SGT-MOS device processing control method and system based on internet of things

Abstract

The application discloses an SGT-MOS device processing control method and system based on an Internet of Things, and relates to the technical field of semiconductor power devices. The method comprises the following steps: performing structural defect feature identification on an SGT-MOS device, and establishing a defect space structure; performing Internet of Things equipment matching mapping based on the defect space structure, determining an equipment matching relationship, deploying sensors on the defect space structure by utilizing the equipment matching relationship, and collecting defect sensing data; fusing and analyzing the defect sensing data and conventional sensing data, identifying dynamic processing regulation data, and generating processing regulation instructions. The technical problems of low yield and poor performance consistency caused by the difficulty in timely sensing and accurate regulation of structural defects in the existing SGT-MOS device processing are solved, the technical effect of realizing defect-driven dynamic process parameter optimization through Internet of Things multi-source data fusion is achieved, and the yield and electrical performance consistency of device processing are improved.

Description

Technical Field

[0001] This invention relates to the field of semiconductor power device technology, and more specifically to an IoT-based SGT-MOS device fabrication control method and system. Background Technology

[0002] With the rapid development of power electronics technology, SGT-MOS devices, due to their excellent conduction characteristics, high withstand voltage, and low on-resistance, have become an important development direction for next-generation high-efficiency power devices. These devices have broad application prospects in new energy vehicles, power electronics, smart grids, and high-frequency communication equipment. However, the manufacturing process of SGT-MOS devices is complex, involving multiple key steps such as trench etching, gate oxide growth, polysilicon filling, and metallization interconnection. The processing accuracy and structural consistency directly affect the electrical performance and reliability of the devices. In actual production, due to the influence of equipment operating conditions, environmental factors, and material fluctuations, microstructural defects are prone to occur inside the devices, such as trench morphology deviations, uneven gate oxide thickness, and polysilicon voids. These defects are difficult to identify and respond to in real time using traditional detection methods, leading to decreased product yield and increased performance dispersion. Summary of the Invention

[0003] This application provides an IoT-based SGT-MOS device fabrication control method and system, which solves the technical problems of low yield and poor performance consistency caused by the difficulty in timely detection and precise control of structural defects in the existing SGT-MOS device fabrication process.

[0004] The first aspect of this application provides a method for controlling the fabrication of SGT-MOS devices based on the Internet of Things, the method comprising:

[0005] Structural defect features of the SGT-MOS device are identified to establish a defect space structure; IoT device matching and mapping are performed based on the defect space structure to determine device matching relationships; sensor deployment is carried out on the defect space structure using the device matching relationships to collect defect sensing data; the defect sensing data and conventional sensing data are fused and analyzed to identify dynamic processing control data and generate processing control instructions.

[0006] A second aspect of this application provides an IoT-based SGT-MOS device fabrication control system, the system comprising:

[0007] Defect identification module: identifies structural defect features of the SGT-MOS device and establishes a defect space structure; Data acquisition module: performs IoT device matching mapping based on the defect space structure, determines device matching relationships, uses the device matching relationships to deploy sensors on the defect space structure, and collects defect sensing data; Fusion and analysis module: performs fusion and analysis of the defect sensing data and conventional sensing data, identifies dynamic processing control data, and generates processing control instructions.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] First, structural feature analysis is performed on the SGT-MOS device to identify potential defect types and distribution characteristics. Then, IoT technology is used to intelligently match various sensors with the defect spatial structure, determining the correspondence between different devices and defect areas. Based on this relationship, appropriate sensor deployment is implemented to achieve real-time monitoring and data acquisition of key processing steps. Next, the collected defect-related sensor data is fused and analyzed with conventional process monitoring data. Through data correlation and dynamic modeling, key parameters requiring adjustment during processing are identified, and corresponding processing control commands are generated to achieve real-time optimization and intelligent adjustment of the SGT-MOS device processing. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic flowchart of the IoT-based SGT-MOS device fabrication and control method provided in the embodiments of this application.

[0012] Figure 2 A schematic diagram of the IoT-based SGT-MOS device fabrication and control system provided in this application embodiment.

[0013] Figure labeling: Defect identification module 11, data acquisition module 12, fusion analysis module 13. Detailed Implementation

[0014] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0015] Example 1, as Figure 1 As shown, this application provides an IoT-based SGT-MOS device fabrication control method, the method including:

[0016] Structural defect features of the SGT-MOS device are identified, and a defect spatial structure is established.

[0017] In this embodiment, the structural features of the SGT-MOS device sample to be processed are first identified and collected. These structural features include, but are not limited to, key dimensions and material parameters such as trench aspect ratio, gate oxide thickness, polysilicon fill height, substrate doping concentration, and source metal coverage uniformity. Raw data for each structural feature can typically be obtained using online inspection equipment such as laser thickness gauges, optical interferometers, and scanning electron microscopes. The collected structural features are then matched and analyzed with defect sample data in a historical processing database to obtain historical defect samples corresponding to the structural features. Subsequently, these historical defect samples are combined with the geometric space model of the SGT-MOS device to establish a defect space structure. This defect space structure describes the distribution area, range, and severity of defects within or on the surface of the device in three-dimensional coordinates. This enables spatial visualization of defect features, guiding the selection of sensor types and spatial deployment strategies, and providing data support for subsequent processing control.

