A Review of Data Mining Applications in Semiconductor Manufacturing

Background: E-beam metrologies, both critical dimension scanning electron microscope metrology and defect scan metrology, have been playing a very critical role in gating patterning quality. SEM images can provide rich visual information for engineers to do qualitative and quantitative analyses. However, the low e-beam metrology tool throughput makes it impossible to obtain SEM images for larger area. Monte Carlo-based SEM image simulations or other SEM image simulations require postlithography or postetch pattern three-dimensional structures as prerequisite, and the simulation speed is not sufficiently fast for full chip implementation. Aim: We aim to develop machine learning SEM models with sufficient accuracy and speed for full chip application in semiconductor manufacturing environment. Approach: We have proposed a virtual SEM metrology solution based on U-Net neural network with physics-based feature maps as model input. With information in aerial image space encoded properly, SEM images of both postlithography and postetch can be predicted accurately enough for practical applications using our proposed virtual SEM metrology models. Equipped with GPUs, the machine learning-based SEM image models are fast enough to make it possible to realize full chip SEM generation from post-OPC data. Results: Our machine learning SEM image models can predict SEM images with normalized cross correlation around 0.95 in reference to ground truth SEM images, each SEM image (512 × 512 in size) takes about 800 ms using single CPU, and the speed can be accelerated to about 10 ms with single GPU. Conclusions: Using U-Net structure and physics-based feature maps as model inputs, machine learning-based SEM image models can be developed. The models are sufficiently accurate and fast to find their applications in semiconductor manufacturing, and they can be used as independent model for OPC data verification or generate SEM images as reference for SEM defect scan metrology.

ABSTRACTThe evolution of the excimer laser from a research tool in the 1970's to the industrial models of the 1980's has opened up numerous applications in semiconductor manufacturing, materials processing, and biological and medical research. Present production-capable, computercontrolled excimer systems with handling equipment enabling efficient automation can now operate for up to 50 million shots before undergoing scheduled maintenance, with fluences remaining within ±10% in most cases. Thus, although many uses for the excimer laser are still researchoriented, production applications are rapidly emerging.

The what-if analysis process is essential in symbiotic simulation systems. It is responsible for creating a number of alternative what-if scenarios and evaluating their performance by means of simulation. Most applications use a reactive approach for triggering the what-if analysis process. In this paper we describe a preventive triggering approach which is based on the detection of a future critical condition in the forecast of a physical system. With decreasing probability of a critical condition, using preventive what-if analysis becomes undesirable. We introduce the notion of a G-value and explain how this metric can be used to decide whether or not to use preventive what-if analysis. In addition, we give an example for a possible application in semiconductor manufacturing.

A noncontact temperature measurement technique, diffraction-order analysis, based on monitoring the change in diffraction angle from a grating as a result of thermal expansion, is described and results for both Si and GaAs are presented. Two incident beams are used to provide a differential measurement, independent of wafer tilt. Image processing techniques are used to calculate the relative temperature in near real time from the optical signals. Good agreement between optical and the thermocouple temperature measurements is obtained, with an accuracy and precision of ±0.3 °C demonstrated over a 20–600 °C temperature range for a GaAs sample. Analysis of the effects of all six rigid-body motions of the wafer on the measurement is presented. The measurement is independent of all translational motions; rotational motions (pitch, roll, and yaw) can all be monitored with the same measurement scheme and the temperature measurement corrected for their effects. In many applications in semiconductor manufacturing, wafers are rotated to ensure uniformity. The diffraction-order analysis technique is demonstrated for a rotating wafer with a 3σ precision of 1.95 °C.

Perhydropolysilazane (PHPS) is an important polymeric precursor for producing silica-related inorganic materials through versatile solution-based processing, and has found broad applications in semiconductor manufacturing and microelectronic encapsulation. However, owing to its complex linear and branched architecture, high sensitivity to H2O and O2, and the inherently amorphous nature of the resulting silica or silicon oxynitride, the conversion mechanism of PHPS to SiOx/SiOxNy remains poorly understood. This knowledge gap significantly hampers the rational design and optimization of this promising methodology for the development of high-performance inorganic functional materials. Here, we developed a new Si/N/O/H ReaxFF parameter set capable of accurately describing the thermal conversion of PHPS under various O2/H2O atmospheres. ReaxFF molecular dynamics (MD) simulations reveal distinct atomistic pathways for O2 and H2O: O2 accelerates Si-O bond formation but promotes the growth of a SiOx surface layer that impedes further O2 diffusion, whereas H2O facilitates NH3 release and Si-N bond cleavage. Coordination analysis shows that Si-N tetrahedra are highly sensitive to the atmosphere, with higher O2 levels favoring Si atoms bonded to four N atoms. These findings demonstrate that the Si-O-N composition of PHPS-derived SiOx/SiOxNy can be tuned by controlling the atmosphere during thermal conversion, offering practical guidance for processing optimization and material performance enhancement.

