Application of neural classification in ellipsometry for robust thin-film characterizations

Nowadays, ellipsometry is a widely used technique for thin-film analyses among the existing methods. It offers a rapid, accurate and non-destructive control. The main task in this technique remains in the inverse problem which goal is to extract the interesting characteristics of the sample from the ellipsometric measurement. This is a purely mathematical step and the common algorithms used to achieve it are based on the gradient method. These latter proceed iteratively and hence require a well-chosen initial solution in order to converge towards a global minimum corresponding to the real physical solution. The main limitation of these algorithms is the risk to slip into a local minimum, leading to an erroneous solution. In this paper, we propose an original method based on neural classification in order to give a first estimation about the location of the solution in the multi-parameter space. This operation generally takes place before the characterization step itself. In this work, the method is implemented experimentally to estimate the thickness range of photoresist thin films deposited on a glass substrate.