A method and system for intelligent generation of MEMS processes
By using a MEMS process intelligent generation method, which combines local process constraints and prior knowledge, a MEMS process route that meets manufacturing conditions is automatically generated. This solves the problems of long R&D cycles and low design success rates in existing technologies, and achieves efficient process planning and verification.
Patent Information
- Application Number
- CN202611279023.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-21
- Publication Date
- 2026-11-10
AI Technical Summary
The existing MEMS process design and verification relies on human experience, resulting in long R&D cycles, difficulties in process transfer, low design success rates, and difficulty in meeting process constraints in heterogeneous integration scenarios of IC and MEMS, making it difficult to meet actual production needs.
A MEMS process intelligent generation method is adopted, which generates a process route that meets local device constraints through target device physical topology analysis, local process constraint matrix construction, process structure causal model construction, prior manufacturing knowledge structuring, multi-objective heuristic fitness function evaluation, and DRC verification and diagnosis.
Shorten the process planning cycle, improve R&D efficiency in the early stages of design, reduce trial and error costs, increase the success rate of the first tape-out, and ensure that the generated solution meets actual manufacturing conditions.
Smart Images

Figure CN123069890A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of microelectromechanical systems (MEMS), specifically relating to a method and system for intelligent generation of MEMS processes. Background Technology
[0002] Micro-Electro-Mechanical Systems (MEMS) are cutting-edge technologies integrating microelectronics, mechanics, materials science, and other disciplines. Essentially, they involve realizing desired functions such as sensing, actuation, or radio frequency on a chip as movable three-dimensional mechanical structures and electrical systems using similar micro- and nano-manufacturing processes. Semiconductor processes are the core means of manufacturing MEMS devices. Their design is a highly complex process, requiring not only architectural and electrical design but, more importantly, the translation of functionality into a feasible, reliable, and production-line-compatible semiconductor manufacturing process. As MEMS devices evolve towards higher performance, smaller size, and heterogeneous integration with integrated circuits (ICs), the complexity of their process combinations and design freedom increase dramatically, creating an urgent need for efficient, intelligent design automation tools that are deeply compatible with manufacturing processes.
[0003] In the current semiconductor process design flow, process design and verification are the core R&D bottlenecks. Existing process design and verification heavily rely on the designer's personal experience. Researchers primarily depend on engineers combining their experience with Electronic Design Automation (EDA) tools for semi-automated process construction and verification, requiring significant manual intervention in parameter iteration and trial and error. This approach suffers from long development cycles, difficulties in process migration, and low design success rates. More seriously, the verification of design solutions and specific production line process constraints (such as equipment limits, material compatibility, and thermal budgets) is often delayed, causing manufacturability issues to surface only in later development stages or even during tape-out, resulting in high rework costs and lengthy development cycles. Especially in IC and MEMS heterogeneous integration scenarios, complex constraints exist between different material systems, such as thermal budget limitations, cross-contamination, stress mismatch, and process compatibility, making it difficult for traditional methods to quickly generate manufacturable process solutions that meet local experimental conditions.
[0004] In recent years, large language models have made rapid progress in the field of text understanding and generation. However, existing methods mostly remain at the level of general text generation. Due to a lack of deep understanding of the physical mechanisms of micro- and nano-manufacturing and specific process constraints, the generated solutions often fail to meet the high-precision requirements of actual production. For example, the solution disclosed in CN202511417050, although it achieves process flow generation based on large models, is mainly aimed at general or macro-manufacturing processes. It has not established a closed-loop interaction with MEMS-specific process rule bases and DRC feedback mechanisms, making it difficult to directly use in the actual process design of MEMS.
[0005] Therefore, there is an urgent need to develop a technical solution that can automatically generate MEMS process routes and perform intelligent optimization by combining local equipment constraints, process rules and process knowledge base. Summary of the Invention
[0006] The purpose of this invention is to provide a method and system for intelligent generation of MEMS processes, which can solve the defects of existing methods such as the disconnect between the design and manufacturing stages and the inability of the generated solutions to meet actual needs.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] A method for intelligent generation of MEMS processes includes the following steps:
[0009] S1. Target Device Physical Topology Analysis and Data Initialization: Receive the structural information of the target MEMS device, analyze and extract the hierarchical topology relationship and geometric dimension parameters of the target device, and construct the target physical structure parameter set;
[0010] S2. Construction of local process constraint matrix: Collect equipment hardware information and physical environment configuration parameters of local manufacturing production line, and construct a local process constraint matrix that represents the absolute physical limit of the production line;
[0011] S3. Construction of a causal model of process structure based on physical constraint pruning: Extract historical standard process flow monitoring logs and corresponding device physical morphology detection tags, construct a causal model of process structure, and generate a localized DRC rule base;
[0012] S4. Prior Manufacturing Knowledge Structuring and Initial Screening: Using a natural language processing model, unstructured text in the MEMS process domain is semantically parsed to extract process step definitions, material systems, process parameters, and equipment dependencies. Based on the extraction results, a process entity-relationship graph is constructed to form a prior process knowledge base. According to user input, a similarity search is performed in the prior process knowledge base to recall candidate process module sequences with high structural matching degree with the input.
