Whole-process intelligent management and control system and method for SiC device production
By constructing a full-process timing data model and optimizing multiple algorithms, the problem of data fragmentation in SiC device production was solved, and the unification of data correlation and production timing was achieved, thereby improving production efficiency and quality.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-24
- Publication Date
- 2026-05-12
AI Technical Summary
In existing technologies, upstream and midstream data in SiC device production are fragmented and lack a unified integration framework. This makes it impossible to accurately capture the fluctuation patterns of process duration and data drift characteristics. A single algorithm is also insufficient to take into account delivery time, process priority, and timing stability.
By integrating core data from upstream and midstream processes, a full-process time-series data model is constructed. A combination of genetic algorithms and analytic hierarchy process is used for multi-objective optimization, a scheduling scheme is built, and time-series alignment is achieved through a graph-structured time-series data model.
It achieves the correlation of data throughout the entire process and the unification of production timing, accurately captures the production timing pattern, takes into account multiple objectives for optimization, ensures that the scheduling scheme fits the production constraints, and improves the efficiency and quality of SiC device production.
Smart Images
Figure CN122022435A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of process control technology, specifically to an intelligent control system and method for the entire process of SiC device production. Background Technology
[0002] As a third-generation wide-bandgap semiconductor material, SiC possesses excellent properties such as a large bandgap, high thermal conductivity, high breakdown electric field strength, and fast electron saturation drift velocity. Its application demand continues to rise in high-end equipment fields such as new energy vehicles, photovoltaic inverters, rail transportation, and smart grids. The manufacturing process of SiC devices is characterized by a long process chain, high process complexity, and strong parameter correlation. It corely encompasses two major stages: upstream substrate preparation and midstream device manufacturing. The process parameters and production sequence of each stage directly determine the performance and yield of SiC devices.
[0003] Existing technologies primarily focus on data collection for single stages of SiC device production, such as upstream substrate preparation or midstream device manufacturing. This results in fragmented data flow between upstream and downstream processes, lacking a unified integration framework. Current technologies often only superficially analyze production data, failing to provide detailed breakdowns of time-series data into trends, periods, and residuals. This makes it difficult to accurately capture fluctuations in process duration, data drift characteristics, and the range of random interference. Furthermore, existing technologies often employ single algorithms to construct scheduling schemes, lacking multi-algorithm collaborative optimization mechanisms. A single algorithm struggles to simultaneously address multiple objectives such as delivery time, process priority, and timing stability, easily leading to a disconnect between the scheduling scheme and actual production constraints. Summary of the Invention
[0004] The purpose of this invention is to provide an intelligent control system and method for the entire process of SiC device production, in order to solve the problems raised in the prior art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: Firstly, this application provides a method for intelligent control of the entire process of SiC device production, including the following steps: Integrate core data from various processes in the upstream and midstream to construct a full-process time-series data model; The entire process time series data model is decomposed to separate the data sequences, including trend sequence, periodic sequence and residual sequence; batch parameters are collected and quantified; Combining sequence and batch parameters, an initial batch scheduling framework is constructed based on a genetic algorithm to obtain a scheduling scheme; the analytic hierarchy process (AHP) is introduced to optimize the scheduling scheme to obtain an optimized scheduling scheme; the optimized scheduling scheme is then time-series aligned with a graph-structured time-series data model to obtain a scheduling table. The entire process of SiC device production is controlled based on a scheduling table.
[0006] In conjunction with the first aspect, in the first embodiment of the first aspect of this application, the integration of core data from each process in the upstream and midstream to construct a full-process time-series data model includes: Key data for upstream processes include the purity of substrate raw materials. Substrate crystal integrity and substrate slice thickness tolerance The core data for the midstream process includes the uniformity of the epitaxial layer thickness. Photolithographic pattern alignment accuracy and etch depth consistency ;Will Related , Related , Related , Related , Related , Related , Related ; Define the node system, edge system, and attribute system, construct a graph structure, and integrate the entire process data based on the graph structure to obtain a full-process time-series data model.
[0007] In conjunction with the first aspect, in the second embodiment of the first aspect of this application, the definition of the node system, edge system, and attribute system, the construction of a graph structure, and the integration of full-process data based on the graph structure to obtain a full-process time-series data model include: The node system is defined using process nodes and data nodes. Upstream process nodes include raw material purification. Crystal growth and slicing The midstream process nodes include epitaxial growth. Photolithography and etching process Upstream data nodes include , and The midstream data nodes include , and ; The edge system is defined using temporal edges, attribution edges, and association edges; among them, temporal edges connect adjacent process nodes, attribution edges connect data nodes with the corresponding generated process nodes, and association edges include intra-process association edges and cross-process corresponding association edges. In the attribute system, the process node attributes include process ID, standard execution duration, and preceding process ID; the data node attributes include data ID, collection timestamp, value range, and collection device ID; the time sequence side is marked with time sequence priority, the affiliation side is marked with data collection timing, and the association side is marked with association confidence. Construct a graph structure, instantiate nodes and mount data, align the collection timestamps of each data node with the execution time windows of the corresponding process nodes, and sort the entire process data according to the production time sequence through the time sequence priority attribute of the time sequence edge, so that the graph structure has both data association logic and production time sequence logic, forming a full-process time sequence data model.
[0008] In conjunction with the first aspect, in the third embodiment of the first aspect of this application, the step of decomposing the entire process time series data model and separating the data sequences, including trend sequence, periodic sequence, and residual sequence, includes: Based on the full-process time-series data model, time-series datasets are extracted, including process node time-series datasets and data node time-series datasets, and the time granularity is unified and the dimensions are aligned. A sliding window fitting logic is employed, setting a fixed-length time window for the time series data. Linear or nonlinear fitting is performed on the data within the window, and the fitting results of all windows are merged to form a continuous trend term sequence. The extracted trend term is removed from the time series dataset to obtain detrended time series data. Based on the inherent cycle time of the production process, the cycle length is determined. Cycle pattern matching is performed on the detrended time series data to extract recurring fluctuation components, forming a cycle term sequence. Using the original time series data as a benchmark, the trend term sequence and the cycle term sequence are subtracted, and the resulting difference is used as the residual term sequence.
