IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 30, NO. 4, NOVEMBER 2017 345 Exploit the Value of Production Data to Discover Opportunities for Saving Power Consumption of Production Tools Chih-Min Yu, Chen-Fu Chien, Member, IEEE, and Chung-Jen Kuo Abstract—Semiconductor industry is both technology and energy intensive. There is a critical need to develop effective ways for energy saving to support smart and green production. This paper aims to develop data mining approach based on neural networks to exploit the value of production data and derive improvement directions for energy saving. In particular, the power consumption per wafer processed step (kilowatt hour per move, kwh/move) of individual production tool sets can be estimated, in which the relationships between kwh/move and 19 individual input factors, including “lot size,” “process time,” “uptime,” “usable machine,” “Q-time constrain,” and “sampling rate” are derived. An empirical study was conducted in a leading wafer fab and the results have shown practical viability of the proposed approach to discover effective opportunities for saving 17.21% power consumption by production tool sets. Index Terms—Energy management, power saving, data mining, neural network application, manufacturing intelligence, big data analytics, semiconductor fab. I. I NTRODUCTION NERGY conservation has become a critical issue in the recent years. Furthermore, energy consumption is a critical driver for economic growth, yet it is one of key factors for climate change . Therefore, the governments have strived to enable the policy of energy conservation and reduce carbon emission. Semiconductor industry is both technology and energy intensive, in which effective ways are needed to reach energy saving goal . Semiconductor manufacturing is one of the most complicated manufacturing systems that contain hundreds of process steps, in which the data is generated when wafers move from step to step in a wafer fabrication facility (fab) , , which involves highly complex and lengthy wafer fabrication E Manuscript received May 30, 2017; revised July 21, 2017 and August 19, 2017; accepted September 5, 2017. Date of publication September 11, 2017; date of current version October 27, 2017. This work was supported by the Ministry of Science and Technology, Taiwan, under Grant MOST 105-2218-E-007-028, Grant MOST 105-2218-E-007-027, and Grant MOST-105-2622-8-007-002-TM1. (Corresponding author: Chih-Min Yu.) C.-M. Yu is with the Department of Industrial Engineering and Engineering Management, National Tsing Hua University, Hsinchu 30013, Taiwan, and also with YouThought Corporation, Hsinchu 30050, Taiwan (e-mail: email@example.com). C.-F. Chien is with the Department of Industrial Engineering and Engineering Management, National Tsing Hua University, Hsinchu 30013, Taiwan (e-mail: firstname.lastname@example.org). C.-J. Kuo is with YouThought Corporation, Hsinchu 30050, Taiwan (e-mail: email@example.com). Digital Object Identifier 10.1109/TSM.2017.2750712 processes with 300–500 process steps and a large number of interrelated variables . In addition, the semiconductor industry not only requires a high standard on the stability and quality of electricity supply, but also has obviously consumed more electricity than other industries like the steel or petrochemical. If the high-tech industries cannot put more efforts on energy conservation, aside from paying high electricity expenses, they could even face trade sanctions. Thus, a set of OPE (Overall Power Energy Effectiveness) indicators were constructed for Fab energy efficiency assessment and energy saving . Data mining and big data analytics approaches have been developed to extract potentially useful information and manufacturing intelligence from massive data in various problems in different fields such as bioinformatics , marketing trend and demand forecast , sales forecast , human resource management , financial engineering , health and medical decision making , fault location , and manufacturing , . In particular, data mining and manufacturing intelligence approaches have been implemented for semiconductor manufacturing including demand forecast , human resource management , yield enhancement , , wafer bin map analysis , , cycle time reduction , , and fault location . In addition, various statistical analysis approaches were developed to model the relation between the responses and the factors for the traditional manufacturing system . For a semiconductor wafer fab, the facility system accounts for 56% of the total power consumption , . Consequently, most of previous works focus on suppression of power consumption by the facility system. However, little research has been done to save power consumption by process tools that account for 41% of power consumed in the fab. To fill the gap, this study aims to develop a data mining framework that integrates neural networks (NNs) model to exploit the value of production big data and equipment data to discover opportunities for saving power consumption from production tools through manufacturing management. An empirical study was conducted for validation. II. M ETHODOLOGY Data mining has been applied in various fields, in which NNs have been employed for many data mining and machine learning techniques . Data mining problems are generally c 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. 0894-6507 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. 346 Fig. 1. IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 30, NO. 4, NOVEMBER 2017 Multi-layers feed-forward neural network with full connected. categorized as four types of tasks: 1) regression; 2) classification; 3) clustering; and 4) association . Classification consists of examining the attributes of an object and assigning it to one of the predefined categorical classes. Clustering is the process of dividing a dataset into several clusters so as to minimize the intra-class similarity and maximize the inter-class similarity. Association is the discovery of association rules showing the conditions that specific attribute values occur simultaneously. Data mining has been employed as manufacturing intelligence approach to extract useful information and derived patterns from manufacturing data to support related decisions. For example, Hsu and Chien  developed a hybrid approach that integrates spatial statistics and Adaptive Resonance Theory (ART) NNs to efficiently extract patterns from wafer bin maps associating with defects for effective trouble shooting and yield enhancement. NNs are data mining techniques often employed for regression analysis that have inherent capability to map nonlinear relationship between the input and output factors , . Since there are nonlinear relationships between input factors and the kwh/move, NNs for regression are employed for modeling the relationships of related factors and the kwh/move levels of individual tool sets in a semiconductor fab. To investigate which approach will be most suitable for evaluating the relation between input factors and kwh/move of tool sets, we compared four representative NNs for regression task including back-propagation neural networks (BPNN), recurrent neural networks (RNN), radial basis function neural networks (RBFNNs), and support vector regression (SVR) as follows. A. Back-Propagation Neural Networks (BPNN) BPNN are multilayer feed-forward NNs with an input layer, an output layer and some hidden layers between the input and output layers as illustrated in Fig. 1. B. Recurrent Neural Networks (RNN) Comparing to BPNN that are confined to static mapping, RNN contain feedback connections and can model the evolution of dynamic systems, as showed in Fig. 2. Fig. 2. Multi-layers recurrent neural network with dynamic connected. C. Radial Basis Function Neural Networks (RBFNNs) The major advantage of RBFNN over BPNN is the fast convergence . The hidden layer neuron in RBFNN generally uses Gaussian basis function to transfer the inputs. The training in the hidden layer acts in an unsupervised way, unlike the supervised way in BPNN . D. Support Vector Regression (SVR) SVR is a novel neural network algorithm based on the concept of support vector machine classifier . The SVR learning minimizes an upper bound on the generalization error, as opposed to typical learning algorithms that minimize the error on the training data . SVR has several advantages over other typical NNs, such as a global and unique solution and fast convergence , . III. E MPERICAL S TUDY A. Data Preparation The data used for this study have been collected from 48 production