Deep neural network and meta-learning-based reactive sputtering with small data sample counts

Although several studies have focused on the application of deep-learning techniques in manufacturing processes, the lack of relevant datasets remains a major challenge. Hence, this paper presents a meta-learning approach to resolve the few-shot regression problem encountered in manufacturing applications. The proposed approach is based on data augmentation using conventional regression models and optimization-based meta-learning. The resulting deep neural network can be employed to optimize the reactive-sputtering process used in the fabrication of thin, compounded films of titanium and nitride. The performance of the proposed meta-learning approach is compared to the conventional regression models, including support vector regression, Bayesian ridge regression, and Gaussian process regression, which exhibit state-of-the-art performance for regression over small data sample counts. The proposed meta-learning approach outperformed the baseline regression models when tested by varying the training sample counts from 5 to 40, resulting in a decrease in the root mean square error to 74.6% of that observed in the conventional models to predict the stoichiometric ratio of the film produced during the reactive sputtering process. This is remarkable because regression performed over a small number of data is usually considered unsuitable for deep-learning approaches. Therefore, this approach exhibits considerable potential for usage in different manufacturing applications because of its capability to handle a range of dataset sizes.