[0009] In conjunction with the first aspect, in the fourth embodiment of the first aspect of this application, the step of constructing an initial batch scheduling framework based on a genetic algorithm by combining sequence and batch parameters to obtain a scheduling scheme includes: Extracting sequence features specifically involves: in the trend sequence, extracting the long-term trend of execution time for each process and the numerical drift trend of each core data point, which serve as the basis for adjusting process time consumption benchmarks; in the periodic sequence, extracting the periodic fluctuation amplitude of each process based on the inherent cycle time of the production process, which serves as the constraint boundary for process time consumption fluctuations; in the residual sequence, statistically analyzing the variance range of residuals for each process node and data node, and selecting stable time intervals with residuals less than a threshold, which serve as a reference for priority time windows in batch scheduling; and establishing the correspondence between sequence features and batch parameters. A real-number encoding method is adopted to ensure a one-to-one correspondence between chromosomes and batch scheduling schemes; constraints of the genetic algorithm are defined; a multi-objective weighted function is constructed by fusing sequence features and batch parameters with the goal of maximizing fitness; an initial population is generated, and selection, crossover, and mutation operations are performed; an iteration termination condition is set, and the chromosome with the highest fitness value in the final population is used as the scheduling scheme.
[0010] Specifically, real-number coding is used, with each chromosome corresponding to a complete batch scheduling scheme. The coding length is determined by the number of batches and the number of processes. It is designed with 6 processes (P1-P6) for each batch and 2 time parameters (start time and end time) for each process. That is, the chromosome length is the total number of batches × 6 × 2, ensuring that each scheduling detail has a corresponding gene fragment.
[0011] Chromosomes are grouped by batch and process, with each group consisting of two consecutive real genes corresponding to the start and end times of a specific process within a particular batch. When generating gene fragments, it is necessary to ensure that the start time of a single process is earlier than its end time, and that the process duration falls within the range of the trend term correction baseline ± the periodic term fluctuation. Invalid codes that violate the basic timing and duration constraints are directly filtered out to ensure the feasibility of the initial codes.
[0012] Furthermore, regarding timing constraints, all batches must follow the graph structure timing edge logic, meaning that within the same batch, P2 can only be started after P1 is completed, P4 can only be started after P3 is completed, and so on. The actual duration of each process must not exceed the superposition range of the baseline duration corrected by the trend term and the fluctuation amplitude of the periodic term, to avoid deviating from the production timing pattern. Regarding delivery constraints, the end time of P6 (the final process) in each batch must not be later than the planned delivery time of that batch, strictly adhering to the delivery requirements in the batch parameters to prevent overdue payments. Regarding dependency constraints, based on cross-process data association relationships (S1→M1, S2→M2, S3→M3), upstream data nodes (S1-S3) in the same batch must complete data collection and their values must be within the specified range before the corresponding midstream processes (P4-P6) can be started, ensuring process dependency compliance. Regarding priority constraints, batch combinations with higher process connection priority scores are allocated continuous time windows first; batches with higher delivery urgency weights are given priority to occupy stable time intervals where the residual variance is less than the threshold, aligning with the batch parameters and residual term sequence characteristics. For stability constraints, the fluctuation range of the duration of each process in a batch must conform to the periodicity of the terms and avoid exceeding the periodic fluctuation range; at the same time, the impact of the residual terms should be controlled within the allowable threshold to ensure the timing stability of the scheduling scheme.
[0013] Furthermore, focusing on maximizing the fitness value, a multi-dimensional weighted evaluation function is constructed by integrating sequence features and batch parameters to balance different optimization objectives. The core dimensions of the function cover three major dimensions: delivery compliance, process priority matching, and time sequence stability. Delivery compliance measures the fit between the schedule and the batch plan delivery time and urgency level; process priority matching measures the fit between the schedule and the process connection priority matrix; and time sequence stability measures the adaptability of the schedule to the sequence features of trend items, periodic items, and residual items. The weight allocation logic sets the weights for each dimension based on production needs. If delivery is prioritized, the weight of delivery compliance is increased; if process stability is prioritized, the weight of time sequence stability is increased, ensuring that the function accurately matches the actual production priority. The sum of the weights for each dimension is 1. The higher the fitness value, the better the corresponding scheduling scheme.
[0014] Furthermore, based on the set population size and coding rules, several chromosomes are randomly generated, with each chromosome corresponding to an initial scheduling scheme. After generation, hard constraints are verified one by one, invalid individuals that violate timing, delivery, and dependency constraints are removed, and new feasible individuals are added to ensure that the initial population consists of scheduling schemes that meet the basic constraints, while preserving population diversity.
[0015] The tournament selection method is adopted. Each time, 3-5 individuals are randomly selected from the population to form a candidate group. The individual with the highest fitness value is selected to enter the next generation of the population. This process is repeated until the size of the next generation population reaches the set value. This operation preserves the high-quality scheduling scheme and guides the population to evolve in a better direction.
[0016] A single-point crossover strategy is adopted, with the crossover probability controlled between 0.6 and 0.8. First, a crossover point is randomly selected, and the corresponding gene segments of the two parent chromosomes are exchanged to generate offspring chromosomes. After crossover, the hard constraints are checked one by one for each generation. For gene segments that violate the constraints, the baseline is corrected by combining the trend term and the fluctuation amplitude of the period term to fine-tune the time parameters, ensuring that all offspring are feasible solutions.
[0017] A small-amplitude mutation strategy is adopted, with the mutation probability controlled between 0.01 and 0.05: gene segments in chromosomes are randomly selected, and the time values are finely adjusted within the range of the trend term correction baseline ± the fluctuation range of the periodic term, or the time windows of the same process in two batches are swapped; after mutation, hard constraints are also checked, invalid genes are corrected, premature maturity of the population is avoided, and population diversity is maintained.
