Improved impedance matching speed with gradient descent for advanced RF plasma system

Abstract

Plasma processes are widely used in the fabrication of integrated circuits for the smaller, faster, more powerful circuits demanded by new technology. High-performance semiconductor equipment is required for dynamic random-access memory and three-dimensional NAND flash memory in semiconductor manufacturing. To enhance the speed of RF impedance matching, this paper proposes a method for calculating the load impedance of the plasma chamber. To calculate the target value of the variable components, we employ the input impedance of the impedance-matching unit and the complex impedances of the variable components. In addition, the internal parameters are automatically tuned through the gradient-descent method of machine learning, which improves the accuracy of impedance matching. The new matching method achieved a fast matching time and high efficiency of the matching trajectories. Moreover, it avoids the increase of the reflected power during the matching process, which causes plasma instability, and prohibits significant matching delay in a specific area, which disadvantages the conventional matching algorithm. These advantages are expected to significantly expand the development of new plasma processes from that of conventional impedance matching.