tool sets of an 8-inch wafer foundry fab in Singapore; and the duration for data collection is 120 days. The tool set means a group of machine which processes same working step in semiconductor manufacturing factory, included each kinds of IC manufacturing process ones, like Metal Lithography, Metal Etching, Metal Sputter, Tungsten chemical machine polish (W CMP), and etc. Without loss of generality, the data used for this research has been transformed to protect proprietary information of this case company. Take a tool set for example, Fig. 3 shows that the average kwh/move decreases nonlinearly as utilization gets higher. Hence, power consumption will be less if the variation of utilization is smaller given the same average utilization level. Similarly, Fig. 4 implies that the larger the lot size, the less the kwh/move. In addition, kwh/move will be less if the variation of lot size is smaller. The factors mentioned above were for estimating the kwh/move levels of individual tool sets. By referring to Kuo et al.  and domain knowledge of the case fab, this study defined 19 factors that might affect YU et al.: EXPLOIT VALUE OF PRODUCTION DATA TO DISCOVER OPPORTUNITIES FOR SAVING POWER CONSUMPTION 347 TABLE II C OMPARISON OF MAPE FOR D IFFERENT NN S Fig. 3. Kwh per move vs. tool utilization. Table I means “Engineering lots” which lot owner belongs to engineering department instead of production lots belonged to manufacturing one. In addition, “COV” and “STDEV” mean “Coefficient of Variance” and “Standard Deviation” respectively. B. Neural Networks (NNs) Model Construction Fig. 4. Kwh per move vs. mean lot size. TABLE I T HE I NPUT FACTORS FOR E STIMATING KWH /M OVE the power consumption of individual tool sets in a semiconductor wafer fab, as shown in Table I. Among them, 12 factors are related to mean value while the other 7 factors are related to variation, covered by the 15 input factors defined by Kuo et al. . The terminology of “ENG” in In this study, NNs are constructed to analyze the relationships between related input factors and kwh/move of individual tool set in a fab. 4 learning algorithms for NNs are evaluated. And for each algorithm, different combinations of network architecture and parameters were experimented. We used the package of NeuroSolutions 5 (by Neurodimensions, Inc.) to develop the models of BPNN, RNN and RBFNN. The package of STATISTICA 7 (by STATSOFT, Inc.) was applied for SVR. The learning algorithms are evaluated with the least mean absolute percent error (MAPE) for testing data, as shown in Table II. As a result, the best NNs method is backpropagation neural networks using Levenberg-Marquardt (LM) learning algorithm with number of hidden layer = 1 and number of hidden neurons = 3. Sensitivity analysis based on the trained NNs are conducted to estimate the saved power consumption in the fab per a specific percent, e.g., 10%, of improvement on each factor, as shown in Table III. As for the relation to kwh, the sign, “+”, of relation between kwh/move and input factors means positive relation, kwh/move will be higher with incremental input factors. On the contrary, the sign, “−”, represents negative relations between kwh/move and input factors. This table also lists the managerial implications for improving individual factors so as to take actions for saving power consumption. However, in semiconductor manufacturing factory, there will be some trade-off on other purpose occurred. Take “mean percentage of super hot lots” as example, if emphasizing on power consumption saving prior to speeding delivery, the lead time of super hot lots must be impacted because they are not allowed to be waited with empty load ports in next process 348 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 30, NO. 4, NOVEMBER 2017 TABLE III R ELATION B ETWEEN KWH AND I NPUT FACTORS , AND T HEIR M ANAGERIAL I MPLICATIONS TABLE IV P RIORITIZED FACTORS FOR I MPROVEMENT TABLE V P RIORITIZED T OOL S ETS FOR THE FACTOR OF “M EAN P ERCENTAGE OF U SABLE M ACHINES ” be obtained by what-if analysis based on the room for factor improvement for each factors. Table IV shows that the case fab might save 17.21% power consumption if it aligned all the factors to the performance of the top 25 percentile of benchmark. Similarly, the tool sets can also be