Wafer manufacturing scheduling trigger decision system considering multi-level relations

Abstract

The present application relates to a kind of wafer manufacturing scheduling trigger decision system considering multilevel relationship, information collection module accepts the multidimensional heterogeneous data from system;Information processing module encodes and pre-processes the data collected;Multilevel relationship perception module carries out relationship modeling to the data processed, constructs the multilevel relationship model including physical layer, process layer and scheduling layer;Scheduling trigger decision module carries out scheduling decision and trigger to multilevel relationship model in combination with deep reinforcement learning, realizes predictive trigger, event-driven trigger and multilevel collaborative trigger;GUI result output module visualizes the scheduling result and system state and shows.Solve the problem that existing wafer manufacturing scheduling trigger decision system is difficult to meet the high complexity requirement of wafer manufacturing, the present application can effectively handle the complex scheduling problem in wafer manufacturing process, improve production efficiency and equipment utilization, reduce production cycle time, adapt to dynamic changing manufacturing environment.

Description

TECHNICAL FIELD

[0001] The present application relates to semiconductor manufacturing, intelligent manufacturing and artificial intelligence technology, in particular to a wafer manufacturing scheduling trigger decision system considering multi-level relationships. BACKGROUND

[0002] As the core link of the semiconductor industry chain, the workshop of wafer manufacturing has the complex characteristics of "multi-process (hundreds of processes), multi-device (thousands of devices), and multi-constraint (cleanliness / temperature / timing)", and dynamic events such as device failure, process fluctuation, and order change frequently occur in the production process, resulting in complex wafer manufacturing conditions and difficulty in determining the scheduling trigger timing. The traditional scheduling trigger decision method faces the following challenges:

[0003] Single relationship modeling dimension: existing methods focus on the binary relationship between workpieces and machines, ignoring the cross-layer interaction of the physical layer (device layout), process layer (process dependence), and scheduling layer (resource competition), for example, the device location will affect the workpiece transit time, and then restrict the process order decision.

[0004] Insufficient reliability of scheduling trigger decision: traditional methods use fixed time intervals or simple event-triggered scheduling strategies, lacking predictive analysis of system state; in addition, traditional trigger decision models are difficult to embed domain knowledge such as process rules (such as lithography alignment accuracy requirements), device physical properties (such as diffusion furnace temperature curve), making scheduling decisions inefficient.

[0005] In recent years, manufacturing scheduling methods based on graph neural networks have gradually emerged, however, the existing solutions have not solved the key problems of multi-level relationship modeling, intelligent trigger mechanism, and physical constraint embedding, making it difficult to meet the high complexity requirements of wafer manufacturing. SUMMARY

[0006] In view of the problems of single relationship modeling dimension, insufficient reliability of scheduling trigger decision, and difficulty in meeting the high complexity requirements of wafer manufacturing of the existing wafer manufacturing scheduling trigger decision system, the present application provides a wafer manufacturing scheduling trigger decision system considering multi-level relationships, thereby analyzing the multi-level relationships between each processing area and device based on considering the complex space-time coupling relationship of the wafer manufacturing workshop, and providing a decision support tool for the optimization control of the semiconductor manufacturing system.

[0007] The technical solution of the present application is:

[0008] The wafer manufacturing scheduling trigger decision system considering multi-level relationship comprises an information collection module, an information processing module, a multi-level relationship perception module, a scheduling trigger decision module and a GUI result output module; wherein the information collection module accepts multi-dimensional heterogeneous data from the system, including device state data, process flow data and material state data; the information processing module encodes and pre-processes the data collected by the information collection module, including process state encoding, device state encoding and material state encoding; the multi-level relationship perception module models the relationship of the data processed by the information processing module, and constructs a multi-level relationship model comprising a physical layer, a process layer and a scheduling layer; the scheduling trigger decision module makes scheduling decisions and triggers in combination with the multi-level relationship model and deep reinforcement learning, realizes predictive triggering, event-driven triggering and multi-level collaborative triggering; and the GUI result output module visually displays the scheduling results and system state generated by the scheduling trigger decision module.

