Integration of Artificial Intelligence for Real-Time Fault Detection in Semiconductor Packaging

Semiconductor manufacturing comprises hundreds of consecutive unit processes. A single misprocess could jeopardize the whole manufacturing process. In current manufacturing environments, data monitoring of equipment condition, wafer metrology, and inspection, etc., are used to probe any anomaly during the manufacturing process that could affect the final chip performance and quality. The purpose of investigation is fault detection and classification (FDC). Various methods, such as statistical or data mining methods with machine learning algorithms, have been employed for FDC. In this paper, we propose an artificial immune system (AIS), which is a biologically inspired computing algorithm, for FDC regarding semiconductor equipment. Process shifts caused by parts and modules aging over time are main processes of failure cause. We employ state variable identification (SVID) data, which contain current equipment operating condition, and optical emission spectroscopy (OES) data, which represent plasma process information obtained from faulty process scenario with intentional modification of the gas flow rate in a semiconductor fabrication process. We achieved a modeling prediction accuracy of modeling of 94.69% with selected SVID and OES and an accuracy of 93.68% with OES data alone. To conclude, the possibility of using an AIS in the field of semiconductor process decision making is proposed.

In the era of nanotechnology, variability detection and control in the manufacturing process is vital in delivering products at high yields and at optimal cost. At Intel, DFM techniques are utilized to combat variability at all stages of product delivery-design, tape-out, mask generation, fab processing and assembly. Statistical process control (SPC), advanced process control (APC) and fault detection and classification (FDC) are examples of key capabilities used in the fab for variability detection, containment and correction on processing equipment and manufacturing process. This paper will focus on adaptive metrology sampling (AMS) and its role in variability detection and control during fab processing. AMS includes lot, wafer and field level metrology sampling strategies that varies the amount of sampling based on previous observations which include fab events and/or process control feedback. Adaptive sampling if implemented appropriately will improve precision in excursion detection, reduce detection time, minimize the quantity of silicon at risk and yield losses from such excursions. This paper will illustrate key strategies and techniques that should be considered for a successful AMS implementation. This includes details on system architecture, data inputs, sample material selection rules, and metrology capacity management strategy.

Test of covariance changes without a large sample and its application to fault detection and classification

The proposed paper presents an equipment level FDC system for the optimization and anomaly detection of semiconductor equipment. Through the equipment level FDC(Fault Detection and Classification) system, various data in the equipment can be converted into meaningful and accurate analysis data through context mapping to facilitate analysis of the management and condition of the equipment. In addition, it is possible to proactively identify and respond to problems at the equipment level before identifying and responding to problems on the host by processing data of diverse equipment in real time.

Kernel approaches for fault detection and classification in PARR-2

This research proposes an application of generative adversarial networks (GANs) to solve the class imbalance problem in the fault detection and classification study of a plasma etching process. Small changes in the equipment part condition of the plasma equipment may cause an equipment fault, resulting in a process anomaly. Thus, fault detection in the semiconductor process is essential for success in advanced process control. Two datasets that assume faults of the mass flow controller (MFC) in equipment components were acquired using optical emission spectroscopy (OES) in the plasma etching process of a silicon trench: The abnormal process changed by the MFC is assumed to be faults, and the minority class of Case 1 is the normal class, and that of Case 2 is the abnormal class. In each case, additional minority class data were generated using GANs to compensate for the degradation of model training due to class-imbalanced data. Comparisons of five existing fault detection algorithms with the augmented datasets showed improved modeling performances. Generating a dataset for the minority group using GANs is beneficial for class imbalance problems of OES datasets in fault detection for the semiconductor plasma equipment.

Tremendous amounts of equipment parameters and sensor values were generated during modern semiconductor wafer manufacturing. These process data were utilized for early diagnosis of anomalies to prevent subsequent yield loss. However, process data from different steps and machines were highly customized that it is of great difficulties for traditional fault detection and classification (FDC) analysis to find a universal model to identify process excursions. In this paper, we present a neural network method with deep convolutional variational autoencoder structures which used reconstruction error as abnormal scores. This method exhibited a more reliable detection precision than FDC method and traditional anomaly detection algorithms. Furthermore, the reconstruction error of different sensors could be used to identify the root cause of abnormal processes. It was found that similarly structured networks could be applied to different processing steps, which enable this method to be accepted as a standard method to process data.