[0018] A dual termination condition is set: first, the number of iterations reaches a preset upper limit; second, the fluctuation of the optimal fitness value of the population for 10-20 consecutive generations is less than a set threshold. Iteration stops when either condition is met. After iteration terminates, the chromosome with the highest fitness value is selected from the final population and decoded into its corresponding batch scheduling scheme, which is the initial scheduling scheme.
[0019] In conjunction with the first aspect, in the fifth embodiment of the first aspect of this application, the constraints defining the genetic algorithm include: The constraints of genetic algorithms include hard constraints and soft constraints; hard constraints mean that they must be satisfied, while soft constraints mean that the optimization goal is oriented. The hard constraints of genetic algorithms include timing constraints, delivery constraints, and dependency constraints. The timing constraints specifically require that batch processes follow the timing edge logic of the graph structure, subsequent processes cannot be started until the preceding process is completed, and the process duration cannot exceed the superposition range of the trend term correction benchmark and the periodic term fluctuation amplitude. The delivery constraints specifically require that the end time of all processes in a batch cannot be greater than the planned delivery time. The dependency constraints specifically require that based on the cross-process association relationship of the associated edges, the corresponding midstream process is started after the upstream data node of the same batch has completed the collection and the value meets the standard. The soft constraints of genetic algorithms include priority constraints and stability constraints. The priority constraint specifically means that batch combinations with higher process connection priority are allocated continuous time windows first, and batches with higher delivery urgency weights are given priority to occupy the time window where the residual is stable. The stability constraint specifically means that the fluctuation range of batch process duration must conform to the periodic term pattern, and the impact of residual term must be controlled within the allowable range.
[0020] In conjunction with the first aspect, in the sixth embodiment of the first aspect of this application, the step of introducing the analytic hierarchy process (AHP) to optimize the scheduling scheme and obtain an optimized scheduling scheme includes: With optimizing the production batch scheduling scheme for SiC devices as the core, and combining genetic algorithm constraints and sequence characteristics, a target layer, criterion layer, sub-criterion layer, and scheme layer are constructed, with each layer strongly correlated with previous data. This ensures that each layer corresponds to the full-process time-series data model and genetic algorithm constraints. Based on actual production needs and previous data characteristics, a nine-level scaling method is used to construct judgment matrices for the criterion layer versus the target layer, the sub-criterion layer versus the criterion layer, and the scheme layer versus the sub-criterion layer. Weight calculations and consistency checks are performed. Based on normalized weights, the priority ranking of each sub-criterion layer element relative to its corresponding criterion layer element, and the priority ranking of each scheme layer element relative to its corresponding sub-criterion layer element are determined. Combining the weights of each level, the comprehensive weight of each candidate scheme in the scheme layer relative to the target layer is calculated, and the scheme with the highest comprehensive weight is selected as the optimized scheduling scheme.
[0021] In conjunction with the first aspect, in the seventh embodiment of the first aspect of this application, the step of aligning the optimized scheduling scheme with the graph-structured time-series data model to obtain a scheduling table includes: Based on the graph structure and time-series edge logic, verify the sequence and duration of the scheduled processes; based on the attribution edges, align the data node collection timestamps with the corresponding process time windows; based on the association edges, verify the rationality of the time sequence connection of cross-process data; investigate time sequence conflicts, correct deviations according to hard constraints, control the impact of residual terms within the allowable range, and ensure the compliance of the aligned solution; summarize the aligned time windows, process connection relationships, delivery thresholds, and stability indicators to generate a scheduling table.
[0022] In conjunction with the first aspect, in the eighth embodiment of the first aspect of this application, the full-process control of SiC device production based on the scheduling table includes: The scheduling table is split into upstream and downstream processes and synchronized to each workstation. Graph structure temporal edges and related edge constraints are bound to ensure alignment of the execution end with the control benchmark. Based on the graph structure's belonging edges, execution data and data node monitoring values for each process node are collected in real time and compared with the scheduling table's temporal window. Trend item sequences, periodic item sequences, and residual item sequences are tracked synchronously to verify the compliance of process connections and cross-process data associations. When data deviates from the scheduling threshold or there is a temporal logic conflict, an alert is triggered based on the confidence level of the related edges. After each batch of processes is completed, the actual data is compared with the scheduling table, and the graph structure temporal model parameters are updated.
[0023] Secondly, this application provides a fully intelligent control system for the entire process of SiC device production, including: Model building module: includes: a data integration unit that integrates core data from upstream and midstream processes, and a model building unit that builds a full-process time-series data model; Model decomposition and parameter quantization module: includes: a model decomposition unit that decomposes the entire time series data model, separating the data sequences, including trend sequence, periodic sequence and residual sequence; and a parameter quantization unit that collects batch parameters and quantizes them. The scheduling table generation module includes: a scheduling scheme construction unit that combines sequence and batch parameters to construct an initial batch scheduling framework based on a genetic algorithm to obtain a scheduling scheme; a scheduling scheme optimization unit that introduces the analytic hierarchy process (AHP) to optimize the scheduling scheme to obtain an optimized scheduling scheme; and a scheduling table generation unit that aligns the optimized scheduling scheme with a graph-structured time-series data model to obtain a scheduling table. The end-to-end management module includes: an end-to-end management unit that manages the entire SiC device production process based on a scheduling table.
[0024] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention integrates core data from the entire upstream and downstream process, defines nodes, edges, and attribute systems based on graph structures, clarifies data relationships within and across processes, aligns data collection timestamps with process execution time windows, and forms a full-process time-series data model that combines data association logic and production time-series logic, thus solving the problems of data fragmentation and missing associations in existing technologies.
[0025] 2. This invention precisely decomposes the time series data model into trend terms, periodic terms, and residual terms to accurately capture the production time series patterns; at the same time, it standardizes the batch parameter quantification process, establishes the correspondence between time series characteristics and batch parameters, and makes the scheduling basis take into account long-term trends, periodic fluctuations, random interference, and batch personalized needs.
