Modular neural networks (MNNs) are investigated as a tool for modeling process behavior and fault detection and classification (FDC) using tool data in plasma etching. Principal component analysis (PCA) is initially employed to reduce the dimensionality of the voluminous multivariate tool data and to establish relationships between the acquired data and the process state. MNNs are subsequently used to identify anomalous process behavior. A gradient-based fuzzy C-means clustering algorithm is implemented to enhance MNN performance. MNNs for eleven individual steps of etch runs are trained with data acquired from baseline, control (acceptable), and perturbed (unacceptable) runs, and then tested with data not used for training. In the fault identification phase, a 0% of false alarm rate for the control runs is achieved.
To achieve timely and accurate fault detection, neural network-based time series modeling is applied to a reactive ion etching (RIE) process using an in-situ plasma sensor called optical emission spectroscopy (OES). OES is a wellestablished method of etch endpoint detection, but the large volume of data generated by this technique makes further analysis challenging. To alleviate this concern, principal component analysis (PCA) is adopted for dimensionality reduction of a voluminous OES data set, and the reduced data set is utilized for time series modeling and malfunction identification using neural networks. Four different RIE subsystems (RF power, chamber pressure, and two gas flow systems) were considered, and multiple degrees of potential faults were tested. The time series
neural networks (TSNNs) are trained to forecast future process conditions, and those forecasts are compared to established baselines. Satisfying results are achieved, demonstrating the potential of this technique for real-time fault detection and diagnosis.