Elsevier

Information Sciences

Volume 179, Issue 6, 1 March 2009, Pages 844-850
Information Sciences

Recognizing yield patterns through hybrid applications of machine learning techniques

https://doi.org/10.1016/j.ins.2008.11.008Get rights and content

Abstract

Yield management in semiconductor manufacturing companies requires accurate yield prediction and continual control. However, because many factors are complexly involved in the production of semiconductors, manufacturers or engineers have a hard time managing the yield precisely. Intelligent tools need to analyze the multiple process variables concerned and to predict the production yield effectively. This paper devises a hybrid method of incorporating machine learning techniques together to detect high and low yields in semiconductor manufacturing. The hybrid method has strong applicative advantages in manufacturing situations, where the control of a variety of process variables is interrelated. In real applications, the hybrid method provides a more accurate yield prediction than other methods that have been used. With this method, the company can achieve a higher yield rate by preventing low-yield lots in advance.

Introduction

In the manufacturing of semiconductors, final products are fabricated through several hundred processes which are highly automated and dramatically interdependent. Most manufacturing processes in use today are complexly intertwined and become infinitesimal when using nanometer-scale technology.

For those manufacturers or engineers, the yield is considered a very important factor that has to be monitored and controlled. Yield is defined as the ratio of normal products to finished products. Yield management in the semiconductor industry is understood as a comprehensive analytical system which has the characteristics of a complex system. A complex system has many independent component variables that interact with each other in many complicated ways. Therefore, it is considered difficult to predict and control.

Yield in semiconductor fabrication is strongly affected by several factors, including particles or contaminants on the wafer, substances in the manufacturing instruments, manufacturing process parameters, process engineers’ attitudes, and the design of semiconductors.

Semiconductor companies can achieve a certain degree of yield by applying statistical process controls and 6-sigma to a semiconductor. Yield enhancement employing statistical measurements, however, has difficulty in preventing low-yield lots effectively in advance. This is because manufacturing process variables which affect changes in the yield have a non-linear complex relationship with the yield. Due to this interactive effect among several variables, manufacturers find it difficult to pinpoint problems in time, when small changes in a relationship between process parameters can cause changes in the yield.

Thus, other intelligent techniques are needed in order to detect the main process variables which seriously affect changes in the yield. This study has developed a hybrid yield prediction system in the semiconductor industry, called HYPSSI, as a complement to the existing statistical approach. This system is based on a hybrid application of machine learning techniques to depict multiple process variables concerned with effectively predicting the production yield in semiconductor manufacturing.

HYPSSI adopts neural networks (NNs) and case-based reasoning (CBR) which can be directly applied to prediction purposes. However, CBR suffers from feature weighting; when it measures the distance between cases, some features should be weighted differently. Many feature-weighted variants of the k-Nearest Neighbor (k-NN) have been proposed to assign higher weights to more relevant features for case retrieval purposes [2], [37]. Though those variants have been reported as improving their retrieval accuracy regarding some tasks, few have been used in conjunction with neural networks to predict yield performance in semiconductor manufacturing.

In order to weigh features and guide CBR, HYPSSI adopts four feature-weighting methods: sensitivity, activity, saliency, and relevance. Each method calculates the degree of each feature’s importance by using the connection weights and activation patterns of the nodes in the trained neural network.

In order to validate this hybrid approach within the semiconductor industry, HYPSSI was applied to the International Semiconductor Company, which has been ranked one of the top manufacturers in the world. After comparing this hybrid method with other methods that have been used, this paper shows the hybrid method provides a more accurate yield prediction.

This paper is organized as follows: Section 2 reviews various approaches used in providing yield management applied to semiconductor manufacturing. This section focuses on hybrid applications combining machine learning techniques. Section 3 describes the methodology of the hybrid yield prediction system in the semiconductor industry, called HYPSSI. Experimental results are presented in Section 4 to validate the system. Finally, this paper is concluded by briefly summarizing the study and the direction of future research.

Section snippets

Research methods applied to semiconductor manufacturing

In a high-tech industry such as semiconductor manufacturing, yield improvement is increasingly important as advanced fabrication technologies are complicated and many interrelated factors affect the yield of fabricated wafers. A few studies have aimed to improve yield performance and to reduce operating and capital investment cost in the semiconductor industry.