[0026] 3. This invention employs a combination of genetic algorithm to construct the initial framework and hierarchical analysis method for optimization, taking into account both multi-objective optimization and fine-grained priority matching, thus solving the problem of insufficient optimization by a single algorithm. At the same time, by optimizing the timing alignment between the scheduling scheme and the graph structure timing model, timing conflicts are verified and corrected, ensuring that the scheduling scheme conforms to the production constraints of the entire process. Attached Figure Description
[0027] Figure 1 This is a schematic diagram of the steps of the intelligent control method for the entire process of SiC device production according to the present invention; Figure 2 This is a system structure diagram of the intelligent control system for the entire process of SiC device production according to the present invention. Detailed Implementation
[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0029] Example: Figures 1-2 As shown, the present invention provides a technical solution: like Figure 1 As shown, this application provides a method for intelligent control of the entire process of SiC device production, including the following steps: Step S100: Integrate the core data of each process in the upstream and midstream to construct a full-process time-series data model; Specifically, key data for upstream processes include the purity of substrate raw materials. Substrate crystal integrity and substrate slice thickness tolerance The core data for the midstream process includes the uniformity of the epitaxial layer thickness. Photolithographic pattern alignment accuracy and etch depth consistency ;Will Related , Related , Related , Related , Related, Related , Related ; Define the node system, edge system, and attribute system, construct a graph structure, and integrate the entire process data based on the graph structure to obtain a full-process time-series data model.
[0030] Furthermore, a node system is defined using process nodes and data nodes. Upstream process nodes include raw material purification. Crystal growth and slicing The midstream process nodes include epitaxial growth. Photolithography and etching process Upstream data nodes include , and The midstream data nodes include , and ; The edge system is defined using temporal edges, attribution edges, and association edges; among them, temporal edges connect adjacent process nodes, attribution edges connect data nodes with the corresponding generated process nodes, and association edges include intra-process association edges and cross-process corresponding association edges. In the attribute system, the process node attributes include process ID, standard execution duration, and preceding process ID; the data node attributes include data ID, collection timestamp, value range, and collection device ID; the time sequence side is marked with time sequence priority, the affiliation side is marked with data collection timing, and the association side is marked with association confidence. Construct a graph structure, instantiate nodes and mount data, align the collection timestamps of each data node with the execution time windows of the corresponding process nodes, and sort the entire process data according to the production time sequence through the time sequence priority attribute of the time sequence edge, so that the graph structure has both data association logic and production time sequence logic, forming a full-process time sequence data model.
[0031] In one specific embodiment, taking a batch of SiC device production data as an example, the measured values of the core data of the upstream process are: substrate raw material purity S1=99.998%, substrate crystal integrity S2=98.7%, and substrate slice thickness tolerance S3=±0.02μm; the measured values of the core data of the midstream process are: epitaxial layer thickness uniformity M1=99.1%, photolithography pattern alignment accuracy M2=±0.015μm, and etching depth consistency M3=98.9%. A mapping is established according to the association rules: S1 is associated with S2, S2 is associated with S3, M1 is associated with M2, M2 is associated with M3, and S1 corresponds to M1, S2 corresponds to M2, and S3 corresponds to M3. The initial confidence level of all association relationships is marked as 0.95.
[0032] The process nodes are specifically defined as follows: upstream process nodes P1 (raw material purification, process ID: P001), P2 (crystallization growth, process ID: P002), P3 (slicing processing, process ID: P003); midstream process nodes P4 (epitaxy growth, process ID: P004), P5 (photolithography, process ID: P005), P6 (etching process, process ID: P006). Data nodes are specifically defined as follows: upstream data nodes S1 (ID: D001), S2 (ID: D002), S3 (ID: D003); midstream data nodes M1 (ID: D004), M2 (ID: D005), M3 (ID: D006).
[0033] The temporal edges connect P1→P2→P3→P4→P5→P6, with temporal priorities labeled as 1, 2, 3, 4, and 5 respectively; the belonging edges connect S1-P1 (collection timing: 5 minutes after the end of process P1), S2-P2 (collection timing: 3 minutes after the end of process P2), and S3-P3 (collection timing: 4 minutes after the end of process P3). The collection timings for M1-P4, M2-P5, and M3-P6 are all 5 minutes after the end of their respective processes; among the associated edges, the confidence level of intra-process associated edges is labeled as 0.93, and the confidence level of cross-process associated edges is labeled as 0.95.
[0034] The process node attributes are as follows: P1 standard execution time is 2 hours, and the preceding process ID is empty; P2 standard execution time is 8 hours, and the preceding process ID is P001; P3 standard execution time is 4 hours, and the preceding process ID is P002; P4-P6 standard execution times are 6 hours, 3 hours, and 2 hours respectively, and the preceding process IDs are P003, P004, and P005 respectively. Data node attributes: S1 acquisition timestamp is 2026-01-21 08:05:00, value range is 99.995%-99.999%, acquisition device ID: EQ001; align the S1 acquisition timestamp with the P1 execution time window (2026-01-21 06:00:00-08:00:00), sort the entire process data according to the priority of the time-series edges, complete the graph structure data mounting, and form a full-process time-series data model.
[0035] Step S200: Decompose the entire process time series data model, separate the data sequences, including trend sequence, periodic sequence and residual sequence; collect batch parameters and quantify them; Specifically, based on the full-process time-series data model, time-series datasets are extracted, including process node time-series datasets and data node time-series datasets, and the time granularity is unified and the dimensions are aligned. A sliding window fitting logic is employed, setting a fixed-length time window for the time series data. Linear or nonlinear fitting is performed on the data within the window, and the fitting results of all windows are merged to form a continuous trend term sequence. The extracted trend term is removed from the time series dataset to obtain detrended time series data. Based on the inherent cycle time of the production process, the cycle length is determined. Cycle pattern matching is performed on the detrended time series data to extract recurring fluctuation components, forming a cycle term sequence. Using the original time series data as a benchmark, the trend term sequence and the cycle term sequence are subtracted, and the resulting difference is used as the residual term sequence.
