Preface

Sputtering yield which is a number of atoms removed from material's surface per an incident ion, is one of the most important parameters in various studies such as plasma-material interactions. Measurements of the yields have been performed by many authors for various combinations of substrates and incident ions, and the published values are available from literature. This web page provides sputtering yields at normal incident angle, which are predicted by a machine learning technique called Gaussian process regression (GPR) using those published yields as a learning dataset.

If you are new to use this page, please see Notes first to consider whether the model here is reasonable for your parameter (energy and sputtering yield) range or not.

Prediction

Notes

• If you use the GPR prediction data above in your publication, please refer to this web site
  https://www-camt.eng.osaka-u.ac.jp/hamaguchi/SY/
  as well as the paper
  Kino et al.: Phys. Plasmas 28, 013504 (2021).
• References of learning data
  • S. Tachi and S. Okudaira: J. Vac. Sci. Technol. B 4(2), 459-467 (1986).
  • Y. Yamamura and H. Tawara: Atomic Data and Nuclear Data Tables 62, 149-253 (1996).
  • R. P. Doerner, D. G. Whyte, and D. M. Goebel: J. Appl. Phys. 93(9), 5816-5823 (2003).
  • R. D. Kolasinski, J. E. Polk, D. Goebel, and L. K. Johnson: J. Vac. Sci. Technol. A 25(2), 236-245 (2007).
  • T. Ito and K. Karahashi: Si sputtering yields by halogen (F, Cl, Br) ions (private communications)
  • S. Taira, T. Ito, and K. Karahashi: Mn sputtering yields by noble gas (Ne, Ar) ions (private communications)
• Dataset property

  Number of learning data points	:	5687
  Substrate element range	:	Be - U
  Incident ion element range	:	H - Pb

• Histograms of learning data
  If the parameter range you would like to use is completely off from these learning dataset range, predicted sputtering yields may be much less accurate.
• Prediction model development
  The prediction model has been developed based on Gaussian process regression (GPR), which is a non-parametric regression method. The code of our model is written in python with GPy for GPR. The RBF kernel is used as a kernel function, and maximum likelihood estimation method and 10-fold cross validation (CV) have been applied for tuning of hyperparameters.

  Descriptors of the created model consist of atomic numbers and masses of substrate and incident ion elements, and an incident ion energy, additionally melting temperature of a substrate is also one of them.

  The observed-predicted plot of the model for all learning data is shown here. The prediction ability in the range of the dataset is quite nice, such as that R² coefficient is 0.98 and the root mean square error (RMSE) is 0.15 for test of CV with the best parameters.

  Regression has been done for all dataset at once, so this model can predict sputtering yields even for combinations of no data by using similar conditions of descriptors.

• Versions

  Environment on development
  python	:	3.8.3
  GPy	:	1.9.9
  numpy	:	1.19.5
  prediction model	:	0.0.3 (developed on 2022/12/16)
  web page	:	0.2.3 (last updated on 2023/10/25)

• Recommended screen width: 1024px and more.

Contact us

Kazumasa Ikuse, Ph.D.
ikuse☆ppl.eng.osaka-u.ac.jp (Please replace "☆" in the e-mail address with "@".)

We would appreciate it if you could share your sputtering yield data of single-element substrates sputtered by single-element incident ions, not used as training data in this model yet. Your data will be added with proper references.