This paper proposes a new Run-to-Run (R2R) control framework based on deep deterministic policy gradient (DDPG) for the mixed-product production mode in semiconductor manufacturing. The DDPG algorithm is particularly developed to configure a deep reinforcement learning environment well suited to mixed-product production modes. To address the challenges posed in deep reinforcement learning, three enhanced mechanisms have been developed to improve the training of the proposed DDPG model for mixed-product R2R applications. These mechanisms include a piece-wise reward function, training with dynamic targets, and the new recall principle. It is demonstrated from the comprehensive simulation results that the proposed R2R control framework outperforms five noted mixed-product R2R control algorithms in the literature. The research outcome of this paper signifies a promising viability of deep reinforcement learning for highly complex and dynamic environments with continuous action spaces in the mixed-product R2R practice.

This paper proposes a novel self-supervised clustering framework for automatic defect annotation in semiconductor manufacturing, aiming to reduce the heavy reliance on expert-labeled data. Building upon the Semantic Clustering by Adopting Nearest-neighbors (SCAN) model, we introduce a new SCAN methodology that integrates three key innovations: a redesigned pretext task, an enhanced SCAN model with neighbor loss, and a self-labeling mechanism. The new pretext task replaces the traditional ResNet18 backbone with MobileNetV2 and incorporates focal loss to address class imbalance, enabling more effective feature extraction from unlabeled wafer maps. The SCAN model is further refined through a neighbor loss function that leverages high-confidence samples to iteratively correct pseudo-labels, improving clustering accuracy. Additionally, we apply strong augmentation techniques and a confidence-based thresholding strategy to enhance model robustness. Experimental evaluations on the WM-811K and MixedWM38 datasets demonstrate that the proposed unsupervised clustering framework achieves a notable accuracy improvement (from 36.5% to 57.7% and 36.09% to 57.69%, respectively) compared to the original SCAN, with further gains up to 66.20% and 64.80% using high-confidence sample filtering. The results confirm that our approach significantly enhances classification performance in unsupervised settings and offers a practical solution for scalable, automated defect labeling in intelligent manufacturing environments.

Virtual metrology enables semiconductor manufacturers to continuously monitor and predict the quality and performance of semiconductor production during the manufacturing process. This kind of real-time monitoring tool is essentially important for maintaining process stability and ensuring consistent product quality. In this paper, a new rolling-window and production-maintenance-based approach is proposed on a weekly basis to establish the virtual metrology system. The proposed thickening and dewatering framework serves as the central core of the virtual metrology model, enabling real-time information and feedback to aid in making timely adjustments to the processing tool within machinery limitations. A meticulous approach to addressing each component within this innovative framework has been instrumental in achieving exceptional VM performance. The training dataset is routinely updated as the production maintenance is taking place to facilitate model retraining. A practical challenge to the domain engineer under certain circumstances where the process parameter considered in the virtual metrology model is realistically run out of specification is also rigorously investigated in this research. For the evaluation purposes, a real-world industrial case of the chemical vapor deposition process in semiconductor manufacturing is presented to illustrate the proposed approach. The proposed virtual metrology approach shows a substantial improvement in mean absolute error with 39.99%, 26.26%, and 30.75% for three different recipes, respectively, as compared to the base model currently implemented on the operation site.

The current thin film transistor liquid crystal display (TFT-LCD) panel industry evolves in transition from volume production to customized production service as a major competition strategy. Owing to the highly mixed production type of small batch sizes and high varieties, metrology data are usually collected too infrequently for building accurate virtual metrology (VM) models. In this paper, a new product-to-product (P2P) VM model to predict photoresist spacer heights in the color filter process of the array sector is developed. First, a position-based random forests model is proposed for the base product. Second, an ensemble model combining the base model and a biased estimation random forests model for the new product of small batch size is proposed to perform P2P VM between different products on the same tool. The real datasets of photoresist spacer heights for different products are used to illustrate the accuracy of the new P2P VM framework in the experimental study. Note to Practitioners—As global competition intensifies, production management needs to be adjusted to adapt swiftly to changes in the time-varying and competitive market. For manufacturing firms, how to operate the manufacturing system effectively and efficiently plays a crucial role in quality and yield improvements. The color filter process in the array sector is of critical importance in the TFT-LCD practice since it essentially dictates the quality level of subsequent processing steps in the cell and assembly sectors. Virtual metrology is a state-of-the-art measure that can be used for monitoring and controlling purposes to maintain the yield and productivity of the color filter process in the highly mixed production environment. An immediate challenge to be faced by the data scientists is how to build accurate virtual metrology models for a variety of product lines separately in color filter manufacturing. A new proposal of product-to-product virtual metrology is first addressed between products, enabling to build appropriate virtual metrology models for new products with small-sized production batches.

