Hung-Kai Wang

With the rapid expansion of smart display applications, LED manufacturing has increasingly relied on surface mount technology (SMT) to achieve high-density and high-precision component placement. However, the highmix, low-volume characteristics of LED production substantially increase the complexity of SMT scheduling and hinder efficient production planning. This study proposes a hybrid fuzzy genetic algorithm with Tabu search (HFGATS) to address the SMT scheduling problem (SMTSP), which is modeled as a flow shop scheduling problem with several real-world constraints, such as batch splitting, machine availability, material availability, designated machine rule, and sequence-dependent setup time constraints. Of the constraints considered, batch splitting poses the most critical limitation. In the LED industry, certain products involve large order volumes, necessitating batch splitting in advance to prevent excessive concentration of production on a single machine. However, batch splitting inevitably increases the number of machine setups. Therefore, decisions regarding batch splitting strongly affect the overall effectiveness of a scheduling system. The proposed framework integrates a fuzzy logic controller to adaptively tune the parameters of the genetic algorithm, and Tabu search is incorporated to enhance local search capability. To validate the proposed HFGATS, an empirical study was conducted using real production data collected from the SMT station of a Taiwanese LED manufacturer. The HFGATS obtained optimal solutions for small-scale problems; for large-scale problems, it outperformed traditional dispatching rules and general metaheuristic algorithms by minimizing total tardiness. This improvement can enhance the overall on-time delivery rate of the factory, thereby increasing customer satisfaction and improving the factory's profitability.

Semiconductor manufacturing is an energy-intensive process with long cycle times (CTs) and high carbon emissions. This study developed a predictive optimization framework aimed at improving the operational efficiency and environmental performance of semiconductor manufacturing. In this framework, a bidirectional long short-term memory model with attention mechanisms is used to forecast CT trends, and Gauss–Newton regression is used to capture the nonlinear relationship between CT and work in progress. The CT predictions and the captured relationship are then input to a mixed-integer nonlinear programming model to generate optimal lot release strategies. In evaluation experiments, the proposed framework substantially reduced the average CT across three 12-inch wafer fabs in Taiwan. From a life-cycle perspective, shortening CT leads to lower energy usage per wafer, shorter machine idle time, and improved capacity utilization, resulting in measurable decreases in resource consumption and emissions during semiconductor manufacturing. This study contributes to the achievement of net-zero manufacturing by embedding data-driven sustainability into production planning and highlights the synergy between smart scheduling and life-cycle efficiency in semiconductor operations.

As semiconductor fabrication technologies advance and critical dimensions (CD) shrink, process stages become increasingly complex, placing more stringent demands on scheduling throughput optimization and resource allocation. Simultaneously, environmental sustainability has emerged as an industry imperative: manufacturers must boost productivity, conserve energy, and control carbon emissions. This study focuses on the practical the semiconductor ion implantation stage scheduling problem (IISP), accounting for machine setup times, availability, and job priorities under multiple constraints, and proposes an intelligent scheduling framework based on reinforcement learning.

This study integrates double Q-learning with invalid action masking (DQIAM) method, which employs two Q-value estimators alongside an action-masking mechanism to filter out infeasible choices. The composite reward function minimizes the makespan, setup count, and scrap quantity, improving scheduling performance and production flexibility. To validate the performance of the proposed DQIAM approach, this study compares DQIAM with the shortest processing time (SPT) heuristic, genetic algorithm (GA), particle swarm optimization (PSO), standard Q-learning (QRL), and invalid action masking QRL (IQRL). The results show that DQIAM outperforms all benchmark methods on key performance indicators, achieving reduced completion times, fewer setup operations, and lower scrap quantities while providing stable and adaptive scheduling decisions.