[0018] Furthermore, structural defect feature identification is performed on the SGT-MOS device to establish a defect spatial structure, including:

[0019] Identify the structural features of the SGT-MOS device and collect historical defect samples corresponding to the structural features; perform defect feature and defect probability identification analysis on the historical defect samples corresponding to the structural features to generate structural defect feature parameters; project the structural defect feature parameters into the SGT-MOS device spatial structure to establish the defect spatial structure.

[0020] Preferably, firstly, by combining design specifications and actual test data, the structural features of the SGT-MOS device to be analyzed are identified to obtain the structural characteristics of the SGT-MOS device. Then, based on the identified structural features, defect sample data for each structural feature is retrieved from a historical processing database. This defect sample data contains defect records for each structural feature formed under different process parameters, equipment conditions, or material batches. Based on these defect records, a historical defect sample set corresponding to the structural feature can be constructed. Next, multi-dimensional feature analysis is performed on the historical defect sample set. In this process, defect image feature extraction algorithms, such as convolutional neural networks pre-trained through forward propagation, loss calculation, backpropagation, and parameter optimization, are used to analyze the historical defect sample set, quantifying the morphological features of each defect to form defect feature parameters. These defect feature parameters typically include the defect's size, shape coefficient, contour complexity, spatial location parameters, and grayscale of the defect area. Simultaneously, structural features are set as input node variables in the Bayesian network, and defect types are set as output node variables. Combined with expert knowledge and data-driven algorithms, such as the K2 algorithm, the dependencies between nodes are determined. Conditional probability distributions of each node under its parent node conditions are calculated using maximum likelihood estimation or Bayesian parameter estimation. For example, a quantitative result such as "when the trench aspect ratio is greater than 2.5 and the etching rate is higher than a threshold, the probability of trench collapse is 0.72" can be obtained. After training, the Bayesian network is used to perform defect probability inference on the identified structural features, obtaining the defect probabilities of different defect types under a given structure. By summarizing these defect probabilities with the previously obtained defect features, a structural defect feature parameter can be obtained. Then, based on the device's three-dimensional geometric coordinate system, each structural defect feature parameter is mapped to the corresponding physical region in the SGT-MOS device spatial structure model, forming a defect spatial structure with spatial resolution. This defect spatial structure reflects the true spatial distribution characteristics of internal and surface defects of the device, providing a foundation for subsequent IoT device matching mapping and sensing deployment.

[0021] Furthermore, historical defect samples corresponding to structural features are collected, including:

[0022] Matching the structural features with historical defect size ranges yields a set of defect samples with matching sizes and a set of defect samples with missing sizes, wherein the set of defect samples with missing sizes has a corresponding set of defect samples with matching sizes in its neighborhood. Based on the missing size and structural feature processing response coefficients corresponding to the set of defect samples with missing sizes, sample operations are performed on the set of defect samples with matching sizes in the neighborhood to supplement the defect sample set, thus obtaining a supplementary defect sample set. The set of defect samples with matching sizes and the supplementary defect sample set are then mapped using structural features to obtain the historical defect samples.

[0023] Optionally, the structural features of the SGT-MOS device are first divided into size ranges. Specifically, based on design specifications and process tolerances, the main structural features of the device, such as trench aspect ratio, gate oxide thickness, and polysilicon fill height, are divided into several continuous historical defect size ranges. Each historical defect size range corresponds to a relatively stable processing feature region, and corresponding defect records are stored in a historical processing database. Then, the currently identified structural feature parameters are matched one by one with the historical defect size ranges to obtain corresponding historical defect samples. For portions where corresponding size range defect records can be found in the historical processing database, a matching size defect sample set is extracted. For portions lacking sufficient samples or with incomplete records within certain size ranges, they are marked as missing size defect sample sets. The neighborhood matching size defect sample sets of these missing size defect sample sets are then matched using the missing size and structural feature processing response coefficients. Subsequently, to ensure data integrity and the reliability of statistical analysis, a feature-level transfer operation is performed on the neighboring matching-size defect sample set based on the missing dimensions and corresponding structural feature processing response coefficients of the matching-size defect sample set, generating a supplementary defect sample set. This supplementary defect sample set maintains statistical distribution consistency with the real defect samples. Finally, the matching-size defect sample set and the supplementary defect sample set are fused and mapped. That is, by establishing a unified structural feature index system, the sample features of different size ranges are normalized and correlated, mapping all sample data to the structural feature space of the SGT-MOS device, thereby obtaining a complete set of historical defect samples. Through this step, multi-dimensional completion and accurate mapping of defect sample data are achieved, effectively solving the model bias problem caused by discontinuity and missing data in traditional defect samples, and ensuring the accuracy and integrity of the defect space structure.

