Novel genetic fitting algorithms and statistical error analysis methods for X-ray reflectivity analysis

Jouni Tiilikainen
Abstract & Cover

This thesis deals with x-ray reflectivity (XRR) analysis. XRR is a very accurate technique for the metrology of thin films but the analysis of measurements has been difficult thus limiting every day material research. In this thesis, novel genetic algorithms (GAs) for XRR curve fitting and statistical error analysis methods are developed. The XRR analysis utilizes very accurate Parratt's formalism combined with Nevot–Croce interface roughness. The analysis concentrates on the atomic layer deposited materials by using models mimicking their properties. The properties of GAs are studied using aluminium oxide/zinc oxide nanolaminate models. Models of aluminium oxide layers on silicon substrate are used in the case of the error analysis.
The demonstrated novel GAs are utilizing the rotation of coordinates during the crossover phase to reduce interparameter dependencies. The new basis is formed from the eigenvectors of Hessian and statistical covariance matrices. The crossover is performed in the rotated coordinates and the new combinations are transformed back to the original coordinates. It is shown that the coordinate rotation improves the convergence properties of GAs in complex XRR curve fitting problems and a statistical approach is more powerful than the Hessian matrix method. Furthermore, a GA using independent component analysis gives additional robustness to the curve fitting by utilizing a nonorthogonal linear transformation technique.
The interdependency of XRR parameters is studied using fitness landscapes. The fitness landscape analysis utilizes subspace projection of the original parameter space where the projection is done using an experimental model. The work reveals that the error in the determined mass density can compensate the error in surface roughness thus diminishing the accuracy of both of these parameters. This result is also verified later with other methods.
The effect of Poisson noise on the accuracy of XRR analysis is studied statistically. Thickness determination accuracy of an aluminium oxide layer is ±0.09 nm with 99% confidence in the studied case which represents the lower limit for the error. Here the analysis assumed a perfect fit to the measurement. The upper error is achieved by taking into account a nonideal fit by separating the effect of noise from the fitness value. In a case of the studied measurement, the determined thickness error is ±0.12 nm with 99% confidence.

Source of Information
FinALD40 exhibition material,
Helsinki University of Technology, Faculty of Electronics, Communications and Automation, Department of Micro and Nanosciences
(Espoo, Finland)
External Link
Read Thesis
linkedin invite