[0013] S5. Process route optimization based on multi-objective heuristic fitness function: The local process constraint matrix is used as the absolute boundary of the search space. The path search algorithm is used to reorganize the recalled candidate process modules. All non-boundary candidate process routes are evaluated and ranked for recommendation.
[0014] S6. Customized Process Scheme Editing: Select a basic process route from the multiple process routes recommended in S5, and then customize and improve it according to the specific performance requirements of the target device;
[0015] S7. DRC Verification and Diagnosis Based on Causal Graph and Counterfactual Inference: Input the personalized process scheme into the structural causal model constructed in S3 for inspection and verification, and output the inspection and verification results.
[0016] Furthermore, the dimensions of the local process constraint matrix in S2 include equipment type, process temperature range, material license list, cross-contamination rules, process window parameters, and maximum thermal budget limit.
[0017] Furthermore, the construction of the structural causal model in S3 specifically includes the following sub-steps:
[0018] S3.1 Initialize the parameter causal network to form a fully connected network;
[0019] S3.2 transforms the local process constraint matrix constructed in S2 into a physical constraint matrix of a directed acyclic graph, and performs structural pruning based on the physical constraint matrix to remove process parameter-related edges that exceed the physical thermodynamic and geometric limits.
[0020] S3.3 For the causal network after pruning, conditional independence tests are performed using historical monitoring logs, the conditional mutual information between the retained nodes is calculated, and the directional constraints of the causal edges are executed according to the temporal evolution characteristics of the semiconductor process to generate the final structural causal graph characterizing the causal relationship between process parameters, physical intermediate states and final morphological failure.
[0021] S3.4 Based on the final structural causal graph, calculate the structural causal effect value of parameter fluctuations at each process node on topology failure, and combine it with the high-yield production dataset to calculate and generate a physically safe parameter boundary manifold in the multidimensional parameter space, which serves as an early warning benchmark for process design rule checking DRC.
[0022] Furthermore, the localized DRC rule base in S3 includes local device limit parameters and physical / chemical conflict rules.
[0023] Furthermore, in step S5, candidate process routes are specifically evaluated based on a multi-objective heuristic fitness function. The calculation rules for the multi-objective heuristic fitness function include:
[0024] (1) Topology mapping fitness: Constructing target topology subgraphs based on the core physical structure units of the target device. Extracting the feature structures generated from candidate process routes to construct candidate topology subgraphs The fitness of the topology mapping is calculated using a subgraph isomorphism algorithm. :
[0025]
[0026] in, Let be the cardinality of the feature node set of the topological subgraph. The penalty coefficient is... The number of missing core physical structure units;
[0027] (2) Spatial drift of physical parameters: Extract the center vector of ideal parameters corresponding to the performance index of the target device from the local process constraint matrix described in S2. And obtain the current set parameter vector of the candidate process route. The drift cost of the parameter vector within the constrained space is calculated using Mahalanobis distance. :
[0028]
[0029] in, The inverse covariance matrix characterizing the tolerance distribution of various process parameters;
[0030] (3) Manufacturing complexity fitness: Calculate the global complexity cost of a single process route. The formula is:
[0031]
[0032] in, For process depth, The quantity of the i-th special process / material used in the scheme. For the number of type j stringent process conditions in the scheme, the global cost is... Perform range-based normalization mapping to generate a fitness score for manufacturing complexity. ;
[0033] Based on the above factors, the multi-objective heuristic fitness function is calculated as follows:
[0034]
[0035] in , , Let each factor have its corresponding weight, and satisfy the following condition: .
[0036] Furthermore, the specific mechanism for checking and verifying in S7 includes:
[0037] (1) Causal path anomaly detection: Map the node sequence of the process scheme to the structural causal model, and detect whether there is a directed path leading to the physical structure failure node by graph traversal. If there is, and the parameter setting value on the path breaks through the parameter boundary manifold of the physical safety, it is determined to be a physical rule conflict and an blocking error signal is output.
