Partial X-ray photoelectron spectroscopy to constructing neural network model of plasma etching surface

Abstract

A new model to control plasma processes was constructed by combining a backpropagation neural network (BPNN) with X-ray photoelectron spectroscopy (XPS). This technique was evaluated with the data collected during the etching of silicon carbide films at NF 3 inductively coupled plasma. The etching characteristics modeled were the etch rate and surface roughness measured by scanning electron microscope and atomic force microscopy, respectively. For systematic modeling, the etching was characterized by means of 2 4 full factorial experiment plus one center point. The BPNN was trained by the training data composed of XPS spectra corresponding to five major peaks. Prediction performance of trained BPNN model was tested with a test data set, not belonging to the training data. In modeling surface roughness, pure XPS model yielded an improvement of about 24% over PCA-XPS (99% data variance) model. For the etch rate data, the improvement was more than 40% irrespective of the data variances. These results indicate that non-reduced XPS spectra are more effective in constructing a prediction model. XPS models can be utilized to diagnose or control plasma processes.