prioritized for each factor. As shown in Table V, the top 21% tool sets (10 / 48 tool sets) account for 70.3% (3.10%/4.41%) the impact to total kwh of the fab for the factor “Mean percentage of usable machines”, which means the percentage of released tools in this semiconductor processed step. V. C ONCLUSION machines if other available lots could be processed at that time. The decision making could be judged by top manager depended on different focus points. IV. R ESULT AND D ISCUSSION Due to limited resources, the fab needs to prioritize the factors so as to save more power consumption with less effort. As shown in Table IV, the room for factor improvement is defined by the gap between the as-is factor value of the fab and the factor value of the top 25 percentile of all the benchmark fabs that is provided by YouThought corporation. Then the Impact percent to total kwh of the Fab can The climate change has caused increasing natural disasters all over the world in recent years. Semiconductor manufacturing that is energy intensive should strive for green and smart production. This study has validated the proposed data mining framework based on NNs to estimate the kwh/move of individual tool sets, and to analyze the relationships between kwh/move and individual input factors. An empirical study was conducted by using the equipment data of a real fab. The results have validated that the proposed approach can discover useful opportunities for saving power consumption by production tools. Future research can be done to employ the proposed approach for analyzing big data automatically collected in wafer fab to derive effective directions for saving consumption of other resources for total resource management , . Similar approach can be extended to other high energyconsumption manufacturing such as TFT-LCD and solar cells fabrication. On the other hand, saving power and other productivity improvement share common points in semiconductor manufacturing factory as well. Some overall resource efficiency improvement activities will contribute the benefits for both productivity and power saving . For example, the following actions can be employed, 1) To reduce the variation of mean available time and process time of process machine; 2) To enlarge the mean wafer lot size and batch size to eliminate the frequency of venting and pumping while load and unload lots; 3) To enhance the YU et al.: EXPLOIT VALUE OF PRODUCTION DATA TO DISCOVER OPPORTUNITIES FOR SAVING POWER CONSUMPTION same recipe rate to avoid higher setup time while changing recipe to investigate productivity improvement directions. More studies can be done to compare with other productivity improvement efforts, while considering realistic manufacturing contexts such as under super hot lot, to examine the trade-off among the factors of power saving, quality, and productivity. R EFERENCES  A. Battaglini, J. Lilliestam, A. Haas, and A. Patt, “Development of SuperSmart grids for a more efficient utilisation of electricity from renewable sources,” J. Clean. Prod., vol. 17, no. 10, pp. 911–918, 2009.  C.-F. Chien, J.-T. Peng, and H.-C. Yu, “Building energy saving performance indices for cleaner semiconductor manufacturing and an empirical study,” Comput. Ind. Eng., vol. 99, pp. 448–457, Sep. 2016.  C.-F. Chien and S.-C. Chuang, “A framework for root cause detection of sub-batch processing system for semiconductor manufacturing big data analytics,” IEEE Trans. Semicond. Manuf., vol. 27, no. 4, pp. 475–488, Nov. 2014.  C.-F. Chien, Y.-J. Chen, C.-Y. Hsu, and H.-K. Wang, “Overlay error compensation using advanced process control with dynamically adjusted proportional-integral R2R controller,” IEEE Trans. Autom. Sci. Eng., vol. 11, no. 2, pp. 473–484, Apr. 2014.  C.-F. Chien, C.-Y. Hsu, and P. Chen, “Semiconductor fault detection and classification for yield enhancement and manufacturing intelligence,” Flexible Services Manuf. J., vol. 25, no. 3, pp. 367–388, 2013.  K.-S. Lin and C.-F. Chien, “Cluster analysis of genome-wide expression data for feature extraction,” Expert Syst. Appl., vol. 36, no. 2, pp. 3327–3335, 2009.  C.-F. Chien, Y.-J. Chen, and J.-T. Peng, “Manufacturing intelligence for semiconductor demand forecast based on technology diffusion and product life cycle,” Int. J. Product. Econ., vol. 128, no. 2, pp. 496–509, 2010.  