基于神经网络的CMP过程智能R2R预测控制

Neural Network Based Intelligent R2R Predictive Control to CMP Process

针对化学机械研磨(CMP)过程非线性、时变和不易在线测量的特性,提出了基于径向基函数(RBF)神经网络和微粒群(PSO)算法的CMP过程run-to-run(R2R)预测控制器NNPR2R。首先通过样本数据用减聚类算法和最小二乘法构建CMP过程的RBF神经网络预测模型,解决了复杂CMP过程难以建立精确数学模型的难题和提高了预测模型的精度。然后通过PSO算法滚动优化求取控制律,解决了基于导数的优化技术易于陷入局部最优的问题并提高了控制精度。仿真结果表明,CMP过程NNPR2R控制器的性能优于常规的EWMA方法,有效抑制了过程漂移和减小了不同批次间产品的差异,显著降低了材料去除率(MRR)的均方根误差。

For chemical mechanical polishing （CMP） process characteristics of nonlinear, time- varying and not being in-situ easily measured, CMP process neural network run-to-run （R2R） predictive controller named NNPR2R was proposed. Radial basis function （RBF） neural network predictive model about CMP was constructed by subtractive clustering algorithms and least squares method, thus the difficult problem of constructing accurate mathematical model of complicated CMP process was solved and the prediction accuracy was improved. Control law was calculated by particle swarm optimization （PSO） rolling optimization, therefore the problem that the derivative-based optimization technology was easy to fall into local optimum was solved and the control precision was improved. Simulation results illustrate that the performance of NNPR2R controller is better than that of EWMA, process drifts and shifts are suppressed significantly, variation in various runs of products is reduced, and the root mean squared error for material removal rate （MRR） is brought down substantially.