اطلاعات تولید برای پیش بینی تقاضای نیمه رسانا بر اساس نفوذ فن آوری و دوره عمر محصول

Semiconductor industry is capital intensive in which capacity utilization significantly affect the capital effectiveness and profitability of semiconductor manufacturing companies. Thus, demand forecasting provides critical input to support the decisions of capacity planning and the associated capital investments for capacity expansion that require long lead-time. However, the involved uncertainty in demand and the fluctuation of semiconductor supply chains make the present problem increasingly difficult due to diversifying product lines and shortening product life cycle in the consumer electronics era. Semiconductor companies must forecast future demand to provide the basis for supply chain strategic decisions including new fab construction, technology migration, capacity transformation and expansion, tool procurement, and outsourcing. Focused on realistic needs for manufacturing intelligence, this study aims to construct a multi-generation diffusion model for semiconductor product demand forecast, namely the SMPRT model, incorporating seasonal factor (S), market growth rate (M), price (P), repeat purchases (R), technology substitution (T), in which the nonlinear least square method is employed for parameter estimation. An empirical study was conducted in a leading semiconductor foundry in Hsinchu Science Park and the results validated the practical viability of the proposed model. This study concludes with discussions of the empirical findings and future research directions.

Semiconductor industry is capital intensive, in which most chip makers focus on core competence of wafer fabrication to enhance the effectiveness of capital investments for technology migration and capacity expansion requiring long lead-time (Wu and Chien, 2008). Indeed, corporate manufacturing strategic decisions involve the interrelated elements including pricing strategies (P), demand forecast and demand fulfillment planning (D), capacity planning and capacity portfolio (C), capital expenditure (C), and cost structure (C), which will affect the overall return (R) of a company, as illustrated in the PDCCCR conceptual framework of Fig. 1 (Chien, 2009). Thus, semiconductor manufacturing companies have to forecast future demands to provide the basis for manufacturing strategic decisions including new fab construction, technology migration, capacity transformation and expansion, tool procurement, and outsourcing (Cakanyildirim and Roundy, 2002 and Chou et al., 2007). Given demand uncertainty and forecast errors, companies often carry a safety stock in terms of the days of in the semiconductor supply chain. As shown in the Bullwhip Effect (Lee et al., 1997), the variations are amplified as moving upstream in the supply chain. Thus, it is critical for high-tech industry to develop flexible forecasting systems that allow them to quickly respond to mitigate the negative impacts of the Bullwhip Effect involved in the supply chain to maintain robust demand fulfillment strategies. However, the demand fluctuation due to shortening product life cycle and increasing product diversification in the consumer electronics era make the demand forecast problem increasingly difficult and complicated. Demand forecast errors cause either inefficient capacity utilization or capacity shortage that will significantly affect the capital effectiveness and profitability of semiconductor manufacturing companies.In practice, most companies forecast the demand by combining regional sales inputs from various customers and then adjusted it with their domain knowledge and market insights. However, sales inputs tend to be biased by the customers and the Bullwhip Effect in the supply chain. Different forecasting methods have been applied in different areas. Most of the existing demand forecasting studies employ time series methods (Hamilton, 1994). However, these methods have difficulty for expressing the adoption process of new products. In addition, forecasting methods that are designed for a single generation cannot consider inter-generational substitution involved in semiconductor industry. Driven by Moore’s Law, the semiconductor industry has continued technology migrations and wafer size enlargement to maintain technology innovation and cost reduction per transistor to penetrate into other segments for component substitution and thus achieve unparalleled growth (Leachman et al., 2007). Technology diffusion models are applied primarily to consumer durables (Meade and Islam, 2006), while little research has been done for semiconductor product demand forecast. Focused on realistic needs for manufacturing intelligence, this study aims to construct a manufacturing intelligence framework to derive the SMPRT model based on product life cycle and technology diffusion theories, for forecasting semiconductor product demand, in which the seasonal factor (S), market growth rate (M), price (P), repeat purchases (R), technology substitution (T) are considered from historical data. Manufacturing intelligence aims to extract useful information and derived patterns from production and supply chain data to support manufacturing strategic decisions (Kuo et al., 2010). To estimate the validity of this approach, an empirical study was conducted in a leading semiconductor company, in which realistic data of 36 quarters were employed to derive the parameters in the proposed model using the nonlinear least square method and then other testing data was used to examine the forecast accuracy. The results showed the practical viability to employ the proposed model for demand forecast with little error as the basis to enhance the decision quality for capacity planning to reduce the risks of capacity shortage or surplus. The remainder of this study is organized as follows. Section 2 reviews the related models as the fundamental. Section 3 proposes a demand forecast framework for semiconductor product and develops a multi-generation diffusion model considering essential characteristics of the present problem. Section 4 validates the proposed model with empirical data from a leading semiconductor company in Taiwan. Section 5 concludes with discussions of empirical findings and future research directions.

Demand forecast is critical for supply chain management and capacity planning to thus determine appropriate capital expenditure and enhance capacity utilization and capital effectiveness. This study developed a framework for forecasting semiconductor product demand and proposed a multi-generation diffusion model (i.e., SMPRT model) that incorporated the seasonal factors, market growth rate, price, repeat purchases, and technology substitution to forecast the demand of semiconductor products, in which the nonlinear least square method was employed to estimate the parameters. An empirical study using real data was conducted in a leading foundry to validate the proposed model to forecast the demands of semiconductor products. The results of empirical study have shown significant goodness of fit and validated practical viability of this approach to accurately forecast the demands for semiconductor products. This approach can provide valuable forecasting information for support capacity planning decisions and manufacturing strategies. Future research can be done to examine the model assumptions in various settings. The SMPRT model does not consider the leapfrogging via which the customers skip to the last generation when switching from any previously purchased generation. For matured Technology A which is not fitted well by the proposed SMPRT due to the Long-tail effect (Anderson, 2006), future research can be done to effectively expand the multi-generation diffusion model to forecast the sales of the matured technologies. Furthermore, the SMPRT model incorporates some factors. However, the SMPRT model still does not incorporate other factors that influence product demand. Future studies can consider other factors such as market share, capacity restriction, and cycle effect in the diffusion model. In addition, an analytical method can be developed to extract the parameters of the newer generation based on the estimated parameters of older ones.