[0036] In one specific embodiment, based on the constructed full-process time-series data model, the process node time-series dataset (including the hourly execution duration records of each process from P1 to P6) and the data node time-series dataset (including the hourly monitoring values of S1 to M3) are extracted. The time granularity is unified with "hours", and all data dimensions are aligned to the production timeline from 06:00 on 2026-01-21 to 06:00 on 2026-01-22, ensuring that the process duration data and the corresponding data node monitoring values are accurately matched in the same time dimension.
[0037] A sliding window with a fixed length of 4 hours was used to perform linear fitting on the aligned time series data. For example, the fitted values for the duration within the P2 process (crystallization growth) window were 8.0h, 8.02h, 8.03h, and 8.05h, respectively, while the fitted value for the S1 purity remained in the range of 99.998%-99.9982%. The fitting results of all windows were merged to form a continuous trend sequence, in which the process duration showed a slow increasing trend, and the drift of the core data values was controlled within ±0.0002% / hour.
[0038] After removing trend items from the time-series data, the cycle length was determined to be 8 hours based on the inherent production rhythm of 3 shifts per day (8 hours per shift). Cycle pattern matching was performed on the detrended data to extract recurring fluctuation components: P1 process duration fluctuation ±0.1h for the morning shift (06:00-14:00), ±0.07h for the afternoon shift (14:00-22:00), and ±0.08h for the evening shift (22:00-06:00). Based on this, a cycle item sequence for each process was formed, with the core data fluctuation cycle completely consistent with the shift rhythm.
[0039] Based on the original time series data, the corresponding trend and periodic values were subtracted to obtain the residual sequence. Statistical analysis showed that the residual time for each process was within ±0.05h, the residuals for S1-S3 were within ±0.0001%, and the residuals for M1-M3 were within ±0.001μm / ±0.01%. Stable time series intervals with residuals less than 0.03h (process), 0.00008% (purity), and 0.0008μm (accuracy) were selected, concentrated in the middle 4-hour period of each shift.
[0040] The parameters of three production batches were collected and quantified as follows: Batch 1 has a planned delivery time of 10:00 on January 22, 2026, an urgency level weight of 4 (levels 1-5), and a process connection priority score of 8 (points 1-10); Batch 2 and 3 have delivery times of 16:00 on January 22, 2026 and 08:00 on January 23, 2026, respectively, with urgency level weights of 3 and 2, and process connection priority scores of 7 and 6, respectively. The batch parameters have been quantified.
[0041] Step S300: Combining sequence and batch parameters, construct an initial batch scheduling framework based on a genetic algorithm to obtain a scheduling scheme; introduce the analytic hierarchy process (AHP) to optimize the scheduling scheme to obtain an optimized scheduling scheme; align the optimized scheduling scheme with the graph-structured time-series data model to obtain a scheduling table; Specifically, sequence features are extracted as follows: In the trend sequence, the long-term trend value of the execution time of each process and the numerical drift trend of each core data are extracted as the basis for adjusting the process time baseline; in the periodic sequence, based on the inherent cycle time of the production process, the periodic fluctuation amplitude of each process is extracted as the constraint boundary for process time fluctuation; in the residual sequence, the variance range of the residuals of each process node and data node is calculated, and stable time intervals with residuals less than the threshold are selected as the reference for priority time windows in batch scheduling; and the correspondence between sequence features and batch parameters is established. A real-number encoding method is adopted to ensure a one-to-one correspondence between chromosomes and batch scheduling schemes; constraints of the genetic algorithm are defined; a multi-objective weighted function is constructed by fusing sequence features and batch parameters with the goal of maximizing fitness; an initial population is generated, and selection, crossover, and mutation operations are performed; an iteration termination condition is set, and the chromosome with the highest fitness value in the final population is used as the scheduling scheme.
[0042] Furthermore, the constraints of genetic algorithms include hard constraints and soft constraints; hard constraints represent requirements that must be met, while soft constraints represent optimization goal orientation. The hard constraints of genetic algorithms include timing constraints, delivery constraints, and dependency constraints. The timing constraints specifically require that batch processes follow the timing edge logic of the graph structure, subsequent processes cannot be started until the preceding process is completed, and the process duration cannot exceed the superposition range of the trend term correction benchmark and the periodic term fluctuation amplitude. The delivery constraints specifically require that the end time of all processes in a batch cannot be greater than the planned delivery time. The dependency constraints specifically require that based on the cross-process association relationship of the associated edges, the corresponding midstream process is started after the upstream data node of the same batch has completed the collection and the value meets the standard. The soft constraints of genetic algorithms include priority constraints and stability constraints. The priority constraint specifically means that batch combinations with higher process connection priority are allocated continuous time windows first, and batches with higher delivery urgency weights are given priority to occupy the time window where the residual is stable. The stability constraint specifically means that the fluctuation range of batch process duration must conform to the periodic term pattern, and the impact of residual term must be controlled within the allowable range.
[0043] Furthermore, focusing on optimizing the production batch scheduling scheme for SiC devices, and combining genetic algorithm constraints and sequence characteristics, a target layer, criterion layer, sub-criterion layer, and scheme layer are constructed, with each layer strongly correlated with previous data. This ensures that each layer corresponds to the full-process time-series data model and genetic algorithm constraints. Based on actual production needs and previous data characteristics, a nine-level scaling method is used to construct judgment matrices for the criterion layer versus the target layer, the sub-criterion layer versus the criterion layer, and the scheme layer versus the sub-criterion layer. Weight calculations and consistency checks are performed. Based on normalized weights, the priority ranking of each sub-criterion layer element relative to its corresponding criterion layer element, and the priority ranking of each scheme layer element relative to its corresponding sub-criterion layer element are determined. Combining the weights of each level, the comprehensive weight of each candidate scheme in the scheme layer relative to the target layer is calculated, and the scheme with the highest comprehensive weight is selected as the optimized scheduling scheme.