[0024] Furthermore, before obtaining a sample set of defects with matching sizes by matching the structural features with historical defect size ranges, the process includes:

[0025] Structural feature data and corresponding defect data of SGT-MOS devices during historical processing are collected to form an initial historical defect sample set. Processing response coefficient analysis is performed on the structural features to identify their sensitivity to processing fluctuations and one or more key dimensions strongly correlated with defect incidence. These key dimensions include, but are not limited to, trench aspect ratio, gate oxide thickness, and polysilicon fill height. Based on these key dimensions, the design specification space of the SGT-MOS devices is divided into multiple continuous size intervals. The initial historical defect sample set is then classified according to these size intervals, resulting in multiple size interval samples, each with a structural feature processing response coefficient label.

[0026] Optionally, process records and inspection results of historical processing batches are first extracted from the historical processing database, including SGT-MOS device structural feature data for each process step and corresponding defect data, forming an initial historical defect sample set containing multiple batches and multiple process conditions. Then, statistical modeling and data fitting methods are used to calculate the sensitivity of each structural feature to processing fluctuations. Specifically, a linear regression model is constructed based on each structural feature and its corresponding defect incidence rate. The historical structural feature parameters and historical defect incidence rates from the initial historical defect sample set are then input into this linear regression model for fitting, and the intercept and slope in the linear regression model are obtained using the least squares method. By dividing the final fitted slope by the maximum slope generated during the fitting process, the processing response coefficient of the structural feature parameter is obtained, which is used to quantify the sensitivity of the structural feature parameter to defects. Afterwards, the calculated processing response coefficient is compared with a set threshold to identify one or more key dimensions highly correlated with the defect incidence rate. These key dimensions include, but are not limited to, trench aspect ratio, gate oxide thickness, and polysilicon fill height. Then, based on the device design specifications and process tolerance ranges, the value ranges of each key dimension in the design specification space of the SGT-MOS device are divided into equal-step or non-linear segments to form multiple continuous dimension ranges. Each dimension range corresponds to a set of structural feature parameter value ranges to distinguish different processing feature regions. For example, the trench depth-to-width ratio can be divided into ranges such as 1.5–2.0, 2.0–2.5, and 2.5–3.0, and the gate oxide thickness can be divided into ranges such as 10–15nm and 15–20nm. Finally, based on the dimension ranges, the corresponding structural feature values ​​are extracted for each sample data in the initial historical defect sample set, and the dimension range in which it belongs is determined. Samples falling into the same range are grouped into the same category, forming multiple dimension range sample sets. At the same time, a processing response coefficient label is added to each interval sample set to describe the process sensitivity and defect correlation strength of the structural features within that dimension range. Through the above steps, a multi-interval historical defect sample system containing structural features, defect features, and processing response information is established, realizing the structured and hierarchical management of processing data and providing mathematical support and physical reference for subsequent sample supplementation and spatial mapping.

[0027] Furthermore, based on the missing dimensions and structural feature processing response coefficients corresponding to the missing size defect sample set, sample operations are performed on the neighborhood-matched size defect sample set to obtain a supplementary defect sample set, including:

[0028] Based on the missing size and structural feature processing response coefficient, the size range and process sensitivity threshold of the target supplementary region are determined; based on size similarity and process response similarity, a dual matching is performed to select a neighborhood matching size defect sample set, and feature-level transfer operation is performed on the neighborhood matching size defect sample set; based on the features after the transfer operation, the supplementary defect sample set is generated, and the supplementary defect sample set contains a complete process-defect mapping relationship.

[0029] Optionally, for the missing size range in the historical defect sample set, the processing response coefficients of the corresponding structural features are extracted. Then, based on the trend of response coefficient changes between the missing size and the size of the neighboring range, the size range of the target supplementary region is determined. For example, when the missing range is a gate oxide thickness of 14–15 nm, the neighboring ranges of 12–14 nm and 15–17 nm can be selected as supplementary reference ranges. Simultaneously, based on the statistical results of historical samples, a process sensitivity threshold is set to filter samples that should be excluded when the difference in response coefficients is too large, ensuring the rationality of data supplementation and the convergence of the model. Subsequently, within the target supplementary region, the size similarity between the size of each historical defect sample and the mean size of the missing size range is calculated using Euclidean distance. The process response similarity of the processing response coefficients of the structural features is also calculated using Euclidean distance. Then, the size similarity and process response similarity of each historical defect sample are fused through weighted fusion to obtain a comprehensive similarity score. Samples with a comprehensive similarity score higher than a preset threshold are selected and added to the neighborhood matching size defect sample set. Subsequently, for the selected neighborhood-matched size defect sample set, its process parameters and defect features are normalized and mapped, transforming the original sample features to the theoretical process space corresponding to the target size. Based on this theoretical process space, a weighted correction of the response coefficients is performed to obtain the features after the transfer operation. Then, by resampling the transferred neighborhood-matched size defect sample set, a supplementary defect sample set for the target size range is generated. This supplementary defect sample set contains information such as process parameters, structural features, defect types, defect probabilities, and influence weights under the target size, thus forming a complete process-defect mapping relationship. Through the above process, even when defect samples are incomplete or local size data is missing, transfer compensation using neighborhood samples can be used to achieve adaptive expansion of the sample space, providing a reliable data foundation for the subsequent construction of the defect space structure and the deployment of IoT sensors.

