Diagnosis of plasma processing equipment using neural network recognition of wavelet-filtered impedance matching

A plasma is a key means to deposit or etch thin films in manufacturing integrated circuits. To maintain device quality and throughput, plasma faults should be diagnosed accurately and timely. A method for plasma diagnosis is presented and this consisted of wavelets, neural network, and radio frequency impedance match monitoring system. Plasma faults were simulated with the variations in process parameters. Using the monitor system, impedance matching variables were collected, including two electrical match positions and one reflected power. The collected data were subsequently filtered using discrete wavelet transformation (DWT) and continuous wavelet transformation (CWT). A backpropagation neural network (BPNN) was used to capture causal relationships between fault symptoms and root causes. For either DWT- or CWT-filtered data, fault model was constructed using single BPNN and modular BPNN, composed of multiple single BPNNs. Model performance was examined in terms of the prediction accuracy and fault sensitivity. For single network model, DWT-based model demonstrated improved fault sensitivity compared to CWT-based model. For modular network model, both DWT and CWT were comparable in identifying plasma faults. Meanwhile, modular network model yielded more improved prediction accuracy and fault sensitivity than single network model.