This paper deals with the surveillance of the health state and the prediction of the Remaining Useful Life (RUL) of an operating equipment unit of the semiconductor manufacturing industry. It aims at improving an existing work performed in this domain. For that, a new framework for RUL prediction is proposed based on modeling the behavior of the Health Indicator (HI). The contribution of this framework is the effective combination of the proposed HI extraction and the RUL prediction approaches. The HI extraction approach is mainly based on the Least Absolute Shrinkage and Selection Operator (LASSO) regression model. The RUL prediction approach relies on the adaptive Wiener process in which the similarity principle is introduced. An application of the proposed framework on real industrial data shows an improvement of the RUL prediction accuracy compared to the existing work.

The manufacturing process of semiconductor devices is one of the most complex processes in manufacturing industry. The devices fabrication is performed through hundreds of sequential process steps with different recipes. The level of complexity is also increasing due to the high demands in terms of feature size and number of devices. Maintaining high yield and good quality production are the main objectives of these industries. These objectives can be achieved by adopting efficient maintenance strategies. In this context, a suitable prognostic model is required in order to schedule the maintenance actions. Among the different prognostic approaches, data-driven ones received a lot of attention since they do not require any specific knowledge for modeling these complex processes. Although the advances in data-driven prognostic works, there is a real lack of survey papers that overview and discuss the existing approaches for this industry over the past 10 years. Therefore, this paper presents a systematic overview of data-driven prognostic for semiconductor manufacturing. It investigates the different used methods, the challenges of their application and the unexplored research areas.

For decades, manufacturers have been collecting and storing high amounts of data with the aim of better controlling and managing their processes. With the vast amount of information and hidden knowledge in all of these data, the challenge for these manufacturers to monitor their equipment units, is the extraction of an appropriate health indicator from these data that illustrates the actual state of their equipment units. In this paper, we are interested in extracting the health indicator of semiconductor equipment where manufacturing is performed by batch. For that, a novel automatic approach named Significant Points combined to the Least Absolute Shrinkage and Selection Operator (SP-LASSO) is proposed. This approach is mainly based on LASSO regression model. Its accuracy is illustrated by numerical application on simulated data.

Over the last few years, with the increasing worldwide competition, semiconductor industries have had to constantly innovate in order to enhance their performance, productivity and minimize the downtime. Monitoring the state of health of their equipment units is important to avoid machine failures and to plan maintenance actions. For that, a novel approach for health indicator extraction named Significant Points combined to the Least Absolute Shrinkage and Selection Operator (SP-LASSO) is proposed in this paper. It deals with the problem of high dimensional data and the specificity of the health indicator in real industrial cases. The proposed method performs feature selection and health indicator extraction and it is mainly based on LASSO. A numerical application on simulated data illustrates the accuracy of this approach.

The performance improvement of alarm systems helps enhance control, operator effectiveness, facility up-times, safety, and eventually reduce losses. During the facility's operation, a large amount of alarm system data are collected and historized. Valuable information can be extracted from the alarms history and used to enrich the facility's control and performance, which prompted the development of alarm management methods and resulted in a need to transform such information into useful visual forms. These visual forms need to help facility operators and engineers understand alarm system behavior and facilitate decision making. This paper presents a framework to monitor and evaluate the alarm system's performance in a semiconductor manufacturing facility. We provide a mathematical formulation of elaborated calculations in this context. We also demonstrate the proposed framework's effectiveness using a real dataset issued from the ST-Rousset semiconductor manufacturing facility.

This paper proposes a data-driven framework for Remaining Useful Life (RUL) prediction, based on the Brownian Motion model (BM) and the similarity principle, for an operating system given its health indicator. It addresses the issues of noisy and limited run-to-failure (R2F) data. The Percentile filtering is used to extract, from the R2F data, 100 monotonic profiles used as references in the modeling and the RUL prediction. Then, the similarity is computed between these references and the Health Indicator (HI) of the operating system. Fitting the most similar reference into the BM improves the RUL prediction. A numerical application using simulated data justifies the accuracy of this approach.

This paper proposes a data-driven framework for Remaining Useful Life (RUL) prediction of an operating equipment unit in the case of noisy and limited data. It consists of introducing the Brownian Motion model (BM) in the similarity framework and computing the RUL based on a collection of generated HIs (Health Indicators). For that, the percentile filter is used to pre-process the HIs by generating a collection of profiles from the operating equipment unit’s HI and a set of references from a given R2F (Run-to-Failure) indicator. Then, the similarity is computed between each profile of this collection and the references in order to pick the most similar reference to each profile for modeling and RUL prediction. The final RUL is calculated as a weighted aggregation of the obtained RULs of the profile collection. A numerical application using simulated data illustrates the accuracy of this approach.