[0030] Furthermore, performing feature-level transfer operations on the neighborhood-matching size defect sample set includes:

[0031] The process parameter features are dimensionally normalized and mapped, the process parameters of the neighboring samples are adjusted to the theoretical process space corresponding to the target size, and the defect features in the theoretical process space are weighted and corrected by response coefficients. Based on the difference in processing response coefficients between the target size and the neighboring sample sizes, the defect probability and severity are corrected to obtain the features after the transfer operation.

[0032] Optionally, for each sample in the neighborhood matching size defect sample set, its process parameter features are extracted. These process parameter features include, but are not limited to, etching power, gas flow rate, process temperature, reaction pressure, deposition rate, oxidation time, and polysilicon fill rate. Since the calibration range of process parameters differs across size ranges, a size normalization coefficient is first defined. This coefficient is the ratio of the target size to the size of the neighborhood sample. By multiplying this size normalization coefficient by the process parameters of the neighborhood samples, the process parameters of all neighborhood samples can be adjusted to the theoretical process space corresponding to the target size, achieving a unified representation of cross-size process conditions. Subsequently, the difference between the processing response coefficient of the target size and the processing response coefficient of the neighborhood sample size is multiplied by an empirical adjustment coefficient, and then added to 1 to obtain the influence weight. Based on the calculated influence weight, defect features in the theoretical process space, such as defect size and shape coefficient, are weighted and corrected. Simultaneously, the defect probability and severity are also corrected based on the influence weight, thereby improving the defect feature offset caused by the difference between size and process response, making the mapped features more consistent with the physical laws under the target size. Finally, after size normalization and response coefficient weighting correction, the features corresponding to the target size range after migration operation are generated, providing data support for constructing a complete process-defect mapping.

[0033] Based on the defect space structure, IoT device matching and mapping are performed to determine device matching relationships. The device matching relationships are then used to deploy sensors on the defect space structure and collect defect sensing data.

[0034] In one embodiment, after establishing the defect space structure, IoT device matching and mapping is performed based on this structure to determine the matching sensor type, acquisition accuracy, and acquisition range for the corresponding structure within the defect space, forming a device matching relationship. For example, a high-resolution laser displacement sensor is matched for trench collapse defects; a high-sensitivity current / voltage detection sensor is matched for gate oxide breakdown defects; and an acoustic or infrared imaging sensor is matched for polysilicon void filling defects. Subsequently, the deployment positions are optimized based on this device matching relationship, and sensing deployment is carried out on the defect space structure according to the optimization results. This allows corresponding sensors to be installed on the processing equipment and related detection nodes, establishing a multi-source data acquisition channel. Then, the timestamps of each sensor are automatically synchronized using this multi-source data acquisition channel to ensure that the defect data and process data correspond accurately in the time dimension. Real-time monitoring data related to the defect space structure, such as local temperature distribution, stress changes, abnormal electric field or current signals, and processing equipment operating status data, are then synchronously acquired according to the set sampling strategy to form defect sensing data, providing input for subsequent processing control and fusion analysis stages.

[0035] Furthermore, based on the defect space structure, IoT device matching and mapping are performed to determine device matching relationships. The device matching relationships are then used to deploy sensors on the defect space structure, including:

[0036] Based on the defect space structure, IoT device matching and mapping are performed to determine the matching sensor device type, acquisition accuracy, and acquisition range of the corresponding structure in the defect space, thereby obtaining the device matching relationship. According to the spatial coverage relationship of the device matching relationship and the consistency of the matching sensor device type, the deployment location is optimized to obtain the sensor deployment strategy, and the sensor is deployed in the defect space structure.

[0037] Preferably, based on the established defect spatial structure of SGT-MOS devices, the defect types in different spatial regions are identified. For each type of defect, such as trench collapse, gate oxide breakdown, and polysilicon voids, the types of physical quantities to be monitored are determined, such as temperature, pressure, stress, electric field strength, displacement, current, voltage, vibration frequency, and gas composition concentration. Based on these monitoring requirements, sensor type information capable of measuring the corresponding physical quantities is retrieved from the IoT device resource library, such as thermocouples and infrared temperature sensors for temperature monitoring, capacitive and piezoelectric stress sensors for stress and strain monitoring, and current / voltage sampling modules for electrical characteristic monitoring. Subsequently, combining the monitoring accuracy requirements and range of each defect, the required acquisition accuracy level and acquisition spatial range of the sensing equipment for each type of defect are determined. For example, for trench collapse defects requiring resolution of nanometer-level morphological changes, a laser interferometric displacement sensor with an acquisition accuracy of less than 10 nm is matched; for gate oxide breakdown defects, a high-sensitivity current sensor with an acquisition accuracy of 10 nm is matched. -9 Level A. By establishing a one-to-one correspondence between defect characteristic parameters and sensor performance parameters, a device matching relationship is established in the defect space, achieving a functional mapping between defects and equipment. Then, based on the device space coverage characteristics of the device matching relationship and the type of matching sensor equipment, an optimization model with maximizing monitoring coverage benefits as the core objective determines the optimal deployment location that meets the requirements of physical space, cost, and signal interference, thus forming a sensor deployment strategy. Following this strategy, sensor nodes are installed on the processing equipment or production line. Each node is interconnected via an IoT gateway, enabling automatic device identification and network registration. After deployment, it automatically enters data acquisition mode, continuously monitoring the area corresponding to the defect space structure and collecting defect-related data in real time (such as temperature fluctuations, current anomalies, morphological changes, etc.), forming defect sensing data. This provides the basic input for subsequent processing control and fusion analysis stages.

