Method, system, medium and product for optimizing cmp process based on multi-physics coupling

Abstract

The application provides a CMP process optimization method, system, medium and product based on multi-physical field coupling. The target chip layout is divided by a three-level grid system, and the multi-scale characteristic quantization data of the micro grid unit is extracted. Combined with the physical simulation operation of the CMP multi-physical field full coupling mathematical model, the complex evolution law of the three-dimensional topography of the wafer surface under the advanced process can be fully described, the accuracy and consistency of the simulation result are improved, and the problem of large prediction deviation of the traditional model assumption simplification is overcome. At the same time, the multi-physical effects such as contact mechanics, chemical reaction and thermodynamics are included in the unified modeling framework, the multi-physical field coupling simulation is realized, and the scientific and systematic theoretical support is provided for process optimization. The simulation result data is compensated and corrected by using the machine learning intelligent compensation network, and the final prediction result and the corresponding confidence interval are output, the prediction accuracy is improved, and the prediction reliability is quantified, which provides risk basis for process decision.

Description

Technical Field

[0001] This application relates to the field of semiconductor manufacturing technology, and in particular to a CMP process optimization method, system, medium, and product based on multi-physics coupling. Background Technology

[0002] In integrated circuit manufacturing, the planarity of the wafer surface directly determines the precision of subsequent processes such as photolithography and etching, thus affecting chip performance and yield. Chemical mechanical polishing (CMP), as the core process currently capable of achieving global planarization of the wafer surface, removes excess material from the wafer surface through the synergistic effect of mechanical grinding and chemical etching, enabling the wafer surface to achieve nanometer-level flatness requirements. It has become an indispensable key step in the manufacturing of very large-scale integrated circuits.

[0003] With the continuous development of semiconductor technology, the technology nodes for chip manufacturing are constantly shrinking, and we have now entered the advanced process stage of 7nm and below. This miniaturization of the process places more stringent demands on wafer surface planarization accuracy and defect control levels. Traditional CMP processes and control technologies mainly suffer from the following problems:

[0004] (1) Insufficient modeling accuracy: Traditional CMP process models are mostly based on empirical formulas or simplified physical assumptions, failing to fully consider the complexity of the three-dimensional morphology evolution of wafer surfaces in advanced processes. They cannot accurately predict the formation mechanism and evolution law of process defects such as edge erosion and dish defects. In processes below 7nm, even small morphological changes on the wafer surface can lead to subsequent process failures. The prediction bias of traditional models often results in insufficient polishing accuracy, affecting chip yield.

[0005] (2) Lack of multi-physics coupling: Existing CMP process analysis and control methods often consider mechanical grinding or chemical corrosion effects alone, lacking systematic modeling of the coupling effects of multiple physical fields such as contact mechanics, chemical corrosion, and thermodynamics. This makes it impossible to fully reveal the intrinsic relationship between process parameters and wafer planarization effect and defect generation, resulting in a lack of scientific theoretical support for process optimization.

[0006] (3) Poor real-time adaptability: Existing CMP processes mostly adopt optimization methods based on fixed process rules. However, in actual production, process conditions (such as polishing pad wear, slurry composition changes, wafer initial morphology differences, etc.) are always in dynamic change. Optimization methods based on fixed process rules lack the ability to adapt to dynamic changes in the process and cannot achieve true real-time closed-loop control.

[0007] (4) Insufficient data utilization: With the application of industrial internet and intelligent manufacturing technologies in the semiconductor manufacturing field, massive amounts of process parameter data and testing data are generated during the CMP process. Traditional CMP process control methods have failed to fully explore the value of these data, nor have they effectively combined the physical mechanism of CMP process to form complementary advantages. As a result, the constructed prediction models are often limited to specific process scenarios, have limited generalization ability, and are difficult to adapt to the polishing requirements of different processes and different types of wafers.

[0008] Therefore, there is an urgent need in this field for a CMP process control system that can integrate physical mechanisms and data-driven approaches to achieve accurate multi-scale prediction and adaptive optimization. Summary of the Invention

[0009] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a CMP process optimization method, system, medium and product based on multi-physics coupling, to solve the technical problems of lack of multi-physics coupling and poor real-time adaptability in existing CMP processes.

[0010] To achieve the above and other related objectives, a first aspect of this application provides a CMP process optimization method based on multi-physics coupling, comprising: performing a meshing operation on a target chip layout based on a pre-constructed three-level mesh system to obtain multiple micro-mesh cells, and extracting multi-scale feature quantization data of each micro-mesh cell; performing a physical simulation operation based on a fully coupled CMP multi-physics mathematical model and according to the multi-scale feature quantization data of each micro-mesh cell to generate simulation result data of each micro-mesh cell; performing a distribution prediction operation based on the simulation result data of each micro-mesh cell and a preset machine learning intelligent compensation network to generate final prediction result data of each micro-mesh cell and the confidence interval corresponding to each final prediction result data; and performing a parameter adjustment operation based on the final prediction result data of each micro-mesh cell and the confidence interval corresponding to each final prediction result data to generate a CMP target process parameter control sequence.