S. Thomassey, “Sales forecasts in clothing industry: The key success factor of the supply chain management,” Int. J. Product. Econ., vol. 128, no. 2, pp. 470–483, 2010.  C.-F. Chien and L.-F. Chen, “Using rough set theory to recruit and retain high-potential talents for semiconductor manufacturing,” IEEE Trans. Semicond. Manuf., vol. 20, no. 4, pp. 528–541, Nov. 2007.  A. Sharma and P. K. Panigrahi, “A review of financial accounting fraud detection based on data mining techniques,” Int. J. Comput. Appl., vol. 39, no. 1, pp. 37–47, 2012.  S. Evans, S. J. Leman, C. A. Deters, R. M. Fusaro, and H. T. Lynch, “Automated detection of hereditary syndromes using data mining,” Comput. Biomed. Res., vol. 30, no. 5, pp. 337–348, Oct. 1997.  C.-F. Chien, S.-L. Chen, and Y.-S. Lin, “Using Bayesian network for fault location on distribution feeder,” IEEE Trans. Power Del., vol. 17, no. 13, pp. 785–793, Jul. 2002.  A. Kusiak, “Rough set theory: A data mining tool for semiconductor manufacturing,” IEEE Trans. Electron. Packag. Manuf., vol. 24, no. 1, pp. 44–50, Jan. 2001.  J. A. Harding, M. Shahbaz, T. Srinivas, and A. Kusiak, “Data mining in manufacturing: A review,” J. Manuf. Sci. Eng., vol. 128, no. 4, pp. 969–976, Dec. 2006.  C.-F. Chien, W.-C. Wang, and J.-C. Cheng, “Data mining for yield enhancement in semiconductor manufacturing and an empirical study,” Expert Syst. Appl., vol. 33, no. 1, pp. 192–198, 2007.  D. Braha and A. Shmilovici, “Data mining for improving a cleaning process in the semiconductor industry,” IEEE Trans. Semicond. Manuf., vol. 15, no. 1, pp. 91–101, Feb. 2002.  S.-C. Hsu and C.-F. Chien, “Hybrid data mining approach for pattern extraction from wafer bin map to improve yield in semiconductor manufacturing,” Int. J. Prod. Econ., vol. 107, no. 1, pp. 88–103, 2007.  C.-W. Liu and C.-F. Chien, “An intelligent system for wafer bin map defect diagnosis: An empirical study for semiconductor manufacturing,” Eng. Appl. Artif. Intell., vol. 26, nos. 5–6, pp. 1479–1486, 2013. 349  Q. P. He and J. Wang, “Fault detection using the k-nearest neighbor rule for semiconductor manufacturing processes,” IEEE Trans. Semicond. Manuf., vol. 20, no. 4, pp. 345–354, Nov. 2007.  Energy Use in the Semiconductor Manufacturing Industry. Washington, DC, USA: Environ. Protect. Agency, 1998.  S.-C. Hu and Y. K. Chuah, “Power consumption of semiconductor fabs in Taiwan,” Energy, vol. 28, no. 8, pp. 895–907, Jun. 2003.  K. A. Smith and J. N. D. Gupta, “Neural networks in business: Techniques and applications for the operations researcher,” Comput. Oper. Res., vol. 27, nos. 11–12, pp. 1023–1044, 2000.  U. Fayyad, G. Piatesky-Shapiro, and P. Smyth, “The KDD process for extracting useful knowledge from volumes of data,” Commun. ACM, vol. 39, no. 11, pp. 27–34, 1996.  M. Chambers and C. A. Mount-Campbell, “Process optimization via neural network metamodeling,” Int. J. Prod. Econ., vol. 79, no. 2, pp. 93–100, 2002.  J. F. Briceno, H. El-Mounayri, and S. Mukhopadhyay, “Selecting an artificial neural network for efficient modeling and accurate simulation of the milling process,” Int. J. Mach. Tools Manuf., vol. 42, no. 6, pp. 663–674, 2002.  B. Kim and K. Park, “Modeling plasma etching process using a radial basis function network,” Microelectron. Eng., vol. 77, no. 2, pp. 150–157, 2005.  K.-J. Kim, “Financial time series forecasting using support vector machines,” Neurocomputing, vol. 55, nos. 1–2, pp. 307–319, 2003.  W.-H. Chen, J.-Y. Shih, and S. Wu, “Comparison of support-vector machines and back propagation neural networks in forecasting the six major Asian stock markets,” Int. J. Electron. Financ., vol. 1, no. 1, pp. 49–67, 2006.  E. Byvatov, U. Fechner, J. Sadowski, and G. Schneider, “Comparison of support vector machine and artificial neural network systems for drug/nondrug classification,” J. Chem. Inf. Comput. Sci., vol. 43, no. 6, pp. 1882–1889, 2003.  C.-J. Kuo, C.-F. Chien, and J.-D. Chen, “Manufacturing intelligence to exploit the value of production and tool data to reduce cycle time,” IEEE Trans. Autom. Sci. Eng., vol. 8, no. 1, pp. 103–111, Jan. 2011.  C.-F. Chien, P. Chu, and L. Zhao, “Overall resource effectiveness (ORE) indices for total resource management and case studies,” Int. J. Ind. Eng., vol. 22, no. 5, pp. 618–630, 2015.  