A method and apparatus for digital twin construction for semiconductor devices

Abstract

This invention discloses a method and apparatus for constructing a digital twin for semiconductor devices. The method includes: identifying the target device and its corresponding data acquisition unit and acquiring basic data and energy consumption data; encoding the basic data and energy consumption data based on a parameterized object identifier encoding strategy to form corresponding basic databases and energy consumption databases; and linking the basic databases, energy consumption databases, and model units based on object attribute matching rules to construct a digital twin for the semiconductor device. By employing a parameterized object identifier encoding strategy to encode the basic data and energy consumption data, and combining this with object attribute matching rules to link the basic databases, energy consumption databases, and model units, a digital twin for semiconductor devices is constructed. This achieves digital twin mapping while ensuring the effectiveness of the semiconductor device construction, providing a data foundation for issues such as device energy consumption optimization and device maintenance.

Description

Technical Field

[0001] This invention belongs to the technical field of equipment simulation, specifically relating to a method and apparatus for constructing a digital twin of a semiconductor device. Background Technology

[0002] In recent years, semiconductor manufacturing plants have responded to the national call to develop digitalization and digital twin businesses, with "digital factories" and "lighthouse factories" becoming industry benchmarks and directions. However, digital twin technology has developed relatively late and started slowly in the construction of smart factories, and there are many problems: most semiconductor factory MES systems capture complex and difficult-to-process machine-end data, lacking accurate and effective database management, or can only achieve the most realistic restoration possible, rather than analysis, prediction, optimization and improvement. Digital twin factories are still in the early stages of development.

[0003] Patent application CN120124312A discloses an adaptive digital twin simulation device and method for semiconductor equipment energy consumption. The device includes a data acquisition module, a simulation model construction and operation module, and a parameter update module connected by signals. The data acquisition module acquires production information and process parameter information of semiconductor products, as well as real-time actual processing data. The simulation model construction and operation module constructs a process equipment energy consumption simulation model based on a preset process equipment configuration unit, integrates the semiconductor product production information and process parameter information, runs the process equipment energy consumption simulation model, and provides energy consumption simulation data in real time. The parameter update module is used to update the process parameter information in the process equipment energy consumption simulation model. By setting up the parameter update module to introduce a real-time data feedback mechanism, the process parameter information is dynamically updated, improving the accuracy and real-time performance of the process equipment energy consumption simulation model.

[0004] The aforementioned existing technologies focus on the accuracy of energy consumption simulation for semiconductor equipment, but neglect the construction process of each semiconductor device. If there are problems in the construction of the semiconductor device, the results of the energy consumption simulation will be significantly compromised. Ensuring the construction effectiveness of semiconductor equipment and achieving digital twin mapping on the semiconductor device to provide a data foundation for equipment energy consumption optimization and equipment maintenance are problems that need to be solved. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, this invention provides a method and apparatus for constructing a digital twin for semiconductor devices. The method includes: identifying target devices and specifying data acquisition units corresponding to each target device; acquiring basic data and energy consumption data output by the data acquisition units; encoding the basic data and energy consumption data based on a parameterized object identifier encoding strategy to form corresponding basic databases and energy consumption databases; establishing the relationship between the basic data, energy consumption data, and the target device status to form model units; and linking the basic databases, energy consumption databases, and model units based on object attribute matching rules to construct a digital twin for semiconductor devices.

[0006] By employing a parameterized object identification coding strategy, basic data and energy consumption data are encoded separately to form corresponding basic databases and energy consumption databases. The relationship between basic data, energy consumption data, and target equipment status is established to obtain model units. Finally, based on object attribute matching rules, the basic database, energy consumption database, and model units are linked to construct a digital twin for semiconductor equipment. This achieves digital twin mapping while ensuring the construction effect of semiconductor equipment, providing a data foundation for equipment energy consumption optimization, equipment maintenance, and other issues.

[0007] In a first aspect, the present invention provides a method for constructing a digital twin for a semiconductor device, comprising: Identify the target devices and provide the corresponding data acquisition units for each target device; Acquire basic data and energy consumption data output by the data acquisition unit; Based on the parameterized object identification coding strategy, the basic data and energy consumption data are encoded separately to form the corresponding basic database and energy consumption database. Establish the relationship between basic data, energy consumption data, and target equipment status to form model units; Based on object attribute matching rules, a digital twin for semiconductor devices is constructed by linking the basic database, energy consumption database, and model units.

[0008] Furthermore, the basic data includes equipment identification information, equipment production parameters, and equipment environmental parameters.

[0009] Furthermore, based on the parameterized object identifier encoding strategy, the basic data and energy consumption data are encoded separately, specifically including the following steps: Extract the device identity information of the target device, encrypt it, perform device model mapping processing, and generate a device model index code; The equipment production parameters and equipment environmental parameters are encrypted separately to generate logical control codes; Based on the basic data, a low-dimensional semantic vector is generated. Combined with a pre-built hierarchical and cascaded coding architecture, bitwise operations are performed on the device model index code, logic control code, and low-dimensional semantic vector to form the basic code. Based on energy consumption data, construct a multi-dimensional energy consumption feature vector; Multi-dimensional energy consumption feature vectors are concatenated with basic coding to form a dynamic energy consumption coding with bimodal mapping.

[0010] Furthermore, based on the basic data, a low-dimensional semantic vector is generated. This vector, combined with a pre-built hierarchical and cascaded coding architecture, undergoes bitwise operations to fuse the device model index code, logic control code, and low-dimensional semantic vector, forming the basic coding. Specifically, this includes: Obtain semantic vectors of basic data and the relationships between various basic data, and map them to a low-dimensional continuous vector space to form low-dimensional semantic vectors; The low-dimensional semantic vector is discretized and then encrypted to generate a semantic code. Based on a pre-built hierarchical and cascaded coding architecture, the device model index code, logic control code, and semantic code are shifted respectively. The shifted device model index code, logic control code, and semantic code are superimposed to form the basic code.