[0009] Further, it specifically comprises:

[0010] The information collection module comprises a four-layer metadata architecture of meta-meta model, meta model, metadata and user data; firstly, the meta model instances contained in the product, system and device meta-meta model are described based on ontology, and the metadata involved is described by using the vocabulary, rules and relationships in the meta model instances, and finally the user data is described by the metadata, thereby realizing unified expression of multi-dimensional heterogeneous data of the wafer manufacturing system;

[0011] The information processing module comprises a process state encoding submodule, a device state encoding submodule and a material state encoding submodule; wherein: the process state encoding submodule encodes the state of each process in the production in the order of time dimension; the device state encoding submodule encodes the availability, performance and maintenance state of the device; and the material state encoding submodule encodes the state and position of the material;

[0012] The multi-level relationship perception module adopts a two-stage graph neural network architecture, comprising: a node feature extraction stage using a graph convolution network to extract feature representations of nodes at each level; and a layer relationship modeling stage capturing complex dependency relationships between different levels through a cross-layer message passing mechanism;

[0013] The scheduling trigger decision module models the wafer manufacturing scheduling problem as a Markov decision process, wherein: the state is represented by the multi-level system state encoded by the heterogeneous graph neural network; the action is a scheduling decision and a trigger signal; and the reward function comprehensively considers multiple objectives of capacity, cycle time and device utilization;

[0014] The scheduling trigger decision module trains the scheduling policy network using a proximal policy optimization algorithm to realize end-to-end scheduling decision learning;

[0015] The scheduling triggering decision module adopts three triggering mechanisms: predictive triggering, which is an active scheduling triggering based on state prediction; event-driven triggering, which is a passive scheduling triggering based on key events; and multi-level collaborative triggering, which is a collaborative scheduling triggering based on inter-level relationships.

[0016] The system integrates knowledge learning techniques in the scheduling domain, including: process constraint embedding, which encodes process flow constraints into the graph structure; equipment physical model, which integrates equipment performance and physical characteristics into node features; and spatiotemporal coupling modeling, which explicitly models the spatiotemporal coupling relationship of workpieces flowing between equipment.

[0017] Furthermore, the information collection module is detailed as follows:

[0018] The information collection module is responsible for collecting multi-dimensional heterogeneous data from the wafer manufacturing system, including equipment status data, process flow data, and material status data. This module adopts a four-layer metadata architecture, including: meta-meta-model (ontology): describing basic concepts and defining the ontology of the wafer manufacturing field; meta-model: defining entity types and relationships in the wafer manufacturing field based on the ontology; metadata: describing the attributes and relationships of specific entities; and user data: dynamic data in the actual production process. Through this hierarchical structure, the system can uniformly express the multi-dimensional heterogeneous data of the wafer manufacturing system.

[0019] Furthermore, the information processing module is detailed as follows:

[0020] The process status coding submodule encodes the status of each production step in chronological order. The process status coding can be represented as:

[0021]

[0022] in, E process ( i , t ) indicates process i In time t The encoded vector, P i Indicates process parameters, S i ( t () indicates the process status. T i ( t () indicates process time information. f process For encoding functions;

[0023] The equipment status coding submodule encodes the availability, performance, and maintenance status of the equipment; the equipment status code can be represented as:

[0024]

[0025] wherein, E equipment j , t represents the equipment j at time t , A j t represents the equipment availability, P j t represents the equipment performance, M j t represents the equipment maintenance status, f equipment is an encoding function;

[0026] Material state encoding submodule: encodes the state and position of the material; the material state encoding can be represented as:

[0027]

[0028] wherein, E material k , t represents the material k at time t , S k t represents the material state, L k t represents the material position, Q k t represents the material yield information, f material is an encoding function;

[0029] The encoding result forms a unified state representation as the input of the graph neural network.

[0030] Further, the multi-level relationship perception module is specifically as follows:

[0031] The multi-level relationship perception module constructs a multi-level relationship model based on a heterogeneous graph neural network, including:

[0032] Physical layer: represents physical entities and their relationships, which can be represented as a graph G physical ( V physical , E physical ), wherein,​​​​​​​​V physical is a set of physical entity nodes, E physical is a set of physical relation edges;

[0033] Process layer: represents process steps, process parameters and their dependencies, which can be represented as a graph G process ( V process , E process ) where, V process is a set of process nodes, E process is a set of process dependency edges;

[0034] Scheduling layer: represents scheduling decisions, resource allocation and their constraint relationships, which can be represented as a graph G scheduling ( V scheduling , E scheduling ), V scheduling is a set of scheduling nodes, E scheduling is a set of scheduling constraint edges;

[0035] The system adopts a two-stage graph neural network architecture:

[0036] Node feature extraction stage: use graph convolutional network to extract the feature representation of each layer node; for node v , its feature update can be represented as:

[0037]

[0038] where, denotes the feature representation of node v at the l th layer, N ( v ) denotes the neighbor set of node v , AGGREGATE is an aggregation function, W (l) and b (l) are learnable parameters, σ is an activation function, u denotes the adjacent node of node v ;

[0039] Inter-layer relationship modeling stage: capture the complex dependency relationship between different layers through cross-layer message passing mechanism; for nodes and The inter-layer message passing can be represented as:

[0040]

[0041]

[0042] wherein, m o,p represents the message passed from node v p to node v o , e o,p is the edge feature between nodes, N cross ( v o ) is the cross-layer neighbor set of node v o , φ and Ψ are learnable functions, is the updated feature representation of node v o after inter-layer message passing, is the original feature representation of node v o , is the feature representation of node v p .