Nowadays, more attention has been placed on cost reductions and yield enhancement in the semiconductor industry. During the manufacturing process, a considerable amount of sensor data called status variables identification (SVID) is collected by sensors embedded in advanced machines. This data is a valuable source for data-driven automatic fault detection and diagnosis at an early manufacturing stage to maintain competitive advantages. However, wafer processing times vary slightly from wafer to wafer, resulting in variable-length signal data. The conventional approaches use much condensed data called fault detection and classification (FDC) data made by manually designed feature extraction. Or, recent deep learning approaches assume that all wafers have the same processing time, which is impotent to the variable-length SVID. To detect and diagnose faults directly from the variable-length SVID, we propose a self-attentive convolutional neural network. In experiments using real-world data from a semiconductor manufacturer, the proposed model outperformed other deep learning models with less training time and showed robustness at different sequence lengths. Compared to FDC data, SVID data showed better fault detection performance. Without manually investigating the lengthy sensor signals, abnormal sensor value patterns were found at the time specified by the model.

Fault Detection and Diagnosis (FDD) in complex industrial systems requires methods that can handle uncertain operating conditions, soft thresholds, evolving sensor behavior, and increasing volumes of heterogeneous data. Traditional model-based or rule-driven approaches offer interpretability but lack adaptability, while purely data-driven and Large Language Model (LLM)-based methods often struggle with consistency, traceability, and causal grounding. Dynamic Master Logic (DML) provides a causal and temporal reasoning structure with fuzzy rules that capture gradual drift, soft limits, and asynchronous sensor signals while preserving traceability and deterministic evidence propagation. Building on this foundation, this paper presents a DML-LLM hybrid architecture that integrates targeted LLM inference to interpret unstructured information such as logs, notes, or retrieved documents under controlled prompts that maintain domain constraints. The combined system integrates Bayesian updating, deterministic routing, and semantic interpretation into a unified FDD pipeline. In a semiconductor manufacturing case study, the proposed framework reduced time to detection (TTD) from 7.4 h to 1.2 h and improved the F1 score from 0.59 to 0.83 when compared with conventional Statistical Process Control (SPC) and Fault Detection and Classification (FDC) workflows. Provenance completeness increased from 18% to 96%, while engineer triage time was reduced from 72 min to 18 min per event. These results demonstrate that the hybrid framework provides a scalable and explainable approach to anomaly detection and fault diagnosis in sensor-rich industrial environments.

At present, the FDC (Fault Detection and Classification) system is controlled by the engineers manually setting one-dimensional SPEC according to the process experience. This control mode only controls the vertical data in a rough range, and cannot accurately control the horizontal data. At the same time, due to too many parameters received by the sensor, engineers cannot manually set the SPEC one by one, which may lead to certain risks of omission and error.

Advanced packaging has a challenge on its hands in dealing with reliability issues caused during singulation. Specifically, in the case of Wafer Level Chip Scale packaging (WLCSP) where the sawn die edge is the package edge and is exposed. WLCSP manufactured on the newer nodes are based on Ultra-low K dielectric material. This material is inherently porous and mechanically fragile. During the singulation step, there is a likelihood of creating chipping and delamination. Rudolph have developed a comprehensive approach to detecting defects generated at the laser grooving and saw process steps. There are two recommended approaches. The first approach is an inspection approach that is focused on detecting the defects after it has occurred. The second, a more sophisticated approach is to attempt to pre-empt occurrence of defects. In the first approach-die seal ring inspection, algorithms have been developed around the die seal ring to ensure the die integrity. These algorithms not only include the capability to detect the defect but also measure the defect size, defect position and frequency of defects. All of this is done while minimizing false defects that are typically found due to street and reticle artifacts that are often falsely detected as chipping. This enables customers to perform advanced disposition to ensure the optimal balance between shipping good die and preventing the shipment of potential reliability fails. The second approach — Kerf metrology and Saw control comprises of Fault detection and classification (FDC) to record equipment data during the singulation processes and inspection equipment to detect surface defects such as chips, delamination and cracks. The combination of FDC to monitor equipment performance and the ability to correlate Metrology/Inspection performance data enables customers to not only detect defects but preemptively eliminate them. Poorly singulated dies can now be analyzed all the way back to the process equipment signals. Once these signals are identified they can be monitored to prevent or alert operators of possible excursions. These analytics will help customers define and ramp robust processes, justify consumable costs and ultimately lead to a better understanding of this difficult and challenging process. Finally, specific YMS capabilities exist to align and analyze the collected equipment and process data.

