方维,王宇宇,宋志龙,吕冰海,赵文宏.基于多源数据融合的半导体晶片CMP抛光材料去除率预测[J].表面技术,2024,53(2):150-157, 167.
FANG Wei,WANG Yuyu,SONG Zhilong,LYU Binghai,ZHAO Wenhong.Prediction of CMP Polishing Material Removal Rate of Semiconductor Wafers Based on Multi-source Data Fusion[J].Surface Technology,2024,53(2):150-157, 167 |
| 基于多源数据融合的半导体晶片CMP抛光材料去除率预测 |
| Prediction of CMP Polishing Material Removal Rate of Semiconductor Wafers Based on Multi-source Data Fusion |
| 投稿时间：2023-01-01 修订日期：2023-07-05 |
| DOI：10.16490/j.cnki.issn.1001-3660.2024.02.014 |
| 中文关键词: 化学机械抛光 材料去除率 数据融合 深度自动编码器 BP神经网络预测 |
| 英文关键词:CMP MRR data fusion deep auto-encoder BP neural network prediction |
| 基金项目:国家自然科学基金(U20A20293);浙江省自然科学基金(LD22E050010) |
|
|
| Author | Institution |
| FANG Wei | College of Mechanical Engineering, Zhejiang University of Technology, Hangzhou 310023, China |
| WANG Yuyu | College of Mechanical Engineering, Zhejiang University of Technology, Hangzhou 310023, China |
| SONG Zhilong | College of Mechanical Engineering, Zhejiang University of Technology, Hangzhou 310023, China |
| LYU Binghai | College of Mechanical Engineering, Zhejiang University of Technology, Hangzhou 310023, China |
| ZHAO Wenhong | College of Mechanical Engineering, Zhejiang University of Technology, Hangzhou 310023, China |
|
| 摘要点击次数: |
| 全文下载次数: |
| 中文摘要: |
| 目的 对半导体晶片抛光过程中的工艺参数、耗材使用量、抛光垫状态参数等多源数据预处理后进行数据融合，建立材料去除率(MRR)预测模型，为实现半导体晶片抛光加工工艺的决策和处理奠定基础。方法 研究晶片抛光加工中的数据特点及数据融合需求，提取数据集中每个晶片加工过程中的统计特征并生成新数据集，同时引入邻域特征以应对晶片加工过程中动态因素对材料去除率的影响。提出基于深度自动编码器的多源数据融合及材料去除率预测方法。设计深度自动编码器参数，优化深度自动编码器的损失函数从而增强深度自动编码器对强相关性特征变量的重建。基于深度自动编码器进行多源传感器信号融合，降低数据维度。使用超参数搜索算法优化BP神经网络超参数，利用BP神经网络方法将融合后的数据进行半导体晶片抛光过程中的材料去除率预测。结果 采用PHM2016数据集对模型进行验证，均方误差MSE达到7.862，相关性R2达到91.2%。结论 基于多源数据的融合模型能有效预测MRR，可以对半导体晶片CMP工艺过程的智能决策与控制起到良好的辅助作用。 |
| 英文摘要: |
| The multi-source data, such as process parameters, consumption of consumables and polishing pad state parameters, were preprocessed for multi-source data fusion, and the material removal rate (MRR) prediction model was established to realize the decision of semiconductor wafer polishing process. The data characteristics and data fusion requirements of multiple sources of data in wafer polishing were analyzed. The statistical features of each wafer in the original data set were extracted to create a new data set. In order to address the impact of dynamic factors in wafer processing on the material removal rate, the neighborhood features were introduced. There were multiple sources and multiple modes of data in polishing of semiconductor wafers, such as the contact pressure between the wafer and the vehicle, the flow rate change of the polishing slurry, and the rotational speed of the polishing disc. There was a certain correlation between the data of these modalities, each modality could provide specific information for the other modal data. Processing the data of different modalities in the same way or simply concatenating all modal features could not guarantee the effectiveness of subsequent regression prediction. An efficient method to combine and reduce the dimensionality of these multiple sources and multiple modes of data was required. The material removal rate prediction method with deep automatic encoder for multi-source data fusion was proposed:the deep autoencoder parameters was designed and the loss function of the deep autoencoder was optimized to enhance the reconstruction of strongly correlated characteristic variables by the deep autoencoder. Considering the degree of data completeness of samples, completeness weights were introduced to reduce the impact of incomplete data on model training. Multi-source sensor signal fusion and data dimensionality reduction based on deep autoencoder were performed, and the fused data were fed into a BP neural network for MRR prediction during semiconductor wafer polishing (The hyperparameters of the BP neural network were optimized by the Hyperband hyperparameter search algorithm). The model was verified in the PHM2016 data for comparative analysis with other models (linear regression model, KNN regressor model and BP neural network model). The model in this study was based on an deep auto-encoder to achieve feature fusion and dimensionality reduction, and the processed data were employed in a BP neural network to fit the material removal rate of semiconductor wafers. Compared with the linear regression model (the mean square error MSE reached 31.186), the KNN regressor model (the mean square error MSE reached 26.17 and the correlation R2 reached 54.3%) and the BP neural network model (the mean square error MSE reached 11.445 and the correlation R2 reached 82.4%), the model in this study had a prediction value closer to the actual value in the three scenarios, with the mean square error MSE reached 7.862 and the correlation R2 reached 91.2%. The model based on multi-source data fusion can effectively predict MRR, which plays a benign auxiliary role in intelligent decision-making and control of the CMP process of semiconductor wafers. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
渝公网安备 50010702501715号