[0043] Further, the scheduling trigger decision module is specifically as follows:

[0044] The scheduling trigger decision module combines deep reinforcement learning to make scheduling decisions and triggers. The system models the wafer manufacturing scheduling problem as a Markov decision process:

[0045] State: the multi-level system state encoded by the heterogeneous graph neural network, which can be represented as s t = { , , };

[0046] wherein, is the graph representation of the physical layer at time t , is the graph representation of the process layer at time t , is the graph representation of the scheduling layer at time t ;

[0047] Action: scheduling decision and trigger signal, which can be represented as a t =d t , τ t}, wherein d t is a scheduling decision, τ t is a triggering signal;

[0048] Reward: Multi-objective considering capacity, cycle time, and equipment utilization, which can be expressed as:

[0049]

[0050] wherein, r throughput , r cycle_time , r utilization respectively represent the rewards related to capacity, cycle time, and equipment utilization, r trigger represent the reward related to triggering, w 1, w 2, w 3, w 4 are weight coefficients;

[0051] The system uses a proximal policy optimization algorithm to train the scheduling policy network, and the objective function is:

[0052]

[0053] wherein, r t ( θ ) represents the probability ratio of the new and old policies, represents the advantage function estimate, ε is a clipping parameter, is the expected calculation of the time step t , and clip is a clipping function.

[0054] Further, the mechanism triggered by the scheduling triggering decision module is as follows:

[0055] Predictive triggering: proactive scheduling triggering based on state prediction; the system triggers the scheduling decision by predicting the future state s t+Δt when the predicted state meets the triggering condition:

[0056]

[0057] wherein, f trigger is a triggering evaluation function, θ predictivefor predictive trigger threshold, I is an indicator function;

[0058] Event-driven trigger: passive dispatch trigger based on key events; when the system detects key events , trigger dispatch decisions:

[0059]

[0060] wherein, E critical is a set of key events; Multi-level collaborative trigger: collaborative dispatch trigger based on inter-layer relationships; the system considers the states and relationships of different levels to trigger collaboratively:

[0061]

[0062] wherein, f collaborative is a collaborative trigger evaluation function, θ collaborative is a collaborative trigger threshold.

[0063] Further, the system integrates scheduling domain knowledge learning technology, embedding wafer manufacturing domain knowledge and physical constraints into the model:

[0064] Process constraint embedding: encode process constraints into graph structure; for process constraint c, its embedding can be represented as:

[0065]

[0066]

[0067] wherein, f constraint is a constraint encoding function, C is a set of process constraints, g embed is a graph embedding function; is the updated process layer graph structure after embedding the process constraints, e c is the embedding vector of the process constraint c after the numbering function processing;

[0068] Device physical model: integrate device performance and physical characteristics into node features; for device node v equipment , the feature enhancement can be represented as:

[0069]

[0070] wherein, P equipment represents the physical characteristics of the device,f physics is a function of physical properties, is a node v equipment is a raw feature representation of the node, is a node v equipment is a new feature representation of the node after feature enhancement;

[0071] Space-time coupling modeling: explicitly modeling the space-time coupling relationship of workpieces flowing between devices; for a workpiece w flowing from a device I to a device J , its space-time coupling relationship can be represented as:

[0072]

[0073] wherein, L I and L J denote the positions of the devices I and J , denotes the flow time of the workpiece w from the device I to the device J , f spatiotemporal is a function of space-time coupling.

[0074] Further, the workflow of the system is as follows:

[0075] Data collection and preprocessing: collecting multi-source heterogeneous data from MES, ERP devices and sensors, and performing unified representation through a multi-dimensional state encoding mechanism;

[0076] Multi-level relationship perception: constructing a heterogeneous graph representation of the physical layer, process layer and scheduling layer, and extracting multi-level features and relationships through a two-stage graph neural network;

[0077] Scheduling state evaluation: evaluating scheduling requirements based on the current system state and historical data, calculating key performance indicators and bottleneck analysis;

[0078] Scheduling decision and triggering: generating scheduling decisions through a deep reinforcement learning policy network, and determining the scheduling execution time according to the triggering mechanism;

[0079] Result feedback and optimization: recording the scheduling execution results and system response, and continuously optimizing the scheduling strategy through online learning.

[0080] Further, the basic concepts described in the meta-meta model include products, systems, and devices; the entity types and relationships in the meta model include device types, process types, and material types; the attributes and relationships of entities in the meta data include specific device parameters and specific process requirements; and the dynamic data in the user data includes device states, workpiece positions, and processing progress.