[0038] (2) Flexible warning signal: Check whether the set process parameters deviate from the safe confidence interval generated by the high yield sample clustering;
[0039] (3) Generation of diagnostic suggestions based on counterfactual deduction: When the blocking error signal is triggered, the abnormal process node that caused the conflict is extracted and counterfactual backtracking is performed on the structural causal model; by virtually intervening the parameter variables of the abnormal node in the algorithm model, the causal distribution expectation of the target physical morphology under different hypothetical states, i.e. counterfactual states, is calculated; the optimal intervention amount that makes the counterfactual distribution expectation meet the safety threshold is extracted, and it is transformed into a diagnostic compensation vector. The vector is output along with the error signal, and a differentiated locking and release mechanism is performed on the current process scheme flow status.
[0040] Furthermore, the blocking error signal includes the following dimensions:
[0041] Process conflict rules define inherent logical or partial order contradictions between process steps;
[0042] Production line capacity constraints: Check if the process parameter settings exceed the physical limits of the equipment in the device manufacturing production line.
[0043] Device manufacturing rationality rules are used to detect physical vulnerabilities in the process through topological mapping.
[0044] Furthermore, the differentiated locking and release mechanism is as follows:
[0045] (1) For blocking error signals, a forced state lock is performed on the current process route. The parameter update instruction for the abnormal node must be received, and the automatic retest is triggered after the node data is updated at the bottom layer. The retest logic is executed in a loop until all blocking error signals are cleared before the forced state lock can be released.
[0046] (2) For the flexible warning signal, the current process route is put into a suspended waiting state. If a parameter update instruction is received, the corresponding parameters are adjusted. If a design intent retention instruction is received, a status mask operation is performed on the flexible warning signal to forcibly release the suspended state and authorize the release.
[0047] A MEMS process intelligent generation system is provided to implement any of the above methods, comprising: a bottom data layer, a core engine layer, and a user interaction and control layer; the bottom data layer includes target device data, local production line data, process knowledge data, and process log data; the core engine layer is used to perform topology parsing and data initialization on the target device data, establish a local process constraint matrix for the local production line data, establish a priori process knowledge base through process knowledge data, establish a process structure causal model based on the causal benchmark provided by the process log data, and generate a localized DRC rule base; the core engine layer includes a process route optimization engine and a DRC verification and diagnosis engine; the user interaction and control layer is used to edit and improve the basic route obtained by the process route optimization engine and verify and output the final MEMS process solution through the DRC verification and diagnosis engine.
[0048] Compared with the prior art, the present invention has the following beneficial effects:
[0049] 1. This invention proposes a constraint-driven path search algorithm. The core mechanism of this algorithm lies in using the local process constraint matrix as a hard boundary condition for the search space, eliminating non-compliant nodes in real time. This ensures that every candidate process route generated theoretically satisfies the manufacturability requirement of the target production line. Compared to the traditional fully manual approach, this significantly shortens the process planning cycle and substantially improves the initial R&D efficiency of process design.
[0050] 2. This invention constructs an objective, quantitative, multi-dimensional weighted evaluation model. For multiple generated feasible process paths, candidate solutions are quantitatively scored from three core dimensions: topology mapping fitness, physical parameter spatial drift, and manufacturing complexity fitness. Specifically, step S5 determines the topology mapping fitness based on the mapping between the candidate process path and the core physical structural units of the target device; step S5 calculates the physical parameter spatial drift based on the degree of conformity between the candidate process route and local process constraints; step S5 evaluates the manufacturing complexity fitness based on the manufacturing complexity of the candidate process path, and performs a weighted summation according to preset weights to generate a comprehensive matching index. This model transforms the selection of the best solution from subjective judgment relying on experience to data-driven decision-making based on unified rules, automatically recommending the process route that achieves the optimal balance between technical feasibility and manufacturing cost.
[0051] 3. This invention constructs a dynamically configurable DRC rule base that is deeply integrated with actual production line capabilities. This rule base embeds multi-level verification rules ranging from process logic and equipment limits to parameter rationality. During the interactive editing of process plans, real-time calls to this rule base for synchronous compliance checks can accurately intercept logical conflicts or parameter out-of-bounds errors during the design phase. This effectively avoids the risk of discovering defects only after tape-out, significantly reduces trial-and-error costs, and improves the first-time tape-out success rate.
