destructive, accurate measurement of each layer thickness of such hundreds-layers structure has not been possible, which sets a serious bottleneck in the future scaling of 3D devices.
In a new paper published in Light: Advanced Manufacturing, a team of engineers from Korea Advanced Institute of Science and Technology (KAIST) and Samsung Electronics Co. Ltd., led by Professor Jungwon Kim of KAIST, South Korea, has developed a non-destructive thickness characterization method by combining optical spectral measurements and machine learning. By exploiting the structural similarity between semiconductor multilayer stacks and dielectric multilayer mirrors, spectroscopic optical measurements, including ellipsometric and reflectance measurements, are employed. Machine learning is then used to extract the correlation between spectroscopic measurement data and multilayer thickness. For more than 200 layers of oxide and nitride multilayer stack, the thickness of each layer over the entire stack could be determined with an average of approximately 1.6 Å root-mean-square error.
In addition to the accurate determination of the multilayer thickness under normal fabrication conditions, which is helpful for controlling etching and deposition processes, the research team developed another machine learning model that can detect outliers when layer thicknesses significantly vary from the design target. It used a large number of simulated spectral data for more effective and economical training, and could successfully detect the faulty devices and the exact erroneous layer location in the device.
"The machine learning approach is useful for eliminating measurement-related issues," said Hyunsoo Kwak, a doctoral student at KAIST and first author of the study. "By using noise-injected spectral data as input to the machine learning algorithm, we can eliminate various errors from measurement instruments and changes in material properties under different fabrication conditions," he added.
"This method can be readily applied for the total inspection of various 3D semiconductor devices," said Professor Kim, "which is exemplified by the fact that all the data used in this work were obtained in commercial 3D NAND manufacturing lines of Samsung Electronics."
###