[0044] Furthermore, based on the graph structure and time-series edge logic, the order and duration of the scheduled processes are verified; based on the belonging edges, the data node collection timestamps are aligned with the corresponding process time windows; based on the association edges, the rationality of the time sequence connection of cross-process data is verified; time sequence conflicts are investigated, deviations are corrected according to hard constraints, the impact of residual terms is controlled within the allowable range, and the alignment scheme is ensured to be compliant; the aligned time windows, process connection relationships, delivery thresholds and stability indicators are summarized to generate a scheduling table.
[0045] In one specific embodiment, the P2 process duration increases by 0.02 hours per batch from the trend item sequence, serving as the basis for time consumption benchmark correction. The S1 purity drift is controlled within ±0.0001% per batch. The fluctuation range of each process is extracted from the periodic item sequence: P1 fluctuation is ±0.1 hours for the morning shift, ±0.07 hours for the afternoon shift, and ±0.08 hours for the evening shift, serving as the time constraint boundary. The process residual variance is statistically analyzed from the residual item sequence and found to be ≤0.0025h². The middle 4 hours of each shift are selected as the stable time sequence interval and prioritized for batches with high delivery urgency weight. A correlation is established: the process duration corrected by the trend item is bound to the batch base consumption time, the periodic item fluctuation range is set as the upper limit of time fluctuation, the stable time sequence interval is matched with high-urgency batches, and batches with high process connection priority scores are allocated continuous time windows.
[0046] Using real-number encoding, for 3 batches and 6 processes, the chromosome length is designed to be 3×6×2=36 bits. Every two genes correspond to the start / end time of one process, achieving a one-to-one correspondence between chromosomes and scheduling schemes. Constraints are set: hard constraints require that batch processes follow the sequence P1→P2→P3→P4→P5→P6, the duration of process P2 must be within 8.0h±0.1h, the end time of batch 1 must not be later than 10:00 on January 22, 2026, and P4 can only be started after S3 data collection is completed and the values meet the standards; soft constraints require that batch 1 prioritizes occupying stable time intervals, and batches 1 and 2 are prioritized for continuous window allocation. A multi-objective weighted fitness function is constructed, with a population size of 50. A tournament selection method (k=3), crossover probability of 0.7, and mutation probability of 0.03 are used. After 50 iterations, the chromosome with the highest fitness value is selected as the initial scheduling scheme. This scheme achieves a delivery compliance rate of 98% and a process priority matching rate of 95%.
[0047] The objective layer aims to generate the optimal scheduling scheme. The criterion layer includes delivery compliance, process integration matching degree, and time sequence stability. The sub-criterion layer further subdivides these into delivery time deviation rate, integration priority score, and residual impact rate. The scheme layer selects three candidate scheduling schemes output by the genetic algorithm. A nine-level scaling method is used to construct the judgment matrix. In the criterion layer, the scale of delivery compliance relative to process integration matching degree is 2, and the scale of relative time sequence stability is 3. The calculated consistency ratio CR = 0.06 < 0.1, which meets the consistency requirement. Through hierarchical single sorting and overall sorting, candidate scheme 1 is determined to have a comprehensive weight of 0.42, making it the optimal scheme. This scheme adjusts the P2 process of batch 1 to a stable time sequence interval, reduces the delivery time deviation rate to 1%, and improves process integration efficiency by 8%.
[0048] The optimized scheduling scheme was aligned with the graph structure time-series data model, and validated according to the time-series edge logic. Specifically, for batch 1, P1 process 06:00-08:00 and P2 process 08:00-16:02 meet the time-series requirement of starting P2 after P1 is completed, and the duration is within the range of trend term correction benchmark and period term fluctuation. Data node timestamps were aligned based on the belonging edge: S2 collection timestamp 2026-01-21 16:03 falls within the P2 process time window; cross-process data connection was verified based on the associated edge: S3 collection completion time 2026-01-21 20:04 and P4 start time 2026-01-21 20:05, meeting dependency constraints. A conflict was found where the start time of batch 2 P4 was earlier than the collection completion time of S3; after correction, a delay of 0.5 hours was implemented to control the residual impact within the allowable range. The aligned data was summarized to generate a scheduling table, clarifying the time window, connection relationship, delivery threshold, and stability indicators for each batch process.
[0049] Step S400: Implement full-process control of SiC device production based on the scheduling table.
[0050] Specifically, the scheduling table is split into upstream and downstream processes and synchronized to each workstation. Graph structure temporal edges and related edge constraints are bound to ensure alignment of the execution end with the control benchmark. Based on the graph structure's belonging edges, execution data and data node monitoring values for each process node are collected in real time and compared with the scheduling table's temporal window. Trend item sequences, periodic item sequences, and residual item sequences are tracked synchronously to verify the compliance of process connections and cross-process data associations. When data deviates from the scheduling threshold or there is a temporal logic conflict, an alert is triggered based on the confidence level of the related edges. After each batch of processes is completed, the actual data is compared with the scheduling table, and the graph structure temporal model parameters are updated.
[0051] In one specific embodiment, the generated schedule is split into upstream and downstream processes and synchronized to each workstation from P1 to P6: Workstation P1 receives work instructions for batch 1 (06:00-08:00), batch 2 (08:30-10:30), and batch 3 (11:00-13:00), while workstation P4 receives instructions for the corresponding batches (20:05-02:05, 02:35-08:35, and 09:05-15:05). Simultaneously, graph structure constraints are bound: timing edges require P1 to be completed before P2 can start, and association edges require an association confidence of ≥0.95 between S1 and M1 to proceed, ensuring that the execution baseline of each workstation is completely aligned with the schedule.
[0052] Based on real-time data collection from the attribution edge: the actual execution time of batch 1 at workstation P2 was 08:00-16:01, lasting 8.01 hours, falling within the scheduling window (08:00-16:02); the real-time monitoring value of S2 was 98.72%, with a collection timestamp of 2026-01-21 16:03, precisely aligned with the P2 process window. Synchronously tracking three types of sequences, the P2 duration trend increased by 0.01 hours per batch, the periodicity fluctuated by ±0.07 hours, and the residual was 0.01 hours. Process connections and the cross-process association between S2 and M2 all met compliance requirements.