[0011] Furthermore, the multi-dimensional energy consumption feature vectors are concatenated with the basic encoding to form a dynamic energy consumption encoding with bimodal mapping, specifically including: The multi-dimensional energy consumption feature vector is discretized and then encrypted to provide a timestamp, energy consumption code, and energy quality identifier. Based on a pre-built hierarchical and concatenated coding architecture, the basic code, timestamp, energy consumption code, and energy quality identifier are concatenated to form a concatenated code; In response to the triggering of the mode conversion gateway, and in conjunction with the concatenation code, the mapping relationship between the concatenation code and the simulation mode and the real-time mode is determined respectively; Based on the analysis of conflict perception, combined with the learning agent and reward function, the optimal bit field is given; Based on the optimal bit field, update the hierarchical concatenation coding architecture, update the concatenation code, and adjust the mapping relationship between the concatenation code and the simulation mode and real-time mode.

[0012] Furthermore, based on the analysis of conflict perception, combined with the learned agent and reward function, the optimal bit field is given, specifically including: In response to the analysis of multi-dimensional conflict perception, the utilization rate of bit domain resources is calculated; Construct the encoding state vector and determine the encoding state level; Based on the coding state level, a hierarchical hybrid response model is adopted. By combining the encoded state vector, the action selection is determined; By integrating a multi-objective hierarchical reward function with confidence verification, the optimal bit field is given; The hierarchical hybrid model specifically includes: The primary response sub-model is used to detect changes in the encoded state level and generate trigger instructions. Based on the deep optimization sub-model and combined with the triggering instruction, the changes in the encoding state level are analyzed, and the bit field adjustment range is given. A global planning sub-model is used to find the optimal value within the bit-domain adjustment range to obtain the current bit-domain. The current bit field is validated by combining the validation sub-model, and the optimal bit field is obtained based on the validation results. Multi-objective hierarchical reward functions, specifically including: Based on the response time of each sub-model in the hierarchical hybrid model, and combined with the time decay effect, a time activation term is formed; By integrating the normalized reward values ​​and reward weights of each sub-model in the hierarchical hybrid model, the intra-layer reward sub-items are obtained. Based on the reward triggering thresholds between each sub-model, and combined with the propagation effect of rewards in each sub-model of the hierarchical hybrid model, a global reward sub-item is given; Based on the time activation term, the reward sub-items within the fusion layer and the global reward sub-items, a multi-objective hierarchical reward function is given.

[0013] Furthermore, based on object attribute matching rules, the basic database, energy consumption database, and model units are linked, specifically including the following steps: Establish multi-dimensional attribute matching rules; Execute multi-dimensional attribute matching rules according to priority to construct a holographic data view of the device object; Construct a holographic data association map to link the basic database, energy consumption database, and model units.

[0014] Furthermore, multi-dimensional attribute matching rules are executed according to priority to construct a holographic data view of the device object, specifically including the following steps: Based on priority, attribute matching rules for each dimension are extracted sequentially to obtain a matching rule sequence, which includes primary key association matching rules, fuzzy association matching rules, scenario association matching rules, and function association matching rules. Based on the primary key association matching rules, the basic primary key codes in the basic database, the energy consumption primary key codes in the energy consumption database, and the model primary key codes in the model units are encoded and linked to form a data view primary key chain; Based on the primary key chain of the data view and combined with fuzzy association matching rules, semantic matching is performed on the semantic codes in the basic database to provide basic process information; By using scenario association matching rules, the current spatiotemporal data and basic process information of equipment objects are integrated, and matching is performed in the basic database, energy consumption database and model unit to establish a spatiotemporal chain of data view; Based on the spatiotemporal chain of the data view and the functional association matching rules, the compatibility of the data in the spatiotemporal chain of the data view is judged, and a holographic data view of the device object is constructed.

[0015] Furthermore, the holographic data view of the device object is specifically represented as follows:

[0016] Wherein, Φ(o i ) for device object o i Holographic data view For device object o i The optimal set of basic data in the basic database. For device object o i The optimal set of energy consumption data in the energy consumption database. For device object o i In the optimal model set within the model unit, b′ represents the basic data in the basic database, e′ represents the energy consumption data in the energy consumption database, m′ represents the equipment model within the model unit, and S... total () is the total matching degree function, and argmax() is the maximum value point function.

[0017] In a second aspect, the present invention also provides a digital twin construction apparatus for a semiconductor device, employing a digital twin construction method for a semiconductor device as described in any of the above claims, comprising: The device identification module is used to identify target devices and provide the corresponding data acquisition units for each target device. The data acquisition module is used to acquire basic data and energy consumption data output by the data acquisition unit; The data encoding module is used to encode basic data and energy consumption data based on the encoding strategy of parameterized object identifiers, and form corresponding basic databases and energy consumption databases. The model building module is used to establish the relationship between basic data, energy consumption data, and the status of target equipment to form model units; The twin building unit is used to construct a digital twin for semiconductor devices by linking the basic database, energy database and model unit based on object attribute matching rules.

[0018] The present invention provides a method and apparatus for constructing a digital twin of a semiconductor device, which has at least the following beneficial effects: (1) By adopting a parameterized object identification coding strategy, the basic data and energy consumption data are encoded and processed to form the corresponding basic database and energy consumption database. The relationship between the basic data, energy consumption data and the target equipment status is established to obtain the model unit. Finally, based on the object attribute matching rules, the basic database, energy consumption database and model unit are linked to construct a digital twin for semiconductor equipment. On the basis of ensuring the construction effect of semiconductor equipment, digital twin mapping is realized, providing a data foundation for equipment energy consumption optimization, equipment maintenance and other issues.

