Implementing Supervised and Unsupervised Deep-Learning Methods to Predict Sputtering Plasma Features, a Step toward Digitizing Sputter Deposition of Thin Films

Abstract

The spectral emission data from the plasma glow of various sputtering targets containing indium oxide, zinc oxide, and tin oxide were obtained. The plasma was generated at various power and chamber pressures. These spectral data were then converted into two-dimensional arrays by implementing a basic array-reshaping technique and a more complex procedure utilizing an unsupervised deep-learning technique, known as the self-organizing-maps method. The two-dimensional images obtained from each single-emission spectrum of the plasma mimic an image that can then be used to train a convolutional neural network model capable of predicting certain plasma features, such as impurity levels in the sputtering target, working gas composition, plasma power, and chamber pressure during the machine operation. We show that our single-array-to-2D-array conversion technique, coupled with deep-learning techniques and computer vision, can achieve high predictive accuracy and can, therefore, be fundamental to the construction of a sputtering system’s digital twin.