Process diagnostic and monitoring during production is a fundamental task of the control and alarm system. However, many defected products are still related to various issues of health states of production equipment. Therefore, quality inspection is a crucial step during the manufacturing process, ensuring that a product’s quality is maintained or improved with a reduced or total absence of errors. The final product quality determines whether or not a product unit satisfies its intended use. In this paper, we propose a final quality inspection framework based on alarm events data. In this framework, we first transform the textual alarm data into numeric using binary scoring. Then, we reduce the dimension of the obtained numeric matrix using an appropriate alarms grouping method. After that, we apply the reduced data to learn a classifier and to make a decision. Finally, we compare several machine learning algorithms’ performance in the prediction of scrap-per-lot, namely Decision Tree, Logistic Regression, K-nearest neighbors, Linear Support Vector Machine, and Multi-Layer Perceptron. The results show a satisfactory performance of the compared models that we effectively prove on a dataset collected over the whole semiconductor fabrication facility.

The early information about the health state of the final product quality plays a vital role in the intact management of production. In semiconductor manufacturing, quality control of a too-small number of wafers is routinely carried on specific metrology stations, and the obtained quality measurements are generalized over the entire lot. The unavailability of sufficient product quality information results in a lack of that for a high proportion of products. The latter leads to some overlooked quality problems that might cause a malfunction in the final product. This malfunction is usually conducive to yield loss, resource consumption through its remaining production line steps and also needs a considerable amount of time to be source-identified. This paper proposes a final quality classification data-driven approach using machine learning techniques and alarm events data collected during the production operations. We use the k-mean clustering algorithm to group production lots into clusters based on their passages over equipment. Each cluster has its decision tree classification model elaborated after various information extraction techniques and manipulation applied to alarm event texts. The obtained results show a satisfactory performance demonstrated on a real-world dataset collected over the whole semiconductor fabrication facility.

For decades, Run-to-Run (R2R) controllers have been widely implemented in semiconductor manufacturing. They operate over key process parameters on the basis of the metrological measurements acquired from the process and their deviations from the target setpoints. Conventionally, R2R controllers have been implemented independently of the actual equipment condition, which is obviously affecting the process stability and performance. Therefore, both equipment signals and process states shall be considered to make the R2R controllers more robust to the equipment condition drifts. In this paper, we propose a novel physics-informed framework to integrate the real-time equipment condition, based on the Fault Detection and Classification (FDC) data, into the R2R controllers. By utilizing Dynamic Bayesian Networks (DBN), the implicit relationship structure between metrology measurements, FDC indicators, and R2R regulators can be learned and reviewed explicitly. The structure shall be further reviewed to valid with the existing relationships and expert knowledge. Infeasible causalities on the structure will be constrained via setting up the blacklist at the structure learning stage. The proposed framework consists of the offline modeling stage, which incorporates the process, equipment variables, and the expert knowledge in the structure learning, and the online control stage, which constructs the Structured R2R controller (SRC) based on the relationship structure. As a result, the model is consistent by design with empirically known relationships and fundamental physical laws. The proposed SRC not only optimizes the operation with respect to the target control values but also considers the equipment and process states simultaneously. The effectiveness of SRC and the derivative control strategy are validated through a real dataset of a Chemical-Mechanical Polishing (CMP) process, and two simulated studies. (C) 2020 Elsevier Ltd. All rights reserved.

In this paper we propose an approach combining discrete event simulation and model checking. Indeed, methods like model checking suffer from the state space explosion when the modeled system is complex. Consequently, we propose an approach that allows the model checking procedure to focus on a subset of the total state space that is more likely to contain an erroneous state regarding a property. To do so, a simulation run that allows us to observe some qualitative metrics is performed, and stopped when a predefined "critical state" is reached. This state is then projected as the initial state of the non-deterministic finite automaton used by the model checking procedure. Thus, the search will only explore states reachable from this critical state, limiting state space explosion. We finally illustrate the proposed approach through a Network-On-Chip system, and compare it with the ad hoc classical model checking approach.

Industrial alarm system management and enhancement have recently become a subject of high interest in the industrial domain as well as in the scientific research area. Alarm system performance is the key point to look at for the examination of whether they work correctly. Few research papers so far have focused on the semiconductor manufacturing process area. This paper summarizes some interesting work applied not long ago in other industries and especially those adapted to semiconductor manufacturing. It represents a concise description of semiconductor manufacturing processes, toward an understanding of the problems and challenges encountered. Furthermore, a basic statistical analysis employed to show some of alarm system performance metrics that have used, as shown in some well-known standards, to assess an alarm system performance as well as highlighting misconfigured alarms based on their recorded histories.

This paper deals with a technical issue in the semiconductor industry which is the monitoring of a regulated process subject to various sampling frequencies. A mathematical link between the variance of the fully sampled and the partially sampled process is established based on the sampling rate value. Moreover, a method is proposed in order to adapt the monitoring limits to the sampling rate so that fault detection becomes more efficient. The pertinence of the method is tested using real data provided by a semiconductor foundry.