[0053] In production, batch 2 at workstation P4 actually started at 02:33, 2 minutes earlier than scheduled, and the real-time value of M1 was 99.05%, deviating from the scheduled threshold (99.1%-99.3%). Based on the confidence level of the S1-M1 association edge of 0.95, a level 2 warning was triggered, and the P4 process was immediately suspended. Upon inspection, it was found that the purity of S1 had slightly drifted to 99.9979%. The start time of P4 was fine-tuned to 02:35, and the monitoring parameters of M1 were simultaneously corrected to ensure compliance with dependency constraints and to avoid affecting the rhythm of subsequent processes.
[0054] After batch 1 was completed, the actual data was compared with the schedule: the total time was 14.01 hours, a deviation of 0.01 hours from the schedule; the actual values of S1-S3 and M1-M3 deviated from the scheduling threshold by ≤0.03% / 0.001μm. Based on this data, the graph structure time series model was updated, the increasing coefficient of the trend item of process P2 was corrected to 0.015 hours / batch, the fluctuation range of the cycle item of process P4 was adjusted to ±0.06 hours, and the residual threshold was optimized to 0.04 hours, improving the control accuracy of subsequent batches.
[0055] like Figure 2 As shown, this application provides an intelligent control system for the entire process of SiC device production, including: Model building module: includes: a data integration unit that integrates core data from upstream and midstream processes, and a model building unit that builds a full-process time-series data model; Model decomposition and parameter quantization module: includes: a model decomposition unit that decomposes the entire time series data model, separating the data sequences, including trend sequence, periodic sequence and residual sequence; and a parameter quantization unit that collects batch parameters and quantizes them. The scheduling table generation module includes: a scheduling scheme construction unit that combines sequence and batch parameters to construct an initial batch scheduling framework based on a genetic algorithm to obtain a scheduling scheme; a scheduling scheme optimization unit that introduces the analytic hierarchy process (AHP) to optimize the scheduling scheme to obtain an optimized scheduling scheme; and a scheduling table generation unit that aligns the optimized scheduling scheme with a graph-structured time-series data model to obtain a scheduling table. The end-to-end management module includes: an end-to-end management unit that manages the entire SiC device production process based on a scheduling table.
[0056] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for intelligent control of the entire process of SiC device production, characterized in that, Includes the following steps: Integrate core data from various processes in the upstream and midstream to construct a full-process time-series data model; The entire process time series data model is decomposed to separate the data sequences, including trend sequence, periodic sequence and residual sequence; Collect batch parameters and quantify them; By combining sequence and batch parameters, an initial batch scheduling framework is constructed based on a genetic algorithm to obtain a scheduling scheme; the analytic hierarchy process (AHP) is then introduced to optimize the scheduling scheme, resulting in an optimized scheduling scheme. The optimized scheduling scheme is aligned with the graph structure time series data model to obtain the scheduling table; The entire process of SiC device production is controlled based on a scheduling table.
2. The intelligent control method for the entire process of SiC device production according to claim 1, characterized in that, The integration of core data from upstream and midstream processes to construct a full-process time-series data model includes: Key data for upstream processes include the purity of substrate raw materials. Substrate crystal integrity and substrate slice thickness tolerance The core data for the midstream process includes the uniformity of the epitaxial layer thickness. Photolithographic pattern alignment accuracy and etch depth consistency ;Will Related , Related , Related , Related , Related , Related , Related ; Define the node system, edge system, and attribute system, construct a graph structure, and integrate the entire process data based on the graph structure to obtain a full-process time-series data model.
3. The intelligent control method for the entire process of SiC device production according to claim 2, characterized in that, The defined node system, edge system, and attribute system are used to construct a graph structure. Based on this graph structure, the entire process data is integrated to obtain a full-process time-series data model, including: The node system is defined using process nodes and data nodes. Upstream process nodes include raw material purification. Crystal growth and slicing The midstream process nodes include epitaxial growth. Photolithography and etching process Upstream data nodes include , and The midstream data nodes include , and ; The edge system is defined using temporal edges, attribution edges, and association edges; among them, temporal edges connect adjacent process nodes, attribution edges connect data nodes with the corresponding generated process nodes, and association edges include intra-process association edges and cross-process corresponding association edges. In the attribute system, the process node attributes include process ID, standard execution duration, and preceding process ID; the data node attributes include data ID, collection timestamp, value range, and collection device ID; the time sequence side is marked with time sequence priority, the affiliation side is marked with data collection timing, and the association side is marked with association confidence. Construct a graph structure, instantiate nodes and mount data, align the collection timestamps of each data node with the execution time windows of the corresponding process nodes, and sort the entire process data according to the production time sequence through the time sequence priority attribute of the time sequence edge, so that the graph structure has both data association logic and production time sequence logic, forming a full-process time sequence data model.
4. The intelligent control method for the entire process of SiC device production according to claim 1, characterized in that, The process of decomposing the entire time-series data model and separating the data sequences includes trend sequence, periodic sequence, and residual sequence, including: Based on the full-process time-series data model, time-series datasets are extracted, including process node time-series datasets and data node time-series datasets, and the time granularity is unified and the dimensions are aligned. A sliding window fitting logic is employed, setting a fixed-length time window for the time series data. Linear or nonlinear fitting is performed on the data within the window, and the fitting results of all windows are merged to form a continuous trend term sequence. The extracted trend term is removed from the time series dataset to obtain detrended time series data. Based on the inherent cycle time of the production process, the cycle length is determined. Cycle pattern matching is performed on the detrended time series data to extract recurring fluctuation components, forming a cycle term sequence. Using the original time series data as a benchmark, the trend term sequence and the cycle term sequence are subtracted, and the resulting difference is used as the residual term sequence.