[0081] The beneficial effects of the present application are:

[0082] The complex scheduling problem in wafer manufacturing processes can be effectively handled, production efficiency and equipment utilization can be improved, production cycle time can be reduced, and dynamic changes in the manufacturing environment can be adapted. Specifically as follows:

[0083] 1. By multi-level heterogeneous graph representation, the complex relationships in the wafer manufacturing system are comprehensively captured, and the accuracy and effectiveness of scheduling decisions are improved.

[0084] 2. By a two-stage graph neural network architecture, the complex relationships within and between layers are effectively handled, and the model's expression and learning capabilities are enhanced.

[0085] 3. By an intelligent triggering mechanism, precise scheduling intervention is achieved, unnecessary scheduling calculations and system interference are reduced.

[0086] 4. By scheduling domain knowledge learning, domain knowledge and physical constraints are integrated, and the model's explainability and reliability are improved. BRIEF DESCRIPTION OF DRAWINGS

[0087] Figure 1 The system structure diagram of the present application;

[0088] Figure 2 The meta data model framework diagram of the present application;

[0089] Figure 3 The information encoding schematic diagram of the present application;

[0090] Figure 4 The multi-level heterogeneous graph structure schematic diagram of the present application;

[0091] Figure 5 The scheduling triggering decision flowchart of the present application. DETAILED DESCRIPTION

[0092] The present application will be described in detail below in conjunction with the drawings and specific embodiments. The present embodiment is implemented on the basis of the technical solution of the present application, and gives a detailed implementation manner and specific operation process, but the protection scope of the present application is not limited to the following embodiments.

[0093] The wafer manufacturing scheduling trigger decision system considering multi-level relationship comprises an information collection module, an information processing module, a multi-level relationship perception module, a scheduling trigger decision module and a GUI result output module, wherein the information collection module accepts multi-dimensional heterogeneous data from the system, including device state data, process flow data and material state data; the information processing module encodes and pre-processes the data collected by the information collection module, including process state encoding, device state encoding and material state encoding; the multi-level relationship perception module models the data processed by the information processing module to build a multi-level relationship model comprising a physical layer, a process layer and a scheduling layer; the scheduling trigger decision module makes scheduling decisions and triggers in combination with deep reinforcement learning based on the multi-level relationship model built by the multi-level relationship perception module, to realize predictive triggering, event-driven triggering and multi-level collaborative triggering; and the GUI result output module visually displays the scheduling results and system state generated by the scheduling trigger decision module.

[0094] The information collection module comprises a four-layer metadata architecture of meta-meta model (i.e. ontology), meta model, metadata and user data, which firstly describes the meta model instances contained in the product, system and device meta-meta model based on ontology, then describes the involved metadata by using the vocabulary, rules and relationships in the meta model instances, and finally describes the user data by metadata, thereby realizing unified expression of multi-dimensional heterogeneous data of the wafer manufacturing system.

[0095] The information processing module comprises a process state encoding submodule, a device state encoding submodule and a material state encoding submodule, wherein the process state encoding submodule encodes the state of each process in the production in the order of time dimension; the device state encoding submodule encodes the availability, performance and maintenance state of the device; and the material state encoding submodule encodes the state and position of materials such as wafers and masks.

[0096] The multi-level relationship perception module adopts a two-stage graph neural network architecture, which comprises a node feature extraction stage using a graph convolution network to extract feature representations of nodes in each layer, and a layer relationship modeling stage capturing complex dependency relationships between different levels through a cross-layer message passing mechanism.

[0097] The scheduling trigger decision module models the wafer manufacturing scheduling problem as a Markov decision process, wherein the state is represented by the multi-level system state encoded by the heterogeneous graph neural network, the action is the scheduling decision and the trigger signal, and the reward function comprehensively considers multiple objectives such as capacity, cycle time, device utilization, etc.

[0098] The scheduling trigger decision module trains the scheduling policy network using the proximal policy optimization (PPO) algorithm to realize end-to-end scheduling decision learning.

[0099] The scheduling trigger decision module designs three trigger mechanisms: predictive trigger, active scheduling trigger based on state prediction; event-driven trigger, passive scheduling trigger based on key events; multi-level collaborative trigger, collaborative scheduling trigger based on inter-layer relationship.

[0100] The system integrates scheduling domain knowledge learning technology, including: process constraint embedding, encoding process flow constraints into graph structure; device physical model, integrating device performance and physical characteristics into node features; space-time coupling modeling, explicitly modeling the space-time coupling relationship of workpieces between devices.

[0101] In order to realize the intelligent scheduling of wafer manufacturing system, the wafer manufacturing scheduling trigger decision system considering multi-level relationship is proposed, which is divided into information collection module, information processing module, multi-level relationship perception module, scheduling trigger decision module and GUI result output module five parts, as Figure 1 shown.