Liquid particle counting (LPC) based upon laser light scattering technology continues to find wide applications in semiconductor manufacturing. Because defect levels in semiconductor devices are related to cleanliness of the processing chemicals, both users and vendors of photoresists and auxiliaries have expended efforts toward LPC testing programs. This paper describes Dynachem's approach to solving problems that have made vendor-user correlations difficult and have prevented setting of meaningful specifications. The paper also discusses how statistical process control can be applied to production of photoresists and developers. As demonstrated in the following discussions, the combination of statistical process control and particulate analysis is powerful and should help vendors and users move toward their goals of specification development.

To detect sustainable changes in the production processes, memory-type control charts are frequently utilized. This study is conducted to assess the performance of the Bayesian adaptive exponentially weighted moving average (AEWMA) control chart using ranked set sampling schemes following two different loss functions in the presence of a measurement error for posterior and posterior predictive distributions using conjugate priors. This study is based on the covariate model and multiple measurement methods in the presence of a measurement error (ME). The performance of the proposed Bayesian-AEWMA control chart with ME has been evaluated through the average run length and the standard deviation of the run length. Finally, a real-life application in semiconductor manufacturing was conducted to evaluate the effectiveness of the proposed Bayesian-AEWMA control chart with a measurement error based on different ranked set sampling schemes. The results demonstrate that the proposed control chart, in the presence of a measurement error, performed well in detecting out-of-control signals compared to the existing control chart. However, the median ranked set sampling scheme (MRSS) proved to be better than the other two schemes in the presence of a measurement error.

Commercial quadrupole mass spectrometer (QMS) residual gas analyzers became available in the late 1960s and have been popular in R&D laboratories, but have found limited use in semiconductor manufacturing. To sample at pressures above 10−5 Torr with ppm sensitivity or better (relative to the total process pressure), differential pumping is usually required. The newly available, small, high-pressure QMS sensors can operate as high as 10–20 mTorr without differential pumping, but provide somewhat lower mass resolution and partial pressure sensitivity than a standard QMS. Applications in semiconductor manufacturing include equipment monitoring, process monitoring, and effluent analysis. Equipment monitoring can include qualification after preventative maintenance, rate-of-rise tests, and leak identification and detection. Usually the burning question is, Why won’t the vacuum chamber pump down? Other uses that are not usually considered include the qualification of replaceable parts: sputter cathodes, electrodes, lamps, shields, etc. In process monitoring, the key question is, Is this process running normally? A manufacturing monitor can be useful simply by providing a comparison between a well-behaved high-yield process and a marginal or failing process. Examples are given for physical vapor deposition (sputtering) processes, chemical vapor deposition, and plasma etching. The effluent from chemical vapor deposition and plasma etch processes can be analyzed to measure the efficiency of process gas utilization or to monitor the efficacy of abatement methods used for the removal of global warming gases.

The use of a digital twin as an enabling technology for industry 4.0 provides control systems engineers with novel tools for modelling, designing, and controlling complex systems, providing a deep understanding of the physical asset based not only on its physics but also the real system’s response. It is particularly critical for uniformity temperature control applications, where providing a reasonable model of the system’s diffusion is always affected by the physical behavior of the system’s components required for heating, cooling, or power distribution. In this paper, a digital twin is used to represent a multivariable thermoelectric system employed for temperature uniformity distribution control with potential applications in semiconductor manufacturing. The modelling employs a five-step methodological framework consisting of the stages: target system definition, system description, multiphysics and data-driven simulation, behavioral matching, and implementation to represent the system’s temperature distribution accurately. The temperature distribution is measured using an infrared thermal camera to perform model behavioral matching on heating and cooling temperature uniformity applications. The obtained results indicated that using digital twins not only increases the accuracy of the system’s representation but can also provide the system with novel information that can be leveraged for the design and implementation of smart control systems.

IEEE Transactions on Semiconductor Manufacturing CALL FOR PAPERS for Special Issue on Process-Level Machine Learning Applications in Semiconductor Manufacturing

In the context of modern smart manufacturing, a patent and licensing strategy is formed as a control analytical algorithm for monitoring the technology's readiness for mass production and integrated marketing. New smart manufacturing solutions have a positive impact on the development of most production systems and equipment. Their efficiency is enhanced through vertical and horizontal integration. The proposed technical solutions are based on a method of coding and subsequent identification of the coding element. The method involves applying a special coating (or its technological equivalent) to an object and measuring its thickness. Matching the parameters with the specified code ensures positive identification, while mismatching leads to shutdown or blocking of the equipment or information consumer. This technology has been repeatedly proven in film thickness control applications in solar energy and semiconductor manufacturing. With the advent of multilayer optical discs and recording formats using blue lasers, its importance has increased. Coding at each recording level enables three-dimensional local encryption of information, making the technology particularly relevant for protecting classified and confidential data. In conclusion, it should be noted that new smart manufacturing technologies have a positive impact on the comprehensive development of virtually any production systems and equipment, as well as on their improvement through vertical and horizontal integration.