[0052] 4. This invention provides a robust physical boundary for process design by constructing a localized DRC rule base that includes local equipment limit parameters (such as maximum process temperature and minimum lithography linewidth) and physical / chemical conflict rules (such as etching selectivity and material cross-contamination). It achieves closed-loop verification and matching of process routes based on physical equipment constraints. Overcoming the drawback of easily exceeding physical limits when combining processes based on manual experience, it utilizes a multi-level DRC verification mechanism to perform real-time physical attribute and logical sequence checks on candidate process steps, forcibly filtering out illegal paths and ensuring that the generated process plan fully meets the actual manufacturing physical conditions of the local production line. It also constructs a structured process knowledge base containing specific manufacturing limit parameters. This transforms scattered, unstructured MEMS process experience into an entity-relationship model containing specific process window parameters (such as temperature, pressure, and time range), forming a highly reliable engineering database. Attached Figure Description
[0053] Figure 1 This is a flowchart of the method of the present invention.
[0054] Figure 2 This is an architecture diagram of the system of the present invention. Detailed Implementation
[0055] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0056] like Figure 1 As shown, this invention provides a method for intelligent generation of MEMS processes. By constructing limit definitions for the underlying physical devices, multi-objective heuristic optimization, and a closed-loop verification mechanism based on state evolution topology, an automated process data processing and generation flow is formed. The method specifically includes the following steps:
[0057] Step S1: Target device physical topology analysis and data initialization
[0058] This step aims to receive the structural model, functional description, or two-dimensional layout information of the target MEMS device, analyze and extract the hierarchical topology and geometric parameters of the target device, and construct a set of target physical structure parameters. In practice, the user first configures the substrate in the system's intelligent generation interface, inputting or selecting substrate parameters, including substrate type, substrate size, crystal orientation, resistivity, doping type, etc.; then, the user inputs the functional intent, entering functional keywords describing the target device, such as capacitive pressure sensor or resonant gyroscope.
[0059] Step S2: Construction of Local Process Constraint Matrix
[0060] This step aims to collect the equipment hardware white paper and physical environment configuration parameters of the local manufacturing production line to construct a local process constraint matrix that characterizes the absolute physical limits of the production line. The feature dimensions of the local process constraint matrix include: equipment type, process temperature range, material license list, cross-contamination rules, process window parameters, and maximum thermal budget limit.
[0061] Step S3: Construction of a causal model of the process structure based on physical constraint pruning
[0062] This step aims to extract historical standard process flow monitoring logs and corresponding device physical morphology detection tags to construct a causal model of the process structure. The construction process includes:
[0063] First, initialize the parameter causal network to form a fully connected network;
[0064] Next, the local process constraint matrix constructed in step S2 is transformed into a physical constraint matrix of a directed acyclic graph, and structural pruning is performed based on the physical constraint matrix to remove process parameter related edges that exceed the physical thermodynamic limit and geometric limit.
[0065] Then, for the causal network after pruning, conditional independence tests are performed using historical monitoring logs, the conditional mutual information between the retained nodes is calculated, and the directional constraints of the causal edges are executed according to the temporal evolution characteristics of the semiconductor process, generating the final structural causal map that characterizes the causal relationship between process parameters, physical intermediate states and final morphological failure.
[0066] Finally, based on the final structural causal graph, the structural causal effect value of parameter fluctuations at each process node on topology failure is calculated. Combined with the high-yield production dataset, a physically safe parameter boundary manifold is calculated and generated in the multidimensional parameter space as an early warning benchmark for DRC.
[0067] Step S4: Structuring and Initial Screening of Prior Manufacturing Knowledge
[0068] First, the system uses a natural language processing model to semantically parse unstructured text in MEMS process academic literature databases and professional technical websites, extracting process step definitions, material systems, process parameters, and equipment dependencies, and converting them into process entity-relationship graphs, which are then stored in a priori process knowledge base. Second, based on the parameters input by the user in step S1, a similarity search is performed in the priori process knowledge base to recall candidate process module sequences with high structural matching.