[0102] The information collection module is responsible for collecting multi-dimensional heterogeneous data from the wafer manufacturing system, including device state data, process flow data and material state data. This module adopts a four-layer metadata architecture, as Figure 2 shown, including: (1) Meta-Meta Model (Ontology): Describes basic concepts such as products, systems and devices, and defines the ontology of the wafer manufacturing field. (2) Meta Model: Define entity types and relationships in the wafer manufacturing field based on ontology, such as device type, process type, material type, etc. (3) Metadata: Describe the attributes and relationships of specific entities, such as the parameters of a specific device, the requirements of a specific process, etc. (4) User data: Dynamic data in actual production process, such as device state, workpiece position, processing progress, etc. Through hierarchical structure, the system can uniformly express the multi-dimensional heterogeneous data of wafer manufacturing system, providing a basis for subsequent graph neural network modeling.

[0103] The information processing module encodes and pre-processes the collected data, as Figure 3 shown, including:

[0104] (1) Process state encoding submodule: Encode the state of each process in the production in time dimension as sequence. Process state encoding can be represented as:

[0105]

[0106] Where, E process ( i , t ) represents the encoding vector of process i at time t , Pi represents the process parameter, S i ( t ) represents the process status, T i ( t ) represents the process time information, f process is an encoding function.

[0107] (2) Equipment status encoding submodule: encodes the availability, performance, and maintenance status of the equipment. The equipment status encoding can be represented as:

[0108]

[0109] wherein, E equipment ( j , t ) represents the encoding vector of the equipment j at time t , A j ( t ) represents the equipment availability, P j ( t ) represents the equipment performance, M j ( t ) represents the equipment maintenance status, f equipment is an encoding function.

[0110] (3) Material status encoding submodule: encodes the status and location of materials such as wafers and masks. The material status encoding can be represented as:

[0111]

[0112] wherein, E material ( k , t ) represents the encoding vector of the material k at time t , S k ( t ) represents the material status, L k ( t ) represents the material location, Q k ( t ) represents the material yield information, f material is an encoding function.

[0113] The encoding result forms a unified state representation as input of the graph neural network.

[0114] The multi-level relationship perception module constructs a multi-level relationship model based on a heterogeneous graph neural network, as shown in Figure 4 , including:

[0115] (1) Physical layer: representing physical entities such as equipment, materials, and their relationships, which can be represented as a graph G physical ( V physical , E physical ), where V physical is a set of physical entity nodes, E physical is a set of physical relationship edges.

[0116] (2) Process layer: representing process steps, process parameters, and their dependency relationships, which can be represented as a graph G process ( V process , E process ), where V process is a set of process nodes, E process is a set of process dependency edges.

[0117] (3) Scheduling layer: representing scheduling decisions, resource allocation, and their constraint relationships, which can be represented as a graph G scheduling ( V scheduling , E scheduling ), V scheduling is a set of scheduling nodes, E scheduling is a set of scheduling constraint edges.

[0118] The system adopts a two-stage graph neural network architecture:

[0119] (1) Node feature extraction stage: using a graph convolution network to extract the feature representation of each layer node. For a node v , its feature update can be represented as:

[0120]

[0121] where represents the feature representation of node v in the l layer, N (v represents the neighbor set of node v , AGGREGATE is an aggregation function, W (l) and b (l) is a learnable parameter, σ is an activation function, u represents the adjacent node of node v .

[0122] (2) Inter-layer relationship modeling stage: Capture the complex dependency between different layers through cross-layer message passing mechanism; For nodes and from different layers, the inter-layer message passing can be represented as:

[0123]

[0124]

[0125] wherein, m o,p represents the message passed from node v p to node v o , e o,p is the edge feature between nodes, N cross ( v o ) is the cross-layer neighbor set of node v o , φ and Ψ is a learnable function, is the updated feature representation of node v o after inter-layer message passing, is the original feature representation of node v o , is the feature representation of node v p .

[0126] The scheduling trigger decision module combines deep reinforcement learning to make scheduling decisions and triggers, as shown in Figure 5 . The system models the wafer manufacturing scheduling problem as a Markov decision process:

[0127] (1) State: The multi-level system state encoded by the heterogeneous graph neural network, which can be represented as s t = { , , };

[0128] wherein, is t a graphical representation of the physical layer at time t, is t a graphical representation of the process layer at time t, is t a graphical representation of the scheduling layer at time t.

[0129] (2) Action: scheduling decision and trigger signal, which can be represented as a t = d t , τ t}, wherein d t is the scheduling decision, τ t is the trigger signal.

[0130] (3) Reward: considering multiple objectives such as capacity, cycle time, equipment utilization, etc., which can be represented as:

[0131]

[0132] wherein, r throughput , r cycle_time , r utilization represent the rewards related to capacity, cycle time, and equipment utilization, respectively, r trigger represents the reward related to triggering, w 1, w 2, w 3, w 4 are weight coefficients.