[0019] (2) Through conflict perception analysis, the state changes of physical devices are perceived and the hierarchical cascaded coding architecture and the mapping relationship between the cascaded code and the simulation mode and real-time mode are automatically updated, thus opening up the closed loop of "offline simulation-online monitoring-predictive maintenance" and realizing the autonomous evolution and continuous optimization of the dynamic bit length, elastic bit field of the coding architecture and the mapping relationship between the cascaded code and the simulation mode and real-time mode. Attached Figure Description

[0020] Figure 1 A flowchart illustrating a method for constructing a digital twin of a semiconductor device, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the encoding process provided in an embodiment of the present invention; Figure 3 A schematic diagram illustrating the process of determining the basic code provided in an embodiment of the present invention; Figure 4 A schematic diagram of the process for forming a dynamic energy consumption code for a dual-modal mapping provided in an embodiment of the present invention; Figure 5 A model architecture diagram of the hierarchical hybrid model provided in the embodiments of the present invention; Figure 6 A schematic diagram illustrating the process of constructing a holographic data view of a device object according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of the basic database provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of the model unit provided in an embodiment of the present invention; Figure 9 A flowchart illustrating the linking of the basic database, energy consumption database, and model unit provided in this embodiment of the invention; Figure 10 This is a structural block diagram of a digital twin construction device for semiconductor devices provided in an embodiment of the present invention.

[0021] Among them, 201 is the equipment determination module; 202 is the data acquisition module; 203 is the data encoding module; 204 is the model building module; and 205 is the twin construction unit. Detailed Implementation

[0022] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0024] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0025] like Figure 1 As shown, this embodiment of the invention provides a method for constructing a digital twin for a semiconductor device, with the following specific steps: S101: Identify the target devices and provide the data acquisition units corresponding to each target device.

[0026] First, identify at least one target device within the semiconductor factory. This target device can be any equipment in the factory, such as lithography equipment, etching equipment, or cleaning equipment. Configure data acquisition devices, data acquisition interfaces, and other data acquisition components on the target device. Use these data acquisition devices or interfaces as data acquisition units to collect various data generated by the target device during the semiconductor processing.

[0027] S102: Acquire basic data and energy consumption data output by the data acquisition unit.

[0028] The data acquisition units of each target device can collect basic data and energy consumption data generated by the target device. Basic data includes device identification information, production parameters, and environmental parameters. Device identification information includes static attributes such as device number, name, and type. Production parameters include process parameters (e.g., exposure dose, focus position, overlay accuracy), operating status parameters (e.g., production efficiency, mean time between failures, mean time to repair), and energy consumption parameters (e.g., energy utilization rate). Environmental parameters include temperature, humidity, cleanliness, pressure difference, and fluid quality. Energy consumption data includes energy type and duration parameters.

[0029] S103: An encoding strategy based on parameterized object identifiers is used to encode basic data and energy consumption data respectively, forming corresponding basic databases and energy consumption databases.

[0030] Specifically, based on a parameterized object identifier-based encoding strategy, basic data and energy consumption data are encoded separately, referring to... Figure 2 Specifically, it includes the following steps: Extract the device identity information of the target device, encrypt it, perform device model mapping processing, and generate a device model index code; The equipment production parameters and equipment environmental parameters are encrypted separately to generate logical control codes; Based on the basic data, a low-dimensional semantic vector is generated. Combined with a pre-built hierarchical and cascaded coding architecture, bitwise operations are performed on the device model index code, logic control code, and low-dimensional semantic vector to form the basic code. Based on energy consumption data, construct a multi-dimensional energy consumption feature vector; Multi-dimensional energy consumption feature vectors are concatenated with basic coding to form a dynamic energy consumption coding with bimodal mapping.

[0031] In one specific implementation, device identity information is first extracted from the collected basic data. This device identity information can include a single field or be obtained by concatenating multiple fields, and it can uniquely locate the target device. Next, the device identity information is encrypted, and an association is established between it and a unique device model to complete the device model mapping process. The encrypted device identity information is then integrated into the association with the device model to obtain the device model index code corresponding to the target device. Next, the device production parameters and device environmental parameters are encrypted to generate logical control codes. Based on the obtained device model index code and logical control code, semantic analysis is performed on the basic data to generate low-dimensional semantic vectors. These vectors are then fused using bitwise operations on a pre-built hierarchical and cascaded coding architecture to obtain the basic code. Finally, the multi-dimensional energy consumption feature vectors corresponding to the energy consumption data are concatenated with the basic code to form a bimodal mapping dynamic energy consumption code. The multi-dimensional energy consumption feature vectors are obtained by vector transformation of the energy consumption data.

[0032] The dynamic energy consumption coding of bimodal mapping includes static basic coding and dynamic multi-dimensional energy consumption feature vectors. By decoupling the static model (geometry) and dynamic logic (behavior) through different coding bit fields, it embodies the layered architecture of digital twins.

[0033] Furthermore, based on the basic data, a low-dimensional semantic vector is generated. This vector, combined with a pre-built hierarchical and cascaded coding architecture, undergoes bitwise operations to fuse the device model index code, logic control code, and low-dimensional semantic vector, forming the basic code. (Refer to...) Figure 3 Specifically, it includes: Obtain semantic vectors of basic data and the relationships between various basic data, and map them to a low-dimensional continuous vector space to form low-dimensional semantic vectors; The low-dimensional semantic vector is discretized and then encrypted to generate a semantic code. Based on a pre-built hierarchical and cascaded coding architecture, the device model index code, logic control code, and semantic code are shifted respectively. The shifted device model index code, logic control code, and semantic code are superimposed to form the basic code.

