Statistical Methods for Process Control and Improvement in Advanced Manufacturing

项目负责人：Vijayan N. Nair，密歇根大学Mark H. Hansen，贝尔实验室，朗讯科技项目：先进制造过程控制与改进的统计方法摘要：该研究项目涉及基于空间数据的制造过程建模、监测、诊断和改进方法。该项目的范围超越了传统的Shewhart统计过程控制（SPC）范例，后者主要侧重于过程监控。研究的一个重要部分是使用过程和产品质量数据来开发故障诊断，并将其与过程改进的潜在问题联系起来。这些问题是在过程控制和改进的综合框架的背景下研究的。提出了一种利用缺陷聚类中的空间信息作为过程改进基础的总体策略。该方法由几个部分组成。首先，提出了对空间数据进行常规监测并检测具有显著聚类的目标的过程监测方法。采用马尔可夫随机场小尺度聚类模型对“受控”数据进行表征。获得故障诊断的统计方法（空间模式的特征），然后将模式与过程信息相关联以进行改进。研究了各种方法，包括大规模聚类的参数模型和基于分类的形式化方法。其他几个相关主题，包括从时间和空间过程的有序数据的建模和分析，以及从设计实验的空间数据的分析也进行了研究。这些方法是在集成电路（IC）制造中的晶圆图数据的特定背景下开发和研究的。半导体制造业是美国的主要制造业之一，因此从实际的角度来看，提高工艺和良率的统计方法显然很重要。然而，这里开发的研究问题和方法在本质上是相当通用的，并且对许多其他具有空间数据的制造过程具有普遍的兴趣，包括平板显示器，印刷电路板和汽车车身的制造和组装。这些先进制造业和高技术产业都具有以下特点。由于计算和数据捕获技术的进步，大量的过程数据和生产数据现在被定期收集。这些数据大多具有复杂的结构，以空间对象、图像等形式存在。与此同时，竞争激烈的市场压力将重点放在缩短产品开发周期上。此外，过程/产品设计者是在底层技术的可用主题知识的边界上操作的。在这项技术被充分理解之前，产品就已经开始生产和销售了。正如传统统计过程控制（SPC）范式中通常假设的那样，这些制造过程通常不“稳定”。因此，迫切需要统计方法，利用从过程和产品质量数据中获得的广泛信息，不仅用于过程监控，而且用于过程改进。该研究项目的结果将显著推进先进制造工艺的持续改进方法。

Proposal Number: DMS 9803281 PI: Vijayan N. Nair, University of Michigan Mark H. Hansen, Bell Labs, Lucent Technologies Project: Statistical Methods for Process Control and Improvement in Advanced Manufacturing Abstract: This research project deals with methods for modeling, monitoring, diagnosis, and improvement of manufacturing processes with spatial data. The scope of the project goes beyond the traditional Shewhart's paradigm for statistical process control (SPC) which focuses primarily on process monitoring. A significant part of the research is the use of in-process and product quality data to develop failure diagnostics and to relate these to potential problems for process improvement. These issues are studied in the context of an integrated framework for process control and improvement. An overall strategy is proposed for using the spatial information in defect clustering as the basis for process improvement. The methodology consists of several parts. First, process monitoring methods for routinely monitoring the spatial data and detecting objects with significant clustering are developed. A Markov random field model with small-scale clustering is used to characterize ``in-control'' data. Statistical methods for failure diagnosis (signatures of spatial patterns) are obtained and the patterns are then related to process information for improvement. Various approaches for doing this including parametric models for large-scale clustering and formal methods based on classification are studied. Several other related topics, including modeling and analysis of ordinal data from temporal and spatial processes and the analysis of spatial data from designed experiments are also studied. These methods are developed and studied in the specific context of wafer map data in integrated circuit (IC) fabrication. Semiconductor manufacturing is one of the key manufacturing industries in the US, and hence statistical methods for process and yield improvement are clearly important from a practical viewpoint. However, the research issues and methods developed here are quite generic in nature and are of general interest to many other manufacturing processes with spatial data, including flat panel displays, printed circuit boards, and the manufacture and assembly of auto-bodies. These advanced manufacturing and high-technology industries all share the following features. Massive amounts of in-process and production data are now being collected routinely, made possible by advances in computing and data capture technologies. Much of these data have complex structures, in the form of spatial objects, images and so on. At the same time, competitive market pressures are placing a lot of emphasis on reducing product development cycle time. Moreover, process/product designers are operating on the boundaries of available subject matter knowledge of the underlying technology. Products are being manufactured and marketed before the technology is well-understood. These manufacturing processes are often not ``stable'', as is commonly assumed in the traditional statistical process control (SPC) paradigm. Thus, there is a critical need for statistical methods that exploit the extensive information available from in-process and product quality data not only for process monitoring but also for process improvement. The results from this research project will significantly advance methodology for the continuous improvement of advanced manufacturing processes.