A significant amount of attention appears to be focusing on how to better collect, analyze, and disseminate information. In doing so, technology is commonly and increasingly looked upon as both a tool, and, in some cases, a substitute, for human resources. One such technology that is playing a prominent role in homeland security initiatives is data mining. Similar to the concept of homeland security, while data mining is widely mentioned in a growing number of bills, laws, reports, and other policy documents, an agreed upon definition or conceptualization of data mining appears to be generally lacking within the policy community (Relyea, 2002). While data mining initiatives are usually purported to provide insightful, carefully constructed analysis, at various times data mining itself is alternatively described as a technology, a process, and/or a productivity tool. In other words, data mining, or factual data analysis, or predictive analytics, as it also is sometimes referred to, means different things to different people. Regardless of which definition one prefers, a common theme is the ability to collect and combine, virtually if not physically, multiple data sources, for the purposes of analyzing the actions of individuals. In other words, there is an implicit belief in the power of information, suggesting a continuing trend in the growth of “dataveillance,” or the monitoring and collection of the data trails left by a person’s activities (Clarke, 1988). More importantly, it is clear that there are high expectations for data mining, or factual data analysis, being an effective tool. Data mining is not a new technology but its use is growing significantly in both the private and public sectors. Industries such as banking, insurance, medicine, and retailing commonly use data mining to reduce costs, enhance research, and increase sales. In the public sector, data mining applications initially were used as a means to detect fraud and waste, but have grown to also be used for purposes such as measuring and improving program performance. While not completely without controversy, these types of data mining applications have gained greater acceptance. However, some national defense/homeland security data mining applications represent a significant expansion in the quantity and scope of data to be analyzed. Moreover, due to their security-related nature, the details of these initiatives (e.g., data sources, analytical techniques, access and retention practices, etc.) are usually less transparent.

Data mining has become a major academic research area over the last ten years. However, in the transition from academic prototypes to commercial products there have been few successes with commercial data mining applications failing to make any significant impact in the marketplace. In this paper it is argued that most data mining applications concentrate on the model-building phase of the data mining process and rarely engage the user in other stages. The paper reviews the challenges of producing successful data mining applications and in particular the role of information visualisation. Visualisation in data mining tends to be used to present final results, rather than playing an important part throughout the entire process. We argue that an immersive environment may provide the user with a more suitable interface than is commonly offered. We present a virtual data mining environment which attempts to integrate a data mining application interface and information visualisation in a seamless manner, using the concept of 'liquid data'.

SMA Negeri 7 Kota Bengkulu is one of the State Senior High SchoolsState Senior High School located in Bengkulu City which has a commitment to in developing student potential. So far, data processing of student grades for each subjects have used e-report cards and the report card data can be exported to a Microsoft Excel file. exported into a Microsoft Excel file. However, the student grade data is data is often only stored as an archive without any further analysis to find out to find out information about patterns or trends in student achievement. At where this can help the school in providing guidance and optimizing the academic potential of each student at SMA Negeri 7. optimize the academic potential of each student at SMA Negeri 7 Bengkulu City based on the level of student achievement. Bengkulu City based on student achievement levels. Application of data mining in grouping data grades of SMA Negeri 7 Bengkulu City students using the K-Means Clustering Method is built desktop-based with the Visual Basic .Net programming language. Visual Basic .Net programming language, The application of data mining in grouping data in clustering student grades of SMA Negeri 7 Bengkulu City is carried out per student subject with subject value attributes and attendance percentage. The application of data mining in clustering student grade data of SMA Negeri 7 Bengkulu City can help in knowing the level of achievement of students of SMA Negeri 7 Bengkulu City. Bengkulu City can help in knowing the level of student achievement in each subject and can be used as an evaluation material to improve the level of student achievement in each subject. subjects and can be used as evaluation material to improve the achievement of students who are at low achievement levels. who are at a low level of achievement.

The objective of data mining is to find the useful information from the huge amounts of data. Many researchers have been proposed the different algorithms to find the useful patterns but one of the most important drawbacks they have found that data mining techniques works for single data table. This technique is known as traditional data mining technique. In this era almost all data available in the form of relational database which have multiple tables and their relationships. The new data mining technique has emerged as an alternative for describing structured data such as relational data base, since they allow applying data mining in multiple tables directly, which is known as Multi Relational data mining. To avoid the more number joining operations as well as the semantic losses the researchers bound to use Multi Relational Data Mining approaches. In this paper MRDM focuses multi relational association rule, Multi relational decision tree construction, Inductive logic program (ILP) as well three statistical approaches. We emphasize each MR-Classification approach as well as their characteristics, comparisons as per the statistical values and finally found the most research challenging problems in MRDM.