C.-F. Chien, C.-H. Hu, and Y.-F. Hu, “Overall space effectiveness (OSE) for enhancing fab space productivity,” IEEE Trans. Semicond. Manuf., vol. 29, no. 3, pp. 239–247, Aug. 2016.  C.-F. Chien, C.-W. Chou, and H.-C. Yu, “A novel route selection and resource allocation approach to improve the efficiency of manual material handling system in 200-mm wafer fabs for industry 3.5,” IEEE Trans. Autom. Sci. Eng., vol. 13, no. 4, pp. 1567–1580, Oct. 2016. Chih-Min Yu received the B.S. degree in economics and industrial engineering and the M.S. degree in engineering management from National Tsing Hua University, Hsinchu, Taiwan, in 1996 and 2006, respectively, where he is currently pursuing the Ph.D. degree with the Industrial Engineering and Engineering Management. He joined YouThought Company, big data solution provider for manufacturing, Taiwan, in 2013, where he is the Sales VP. He was a Department Manager with VisEra Technology Company, an IE Principal Engineer with Taiwan Semiconductor Manufacturing Company, and a Manufacturing Supervisor with Macronix International Company for over 14 years. 350 IEEE TRANSACTIONS ON SEMICONDUCTOR MANUFACTURING, VOL. 30, NO. 4, NOVEMBER 2017 Chen-Fu Chien received the B.S. (Phi Tao Phi Hons.) degree with double majors in industrial engineering and electrical engineering from National Tsing Hua University (NTHU), Hsinchu, Taiwan, in 1990, the M.S. degree in industrial engineering, and the Ph.D. degree in operations research and decision sciences from the University of Wisconsin–Madison, Madison, WI, USA, in 1994 and 1996, respectively, and the PCMPCL Training from Harvard Business School, Boston, MA, USA, in 2007. He is a Tsing Hua Chair Professor with NTHU. He is also the Director of the NTHU-Taiwan Semiconductor Manufacturing Company (TSMC) Center for Manufacturing Excellence and the Principal Investigator for the NSC Semiconductor Technologies Empowerment Partners Consortium. From 2002 to 2003, he was a Fulbright Scholar with the University of California at Berkeley, Berkeley, CA, USA. From 2005 to 2008, he had been on-leave as a Deputy Director with Industrial Engineering Division, TSMC. He has received eight invention patents on semiconductor manufacturing and published four books, over 120 journal papers, and a number of case studies in Harvard Business School. His research efforts center on decision analysis, modeling, and analysis for semiconductor manufacturing, manufacturing strategy, and manufacturing intelligence. He was a recipient of the National Quality Award, the Executive Yuan Award for Outstanding Science and Technology Contribution, the Distinguished Research Awards, the Tier 1 Principal Investigator (Top 3%) from the Ministry of Science & Technology, the Distinguished University-Industry Collaborative Research Award from the Ministry of Education, the University Industrial Contribution Awards from the Ministry of Economic Affairs, the Distinguished UniversityIndustry Collaborative Research Award, the Distinguished Young Faculty Research Award from NTHU, the Distinguished Young Industrial Engineer Award, the Best IE Paper Award, the IE Award from the Chinese Institute of Industrial Engineering, the Best Engineering Paper Award, the Distinguished Engineering Professor by Chinese Institute of Engineers in Taiwan, and the 2012 Best Paper Award of the IEEE T RANSACTIONS ON AUTOMATION S CIENCE AND E NGINEERING. He has been invited to give keynote lectures at several conferences, including APIEMS, C&IE, IEEM, IML, and leading universities worldwide. He is an Area Editor of the Flexible Services and Manufacturing Journal, an Editorial Board Member of Computers and Industrial Engineering, and an Advisory Board Member of OR Spectrum. Chung-Jen Kuo received the B.S., M.S., and Ph.D. degrees in industrial engineering from National Tsing Hua University, Hsinchu, Taiwan, in 1989, 2002, and 2010, respectively. He was a Department Manager with Macronix International Company from 1999 to 2003, and a Project Manager with Taiwan Semiconductor Manufacturing Company from 2003 to 2006. In 2007, he found Youthought Company, Taiwan, where he is currently the Chairman and the President. His research interests include data mining, performance evaluation, and production management for semiconductor manufacturing.