[0034] In one specific implementation, the collected basic data and semantic vectors representing the relationships between these basic data points are acquired and mapped to a low-dimensional continuous vector space to form low-dimensional semantic vectors. These low-dimensional semantic vectors are then discretized and encrypted to generate semantic codes. These semantic codes deeply couple device identity information with semantics, providing a data foundation for subsequent intelligent retrieval, alternative recommendations, and process optimization of target devices. In a specific example, if the collected basic data includes A1, A2, A3, and A4, A1, A2, A3, and A4 are converted into semantic vectors, and semantic vectors representing the relationships between these basic data points are generated based on these semantic vectors. Each semantic vector is then mapped to a low-dimensional continuous vector space to generate a corresponding low-dimensional semantic vector. Discretization and encryption of these low-dimensional semantic vectors yield the semantic codes that include the basic data and their relationships.

[0035] The pre-built hierarchical and cascaded encoding architecture includes the distribution of encoding positions and constraints on the number of bits for each encoding. Based on the encoding framework, the device model index code, logic control code, and semantic code are shifted respectively. The shifted device model index code, logic control code, and semantic code are then superimposed to form the basic encoding. In a specific example, the pre-built hierarchical and cascaded encoding architecture includes a 22-bit encoding framework consisting of a 6-bit device model index code in the first layer, an 8-bit logic control code in the second layer, and an 8-bit semantic code in the third layer. The encoding positions of the first, second, and third layers decrease sequentially. First, the device model index code is shifted left by 16 bits, placing it in the high-order bits of the basic encoding. Similarly, the logic control code is shifted left by 8 bits, placing it in the middle of the basic encoding, while the semantic code is in the low-order bits and requires no shifting. After the shifting process, the shifted device model index code, logic control code, and semantic code are bitwise ORed to obtain the 22-bit basic encoding containing the device model index code, logic control code, and semantic code. In other examples, the number of layers and the number of bits per layer in the hierarchical and concatenated coding architecture are adjusted according to the actual situation, without any restrictions.

[0036] By performing bitwise operations to fuse the shifted device model index code, logic control code, and semantic code, the heterogeneous data of the target device is uniformly identified and transformed into a bit pattern space layout, enabling the digital twin system to have a dual retrieval capability of "formal resemblance" and "spiritual resemblance".

[0037] Furthermore, the multi-dimensional energy consumption feature vectors are concatenated with the basic encoding to form a dynamic energy consumption encoding with bimodal mapping, referring to... Figure 4 Specifically, it includes: The multi-dimensional energy consumption feature vector is discretized and then encrypted to provide a timestamp, energy consumption code, and energy quality identifier. Based on a pre-built hierarchical and concatenated coding architecture, the basic code, timestamp, energy consumption code, and energy quality identifier are concatenated to form a concatenated code; In response to the triggering of the mode conversion gateway, and in conjunction with the concatenation code, the mapping relationship between the concatenation code and the simulation mode and the real-time mode is determined respectively; Based on the analysis of conflict perception, combined with the learning agent and reward function, the optimal bit field is given; Based on the optimal bit field, update the hierarchical concatenation coding architecture, update the concatenation code, and adjust the mapping relationship between the concatenation code and the simulation mode and real-time mode.

[0038] In one specific implementation, the multi-dimensional energy consumption feature vector is discretized to obtain time features, energy consumption features, and energy quality features. Based on this, the discretized time features, energy consumption features, and energy quality features are encrypted to obtain feature codes corresponding to each feature, namely, timestamps, energy consumption codes, and energy quality identifiers. By pre-building a hierarchical concatenated coding architecture, the basic codes, timestamps, energy consumption codes, and energy quality identifiers are concatenated to form concatenated codes. Responding to the triggering of the mode conversion gateway and combining the concatenated codes, the simulation mode or real-time mode corresponding to the concatenated codes can be determined. Conflict perception analysis is performed based on the data acquisition volume per unit time by the data acquisition unit. If the data acquisition volume surges, adjustments to the hierarchical concatenated coding architecture are needed. Then, the optimal bit field is given by combining the learning agent and reward function analysis. Based on the optimal bit field, the hierarchical concatenated coding architecture is updated, and the mapping relationship between the concatenated codes and the simulation mode and real-time mode is adjusted. The modality conversion gateway is an externally input command, which can be manually input, triggered by condition settings, or in other forms; there are no limitations on this. The modality conversion gateway can determine the mapping relationship between simulation modes established through concatenation codes or the mapping relationship between real-time modes.

[0039] It is important to understand that the encryption process in the aforementioned steps may employ the same or different encryption methods, such as symmetric encryption, asymmetric encryption, hash functions, and message authentication codes, depending on the specific circumstances. There are no restrictions on this.

[0040] Through conflict perception analysis, the system senses changes in the state of physical devices and automatically triggers updates to the hierarchical and cascaded coding architecture and the mapping relationship between the concatenated code and the simulation mode and real-time mode. This establishes a closed loop of "offline simulation - online monitoring - predictive maintenance," enabling the autonomous evolution and continuous optimization of the dynamic bit length and flexible bit field of the coding architecture, as well as the mapping relationship between the concatenated code and the simulation mode and real-time mode.

