Reducing Thickness Deviation of W-Shaped Structures in Manufacturing DRAM Products Using RSM and ANN_GA

Thickness deviation on a W-shaped structure of word lines in dynamic random access memories (DRAMs) is detrimental to the DRAM manufacturing's yield rate. The case company has suffered from a low yield rate in their DRAM manufacturing due to large thickness deviations on the W-shaped structures. This article proposed a new approach to find an appropriate setting of manufacturing control factors' values to decrease the thickness deviation. In the proposed method, we first used the fractional factorial design to select important control factors related to the thickness deviation. We then used the gradient descent method to find a region of control factors that contained a second-order solution to the problem and adopted the response surface method (RSM) to find the second-order solution. However, the second-order RSM was limited by its second-order relationship between the input control factors and the output thickness deviation. To further reduce the thickness deviation, we used an artificial neural network (ANN) to predict the thickness deviation given a set of control factors' input values. We then used a genetic algorithm (GA) to find a better solution to the problem with the predicted thickness deviations by the ANN. The confirmation experiment showed that the GA had found a better solution than the RSM method. With the proposed GA method, the case company has successfully reduced the thickness deviation from 45.0 to 12.9 Å, saving U.S. $205 000 per year on their DRAM manufacturing cost.