Use of neural network to in situ conditioning of semiconductor plasma processing equipment

A strategy for in situ fault detection of plasma equipment is presented. This was accomplished by combining optical emission spectroscopy (OES), neural network. The OES was used to collect fault spectra, inducted by radio frequency source and bias powers. A fault detection model was constructed by training the backpropagation neural network (BPNN) on the whole OES spectrum representing a normal plasma operation. The trained BPNN model was tested on the test data generated at other powers. The test result indicates that the BPNN model was capable of detecting abnormal plasma caused by a small variation of 1% in the source power. Due to less impact on the plasma properties, the BPNN model reacted sensitively only to a relatively large variation in the bias power. The performance of the BPNN model-based monitoring scheme was further compared to that of identified radicals. Much improved sensitivity of the BPNN model over them was clearly demonstrated for the source power variation. On the other hand, certain radicals yielded much improved detection for the bias power variation. This was manifest as plasma was monitored by means of the CUSUM control chart. In consequence, monitoring BPNN model-based prediction and identified radicals simultaneously is expected to provide an improved detection of plasma processing equipment.