[0038] Furthermore, based on the spatial coverage relationship of the device matching relationship and the consistency of the matching sensor device types, the deployment location is optimized to obtain a sensor deployment strategy, including:

[0039] Based on the severity of the impact of defects in the defect space structure on the device, corresponding monitoring priority weights are configured; one or more sensor data types required for effective identification or monitoring of each structural defect are extracted one by one from the device matching relationship, and the accuracy constraints of each data type are determined; an optimization model with maximizing monitoring coverage benefits as the core objective is constructed using the deployment locations and types of all candidate sensors as optimization variables; the optimization model is solved, and the sensor deployment strategy that maximizes the objective is searched under the constraints of physical space, cost, and signal interference.

[0040] Optionally, within the defect spatial structure, domain experts determine the severity of a defect's impact on the device based on its spatial distribution density, frequency of occurrence in the process, and historical average impact. Corresponding monitoring priority weights are then assigned accordingly. Higher priority weights indicate a greater impact of the defect on the final device performance, requiring priority monitoring. For example, gate oxide breakdown defects could have a weight of 0.9, trench collapse defects 0.8, polysilicon void defects 0.7, and uneven metal coverage 0.5. Subsequently, based on the obtained device matching relationships, one or more sensor data types corresponding to each type of defect are extracted, and the required monitoring accuracy is determined. For example, gate oxide breakdown defects require current and voltage signal acquisition, typically with an accuracy of nA to μA; trench collapse defects require displacement and surface morphology data acquisition, typically with an accuracy of nanometers; and polysilicon void defects require acoustic or infrared reflection characteristic data acquisition, typically with an accuracy of micrometers. Next, using the deployment locations and types of all candidate sensors as optimization variables, an optimization model is constructed with the core objective of maximizing monitoring coverage benefits. This model incorporates an optimization objective function, which is used to weight the spatial coverage rate of each defective area and the monitoring priority weight of that defect. Then, with physical space, cost, and signal interference as constraints, a genetic algorithm is used to solve the optimization model. Specifically, the deployment locations and types of all candidate sensors are first encoded, and the type and location coordinates of each sensor are represented in chromosome form to form an initial population. Then, the fitness value of each individual is calculated using the optimization objective function, which serves as the overall monitoring benefit. Based on this, a selection operation prioritizes retaining individuals with high fitness to ensure the inheritance of excellent deployment schemes. A crossover operation exchanges some deployment parameters among different individuals to explore new combination possibilities. A mutation operation randomly adjusts the position or type parameters of sensors with a certain probability to prevent getting trapped in local optima. These newly generated solutions are then filtered using constraints, and new solutions with conflicting installation locations, excessive costs, or sensor spacing less than the minimum spacing are deleted. In each iteration, the optimal solution is updated according to the objective function. When the value of the objective function exceeds the preset threshold or the maximum number of iterations is reached after multiple generations of optimization, the algorithm converges and outputs the final sensor deployment strategy. This sensor deployment strategy maximizes the defect monitoring coverage and resource utilization while meeting the constraints of space, cost and signal, providing a high-quality data foundation for the dynamic control and intelligent feedback of the subsequent processing.

[0041] The defect sensing data and conventional sensing data are fused and analyzed to identify dynamic processing control data and generate processing control instructions.

[0042] In one embodiment, after obtaining defect sensing data, the defect sensing data and regular sensing data are input into the defect sensing data channel and regular sensing data channel for fusion and analysis. Through multi-granularity spatiotemporal synchronization and hierarchical fusion, the defect sensing data and regular sensing data can be converted into multi-layer spatiotemporal processing state fusion analysis results, which serve as dynamic processing control data. Subsequently, processing state anomalies are identified and located based on this dynamic processing control data, and corresponding processing control instructions are generated based on the location results. These processing control instructions include, but are not limited to, adjusting process parameters, changing equipment operating modes, and triggering partial process pauses, in order to improve product yield and process controllability.

[0043] Furthermore, by fusing and analyzing the defect sensing data and conventional sensing data, dynamic processing control data is identified, and processing control instructions are generated, including:

[0044] A dual-channel data acquisition system is established, comprising a defect sensing data channel and a conventional sensing data channel. The defect sensing data channel obtains monitoring data directly related to defect characteristics through defect sensing data, while the conventional sensing data channel monitors process equipment health status, environmental stability, and process sensing parameters for the normal operation of non-target structures. Multi-granularity spatiotemporal synchronization and hierarchical fusion are performed on the defect sensing data channel and the conventional sensing data channel. The multi-granularity spatiotemporal synchronization includes millisecond-second real-time control granularity, second-minute batch operation control granularity, and hour-day equipment health and process baseline control granularity. The hierarchical fusion dynamically adjusts the decision fusion weights of the two channels based on the processing state context to obtain multi-layer spatiotemporal processing state fusion analysis results. Based on the processing state fusion analysis results, processing state anomalies are identified and located, and corresponding processing control commands are generated for compensating and controlling the anomaly identification and location results.