[0041] Furthermore, based on the analysis of conflict perception, combined with the learned agent and reward function, the optimal bit field is given, specifically including: In response to the analysis of multi-dimensional conflict perception, the utilization rate of bit domain resources is calculated; Construct the encoding state vector and determine the encoding state level; Based on the coding state level, a hierarchical hybrid response model is adopted. By combining the encoded state vector, the action selection is determined; By integrating a multi-objective hierarchical reward function with confidence verification, the optimal bit field is given; The hierarchical hybrid model specifically includes: The primary response sub-model is used to detect changes in the encoded state level and generate trigger instructions. Based on the deep optimization sub-model and combined with the triggering instruction, the changes in the encoding state level are analyzed, and the bit field adjustment range is given. A global planning sub-model is used to find the optimal value within the bit-domain adjustment range to obtain the current bit-domain. The current bit field is validated by combining the validation sub-model, and the optimal bit field is obtained based on the validation results. Multi-objective hierarchical reward functions, specifically including: Based on the response time of each sub-model in the hierarchical hybrid model, and combined with the time decay effect, a time activation term is formed; By integrating the normalized reward values ​​and reward weights of each sub-model in the hierarchical hybrid model, the intra-layer reward sub-items are obtained. Based on the reward triggering thresholds between each sub-model, and combined with the propagation effect of rewards in each sub-model of the hierarchical hybrid model, a global reward sub-item is given; Based on the time activation term, the reward sub-items within the fusion layer and the global reward sub-items, a multi-objective hierarchical reward function is given.

[0042] In one specific implementation, if the multi-dimensional conflict-aware analysis produces anomalies, it is necessary to calculate the bit-field resource utilization rate, i.e., the bit-field resource occupancy in the concatenated code. Based on the bit-field resource occupancy, an encoding state vector is constructed, and the encoding state level is determined. If the bit-field resource occupancy reaches an occupancy threshold, it indicates that the current bit-field allocation may not meet the encoding requirements of the concatenated code, and the bit-field allocation needs to be adjusted. That is, if the encoding state level reaches a preset level, the bit-field allocation needs to be adjusted. A hierarchical hybrid model is then used to analyze the encoding state vector and provide corresponding action selections. Simultaneously, a multi-objective hierarchical reward function and confidence verification are integrated to provide the optimal bit-field.

[0043] Reference Figure 5The hierarchical hybrid model comprises a primary response sub-model, a deep optimization sub-model, a global planning sub-model, and a validation sub-model. The primary response sub-model quickly detects changes in the encoded state level and generates trigger commands. In this example, the primary response sub-model is Contextual Bandit, which can achieve millisecond-level fast responses and lightweight decision-making. Other response models can be used in other examples, without limitation. The deep optimization sub-model receives trigger commands from the previous layer model, analyzes changes in the encoded state level, and provides a bit-domain adjustment range. In this example, the deep optimization sub-model is DQN, which achieves fine-grained adjustment of the bit-domain distribution. Other deep optimization models such as PPO can be used in other examples, without limitation. The global planning sub-model optimizes within the bit-domain adjustment range output by the deep optimization sub-model to obtain the current bit-domain; that is, it seeks an optimal solution from the bit-domain adjustment range as the current bit-domain. In this example, the global planning sub-model is a Bayesian optimization model. Other global planning models such as genetic algorithms can be used in other examples, without limitation. Finally, the validation sub-model validates the current bit-domain, and the optimal bit-domain is obtained based on the validation results. If the verification result is successful, the current bit field is the optimal bit field. Otherwise, if the verification result is unsuccessful, the process returns to the deep optimization sub-model to redetermine the bit field adjustment range, iterating through the deep optimization sub-model → global planning sub-model → verification sub-model until the current bit field passes verification and is determined to be the optimal bit field. In this example, the verification sub-model is the MILP model. In other examples, other verification models such as dynamic programming can be used, and there is no limitation on this.

[0044] The multi-objective hierarchical reward function is specifically expressed as follows:

[0045] Among them, R total For a multi-objective hierarchical reward function, s is the encoded state vector, a is the trigger command, x is the bit field adjustment range, π is the current bit field, t is the timestamp, l is the model hierarchical encoding, n is the number of model hierarchies, and γ is the multi-objective hierarchical reward function. l (t) represents the time decay weight of the l-layer submodel at time t, τ l (t) represents the hierarchical activation indicator of the l-layer sub-model at time t. For time-activated terms, β l The reward weights for the l-layer sub-model. R is the normalized reward value of the l-layer sub-model. l Let θ be the hierarchical reward value of the l-layer sub-model. l The reward trigger threshold for the l-layer sub-model. The propagation reward value for the l-layer sub-model. For sub-items of rewards within the layer, This is a sub-item of the global reward. and This is an indicator function. It takes the value 1 when the condition of the subscript is met, and takes the value 0 when the condition of the subscript is not met.

[0046] S104: Establish the relationship between basic data, energy consumption data, and target equipment status to form model units.

[0047] Specifically, the relationship between basic data, energy consumption data, and target equipment status is established to form model units, which includes the following steps: Based on the basic data, determine the initial model unit corresponding to the target device; Based on the target equipment status, provide the energy consumption data corresponding to the target equipment status; By using the energy consumption data corresponding to the target equipment status, the operating status of each chamber in the initial model unit is determined, and the parameters of the initial model unit are adjusted to give the model unit.

[0048] In one specific implementation, a unique target device and its corresponding initial model unit can be identified based on basic data. Then, energy consumption data can be collected based on the target device's status. By adjusting the parameters of the initial model unit using this energy consumption data, the model unit can be formed.

[0049] S105: Based on object attribute matching rules, link the basic database, energy consumption database and model units to construct a digital twin for semiconductor devices.

[0050] Specifically, based on object attribute matching rules, the basic database, energy consumption database, and model units are linked, which includes the following steps: Establish multi-dimensional attribute matching rules; Execute multi-dimensional attribute matching rules according to priority to construct a holographic data view of the device object; Construct a holographic data association map to link the basic database, energy consumption database, and model units.