The performance of industrial alarm systems has become a subject of very high interest as they strongly contribute to avoiding undesired or abnormal situations during production operation. They are considered as a fundamental part of any production facility and their efficiencies are certainly influencing the final products' quality. Therefore, the increase in process equipment and automation degree have raised, as a consequence, the number of configured alarms to monitor the processes, which, during the operation, results in floods of alarms that decreased the effectiveness of alarm system and increased operator workloads beyond their capacities. The identification of critical and relevant alarms to products quality helps in monitoring simultaneously alarm performance and their impact on the final product. This paper presents an approach based on the AdaBoost algorithm for addressing alarm issues by predicting their risk of final product degradation as a function of their statistical behaviors of activation on product lots during production operation which in turns has used to group alarms. The results show a good performance of this method which has demonstrated on a real dataset collected from a semiconductor fabrication facility.

Virtual metrology (VM) has been widely studied in the semiconductor industry with the purpose of decreasing the cycle time and reducing the expensive metrology measurements. Ideally, a VM model should not only be able to provide accurate predictions but also present an interpretable and rational structure to accommodate fundamental restrictions and relationships that are known to be present in the process. The last aspects have been missing in the VM models proposed hitherto. Therefore, in this article, we propose a novel framework by combining in a single VM model the capability to learn from data with the ability to incorporate the domain knowledge on the process. Thus, the new methodology can use the best of both information sources: data and the subject-matter expert (SME) knowledge. The framework consists of two phases. In the first phase, a Gaussian Bayesian network (GBN) is used to extract the implicit relationships between the metrology and production/process variables. In the second phase, the target response variable is defined, the predictors are selected through the associated Markov blanket, and finally, an empirical model is estimated to accurately predict the response. The proposed framework was tested and its effectiveness was confirmed through a real industrial data from a chemical–mechanical polishing (CMP) process in semiconductor fabrication. The physical meaning of the model obtained was also scrutinized by an SME. Note to Practitioners—Unlike the conventional virtual metrology (VM) techniques that model the x–y relationship based only on process data, the proposed method aims at consolidating the process knowledge from experts into an extensive relational structure. Not only the x–y model is inferred but also the relationships among the predictors are revealed. The structure is illustrated in a form of a connected graph so that the correlation between parameters can be expressed explicitly, which is also compatible with the physical laws. With the clear visualization of the correlation structure of variables, the practitioners are able to utilize the result in different applications. The learned structure can be further integrated into the plant advanced process control system, including VM, process monitoring and diagnosis, and run-to-run (R2R) control.

Physical measurements and inspection tests of products quality in semiconductor manufacturing processes are carried out on specific equipment that often expensive and usually separated from production, which is time-consuming and production throughput decrease. Instead, predicting quality state using the available processes data has now become possible through mathematical models where this enables quick decision making in regards to product health state based on the prognosticated information about product quality. In this context, data-driven techniques based on Multilayer Perceptron Artificial Neural Network (MLP-ANN) have proposed to reveal the relationship between products end quality state and processes alarm events. Within this framework, data transformation and adaptation have discussed and model parameters selection was accomplished using the cross-validation technique. The results show a reasonable performance of the selected model that was effectively proven on a data-set collected over the whole semiconductor fabrication facility.

Process and data are equally important for business process management. Data is especially relevant in the context of automated business processes, and process controlling. In the context of knowledge-intensive domain, data modified by engineers and used by a fully automatic manufacturing system can lead to unpredictable errors. In this paper, a method permitting the discovery of truly enforced business rules used by the process is proposed. These rules will be highlighted using data from the model of the studied process. The proposed methodology is mainly based on a data mining approach. The proposed method has been tested on data from semiconductor industry, in which processes are known to be complex, and the number of business rules is known to be important. The results show that the method is efficient in assisting engineers in process errors detection and practical process improvement.

One of the levers for cycle time and cost reduction in the semiconductor industry is the use of all the available production tools as much as possible. In this paper, a modified Non-threaded Run-to-Run approach is proposed. This approach permits the control of the maximum possible number of equipment simultaneously. The idea is to model the entire system with all possible combinations (threads) in a novel state space representation (SSR) so that we can increase the rank of the controllability matrix. Due to its efficiency, especially with a SSR, a Kalman filter is used as an estimator for the Run-to-Run control purpose. The accuracy of the proposed method is tested using real data derived from the deposition area of STMicroelectronics foundry in Rousset.

Uniform rate sampling is the easiest and most common sampling strategy used in the semiconductor industry. It consists of measuring every N products and is used for both regulated and unregulated steps. However, it is well-know that the sampling frequency has a direct impact on regulator performances making uniform rate sampling ineffective with the regulated steps. This paper proposes an indicator allowing the adaptation of the sampling frequencies taking into account the performances of the regulator and the industrial risk. The developed method offers better regulation performances, less measured products and leads to yield improvement and cycle time reduction. The approach is tested with real data provided by STMicroelectronics.

We present a new chamber matching algorithm, which is completely data-driven and unsupervised, and designed for the semiconductor industry. The behavior of an equipment is classified as different when the shape of the time series given by one of the sensors is significantly different. Shape comparison is performed using linear regression, that authorizes both offset and change of scale. The method detects both the chamber and the sensor in which the fault is present, then helping in activating corrective maintenances. Application results are shown with two examples of real semiconductor industrial failures.