[0045] Preferably, to comprehensively acquire dynamic information during the SGT-MOS device fabrication process, two data acquisition paths are established under the IoT architecture: a defect sensing data channel and a conventional sensing data channel. The defect sensing data channel collects real-time monitoring data directly related to defect characteristics from the defect sensing data, such as local temperature fluctuations, surface displacement, stress distribution, current anomalies, and reflectivity changes. The conventional sensing data channel collects the health status of the process equipment and processing environment parameters, including equipment cavity temperature, vacuum level, gas flow rate, pressure, power output, electrode status, and equipment vibration signals, to reflect the stable operating characteristics of non-defect areas and the fluctuations in the overall processing environment. Subsequently, due to differences in sampling frequency, response delay, and data scale between defect sensing data and conventional sensing data, a multi-granularity spatiotemporal synchronization and hierarchical fusion strategy is adopted to achieve consistent analysis across different time dimensions and spatial scales. Specifically, at the millisecond-second real-time control granularity, high-frequency sampling and time alignment of rapidly changing process parameters are used for real-time feedback control. At the second-minute batch operation control granularity, medium-speed variation characteristics during single-batch processing, such as gas flow stability, cavity pressure fluctuations, and equipment power drift, are monitored for process stability analysis. At the hour-day equipment health and process baseline control granularity, long-term changes, such as equipment aging trends, temperature drift, and contamination accumulation, are monitored for predictive maintenance and process baseline correction. A multi-granularity synchronization mechanism achieves full-cycle data coordination from instantaneous response to long-term trends over time. After data synchronization, the fusion weights of the defect sensing data channel and the conventional sensing data channel are dynamically adjusted based on the current process state context, such as etching, deposition, oxidation, and annealing stages. During high-risk processing stages, such as deep trench etching, the weight of the defect sensing data channel is increased to enhance local anomaly detection; during stable operation stages, such as annealing or cooling, the weight of the conventional sensing data channel is increased to maintain overall health assessment. At the fusion level, the bottom layer employs a time-series filtering and interpolation fusion algorithm, the middle layer uses a statistical model based on weighted feature aggregation, and the top layer achieves multimodal data fusion through an intelligent decision-making model based on neural networks or Bayesian fusion, outputting a comprehensive processing status fusion analysis result. Then, the fusion analysis result is input into an anomaly recognition model pre-trained using a deep neural network through forward propagation, loss calculation, backpropagation, and parameter optimization. The anomaly recognition model uses the learned knowledge to make real-time judgments on the processing status, determining the anomaly type, specific process step, severity index, and physical location corresponding to the anomaly.Finally, based on the anomaly type, specific process step, severity index, and physical location, the built-in processing control database is invoked. The cosine similarity is used to match the processing control strategy that meets the preset similarity threshold. This processing control strategy is then converted into a processing control instruction and sent to the equipment execution end to achieve real-time response and automatic compensation control, thereby improving the stability, repeatability, and product yield of the process.

[0046] In summary, the embodiments of this application have at least the following technical effects:

[0047] First, structural defect features of the SGT-MOS device are identified to establish a defect space structure. Next, IoT device matching and mapping are performed based on the defect space structure to determine device matching relationships. These relationships are then used to deploy sensors within the defect space structure, collecting defect sensing data. Finally, the defect sensing data is fused and analyzed with conventional sensing data to identify dynamic processing control data and generate processing control commands. This solves the technical problem of low yield and poor performance consistency caused by the difficulty in timely detection and precise control of structural defects during the processing of existing SGT-MOS devices. It achieves the technical effect of improving device processing yield and electrical performance consistency through defect-driven dynamic process parameter optimization via multi-source IoT data fusion.

[0048] Example 2, based on the same inventive concept as the IoT-based SGT-MOS device fabrication and control method in the previous examples, such as... Figure 2 As shown, this application provides an IoT-based SGT-MOS device fabrication control system, the system including:

[0049] Defect identification module 11: Identifies structural defect features of the SGT-MOS device and establishes a defect space structure; Data acquisition module 12: Performs IoT device matching mapping based on the defect space structure, determines device matching relationships, uses the device matching relationships to deploy sensors on the defect space structure, and collects defect sensing data; Fusion and analysis module 13: Performs fusion and analysis of the defect sensing data and conventional sensing data, identifies dynamic processing control data, and generates processing control instructions.

[0050] Furthermore, the defect identification module 11 is used to perform the following method:

[0051] Identify the structural features of the SGT-MOS device and collect historical defect samples corresponding to the structural features; perform defect feature and defect probability identification analysis on the historical defect samples corresponding to the structural features to generate structural defect feature parameters; project the structural defect feature parameters into the SGT-MOS device spatial structure to establish the defect spatial structure.

[0052] Furthermore, the defect identification module 11 is used to perform the following method:

[0053] Matching the structural features with historical defect size ranges yields a set of defect samples with matching sizes and a set of defect samples with missing sizes, wherein the set of defect samples with missing sizes has a corresponding set of defect samples with matching sizes in its neighborhood. Based on the missing size and structural feature processing response coefficients corresponding to the set of defect samples with missing sizes, sample operations are performed on the set of defect samples with matching sizes in the neighborhood to supplement the defect sample set, thus obtaining a supplementary defect sample set. The set of defect samples with matching sizes and the supplementary defect sample set are then mapped using structural features to obtain the historical defect samples.