With the rapid development of the semiconductor industry, fault diagnosis is an important task in routine operations to determine the root cause for faults that occur. A tool in manufacturing processes is equipped with a wide variety of sensors that record different types of process data. Wafers being processed are accompanied by a considerable amount of process data as a multivariate time series. Identifying the key process parameters and the corresponding process operations where faults occur can be used to facilitate the tasks of monitoring the process, maintaining the stability of the process, and stabilizing wafer production yield. This article proposes a novel solution procedure for fault diagnosis of wafer acceptance test (WAT) and chip probing (CP) using machine learning (ML). Based on the process flow of wafers and the corresponding process data, a sampling method, called synthetic minority oversampling technique (SMOTE), is first used to augment classification models with an imbalanced process dataset, and the best-practice SMOTE ratio is sought by using four competitive classifiers. By means of principal component analysis (PCA), the original data are transformed into visualizations to explore the distributions of good and bad lots of wafers. Based on the comparison of the four classifiers used in the test, the proposed logistic regression with data augmentation (LR-SMOTE) performs best. The ten most important features identified by using the LR are collected to determine potential failure operations. The identified failures in operations returned by using the proposed solution procedure for fault diagnosis of WAT and CP were confirmed by the engineers who work in the domain. Note to Practitioners—In semiconductor manufacturing practice, during the WAT and CP, domain engineers need to trace upstream the key process parameters and the corresponding key operations in the front-end-of-line (FEOL) process if any faults are detected in the back-end-of-line (BEOL) process. In fact, the FEOL and the BEOL are located in completely separate plants with different clean room standards. Bridging the information gap between these two different processes merits thorough scrutiny. Another challenge posed by fault diagnosis of WAT and CP is that the total number of features (i.e., process parameters) involved in the FEOL can be tens of thousands, leading to an adverse effect of data sparsity on model building.

To monitor process and identify the deviation as early as possible, data-driven methods have been applied for process monitoring and fault detection in semiconductor manufacturing. Although various fault detection and classification models had been discussed in the literature, however, little research has been devoted to feature selection from trace data that is important for process monitoring of natural variation. Additionally, the high-mix production mode with different recipes leads to process dynamic of wafer-to-wafer (W2W) variation which should also be identified for safeguarding false alarms and serving as a warning indicator. Therefore, this paper proposes a data-driven framework to identify the key features with respect to the W2W variation. In particular, the self-organizing map is used to annotate the grade of wafer variation among the in-line metrology data. Subsequently, the adaptive boosting (AdaBoost) is adopted to examine the effectiveness of every feature and its processing times, respectively. To validate the proposed framework, an empirical study from a semiconductor fabrication plant is conducted. The experimental results demonstrate that the key feature identification is of critical importance to build highly capable models for process monitoring. Through the dimensionality reduction technique, it has been illustrated that a smaller set of the identified key features are able to pinpoint the W2W variation of different wafer grades more clearly than the whole set of process features.Note to Practitioners— Process monitoring has become more difficult with the shrinking linewidth in semiconductor manufacturing. The challenges of analyzing equipment sensor or raw trace data for process monitoring in high-mix manufacturing processes are to incorporate subject-matter expert knowledge for setting control limit meticulously, to detect the subtle changes by analyzing the whole trace data profile, and to identify W2W variation for reducing false alarms. This paper proposes a data-driven framework for process monitoring by adopting data-driven approaches without recourse to domain judgement. Experimental results demonstrate that the proposed data-driven framework can effectively identify the key features via sensor readings and corresponding processing times, respectively. The engineers can make use of the extracted features to perform a predictive monitoring on metrology data for detection of potential process deterioration.

Mixed-product production is common practice in semiconductor manufacturing and has attracted considerable academic attention recently. The main purpose of the mixed-product mode is to manufacture a wide variety of products in the same time period, to meet customer demand. The complex and changeable group of products is called “mixed-product production mode.” A specific tool used and a specific product manufactured are together called a “thread.” There is a need for “threaded” Run-to-Run (RtR) control in semiconductor manufacturing. This article develops a new product-based, disturbance-accumulating strategy in threaded RtR control by means of the double exponentially weighted moving average (dEWMA) technique. The main idea is to develop two disturbance-accumulating strategies; one for the intercept update of the same product between cycles and the other for the disturbance update in terms of the previous product between cycles, cope with the transient and transition effects, respectively. A comprehensive simulation study is conducted to compare the performance of the proposed product-based RtR controller with the existing controls specifically designed for the mixed-product mode. The proposed control has great promise as an extremely competitive threaded RtR controller in practice.Note to Practitioners—Run-to-run (RtR) control has been widely applied in batch manufacturing processes to reduce variation and has been identified as a turn-key technology for maintaining quality outputs in semiconductor manufacturing. The mixed-product production mode poses a great challenge of how to adjust process recipes between different products to alleviate initial recipe bias, process shifts, and patterned disturbances from run to run, product to product, and cycle to cycle. The proposed mixed-product RtR controller is developed based on a novel disturbance-accumulating, product-based strategy. This new controller is easy to implement in practice and has been shown to be effective in producing high-quality control outputs.

Fault detection and classification (FDC) is important for semiconductor manufacturing to monitor equipment's condition and examine the potential cause of the fault. Each equipment in the semiconductor manufacturing process is often accompanied by a large amount of sensor readings, also called status variable identification (SVID). Identifying the key SVIDs accurately can make it easier for engineers to monitor the process and maintain the stability of the process and wafer productive yields. This article proposes using the random forests algorithm to analyze the importance of SVIDs of equipment sensors, automatically filters the key SVID by using k-means, and integrates various machine learning methods to verify the key SVIDs and identify key processing time and steps. Upon the key parameters are identified, the key processing time and steps are investigated subsequently. The ensemble models constructed on k-nearest neighbors (kNNs) and naïve Bayes classifiers are presented for classifying wafers as normal or abnormal. Data visualization of multidimensional key SVIDs is performed by using t-distributed stochastic neighbor embedding (t-SNE) to create a graphical aid in FDC for the process engineer. An empirical study is conducted to validate the proposed data-driven framework for fault detection and diagnostic. The experimental results demonstrate that the proposed framework can detect abnormality effectively with highly imbalanced classes and also gain insightful information about the key SVIDs and corresponding key processing time and steps.