A neural network plasma model of semiconductor manufacturing equipment

Plasma variables such as electron density and radicals provide valuable insight into the fundamental physics of rf discharge while contributing to equipment optimization and process simulation. However, modeling precisely plasma variables are difficult due to the extremely complex nature of physical dynamics within a plasma. In this paper, an inductively coupled plasma is modeled using neural networks. Variables modeled include typical electron density, electron temperature and plasma potential, whose data were collected with a Langmuir Probe. The plasma was characterized by a 2/sup 4/ full factorial experimental design with three center point replications. Factors varied in the design are a source power, pressure, chuck holder position, and Cl/sub 2/ flow rate. To test model suitability, experiments corresponding to face-centered arrays were additionally required. Resulting 27 experimental trials were conducted on an inductively-coupled plasma etch system. The R/sup 2/ of model fit are 0.981, 0.874, and 0.932 for electron density, temperature and plasma potential, respectively. Corresponding RMS predictive-errors are 0.716, 0.484 and 1.075, respectively. Neural network was found to accurately model plasma density and potential. Much lessened accuracy in temperature model can be ascribed to its training data least scattered.