[0051] Furthermore, the holographic data view of the device object is specifically represented as follows:

[0052] Wherein, Φ(o i ) for device object o i Holographic data view For device object o i The optimal set of basic data in the basic database. For device object o i The optimal set of energy consumption data in the energy consumption database. For device object o iIn the optimal model set within the model unit, b′ represents the basic data in the basic database, e′ represents the energy consumption data in the energy consumption database, m′ represents the equipment model within the model unit, and S... total () represents the overall matching degree function, argmax() is the maximum value point function, and S b () is the basic attribute matching degree, used to measure the match between b′ and device object o. i The degree of matching of static attributes, S e () represents the energy consumption characteristic matching degree, used to measure the compatibility between e′ and the device object o. i The degree of matching of dynamic energy consumption behavior, S m () represents the model fit, used to measure the fit of m′ to the device object o. i Analyze the degree of fit to the requirements, w b The weighting coefficient w corresponding to the basic attribute matching degree e w represents the weight coefficient corresponding to the energy feature matching degree. m These are the weight coefficients corresponding to the model fit. Each weight coefficient can be set or adjusted according to the actual situation, and there are no restrictions on this.

[0053] In one specific implementation, Φ(·) is the core mathematical expression of the "linking" step in the construction of the digital twin, representing the successfully matched semiconductor device digital twin object o. i From the basic database, energy consumption database, and model units, the most suitable basic data, energy consumption data, and model units are selected to form a unique "holographic data view" for the device, achieving precise binding between physical equipment and digital information. (Equipment object o) i This corresponds to a specific semiconductor device in the physical world, such as an etching machine or a lithography machine. To interact with device object o i The most suitable basic data set (corresponding to the association results of the "basic database"), and the "static identity profile" of the storage device, such as the device model, production parameter configuration, and encoding of environmental adaptation information. To interact with device object o i The most suitable energy consumption data set (corresponding to the association results of the "energy consumption database"), the "dynamic energy consumption profile" of the storage device, such as real-time power consumption curves, energy consumption time-series characteristics of etching / cleaning processes, and energy efficiency level coding. To interact with device object o i The most suitable model unit (corresponding to the association results of the "model unit"), the storage device's "dedicated analysis toolkit", such as the output features of the state mapping model, operating condition prediction model, and anomaly diagnosis model for this device.

[0054] Furthermore, multi-dimensional attribute matching rules are executed according to priority to construct a holographic data view of the device object, referring to... Figure 6Specifically, it includes the following steps: Based on priority, attribute matching rules for each dimension are extracted sequentially to obtain a matching rule sequence, which includes primary key association matching rules, fuzzy association matching rules, scenario association matching rules, and function association matching rules. Based on the primary key association matching rules, the basic primary key codes in the basic database, the energy consumption primary key codes in the energy consumption database, and the model primary key codes in the model units are encoded and linked to form a data view primary key chain; Based on the primary key chain of the data view and combined with fuzzy association matching rules, semantic matching is performed on the semantic codes in the basic database to provide basic process information; By using scenario association matching rules, the current spatiotemporal data and basic process information of equipment objects are integrated, and matching is performed in the basic database, energy consumption database and model unit to establish a spatiotemporal chain of data view; Based on the spatiotemporal chain of the data view and the functional association matching rules, the compatibility of the data in the spatiotemporal chain of the data view is judged, and a holographic data view of the device object is constructed.

[0055] In one specific implementation, attribute matching rules for each dimension are extracted sequentially according to priority: primary key association matching rules, fuzzy association matching rules, scene association matching rules, and function association matching rules, resulting in a matching rule sequence. Based on the primary key association matching rules, the primary key codes in the basic database, the energy consumption primary key codes in the energy consumption database, and the model primary key codes in the model units are linked to form a data view primary key chain. That is, when any one of the primary key codes—basic, energy consumption, or model—is known, the corresponding data information in the other two databases can be determined based on the links between the primary key codes, forming the data view primary key chain. After obtaining the data view primary key chain, semantic matching is performed on the semantic codes in the basic database using fuzzy association matching rules to provide basic process information. Then, using scene association matching rules, the current spatiotemporal data and basic process information of the equipment object are fused, and matching is performed in the basic database, energy consumption database, and model units to establish a data view spatiotemporal chain. Combined with the function association matching rules, the compatibility of the data in the data view spatiotemporal chain is judged, constructing a holographic data view of the equipment object. In this example, the basic primary key is encoded as the device model index code, and the power usage primary key is encoded as the power usage code.

[0056] In a specific example, the basic primary key code of a certain data is "0x0A3FA7". Based on this, data with the same primary key code ("0x0A3FA7") can be searched from the energy consumption database and model unit, and links can be established to form a data view primary key chain. For each data view primary key chain, fuzzy matching is performed in the basic database using semantic codes, combined with fuzzy association matching rules. If the fuzzy matching result is a strong association, the basic process information corresponding to the strong association is retained. If no strong association exists, process development is required. Based on the current spatiotemporal data and basic process information, data matching is performed in the basic database, energy consumption database, and model unit to extract the equipment spatiotemporal parameters under the corresponding process flow of the current spatiotemporal data and basic process information, thus obtaining the data view spatiotemporal chain. For example, the current spatiotemporal data includes the time data 2024-03-15 14:32:18.456 and the spatial coordinates (12.5, 8.3, 0) ∈ BAY-03 region. With the basic process parameter set to "Gate_Etch_Step3", the current state of the equipment is "Processing" and includes energy consumption data. The current spatiotemporal data, basic process parameters, current state, and energy consumption data are linked to form a spatiotemporal chain of the data view. By using functional association matching rules to determine the compatibility of data within the spatiotemporal chain of the data view, a compatibility result under the current spatiotemporal data can be obtained. By integrating the data view primary key chain, basic process information, the data view spatiotemporal chain, and the compatibility result, a holographic data view of the equipment object is obtained.