[0054] Furthermore, the defect identification module 11 is used to perform the following method:

[0055] Structural feature data and corresponding defect data of SGT-MOS devices during historical processing are collected to form an initial historical defect sample set. Processing response coefficient analysis is performed on the structural features to identify their sensitivity to processing fluctuations and one or more key dimensions strongly correlated with defect incidence. These key dimensions include, but are not limited to, trench aspect ratio, gate oxide thickness, and polysilicon fill height. Based on these key dimensions, the design specification space of the SGT-MOS devices is divided into multiple continuous size intervals. The initial historical defect sample set is then classified according to these size intervals, resulting in multiple size interval samples, each with a structural feature processing response coefficient label.

[0056] Furthermore, the defect identification module 11 is used to perform the following method:

[0057] Based on the missing size and structural feature processing response coefficient, the size range and process sensitivity threshold of the target supplementary region are determined; based on size similarity and process response similarity, a dual matching is performed to select a neighborhood matching size defect sample set, and feature-level transfer operation is performed on the neighborhood matching size defect sample set; based on the features after the transfer operation, the supplementary defect sample set is generated, and the supplementary defect sample set contains a complete process-defect mapping relationship.

[0058] Furthermore, the defect identification module 11 is used to perform the following method:

[0059] The process parameter features are dimensionally normalized and mapped, the process parameters of the neighboring samples are adjusted to the theoretical process space corresponding to the target size, and the defect features in the theoretical process space are weighted and corrected by response coefficients. Based on the difference in processing response coefficients between the target size and the neighboring sample sizes, the defect probability and severity are corrected to obtain the features after the transfer operation.

[0060] Furthermore, the data acquisition module 12 is used to perform the following methods:

[0061] Based on the defect space structure, IoT device matching and mapping are performed to determine the matching sensor device type, acquisition accuracy, and acquisition range of the corresponding structure in the defect space, thereby obtaining the device matching relationship. According to the spatial coverage relationship of the device matching relationship and the consistency of the matching sensor device type, the deployment location is optimized to obtain the sensor deployment strategy, and the sensor is deployed in the defect space structure.

[0062] Furthermore, the data acquisition module 12 is used to perform the following methods:

[0063] Based on the severity of the impact of defects in the defect space structure on the device, corresponding monitoring priority weights are configured; one or more sensor data types required for effective identification or monitoring of each structural defect are extracted one by one from the device matching relationship, and the accuracy constraints of each data type are determined; an optimization model with maximizing monitoring coverage benefits as the core objective is constructed using the deployment locations and types of all candidate sensors as optimization variables; the optimization model is solved, and the sensor deployment strategy that maximizes the objective is searched under the constraints of physical space, cost, and signal interference.

[0064] Furthermore, the fusion parsing module 13 is used to perform the following methods:

[0065] A dual-channel data acquisition system is established, comprising a defect sensing data channel and a conventional sensing data channel. The defect sensing data channel obtains monitoring data directly related to defect characteristics through defect sensing data, while the conventional sensing data channel monitors process equipment health status, environmental stability, and process sensing parameters for the normal operation of non-target structures. Multi-granularity spatiotemporal synchronization and hierarchical fusion are performed on the defect sensing data channel and the conventional sensing data channel. The multi-granularity spatiotemporal synchronization includes millisecond-second real-time control granularity, second-minute batch operation control granularity, and hour-day equipment health and process baseline control granularity. The hierarchical fusion dynamically adjusts the decision fusion weights of the two channels based on the processing state context to obtain multi-layer spatiotemporal processing state fusion analysis results. Based on the processing state fusion analysis results, processing state anomalies are identified and located, and corresponding processing control commands are generated for compensating and controlling the anomaly identification and location results.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for SGT-MOS device processing control based on Internet of Things, characterized in that, The method comprises the following steps: performing structural defect feature recognition on the SGT-MOS device to establish a defect space structure; performing Internet of Things device matching mapping based on the defect space structure to determine a device matching relationship, and performing sensing deployment on the defect space structure by using the device matching relationship to collect defect sensing data; performing fusion analysis on the defect sensing data and conventional sensing data to identify dynamic processing control data and generate processing control instructions; performing structural defect feature recognition on the SGT-MOS device to establish a defect space structure, comprising: recognizing the structural features of the SGT-MOS device and collecting historical defect samples corresponding to the structural features; performing defect feature and defect probability recognition analysis on the historical defect samples corresponding to the structural features respectively to generate structural defect feature parameters; projecting the structural defect feature parameters onto the SGT-MOS device space structure to establish the defect space structure; collecting historical defect samples corresponding to the structural features, comprising: matching the structural features with historical defect size intervals to obtain a matching size defect sample set and a missing size defect sample set, wherein the missing size defect sample set has a corresponding neighborhood matching size defect sample set; based on the missing size and structural feature processing response coefficient corresponding to the missing size defect sample set, performing sample operation supplement based on the neighborhood matching size defect sample set to obtain a supplemented defect sample set; performing structural feature mapping on the matching size defect sample set and the supplemented defect sample set to obtain the historical defect samples.