Virtual metrology (VM) is an enabling technology capable of performing virtual inspection on the metrology quality of wafers. Instead of physically acquiring the metrology measurements, VM applies conjecture models on the process data of wafers to estimate the measurements of the targeted metrology variables. Prior works on overlay VM system utilized fault detection and classification (FDC) data as the process data for the conjecture models. Hence, when FDC data are unavailable owing to FDC system enhancement works, FDC-based VM models would be rendered inefficacious. During such events, a competent VM system using a different modeling approach is required to sustain the production line until FDC data resumes availability and FDC-based VM reaches production state. Motivated by a real-world production environment of a 200mm semiconductor manufacturing plant (fab), a novel wafer lot-level modeling approach for overlay VM was proposed in our prior work. Using the proposed modeling, a smart sampling scheme was also designed in the same work. The smart sampling scheme consists of two conjecture tasks, with the first task classifies the overlay quality of the wafers, and the second task estimates the overlay errors of the wafers classified with normal overlay quality. The abnormal ones are diverted to the physical metrology station. In this paper, the implementation of a smart sampling system, C2O, using the designed scheme and its experimental results are presented. The experimental results showed that C2O is capable to achieve a true positive rate (TPR) of 71.34% for the classification task and mean absolute scaled error (MASE) of 9.59 for the regression task. The obtained results set the baseline to measure the efficacy of future enhancement works, which have been enlisted and underway to augment the performance of the system so that its competency meets the requirements of real fab.

This paper discusses several practical issues encountered when applying principal component analysis (PCA) to fault detection and classification (FDC). Those issues presented a challenge to the success of PCA-model based process monitoring. Weighted PCA is proposed to address the issues. Two forms of weighted PCA are discussed: sample-wise and variable-wise. Sample-wise weighted PCA is used to address issues with model updating. By adapting models with process changes, the long-term validity of a PCA model can be maintained. We also propose to deal with drift and shift features separately. Variable-wise weighted PCA is used to incorporate process and sensor knowledge. PCA models built this way require less maintenance and result in better FDC performance. We performed comparison studies on plasma etcher FDC. We have shown that weighted PCA can adapt to process drift, reduce the occurrence of false alarms and make the models easy to maintain.

Applications of Artificial Immune System (AIS) has been widely applied in various engineering, including network security, pattern recognition, combinational optimization, machine learning and fault diagnosis, etc. Fault diagnosis is another AIS application field directly mapped from the theory of immunity after information security and has made certain achievements in research. Real-valued Negative selection algorithms (RNSA) of AIS generate their detector sets based on the points of self data. Self data is regarded as the normal pattern of behavior of the monitored system. This paper provide a new fault detection method based on RNSA of artificial immunity. It can effectively overcome the deficiency of the various fault detection methods of today that cannot implement fault detections because there are only normal samples and not enough fault samples and short of the function of continuous learning. The test result shows that, by increasing a certain number of training samples, the accuracy of fault diagnosis has made great changes. This way has obvious advantage in robustness and accuracy in detection and shows a favorable prospect of application.

Tool behavior modeling and diagnosis is a big challenge in modern semiconductor fabrication, in particular, with high product-mix and complicated technology nodes. Tool condition monitoring has been long conducted by implementing the Fault Detection and Classification (FDC) system and analyzing the large amount of real-time sensor data collected during the process. The tool condition hierarchy developed in the previous work proposed that the excursions can be firstly detected by an overall condition indicator and then intuitively traced down to the level of sensor groups. In this paper, a Run-to-Run (R2R) variation monitoring technique is developed in order to correlate the tool excursions with individual sensors, instead of sensor groups, and thus to close the diagnostic gap in the hierarchy. Therefore, the tool condition can be efficiently monitored by one overall indicator and the detected tool faults can be systematically diagnosed at the sensor level.