[0057] The intelligent coding system based on semiconductor digital twins is a closed-loop technology of "perception-decision-evaluation-execution" and belongs to the "cognitive coding architecture". Its core is to upgrade the traditional "static data structure" into a "dynamic intelligent agent".

[0058] In a specific example, the concatenation code is a 14-bit complete device code. Each segment of the code corresponds to a portion of device information. For instance, the first six bits are the basic code. Based on the basic code, the device identity information page can be obtained, including the Module, Manufacturer, Model, and other device information such as the Module code, Manufacturer code, and Model code. Figure 7 As shown, a 14-digit complete equipment code uniquely points to a specific module in the equipment library, specifying the manufacturer, model, process label, process segment, processing method, wafer load, and robot configuration. For example... Figure 8 As shown, in the model unit, each 3D model has a device code. The first 6 digits of the code contain the Module, manufacturer, and model number, which coincide with the first 6 digits of the 14-digit complete device code. This code can uniquely point to a static 3D model, making it convenient for the dynamic simulation model to identify the static 3D model based on the complete device code and call it in the dynamic simulation model.

[0059] The 14-bit complete device code, as a string, is input into the 3D model determined from the model elements. (See reference...) Figure 9 First, the encoded information is decrypted. The first 6 bits of the code identify the unique 3D model within the unit model. In the PlantSimulation software, this 3D model is called and generated at the origin coordinates (0,0,0). Bits 7-14 of the code determine the core operating parameters and logic of the equipment. Based on the number of bits in the equipment code, the corresponding logic module is called. For example, bits 7-10 decrypt the process and operation segment information to assist in calculating the equipment's processing efficiency (WPH) value; the 11th bit decrypts the processing method information, determining the wafer transfer logic (serial / parallel); the 12th bit decrypts the chamber wafer load information, determining the selection of processing logic (single-chamber module / batch chamber module); and bits 13-14 decrypt the robot type information, determining the selection of robot pick-and-place logic (single-arm single-fork / single-arm multi-fork / double-arm single-fork / double-arm double-fork). When the 14-bit device code is input as a string, the pre-written Python program extracts the 13th-14th digits, number 01. Number 01 corresponds to the single-arm single-fork manipulator, so the single-arm single-fork manipulator module is executed and automatically placed in the 3D model.

[0060] Using a 14-bit device code, a unique device is retrieved from the basic database, energy consumption database, and model units. Key operating parameters and energy consumption curves of the device are then input into the 3D model. For example, the number of loadports and chambers is queried, and mature loadport, foup, and chamber modules are invoked to automatically generate fixed-pitch loadport and chamber chambers. Energy consumption curves output from the energy consumption database are input into the 3D model and processed according to the device's processing and operating cycle. Based on the serial / parallel logic identified and invoked according to the device code, a real-time or simulation mode mapping is performed to execute wafer transfer and processing tasks, resulting in a digital twin for the semiconductor equipment.

[0061] Reference Figure 10 This invention provides a digital twin construction apparatus for semiconductor devices, comprising: The device determination module 201 is used to determine the target device and provide the data acquisition unit corresponding to each target device; Data acquisition module 202 is used to acquire basic data and energy consumption data output by the data acquisition unit; The data encoding module 203 is used to encode basic data and energy consumption data based on the encoding strategy of parameterized object identifiers, and form corresponding basic databases and energy consumption databases. Model building module 204 is used to build the relationship between basic data, energy consumption data and target equipment status to form model units; The twin construction unit 205 is used to construct a digital twin for semiconductor devices by linking the basic database, energy database and model unit based on object attribute matching rules.

[0062] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0063] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for constructing a digital twin for a semiconductor device, characterized in that, Specifically, the steps include the following: Identify the target devices and provide the corresponding data acquisition units for each target device; Acquire basic data and energy consumption data output by the data acquisition unit; Based on the parameterized object identification coding strategy, the basic data and energy consumption data are encoded separately to form the corresponding basic database and energy consumption database. Establish the relationship between basic data, energy consumption data, and target equipment status to form model units; Based on object attribute matching rules, a digital twin for semiconductor devices is constructed by linking the basic database, energy consumption database, and model units.

2. The method for constructing a digital twin for a semiconductor device as described in claim 1, characterized in that, Basic data includes equipment identification information, equipment production parameters, and equipment environmental parameters.

3. The method for constructing a digital twin for a semiconductor device as described in claim 2, characterized in that, The encoding strategy based on parameterized object identifiers encodes both basic data and energy consumption data separately, specifically including the following steps: Extract the device identity information of the target device, encrypt it, perform device model mapping processing, and generate a device model index code; The equipment production parameters and equipment environmental parameters are encrypted separately to generate logical control codes; Based on the basic data, a low-dimensional semantic vector is generated. Combined with a pre-built hierarchical and cascaded coding architecture, bitwise operations are performed on the device model index code, logic control code, and low-dimensional semantic vector to form the basic code. Based on energy consumption data, construct a multi-dimensional energy consumption feature vector; Multi-dimensional energy consumption feature vectors are concatenated with basic coding to form a dynamic energy consumption coding with bimodal mapping.