[0133] The system uses the Proximal Policy Optimization (PPO) algorithm to train the scheduling policy network, and its objective function is:

[0134]

[0135] wherein, r t ( θ ) represents the probability ratio of new and old policies, represents the advantage function estimate, ε is the clipping parameter, is the expected calculation of time step t , clip is the clipping function.

[0136] Three triggering mechanisms are designed in the system:

[0137] (1) Predictive triggering: proactive scheduling triggering based on state prediction. The system triggers scheduling decisions in advance by predicting future states s t+Δt when the predicted state meets the triggering condition:

[0138]

[0139] where, f trigger is the triggering evaluation function, θ predictive is the predictive triggering threshold, and I is the indicator function.

[0140] (2) Event-driven triggering: passive scheduling triggering based on key events (such as emergency orders, equipment abnormalities, etc.). When the system detects a key event , triggering scheduling decisions:

[0141]

[0142] where, E critical is the set of key events.

[0143] (3) Multi-level collaborative triggering: collaborative scheduling triggering based on inter-layer relationships. The system considers the states and relationships of different levels to trigger collaboratively:

[0144]

[0145] where, f collaborative is the collaborative triggering evaluation function, θ collaborative is the collaborative triggering threshold.

[0146] The system integrates scheduling domain knowledge learning technology, embedding wafer manufacturing domain knowledge and physical constraints into the model:

[0147] (1) Process constraint embedding: encode process flow constraints into graph structure. For a process constraint c, its embedding can be represented as:

[0148]

[0149]

[0150] where, f constraint is the constraint encoding function, C is the set of process constraints, g embedEmbedding function for graph. is the updated process graph structure after embedding process constraints, e c is the process constraint c Embedding vector after processing by the numbering function.

[0151] (2) Equipment physical model: Integrate equipment performance and physical characteristics into node features. For equipment nodes v equipment The feature enhancement can be represented as:

[0152]

[0153] where, P equipment represents the physical characteristics of the equipment, f physics is the physical characteristics encoding function, is the original feature representation of the node v equipment is the new feature representation after feature enhancement of the node v equipment

[0154] (3) Spatio-temporal coupling modeling: Explicitly model the spatio-temporal coupling relationship of workpieces flowing between equipment; for workpieces w flowing from equipment I to equipment J , the spatio-temporal coupling relationship can be represented as:

[0155]

[0156] where, L I and L J represent the positions of equipment I and J , represents the flow time of workpiece w from equipment I to equipment J , f spatiotemporal is the spatio-temporal coupling encoding function.

[0157] The workflow of the system is as follows:

[0158] (1) Data collection and preprocessing: Collect multi-source heterogeneous data from MES (Manufacturing Information Management System), ERP (Enterprise Resource Planning System), equipment and sensors, and perform unified representation through multi-dimensional state encoding mechanism.

[0159] ​​(2) Multi-level relationship perception: Constructing heterogeneous graph representations of physical, process, and scheduling layers, extracting multi-level features and relationships through a two-stage graph neural network.

[0160] (3) Scheduling state evaluation: Evaluating scheduling needs based on current system state and historical data, calculating key performance indicators and bottleneck analysis.

[0161] (4) Scheduling decision and trigger: Generating scheduling decisions through a deep reinforcement learning policy network, determining scheduling execution timing according to trigger mechanisms.

[0162] (5) Result feedback and optimization: Recording scheduling execution results and system responses, continuously optimizing scheduling strategies through online learning.

