Feedforward neural network methodology to characterize thin films by Electrostatic Force Microscopy.

The contribution of the present paper is in introducing a numerical method to improve the automatic characterization of thin films by increasing the effectiveness of numerical methods that take into account the macroscopic shape of the tip. To achieve this objective, we propose the combination of different feedforward neural networks architectures adapted to the specific requirements of the physical system under study. First, an Adaline architecture is redefined as a linear combination of Green functions obtained from the Laplace equation. The learning process is also redefined to accurately calculate the electrostatic charges inside the tip. We demonstrate that a complete training set for the characterization of thin films can be easily obtained by this methodology. The characterization of the sample is developed in a second stage where a multilayer perceptron is adapted to work efficiently in experimental conditions where some experimental data can be lost. We demonstrate that a very efficient strategy is to use evolutionary algorithms as training method. By the modulation of the fit function, we can improve the network performance in the characterization of thin films where some information is missing or altered by experimental noise due to the small tip-sample working distances. By doing so, we can discriminate the conductive properties of thin films from force curves that have been altered explicitly to simulate realistic experimental conditions.