There are several statistical approaches applied to semiconductor manufacturing. Wang [36] used the lower confidence bound and

Hybrid yield prediction system in semiconductor industry (HYPSSI)

In order to improve the ability of predicting yield accurately, a hybrid yield prediction system was developed in the semiconductor industry (HYPSSI). It is the following hybrid method, combining machine learning techniques, such as back-propagation network (BPN), CBR, and k-NN (see Fig. 1).

HYPSSI consists of four phases: Learning about the relationship between case variables and yield, feature weighting, extracting k similar cases, and weighted averaging of extracted yields. The first phase

Application and evaluation

In order to verify the effectiveness of the hybrid method devised in this paper, it was applied to the production data collected from the manufacturing lots of a Korean semiconductor company: 230 high-yield lots, 230 low-yield lots, and 16 process variables. By definition, a high-yield lot delivers more than 90% yield from a lot and a low-yield lot conveys less than 60% yield from a lot. This definition is still in use by the engineers of the company. Sixteen process variables, including film

Conclusion

Yield management in the semiconductor industry is a very important management practice that has to be monitored and completely controlled. Because manufacturing process variables have a non-linear complex relationship with the yield, manufacturers need an intelligent approach to pinpoint the relationship between process parameters in time.

In this paper, the authors devised and applied HYPSSI, a hybrid method combining BPN and CBR, to predict the yield of the target semiconductor manufacturing

References (42)

  • S.C. Hui et al.

    Data mining for customer service support

    Information and Management

    (2000)
  • K.H. Im et al.

    Case-based reasoning and neural network based expert system for personalization

    Expert Systems with Applications

    (2007)
  • B.-S. Kang et al.

    Hybrid machine learning system for integrated yield management in semiconductor manufacturing

    Expert Systems with Applications

    (1998)
  • K.-S. Kim et al.

    The cluster-indexing method for case-based reasoning using self-organizing maps and learning vector quantization for bond rating cases

    Expert Systems with Applications

    (2001)
  • M. Last et al.

    Discovering useful and understandable patterns in manufacturing data

    Robotics and Autonomous Systems

    (2004)
  • T.W. Liao

    An investigation of a hybrid CBR method for failure mechanisms identification

    Engineering Applications of Artificial Intelligence

    (2004)
  • C.-H. Liu et al.

    An association-based case reduction technique for case-based reasoning

    Information Sciences

    (2008)
  • S.G. Romaniuk et al.

    SC-net: a hybrid connectionist, symbolic system

    Information Sciences

    (1993)
  • K.M. Saridakis et al.

    Case-DeSC: a system for case-based design with soft computing techniques

    Expert Systems with Applications

    (2007)
  • C.K. Shin et al.

    Memory and neural network based expert system

    Expert Systems with Applications

    (1999)
  • Z. Sun et al.

    Case base building with similarity relations

    Information Sciences

    (2004)
  • Cited by (31)

    • Data-driven optimization of accessory combinations for final testing processes in semiconductor manufacturing

      2022, Journal of Manufacturing Systems
      Citation Excerpt :

      In such circumstances, different regression models from machine learning should be examined so as to build a competitive yield prediction model for the optimization of accessory combinations. With regard to industrial applications of machine learning, Jang and Sung [13] devised a hybrid method of incorporating machine learning techniques together to detect high and low yields in semiconductor manufacturing. Using simulation results, the hybrid method gave more accurate yield predictions than other methods.

    • Incorporating adaptability-related knowledge into support vector machine for case-based design adaptation

      2015, Engineering Applications of Artificial Intelligence
      Citation Excerpt :

      So far BPNN does not have a general rule for determining the optimal number of hidden nodes and for selecting the number of epochs. Lee and Ha (2009) utilized Akaike Information Criterion (AIC) to determine the optimal topology of BPNN. AIC, which is as famous as Schwartz Bayesian Criterion, picks up the optimal number of hidden nodes through a heuristic search (Burnham and Anderson, 2002).

    • Single-machine scheduling with two competing agents and learning consideration

      2013, Information Sciences
      Citation Excerpt :

      In traditional scheduling, job processing times are assumed to be fixed and known [17,19,25,26].

    View all citing articles on Scopus
    View full text