[0163] The above-described embodiments only express one embodiment of the present application, which is described in detail and is not considered as a limitation on the scope of the patent. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are within the scope of protection of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A wafer manufacturing scheduling trigger decision system considering multi-level relationships, characterized in that, The system comprises an information collection module, an information processing module, a multi-level relationship perception module, a scheduling trigger decision module and a GUI result output module; the information collection module receives multi-dimensional heterogeneous data from the system, including device state data, process flow data and material state data; the information processing module encodes and pre-processes the data collected by the information collection module, including process state encoding, device state encoding and material state encoding; the multi-level relationship perception module models the data processed by the information processing module to build a multi-level relationship model comprising a physical layer, a process layer and a scheduling layer; the scheduling trigger decision module makes scheduling decisions and triggers based on the multi-level relationship model and deep reinforcement learning, to realize predictive triggering, event-driven triggering and multi-level collaborative triggering; and the GUI result output module visually displays the scheduling results and system state generated by the scheduling trigger decision module; The information collection module comprises a four-layer metadata architecture of meta-meta model, meta model, metadata and user data; firstly, the meta model instances contained in the product, system and device meta-meta model are described based on ontology, and the metadata involved is described by using the vocabulary, rules and relationships in the meta model instances; finally, the user data is described by the metadata, so as to realize unified expression of multi-dimensional heterogeneous data of the wafer manufacturing system; The information processing module comprises a process state encoding submodule, a device state encoding submodule and a material state encoding submodule; the process state encoding submodule encodes the state of each process in the production in the time dimension; the device state encoding submodule encodes the availability, performance and maintenance state of the device; and the material state encoding submodule encodes the state and position of the material; The multi-level relationship perception module adopts a two-stage graph neural network architecture, comprising a node feature extraction stage using a graph convolution network to extract the feature representation of each layer node, and an inter-layer relationship modeling stage capturing the complex dependency relationship between different levels through a cross-layer message passing mechanism; The scheduling trigger decision module models the wafer manufacturing scheduling problem as a Markov decision process, wherein the state is represented by the multi-level system state encoded by the heterogeneous graph neural network, the action is the scheduling decision and the trigger signal, and the reward function considers multiple objectives of capacity, cycle time and device utilization; The scheduling trigger decision module trains the scheduling policy network using a proximal policy optimization algorithm to realize end-to-end scheduling decision learning; The scheduling trigger decision module adopts three triggering mechanisms: predictive triggering, active scheduling triggering based on state prediction; event-driven triggering, passive scheduling triggering based on key events; and multi-level collaborative triggering, collaborative scheduling triggering based on inter-layer relationships; The system integrates scheduling domain knowledge learning techniques, including process constraint embedding, encoding process flow constraints into the graph structure; device physical model, integrating device performance and physical characteristics into node features; and space-time coupling modeling, explicitly modeling the space-time coupling relationship of workpieces between devices. The multi-level relationship perception module constructs a multi-level relationship model based on a heterogeneous graph neural network, including: Physical layer: represents physical entities and their relationships, which can be represented as a graph G physical =( V physical , E physical ), wherein, V physical is a set of physical entity nodes, E physical is a set of physical relationship edges; Process layer: represents process steps, process parameters and their dependencies, can be represented as a graph G process =( V process , E process ), wherein, V process is a set of process nodes, E process is a set of process dependency edges; Scheduling layer: represents scheduling decision, resource allocation and its constraint relationship, which can be represented as a graph G scheduling =( V scheduling , E scheduling ), V scheduling is a set of scheduling nodes, E scheduling is a set of scheduling constraint edges; The system adopts a two-stage graph neural network architecture: a node feature extraction stage: a graph convolutional network is used to extract the feature representation of each layer of nodes; for the nodes v The feature update of which can be represented as: wherein, represents a node v In a first l characteristic representation of a layer, N v represents a set of neighbors of a node v AGGREGATE is an aggregation function, W (l) and b (l) is a learnable parameter, σ is an activation function, u represents a neighboring node of a node v ;​​ Inter-layer relationship modeling phase: The complex dependency between different layers is captured through the cross-layer message passing mechanism; for the nodes from different layers and The inter-layer message passing can be represented as: wherein, m o,p denotes the representation of a node v p passing to a node v o a message, e o,p is an edge feature between nodes, N cross is a set of cross-layer neighbors of a node v o is a set of cross-layer neighbors of a node v o is a set of cross-layer neighbors of a node φ and Ψ is a learnable function, is a representation of a node v o updated through inter-layer message passing, is an original representation of a node v o is an original representation of a node is a representation of a node v p is a representation of a node The scheduling trigger decision module is as follows: The scheduling trigger decision module combines deep reinforcement learning to make scheduling decisions and triggers. The system models the wafer manufacturing scheduling problem as a Markov decision process: State: multi-level system state encoded by the heterogeneous graph neural network, which can be expressed as s t = { , , }; wherein, is t at the moment, the physical layer's graphical representation, is t at the moment, the process layer's graphical representation, is t at the moment, the scheduling layer's graphical representation; Action: dispatching decision and trigger signal, can be represented as wherein is a dispatching decision, is a trigger signal; Reward: Considering multiple objectives such as capacity, cycle time, and equipment utilization, which can be represented as: wherein, r throughput , r cycle_time , r utilization respectively represent capacity, cycle time and equipment utilization related rewards, r trigger represents trigger related rewards, w 1, w 2, w 3, w 4 are weight factors; The system uses a proximal policy optimization algorithm to train the scheduling policy network, and the objective function is: where, r t θ denotes the probability ratio of new versus old policy, denotes the advantage function estimate, ε is the clipping parameter, is the time step t the expected computation, clip is the clipping function;​ Predictive trigger: proactive scheduling trigger based on state prediction; system predicts future state s t+Δt Trigger scheduling decision early when predicted state meets trigger condition: wherein, f trigger is a trigger function, θ predictive is a predictive trigger threshold, I is an indicator function; Multi-level collaborative triggering: collaborative scheduling triggering based on inter-layer relationships; the system considers different levels of state and relationship for collaborative triggering: wherein, f collaborative is a synergistic trigger evaluation function, The information collection module is as follows: collaborative is a synergistic trigger threshold. 2.The wafer manufacturing scheduling trigger decision system considering multi-level relationships according to claim 1, wherein, The information collection module is responsible for collecting multi-dimensional heterogeneous data from the wafer manufacturing system, including device state data, process flow data, and material state data. This module uses a four-layer metadata architecture, including: meta-meta model, i.e. ontology: describes basic concepts and defines the ontology of the wafer manufacturing field; meta model: defines entity types and relationships in the wafer manufacturing field based on the ontology; metadata: describes the attributes and relationships of specific entities; user data: dynamic data in the actual production process; through hierarchical structure, the system can uniformly express the multi-dimensional heterogeneous data of the wafer manufacturing system. The information processing module is as follows:

