中文摘要：

为了解决多输人多输出和产品质量不易在线测量的化学机械研磨（chemical mechanical polishing，CMP）过程R2R（run—to—run）控制的难题，提出了一种基于贝叶斯最小二乘支持向量机（Bayes least squares support vector machine，BIS-SVM）预测模型和克隆选择免疫多目标滚动优化算法的CMP过程多变量R2R预测控制器BSVMPR2R。由LS-SVM和贝叶斯证据框架（Bayes evidence framework，BEF）方法分别构建材料去除率（material removal rate，MRR）和晶圆内非均匀度（within—wafer nonuniformity，WIWNU）的BIS—SVM预测模型，解决了线性预测模型的失配问题；通过预测误差对后续批次过程扰动和漂移进行在线估计实现反馈校正，提高了预测模型精度；将多变量控制问题转化为基于2个预测模型的多目标优化问题，由克隆选择免疫多目标滚动优化算法求解最优控制律提高了控制精度。仿真结果表明，BSVMPR2R控制器的性能优于双指数加权移动平均（double exponential weighted movingaverage，dEWMA）多变量控制器，抑制了CMP过程扰动和漂移的影响，显著降低了MRR和WIWNU的均方根误差。

英文摘要：

In order to solve the R2R（run-to-run） control problem in chemical mechanical polishing （CMP） process with the features of multi-input ＆ multi-output and difficulty of product quality online measurement, a CMP process multivariable predictive R2R controller named BSVMPR2R based on Bayes least squares support vector machine （BLS-SVM） prediction model and the clonal selection immune multi-objective receding horizon optimization algorithm are proposed. LS-SVM and Bayes evidence framework（BEF） methods are used to build the BLS-SVM prediction models of material removal rate （MRR） and within-wafer nonuniformity （ WIWNU）, respectively, which solve the mismatch problem of linear prediction model. The prediction errors are used to online estimate the next run disturbances and drifts, achieve feedback correction and improve the prediction model accuracy. Muhivariable control problem is transformed into multi-objective optimization problem based on the two prediction models, and clonal selection immune multi-objective receding horizon optimization algorithm is used to solve the optimal control law, which improves the control precision. Simulation results illustrate that the performance of BSVMPR2R controller is superior to that of double exponential weighted moving average （dEWMA） multivariable controller, the effects of CMP process disturbances and drifts are restrained, and the RMSEs of MRR and WIWNU are reduced significantly.