[0069] Step S5: Optimization of process route based on multi-objective heuristic fitness function
[0070] This invention uses the local process constraint matrix described in step S2 as the absolute boundary of the search space, and employs a path search algorithm to reorganize the candidate process modules recalled in step S4. For all non-boundary candidate process routes, an evaluation is performed based on a multi-objective heuristic fitness function, and they are recommended in descending order of matching degree. The calculation rules for the multi-objective heuristic fitness function include:
[0071] 1. Topology mapping fitness: Constructing a target topology subgraph based on the core physical structural units of the target device. Extracting the feature structures generated from candidate process routes to construct candidate topology subgraphs The fitness of the topology mapping is calculated using a subgraph isomorphism algorithm. :
[0072]
[0073] in, Let be the cardinality of the feature node set of the topological subgraph. The penalty coefficient is... This represents the number of missing core physical structural units.
[0074] 2. Physical parameter spatial drift: Extract the ideal parameter center vector corresponding to the target device performance index from the local process constraint matrix described in step S2. And obtain the current set parameter vector of the candidate process route. The drift cost of the parameter vector within the constrained space is calculated using Mahalanobis distance. :
[0075]
[0076] in, The inverse covariance matrix is used to characterize the tolerance distribution of various process parameters.
[0077] 3. Manufacturing Complexity Fit: Calculate the global complexity cost of a single process route. The formula is:
[0078]
[0079] in, For process depth, The quantity of the i-th special process / material used in the scheme. This refers to the number of the j-th type of stringent process conditions (such as high temperature, high pressure, and long duration) in the case. This relates to the global cost. Perform range-based normalization mapping to generate a fitness score for manufacturing complexity. .
[0080] Based on the above factors, the multi-objective heuristic fitness function is calculated as follows:
[0081]
[0082] in , , Let each factor have its corresponding weight, and satisfy the following condition: .
[0083] Step S6: Customized Editing of Process Scheme
[0084] The user selects a basic process route from the multiple process routes generated in step S5 and proceeds to the process flow editing and refinement stage. In this stage, the system first retrieves the corresponding standard parameters from the built-in process parameter library to form a complete initial process plan. The user can then manually add or delete steps, reorder or adjust specific parameters in the design area to personalize the process according to the specific performance requirements of their device.
[0085] Step S7: DRC Validation and Diagnosis Based on Causal Graph and Counterfactual Inference
[0086] All customized process schemes must undergo DRC verification until no errors are detected. The DRC verification described in this invention refers to inputting the process scheme into the causal model of the process structure described in step S3 for checking and verification, and outputting blocking error / flexible warning signals and correction suggestion vectors. The specific mechanism includes:
[0087] 1. Causal Path Anomaly Detection: The node sequence of the process scheme is mapped to the structural causal graph. Graph traversal is used to detect whether there are directed paths leading to nodes where the physical structure fails. If such paths exist, and the parameter settings on these paths exceed the parameter boundary manifold of the physical safety, it is determined to be a physical rule conflict, and the system generates a blocking error signal.
[0088] 2. Generation of diagnostic suggestions based on counterfactual inference: When the blocking error signal is triggered, the system extracts the abnormal process node that caused the conflict and performs counterfactual backtracking inference on the structural causal model; by virtually intervening in the parameter variables of the abnormal node in the algorithm model, the causal distribution expectation of the target physical morphology under different hypothetical states (i.e., counterfactual states) is calculated; the optimal intervention amount that makes the counterfactual distribution expectation meet the safety threshold is extracted, and it is transformed into a diagnostic compensation vector (i.e., specific modification suggestion parameters), and the vector is output along with the error signal.
[0089] 3. Rule-based classification mapping and differentiated signal output: Based on the above spectrum detection and diagnosis results, the system presents the DRC verification results in the form of rule-based prompt signals, specifically divided into the following dimensions:
[0090] (1) Process conflict rules (blocking error signals). Define the inherent logical or partial order contradictions between process steps (such as detecting a necessary preceding node for the growth of the adhesion layer before a certain material deposition node in the causal graph).
[0091] (2) Production line capacity constraints (blocking error signals). Check if the process parameter settings exceed the physical limits of the device manufacturing production line (e.g., the set parameters are lower than the minimum linewidth of the photolithography or higher than the maximum process temperature limit of the equipment).
[0092] (3) Device manufacturing rationality rules (blocking error signals). Physical defects in the process are detected through topological mapping (such as missing key steps, unsupported three-dimensional structure evolution, or material lattice incompatibility).
[0093] (4) Parameter rationality rules (flexible warning signals). Check whether the set process parameters (such as temperature, time, thickness) deviate from the safe confidence interval generated by the high-yield sample clustering. Such deviations do not exceed the absolute physical limits and do not necessarily lead to manufacturing failure, but may affect device performance or yield. The system generates a warning signal to prompt verification.