3. The wafer fabrication scheduling trigger decision system considering multi-level relationships according to claim 1, wherein, Process state encoding submodule: encoding the state of each process in the production in the time dimension; process state encoding can be represented as: Device state encoding submodule: encoding the availability, performance, and maintenance state of the device; device state encoding can be represented as: wherein E process ( i , t ) denotes a process i at time t the encoding vector, P i denotes a process parameter, S i ( t ) denotes a process state, T i ( t ) denotes process time information, f process is an encoding function; Material state encoding submodule: encoding the state and location of the material; material state encoding can be represented as: wherein, E equipment ( j , t ) denotes a device j at a time t of encoding a vector, A j ( t ) denotes a device availability, P j ( t ) denotes a device performance, M j ( t ) denotes a device maintenance status, f equipment is an encoding function; The encoding result forms a unified state representation as the input of the graph neural network. wherein, E material k t represents a material k at a time t of an encoding vector, S k t represents a material state, L k t represents a material position, Q k t represents a material yield information, f material is an encoding function;​​​​​ The mechanism triggered by the scheduling trigger decision module is as follows:

4. The wafer fabrication scheduling trigger decision system considering multi-level relationships according to claim 1, wherein, The system integrates scheduling domain knowledge learning technology, embedding wafer manufacturing field knowledge and physical constraints into the model: Event-driven trigger: passive dispatching trigger based on key events; when the system detects a key event trigger dispatching decisions: wherein E critical is a set of key events.

5. The wafer fabrication scheduling trigger decision system considering multi-level relationships according to claim 1, wherein, Process constraint embedding: encoding process flow constraints into the graph structure; For process constraint c, its embedding can be represented as: The workflow of the system is as follows: wherein, f constraint is a constraint encoding function, C is a set of process constraints, g embed is a graph embedding function; is an updated process layer graph structure after embedding process constraints, e c is a process constraint c is an embedding vector after processing by the numbering function. Device physical model: Incorporate device performance and physical characteristics into node features; for device nodes v equipment The feature enhancement can be expressed as: wherein, P equipment represents a device physical property, f physics is a physical property encoding function, is a node v equipment original feature representation, is a node v equipment new feature representation after feature enhancement; Space-time coupling modeling: explicitly modeling the space-time coupling relationship of workpieces flowing between devices; for workpieces w flowing from device I to device J , the space-time coupling relationship can be represented as: wherein L I and L J representing a device I and J a position of, representing a workpiece w a flow-through time from a device I to a device J f spatiotemporal is a space-time coupled encoding function.​ 6. The wafer fabrication scheduling trigger decision system considering multi-level relationships according to claim 1, wherein, Data collection and preprocessing: Collecting multi-source heterogeneous data from MES, ERP devices and sensors, and uniformly representing through multi-dimensional state encoding mechanism; Multi-level relationship perception: Constructing a heterogeneous graph representation of the physical layer, process layer and scheduling layer, and extracting multi-level features and relationships through a two-stage graph neural network; Scheduling state evaluation: Evaluating scheduling requirements based on current system state and historical data, calculating key performance indicators and bottleneck analysis; Scheduling decision and triggering: generating scheduling decisions through a deep reinforcement learning policy network, and determining scheduling execution time according to the triggering mechanism; Result feedback and optimization: Recording scheduling execution results and system responses, and continuously optimizing scheduling strategies through online learning. ​ 7. The wafer fabrication scheduling trigger decision system considering multi-level relationships according to claim 2, wherein, The basic concepts described in the meta-meta model include products, systems and equipment; the entity types and relationships in the meta model include equipment types, process types, material types; the attributes and relationships of entities in the meta data include parameters of specific equipment, requirements of specific processes; the dynamic data in the user data includes equipment status, workpiece position, processing progress.

Images