5. The intelligent control method for the entire process of SiC device production according to claim 1, characterized in that, The combination of sequence and batch parameters, based on a genetic algorithm, constructs an initial batch scheduling framework to obtain a scheduling scheme, including: Extracting sequence features specifically involves: in the trend sequence, extracting the long-term trend of execution time for each process and the numerical drift trend of each core data point, which serve as the basis for adjusting process time consumption benchmarks; in the periodic sequence, extracting the periodic fluctuation amplitude of each process based on the inherent cycle time of the production process, which serves as the constraint boundary for process time consumption fluctuations; in the residual sequence, statistically analyzing the variance range of residuals for each process node and data node, and selecting stable time intervals with residuals less than a threshold, which serve as a reference for priority time windows in batch scheduling; and establishing the correspondence between sequence features and batch parameters. A real-number encoding method is adopted to ensure a one-to-one correspondence between chromosomes and batch scheduling schemes; constraints of the genetic algorithm are defined; a multi-objective weighted function is constructed by fusing sequence features and batch parameters with the goal of maximizing fitness; an initial population is generated, and selection, crossover, and mutation operations are performed; an iteration termination condition is set, and the chromosome with the highest fitness value in the final population is used as the scheduling scheme.
6. The intelligent control method for the entire process of SiC device production according to claim 5, characterized in that, The constraints defining the genetic algorithm include: The constraints of genetic algorithms include hard constraints and soft constraints; hard constraints mean that they must be satisfied, while soft constraints mean that the optimization goal is oriented. The hard constraints of genetic algorithms include timing constraints, delivery constraints, and dependency constraints. The timing constraints specifically require that batch processes follow the timing edge logic of the graph structure, subsequent processes cannot be started until the preceding process is completed, and the process duration cannot exceed the superposition range of the trend term correction benchmark and the periodic term fluctuation amplitude. The delivery constraints specifically require that the end time of all processes in a batch cannot be greater than the planned delivery time. The dependency constraints specifically require that based on the cross-process association relationship of the associated edges, the corresponding midstream process is started after the upstream data node of the same batch has completed the collection and the value meets the standard. The soft constraints of genetic algorithms include priority constraints and stability constraints. The priority constraint specifically means that batch combinations with higher process connection priority are allocated continuous time windows first, and batches with higher delivery urgency weights are given priority to occupy the time window where the residual is stable. The stability constraint specifically means that the fluctuation range of batch process duration must conform to the periodic term pattern, and the impact of residual term must be controlled within the allowable range.
7. The intelligent control method for the entire process of SiC device production according to claim 1, characterized in that, The introduction of the analytic hierarchy process (AHP) to optimize the scheduling scheme yields an optimized scheduling scheme, including: With optimizing the production batch scheduling scheme for SiC devices as the core, and combining genetic algorithm constraints and sequence characteristics, a target layer, criterion layer, sub-criterion layer, and scheme layer are constructed, with each layer strongly correlated with previous data. This ensures that each layer corresponds to the full-process time-series data model and genetic algorithm constraints. Based on actual production needs and previous data characteristics, a nine-level scaling method is used to construct judgment matrices for the criterion layer versus the target layer, the sub-criterion layer versus the criterion layer, and the scheme layer versus the sub-criterion layer. Weight calculations and consistency checks are performed. Based on normalized weights, the priority ranking of each sub-criterion layer element relative to its corresponding criterion layer element, and the priority ranking of each scheme layer element relative to its corresponding sub-criterion layer element are determined. Combining the weights of each level, the comprehensive weight of each candidate scheme in the scheme layer relative to the target layer is calculated, and the scheme with the highest comprehensive weight is selected as the optimized scheduling scheme.
8. The intelligent control method for the entire process of SiC device production according to claim 1, characterized in that, The step of aligning the optimized scheduling scheme with the graph-structured time-series data model to obtain the scheduling table includes: Based on the graph structure and time-series edge logic, verify the sequence and duration of the scheduled processes; based on the attribution edges, align the data node collection timestamps with the corresponding process time windows; based on the association edges, verify the rationality of the time sequence connection of cross-process data; investigate time sequence conflicts, correct deviations according to hard constraints, control the impact of residual terms within the allowable range, and ensure the compliance of the aligned solution; summarize the aligned time windows, process connection relationships, delivery thresholds, and stability indicators to generate a scheduling table.
9. The intelligent control method for the entire process of SiC device production according to claim 1, characterized in that, The full-process management of SiC device production based on the scheduling table includes: The scheduling table is split into upstream and downstream processes and synchronized to each workstation. Graph structure temporal edges and related edge constraints are bound to ensure alignment of the execution end with the control benchmark. Based on the graph structure's belonging edges, execution data and data node monitoring values for each process node are collected in real time and compared with the scheduling table's temporal window. Trend item sequences, periodic item sequences, and residual item sequences are tracked synchronously to verify the compliance of process connections and cross-process data associations. When data deviates from the scheduling threshold or there is a temporal logic conflict, an alert is triggered based on the confidence level of the related edges. After each batch of processes is completed, the actual data is compared with the scheduling table, and the graph structure temporal model parameters are updated.
10. A fully intelligent control system for SiC device manufacturing, using the fully intelligent control method for SiC device manufacturing as described in any one of claims 1-9, characterized in that, include: Model building module: includes: a data integration unit that integrates core data from upstream and midstream processes, and a model building unit that builds a full-process time-series data model; Model decomposition and parameter quantization module: includes: a model decomposition unit that decomposes the entire time series data model, separating the data sequences, including trend sequence, periodic sequence and residual sequence; and a parameter quantization unit that collects batch parameters and quantizes them. The scheduling table generation module includes: a scheduling scheme construction unit that combines sequence and batch parameters to construct an initial batch scheduling framework based on a genetic algorithm to obtain a scheduling scheme; a scheduling scheme optimization unit that introduces the analytic hierarchy process (AHP) to optimize the scheduling scheme to obtain an optimized scheduling scheme; and a scheduling table generation unit that aligns the optimized scheduling scheme with a graph-structured time-series data model to obtain a scheduling table. The end-to-end management module includes: an end-to-end management unit that manages the entire SiC device production process based on a scheduling table.