Modelling and optimization of surface roughness in chemical mechanical polishing based on DNN-GA

Abstract

The surface quality of Lithium Niobate (LiNbO3) has a significant influence on photonics and optoelectronics components. However, the prediction and optimization models of surface roughness were not accurate due to the random parameters. Hence, a prediction model of surface roughness is established based on an improved neural network, and a new method is proposed to optimize the arguments in chemical mechanical polishing (CMP) process. In the model, the structure of the neural network is optimized according to data features rather than choosing networks randomly. To improve the model accuracy, the optimal number of hidden layers is 4 and the corresponding amounts of nodes in each layer are 46, 34, 28, and 33, respectively. ReLU function is chosen as activation function. Subsequently, the relationship between surface roughness and processing parameters is built and the variation process of surface roughness is particularly considered. The accuracy and generalization ability of the model is verified by the experiments with the Mean Absolute Percentage Error (MAPE) of 6.42% and Root Mean Square Error (RMSE) of 0.403. Furthermore, the Genetic Algorithm (GA) method based on the selected model is applied to optimize processing arguments under the target surface roughness value of 0.3 nm. The accuracy of the fusion model is also validated by experiments with an error of 13.3%.