[0094] 4. Hierarchical Process Flow and State Machine Control: The DRC verification terminal outputs the aforementioned blocking error and flexible warning signals, and implements a differentiated locking and release mechanism for the current process flow status.
[0095] For blocking error signals, a forced state lock is applied to the current process route. A parameter update command (i.e., an externally input parameter correction vector) for the abnormal node must be received, and an automated retest must be triggered after the underlying node data is updated. This lock cannot be bypassed; the retest logic is executed repeatedly until all hardware blocking error signals are completely cleared before the forced state lock can be released.
[0096] For flexible warning signals, the current process route is suspended and awaiting execution. An external command receiving port is opened; if a parameter update command is received, the corresponding parameters are adjusted; if a design intent retention command is received, a status mask operation is performed on the flexible warning signal to forcibly lift the suspension and authorize its passage.
[0097] Through repeated iterations of the above-mentioned hierarchical control logic, a final process solution is finally output that absolutely satisfies the underlying manufacturing physical constraints (zero errors) and is compatible with personalized R&D design intentions (warnings can be retained).
[0098] like Figure 2 As shown, this invention also provides a MEMS process intelligent generation system, comprising: a bottom data layer, a core engine layer, and a user interaction and control layer; the bottom data layer includes target device data, local production line data, process knowledge data, and process log data; the core engine layer is used to perform topology parsing and data initialization on the target device data, establish a local process constraint matrix for the local production line data, establish a priori process knowledge base through the process knowledge data, and use the process log data as training samples to construct a process structure causal model, and generate a localized DRC rule base based on the model; the core engine layer includes a process route optimization engine and a DRC verification and diagnosis engine; the user interaction and control layer is used to edit and improve the basic route obtained by the process route optimization engine and verify and output the final MEMS process solution through the DRC verification and diagnosis engine.
[0099] Example:
[0100] The underlying data storage module is used to store and manage target device data, local production line data, process knowledge data, and process log data.
[0101] The topology parsing and data initialization module is used to receive the target device data, parse and extract the hierarchical topology relationship and geometric dimension parameters of the target device, and construct a set of target physical structure parameters;
[0102] The local process constraint matrix module is used to collect equipment hardware information and physical environment configuration parameters from the local production line data, and construct a local process constraint matrix that characterizes the absolute physical limits of the production line.
[0103] The DRC rule base module is used to import the process log data, perform structural pruning and conditional independence tests under the constraints of the local process constraint matrix, construct a process structure causal model that characterizes the causal relationship between process parameters, physical intermediate states and final morphological failure, and generate a localized DRC rule base based on the model.
[0104] The prior process knowledge base module is used to perform semantic parsing on the process knowledge data and construct a process entity-relationship graph to form a prior process knowledge base.
[0105] The process route optimization engine is used to perform similarity retrieval and recall candidate process module sequences in the prior process knowledge base, and uses the local process constraint matrix as the absolute boundary to perform path reorganization and evaluation on the candidate process module sequences using a multi-objective heuristic fitness function to generate a recommended process route set.
[0106] A process scheme editor is used to provide a visual interactive interface to display the recommended process route set, and to receive and respond to selection and fine-tuning instructions for the recommended process route set, and generate a personalized edited process scheme sequence.
[0107] The DRC verification and diagnostic engine is used to input the personalized process scheme sequence into the process structure causal model for inspection and verification, determine whether there is a directed path leading to the physical structure failure node, and when it is determined that there is a directed path and the parameter setting value on the path exceeds the parameter boundary manifold of physical safety, perform counterfactual backtracking inference to generate a diagnostic compensation vector.
[0108] The inspection and verification control unit is used to receive the verification results of the DRC verification and diagnostic engine and the diagnostic compensation vector, and to trigger corresponding differentiated state control based on the verification results, such as error, warning, or pass. Specifically: when an error state is triggered, the current process route is forcibly locked, and the diagnostic compensation vector is output to the process scheme editor to trigger automated re-verification in a loop; when a warning state is triggered, a suspension waiting state is executed, and a state mask operation authorization is performed to release the process when a design intent retention instruction is received; when a pass state is triggered, the final MEMS process scheme is output.
[0109] The MEMS process intelligent generation system provided in this embodiment adopts a layered and decoupled software architecture, which is divided into a bottom-up data layer, a core engine layer, and a user interaction and control layer.
