Plasma Etching Endpointing by Monitoring Radio-Frequency Power Systems with an Artificial Neural Network

We demonstrate that the endpoint of an etching process can be determined by monitoring only values in the plasma tool's radio-frequency (RF) power system, the reflected power from the plasma source and wafer platen power supplies, capacitor values in the RF matchboxes, and the de bias. This is a systems approach that views the wafer as part of the electrical circuit. As the films on the wafer etch away, the effective impedance of both the wafer and the plasma changes. The state of the RF system is fed into the endpointing system, and the impedance change marks the endpoint. The artificial neural network was trained using endpoints called by an operator monitoring the etching with an in situ ellipsometer. The neural network was trained and tested on 51 wafers, and proved to be as accurate as the operator in calling endpoint. The particular example discussed in this paper finds the TiN endpoint during the etching of 0.25 mu m TiN-polysilicon gate stacks in a Lucas Labs helicon high-density plasma etcher. Problems encountered while developing the network are discussed, as are some of the limitations of neural networks.