Using neural networks with a linear output neuron to model plasma etch processes

A backpropagation neural network (BPNN) has been applied to model various plasma processes. As a neuron activation function, the BPNN typically uses either a bipolar or unipolar sigmoid function in both hidden and output layers. In this study, a BPNN with a linear function in the output layer and a bipolar sigmoid function in the hidden layer is introduced and applied to model a plasma etch process. The gradients related to the functions were experimentally optimized. The BPNN with a linear function in the output layer demonstrated an improved prediction over the BPNNs with other function combinations. By optimizing the gradients, its prediction accuracy was further significantly increased as compared to nonoptimized models. The process modeled is a magnetically enhanced reactive ion etch (MERIE) process and etch responses modeled include an etch rate, Al(Si) selectivity to photoresist, anisotropy and bias in critical dimension. The process was characterized by a 2^6$fractional factorial experiment