4. The method for constructing a digital twin for a semiconductor device as described in claim 3, characterized in that, Based on the basic data, a low-dimensional semantic vector is generated. This vector is then combined with a pre-built hierarchical and cascaded coding architecture to perform bitwise operations on the device model index code, logic control code, and low-dimensional semantic vector, forming the basic coding. Specifically, this includes: Obtain semantic vectors of basic data and the relationships between various basic data, and map them to a low-dimensional continuous vector space to form low-dimensional semantic vectors; The low-dimensional semantic vector is discretized and then encrypted to generate a semantic code. Based on a pre-built hierarchical and cascaded coding architecture, the device model index code, logic control code, and semantic code are shifted respectively. The shifted device model index code, logic control code, and semantic code are superimposed to form the basic code.

5. The method for constructing a digital twin for a semiconductor device as described in claim 3, characterized in that, The multi-dimensional energy consumption feature vector is concatenated with the basic encoding to form a dynamic energy consumption encoding with bimodal mapping, specifically including: The multi-dimensional energy consumption feature vector is discretized and then encrypted to provide a timestamp, energy consumption code, and energy quality identifier. Based on a pre-built hierarchical and concatenated coding architecture, the basic code, timestamp, energy consumption code, and energy quality identifier are concatenated to form a concatenated code; In response to the triggering of the mode conversion gateway, and in conjunction with the concatenation code, the mapping relationship between the concatenation code and the simulation mode and the real-time mode is determined respectively; Based on the analysis of conflict perception, combined with the learning agent and reward function, the optimal bit field is given; Based on the optimal bit field, update the hierarchical concatenation coding architecture, update the concatenation code, and adjust the mapping relationship between the concatenation code and the simulation mode and real-time mode.

6. The method for constructing a digital twin for a semiconductor device as described in claim 5, characterized in that, Based on the analysis of conflict perception, combined with the learned agent and reward function, the optimal bit field is given, specifically including: In response to the analysis of multi-dimensional conflict perception, the utilization rate of bit domain resources is calculated; Construct the encoding state vector and determine the encoding state level; Based on the coding state level, a hierarchical hybrid response model is adopted. By combining the encoded state vector, the action selection is determined; By integrating a multi-objective hierarchical reward function with confidence verification, the optimal bit field is given; The hierarchical hybrid model specifically includes: The primary response sub-model is used to detect changes in the encoded state level and generate trigger instructions. Based on the deep optimization sub-model and combined with the triggering instruction, the changes in the encoding state level are analyzed, and the bit field adjustment range is given. A global planning sub-model is used to find the optimal value within the bit-domain adjustment range to obtain the current bit-domain. The current bit field is validated by combining the validation sub-model, and the optimal bit field is obtained based on the validation results. Multi-objective hierarchical reward functions, specifically including: Based on the response time of each sub-model in the hierarchical hybrid model, and combined with the time decay effect, a time activation term is formed; By integrating the normalized reward values ​​and reward weights of each sub-model in the hierarchical hybrid model, the intra-layer reward sub-items are obtained. Based on the reward triggering thresholds between each sub-model, and combined with the propagation effect of rewards in each sub-model of the hierarchical hybrid model, a global reward sub-item is given; Based on the time activation term, the reward sub-items within the fusion layer and the global reward sub-items, a multi-objective hierarchical reward function is given.

7. The method for constructing a digital twin for a semiconductor device as described in claim 1, characterized in that, Based on object attribute matching rules, the basic database, energy consumption database, and model units are linked, specifically including the following steps: Establish multi-dimensional attribute matching rules; Execute multi-dimensional attribute matching rules according to priority to construct a holographic data view of the device object; Construct a holographic data association map to link the basic database, energy consumption database, and model units.

8. The method for constructing a digital twin for a semiconductor device as described in claim 7, characterized in that, Execute multi-dimensional attribute matching rules according to priority to construct a holographic data view of the device object. The specific steps include the following: Based on priority, attribute matching rules for each dimension are extracted sequentially to obtain a matching rule sequence, which includes primary key association matching rules, fuzzy association matching rules, scenario association matching rules, and function association matching rules. Based on the primary key association matching rules, the basic primary key codes in the basic database, the energy consumption primary key codes in the energy consumption database, and the model primary key codes in the model units are encoded and linked to form a data view primary key chain; Based on the primary key chain of the data view and combined with fuzzy association matching rules, semantic matching is performed on the semantic codes in the basic database to provide basic process information; By using scenario association matching rules, the current spatiotemporal data and basic process information of equipment objects are integrated, and matching is performed in the basic database, energy consumption database and model unit to establish a spatiotemporal chain of data view; Based on the spatiotemporal chain of the data view and the functional association matching rules, the compatibility of the data in the spatiotemporal chain of the data view is judged, and a holographic data view of the device object is constructed.

9. The method for constructing a digital twin for a semiconductor device as described in claim 8, characterized in that, A holographic data view of a device object is specifically represented as follows: ； Wherein, Φ(o i ) for device object o i Holographic data view For device object o i The optimal set of basic data in the basic database. For device object o i The optimal set of energy consumption data in the energy consumption database. Device object o i In the optimal model set within the model unit, b′ represents the basic data in the basic database, e′ represents the energy consumption data in the energy consumption database, m′ represents the equipment model within the model unit, and S... total () is the total matching degree function, and argmax() is the maximum value point function.

10. A digital twin construction apparatus for semiconductor devices, characterized in that, The method for constructing a digital twin for a semiconductor device as described in any one of claims 1-9 specifically includes: The device identification module is used to identify target devices and provide the corresponding data acquisition units for each target device. The data acquisition module is used to acquire basic data and energy consumption data output by the data acquisition unit; The data encoding module is used to encode basic data and energy consumption data based on the encoding strategy of parameterized object identifiers, and form corresponding basic databases and energy consumption databases. The model building module is used to establish the relationship between basic data, energy consumption data, and the status of target equipment to form model units; The twin building unit is used to construct a digital twin for semiconductor devices by linking the basic database, energy database and model unit based on object attribute matching rules.

Images