[0110] During the initialization phase, the target device data in the bottom data layer is input to the topology parsing and data initialization module of the core engine layer, and after parsing, the target physical structure parameter set is output. The local production line data is input to the local process constraint matrix module, which is transformed into a local process constraint matrix to establish the absolute boundary of the search space. At the same time, process knowledge data is input to the prior process knowledge base module to construct a prior knowledge graph. Process log data is used as training samples to construct a process structure causal model, and a localized DRC rule base is generated based on the model.
[0111] During the scheme generation phase, the process route optimization engine jointly retrieves the output data from the aforementioned topology parsing and data initialization module, local process constraint matrix module, and prior process knowledge base module. By performing multi-objective heuristic path search, it outputs multiple candidate basic process routes to the process scheme editor of the user interaction and control layer.
[0112] During the verification and release phase, the process scheme editor sends the user-tuned process scheme sequence to the Design Rule Check (DRC) verification and diagnostic engine. This engine performs full-path anomaly detection based on the causal model of the process structure. The verification results ultimately converge in the inspection and verification control unit, which performs branch control for three states: error, warning, and pass. In the pass state, the data stream is directly sent and output as the final MEMS process scheme; in the warning state, data is suspended and awaits manual masking instructions; in the error state, the system unconditionally intercepts the data and feeds back the diagnostic compensation vector generated by the counterfactual inference of the diagnostic engine to the process scheme editor, locking the scheme and forcibly triggering the automated re-verification loop.
Claims
1. A method for intelligent generation of MEMS processes, characterized in that, Includes the following steps: S1. Target Device Physical Topology Analysis and Data Initialization: Receive the structural information of the target MEMS device, analyze and extract the hierarchical topology relationship and geometric dimension parameters of the target device, and construct the target physical structure parameter set; S2. Construction of local process constraint matrix: Collect equipment hardware information and physical environment configuration parameters of local manufacturing production line, and construct a local process constraint matrix that represents the absolute physical limit of the production line; S3. Construction of a causal model of process structure based on physical constraint pruning: Extract historical standard process flow monitoring logs and corresponding device physical morphology detection tags, construct a causal model of process structure, and generate a localized DRC rule base; S4. Prior manufacturing knowledge structuring and initial screening: Using natural language processing models, semantic parsing is performed on unstructured text in the MEMS process field to extract process step definitions, material systems, process parameters and equipment dependencies, and a process entity-relationship graph is constructed based on the extraction results to form a prior process knowledge base; Based on user input, a similarity search is performed in the prior process knowledge base to recall candidate process module sequences that have a high structural matching degree with the input; S5. Process route optimization based on multi-objective heuristic fitness function: The local process constraint matrix is used as the absolute boundary of the search space. The path search algorithm is used to reorganize the recalled candidate process modules. All non-boundary candidate process routes are evaluated and ranked for recommendation. S6. Customized Process Scheme Editing: Select a basic process route from the multiple process routes recommended in S5, and then customize and improve it according to the specific performance requirements of the target device; S7. DRC Verification and Diagnosis Based on Causal Graph and Counterfactual Inference: Input the personalized process scheme into the structural causal model constructed in S3 for inspection and verification, and output the inspection and verification results.
2. The intelligent generation method for MEMS processes according to claim 1, characterized in that, The dimensions of the local process constraint matrix in S2 include equipment type, process temperature range, material license list, cross-contamination rules, process window parameters, and maximum thermal budget limit.
3. The intelligent generation method for MEMS processes according to claim 1, characterized in that, The construction of the structural causal model in S3 specifically includes the following sub-steps: S3.1 Initialize the parameter causal network to form a fully connected network; S3.2 transforms the local process constraint matrix constructed in S2 into a physical constraint matrix of a directed acyclic graph, and performs structural pruning based on the physical constraint matrix to remove process parameter-related edges that exceed the physical thermodynamic and geometric limits. S3.3 For the causal network after pruning, conditional independence tests are performed using historical monitoring logs, the conditional mutual information between the retained nodes is calculated, and the directional constraints of the causal edges are executed according to the temporal evolution characteristics of the semiconductor process to generate the final structural causal graph characterizing the causal relationship between process parameters, physical intermediate states and final morphological failure. S3.4 Based on the final structural causal graph, calculate the structural causal effect value of parameter fluctuations at each process node on topology failure, and combine it with the high-yield production dataset to calculate and generate a physically safe parameter boundary manifold in the multidimensional parameter space, which serves as an early warning benchmark for process design rule checking DRC.
4. The intelligent generation method for MEMS processes according to claim 3, characterized in that, The localized DRC rule base in S3 includes local device limit parameters and physical / chemical conflict rules.