2. The IoT-based SGT-MOS device processing control method of claim 1, wherein, Before matching the structural features with historical defect size intervals to obtain a matching size defect sample set, comprising: collecting structural feature data and corresponding defect data of SGT-MOS devices in historical processing to form an initial historical defect sample set; performing processing response coefficient analysis on the structural features to identify the sensitivity of the structural features to processing technology fluctuations and one or more key sizes that are strongly related to defect occurrence rates, wherein the key sizes include but are not limited to trench aspect ratio, gate oxide thickness, and polysilicon filling height; based on the key sizes, dividing the design specification space of the SGT-MOS device into multiple continuous size intervals; classifying the initial historical defect sample set into multiple size interval samples according to the size intervals, wherein each size interval has a structural feature processing response coefficient label.

3. The SGT-MOS device processing control method based on Internet of Things according to claim 2, characterized in that, based on the missing size and structural feature processing response coefficient corresponding to the missing size defect sample set, performing sample operation supplement based on the neighborhood matching size defect sample set to obtain a supplemented defect sample set, comprising: determining the size range and process sensitivity threshold of the target supplement area according to the missing size and structural feature processing response coefficient; performing double matching based on size similarity and process response similarity to select a neighborhood matching size defect sample set, and performing feature-level migration operation on the neighborhood matching size defect sample set; based on the features after the migration operation, generating the supplemented defect sample set, wherein the supplemented defect sample set contains complete process-defect mapping relationships.

4. The SGT-MOS device processing control method based on Internet of Things according to claim 3, characterized in that, Performing feature-level migration operation on the neighborhood matching size defect sample set, comprising: Performing size normalization mapping on process parameter features, adjusting process parameters of neighborhood samples to a theoretical process space corresponding to a target size, and performing response coefficient weighted correction on defect features in the theoretical process space, correcting defect probability and severity according to the difference in processing response coefficients between the target size and the neighborhood sample size, to obtain features after migration operation.

5. The IoT-based SGT-MOS device processing control method of claim 1, wherein, Based on the defect space structure, performing Internet of Things device matching mapping to determine the device matching relationship, and using the device matching relationship to perform sensor deployment on the defect space structure, comprising: Based on the defect space structure, performing Internet of Things device matching mapping to determine the matching sensor device type, collection accuracy and collection range of the corresponding structure in the defect space, and obtaining the device matching relationship; According to the spatial coverage relationship of the device matching relationship and the consistency of the matching sensor device type, optimizing the layout position to obtain a sensor deployment strategy, and performing sensor deployment on the defect space structure. 6.The SGT-MOS device processing control method based on Internet of Things according to claim 5, wherein, According to the spatial coverage relationship of the device matching relationship and the consistency of the matching sensor device type, optimizing the layout position to obtain a sensor deployment strategy, comprising: According to the severity of the defect in the defect space structure on the device, configuring a corresponding monitoring priority weight; From the device matching relationship, one or more sensor data types required for effective identification or monitoring of each structure defect are extracted one by one, and the accuracy constraint requirements of each data type are determined; Taking all candidate sensor layout positions and types as optimization variables, an optimization model with the core target of maximizing monitoring coverage benefits is constructed; Solving the optimization model, searching for a sensor deployment strategy that maximizes the target under the constraints of physical space, cost and signal interference.

7. The IoT-based SGT-MOS device processing control method of claim 1, wherein, According to the fusion analysis of the defect sensor data and the conventional sensor data, identifying dynamic processing control data, and generating processing control instructions, comprising: A dual-channel data acquisition system including a defect sensor data channel and a conventional sensor data channel is established, the defect sensor data channel obtains monitoring data directly related to defect features through defect sensor data, and the conventional sensor data channel obtains process sensor parameter data for monitoring process equipment health status, environmental stability and normal operation of non-target structures; Multi-granularity spatiotemporal synchronization and hierarchical fusion are performed on the defect sensor data channel and the conventional sensor data channel, wherein the multi-granularity spatiotemporal synchronization includes millisecond-second level real-time control granularity, second-minute level batch operation control granularity, and hour-day level equipment health and process baseline control granularity; the hierarchical fusion is based on dynamically adjusting the decision fusion weight of the dual channel according to the processing state context to obtain multi-layer spatiotemporal processing state fusion analysis results; According to the processing state fusion analysis results, processing state anomaly identification and positioning are performed, and corresponding processing control instructions are generated for compensation control of the anomaly identification and positioning results.

8. The SGT-MOS device processing control system based on the Internet of Things, characterized in that, A method for implementing the Internet of Things-based SGT-MOS device processing control method of any one of claims 1-7, comprising: The defect identification module is configured to identify structural defect features of the SGT-MOS device, and establish a defect space structure; The data acquisition module is configured to perform Internet of Things device matching mapping based on the defect space structure, determine a device matching relationship, perform sensing deployment on the defect space structure by using the device matching relationship, and acquire defect sensing data; The fusion analysis module is configured to perform fusion analysis on the defect sensing data and conventional sensing data, identify dynamic processing and control data, and generate processing and control instructions.

Images