5. The intelligent generation method for MEMS processes according to claim 1, characterized in that, In S5, candidate process routes are specifically evaluated based on a multi-objective heuristic fitness function. The calculation rules for the multi-objective heuristic fitness function include: (1) Topology mapping fitness: Constructing target topology subgraphs based on the core physical structure units of the target device. Extracting the feature structures generated from candidate process routes to construct candidate topology subgraphs The fitness of the topology mapping is calculated using a subgraph isomorphism algorithm. : in, Let be the cardinality of the feature node set of the topological subgraph. The penalty coefficient is... The number of missing core physical structure units; (2) Spatial drift of physical parameters: Extract the center vector of ideal parameters corresponding to the performance index of the target device from the local process constraint matrix described in S2. And obtain the current set parameter vector of the candidate process route. The drift cost of the parameter vector within the constrained space is calculated using Mahalanobis distance. : in, The inverse covariance matrix characterizing the tolerance distribution of various process parameters; (3) Manufacturing complexity fitness: Calculate the global complexity cost of a single process route. The formula is: in, For process depth, The quantity of the i-th special process / material used in the scheme. For the number of type j stringent process conditions in the scheme, the global cost is... Perform range-based normalization mapping to generate a fitness score for manufacturing complexity. ; Based on the above factors, the multi-objective heuristic fitness function is calculated as follows: in , , Let each factor have its corresponding weight, and satisfy the following condition: .
6. The intelligent generation method for MEMS processes according to claim 1, characterized in that, The specific mechanisms for checking and verifying in S7 include: (1) Causal path anomaly detection: Map the node sequence of the process scheme to the structural causal model, and detect whether there is a directed path leading to the physical structure failure node by graph traversal. If there is, and the parameter setting value on the path breaks through the parameter boundary manifold of the physical safety, it is determined to be a physical rule conflict and an blocking error signal is output. (2) Flexible warning signal: Check whether the set process parameters deviate from the safe confidence interval generated by the high yield sample clustering; (3) Generation of diagnostic suggestions based on counterfactual deduction: When the blocking error signal is triggered, the abnormal process node that caused the conflict is extracted and counterfactual backtracking is performed on the structural causal model; by virtually intervening the parameter variables of the abnormal node in the algorithm model, the causal distribution expectation of the target physical morphology under different hypothetical states, i.e. counterfactual states, is calculated; the optimal intervention amount that makes the counterfactual distribution expectation meet the safety threshold is extracted, and it is transformed into a diagnostic compensation vector. The vector is output along with the error signal, and a differentiated locking and release mechanism is performed on the current process scheme flow status.
7. The intelligent generation method for MEMS processes according to claim 6, characterized in that, The blocking error signal includes the following dimensions: Process conflict rules define inherent logical or partial order contradictions between process steps; Production line capacity constraints: Check if the process parameter settings exceed the physical limits of the equipment in the device manufacturing production line. Device manufacturing rationality rules are used to detect physical vulnerabilities in the process through topological mapping.
8. The intelligent generation method for MEMS processes according to claim 6, characterized in that, The differentiated locking and release mechanism is as follows: (1) For blocking error signals, a forced state lock is performed on the current process route. The parameter update instruction for the abnormal node must be received, and the automatic retest is triggered after the node data is updated at the bottom layer. The retest logic is executed in a loop until all blocking error signals are cleared before the forced state lock can be released. (2) For the flexible warning signal, the current process route is put into a suspended waiting state. If a parameter update instruction is received, the corresponding parameters are adjusted. If a design intent retention instruction is received, a status mask operation is performed on the flexible warning signal to forcibly release the suspended state and authorize the release.
9. A MEMS process intelligent generation system, used to implement the method of any one of claims 1 to 8, characterized in that, include: The underlying data layer, the core engine layer, and the user interaction and control layer; The underlying data layer includes target device data, local production line data, process knowledge data, and process log data; The core engine layer is used to perform topology parsing and data initialization on target device data, establish a local process constraint matrix for local production line data, establish a prior process knowledge base through process knowledge data, establish a process structure causal model based on the causal benchmark provided by process log data, and generate a local DRC rule base; the core engine layer includes a process route optimization engine and a DRC verification and diagnosis engine. The user interaction and control layer is used to edit and improve the basic route obtained by the process route optimization engine, and to verify and output the final MEMS process solution through the DRC verification and diagnosis engine.
Citation Information
Patent Citations
Intelligent machining process generation method based on model driving
CN121477798A