[0011] In some embodiments of the first aspect of this application, the method of generating simulation result data for each microgrid unit by performing physical simulation operations based on the CMP multiphysics fully coupled mathematical model and according to the multi-scale feature quantization data of each microgrid unit includes: calculating the mechanical removal rate component of each microgrid unit based on a preset contact mechanics model and the multi-scale feature quantization data of each microgrid unit; calculating the chemical removal rate component of each microgrid unit based on a preset chemical corrosion model and the multi-scale feature quantization data of each microgrid unit; calculating the thermal effect removal rate component of each microgrid unit based on a preset thermodynamic model and the multi-scale feature quantization data of each microgrid unit; calculating the total material removal rate of each microgrid unit based on the mechanical removal rate component, chemical removal rate component, and thermal effect removal rate component of each microgrid unit and the CMP multiphysics fully coupled mathematical model; and performing physical simulation operations based on the total material removal rate of each microgrid unit to generate simulation result data for each microgrid unit.

[0012] In some embodiments of the first aspect of this application, the CMP multiphysics fully coupled mathematical model includes:

[0013] ;

[0014] in, Indicates the total material removal rate; Indicates the contribution weight of mechanical grinding; Indicates the contribution weight of chemical corrosion; Indicates the weight of the contribution of the thermal effect; This represents the mechanical removal rate component; This represents the chemical removal rate component; This represents the heat removal rate component.

[0015] In some embodiments of the first aspect of this application, the preset machine learning intelligent compensation network includes a feature embedding layer, an attention mechanism layer, a multi-scale feature fusion layer, and a prediction quantization layer; wherein, the method of generating the final prediction result data of each micro-grid unit and the confidence interval corresponding to each final prediction result data based on the simulation result data of each micro-grid unit and the preset machine learning intelligent compensation network includes: the feature embedding layer is used to perform spatial alignment and integration operations based on the simulation result data of the micro-grid unit and the multi-scale feature quantization data of the micro-grid unit to generate simulation feature vector data of the micro-grid unit and the layout multi-scale feature vector data. The attention mechanism layer is used to perform evaluation and focusing operations based on the simulation feature vector data and multi-scale feature vector data of the microgrid unit to generate physical simulation key feature data and multi-scale process-sensitive feature data of the microgrid unit; the multi-scale feature fusion layer is used to perform feature fusion operations based on the physical simulation key feature data and multi-scale process-sensitive feature data of the microgrid unit to generate multi-scale fused feature data of the microgrid unit; the prediction quantization layer is used to perform distribution prediction operations based on the multi-scale fused feature data of the microgrid unit to generate the final prediction result data of the microgrid unit and the confidence interval corresponding to the final prediction result data.

[0016] In some embodiments of the first aspect of this application, the prediction quantization layer includes a temporal memory unit layer, an uncertainty estimation layer, and a Bayesian output layer; the distributed prediction operation includes a temporal correction operation, a prediction quantization operation, and a result correction operation; wherein, the method by which the prediction quantization layer performs distributed prediction operations based on the multi-scale fusion feature data of the microgrid unit to generate the final prediction result data of the microgrid unit and the confidence interval corresponding to the final prediction result data includes: the temporal memory unit layer performs a temporal correction operation based on the multi-scale fusion feature data of the microgrid unit and historical process state data to generate temporal dependency compensation feature data of the microgrid unit; the uncertainty estimation layer performs a prediction quantization operation based on the temporal dependency compensation feature data of the microgrid unit to generate prediction uncertainty feature data of the microgrid unit; and the Bayesian output layer performs a result correction operation based on the prediction uncertainty feature data of the microgrid unit to generate the final prediction result of the microgrid unit and determine the confidence interval corresponding to the final prediction result.

[0017] In some embodiments of the first aspect of this application, the method of generating a CMP target process parameter control sequence by performing parameter adjustment operations based on the final prediction result data of each microgrid unit and the confidence interval corresponding to each final prediction result data includes: constructing a predictive control model based on the final prediction result data of each microgrid unit and the confidence interval corresponding to each final prediction result data, and based on a preset rolling time-domain optimization strategy, thereby generating an initial process parameter control sequence; performing causal association analysis on the final prediction result data of each microgrid unit and the confidence interval corresponding to each final prediction result data based on a preset process knowledge graph, to identify multiple high-risk regions, and determine the process parameter optimization variables and process parameter optimization weights corresponding to each high-risk region; calibrating the initial optimal process parameter control sequence based on the process parameter optimization variables and process parameter optimization weights corresponding to each high-risk region, and performing optimization solving operations on the calibrated initial optimal process parameter control sequence based on a multi-objective optimization algorithm, thereby generating a CMP target process parameter control sequence.

[0018] In some embodiments of the first aspect of this application, the method further includes: controlling the CMP target process parameter sequence according to the CMP target process parameter sequence and acquiring in real time the actual operating data of the CMP equipment and the real measurement data of each microgrid cell to perform parameter update operations on the predictive control model.

[0019] To achieve the above and other related objectives, a second aspect of this application provides a CMP process optimization system based on multi-physics coupling, comprising: a multi-scale feature extraction module, used to perform meshing operations on a target chip layout based on a pre-constructed three-level mesh system to obtain multiple micro-mesh cells, and extract multi-scale feature quantization data of each micro-mesh cell; a multi-physics coupling simulation module, used to perform physical simulation operations based on a fully coupled CMP multi-physics mathematical model and according to the multi-scale feature quantization data of each micro-mesh cell, to generate simulation result data of each micro-mesh cell; an intelligent compensation module, used to perform distribution prediction operations based on the simulation result data of each micro-mesh cell and a preset machine learning intelligent compensation network, to generate final prediction result data of each micro-mesh cell and the confidence interval corresponding to each final prediction result data; and an optimization control module, used to perform parameter adjustment operations based on the final prediction result data of each micro-mesh cell and the confidence interval corresponding to each final prediction result data, to generate a CMP target process parameter control sequence.

[0020] To achieve the above and other related objectives, a third aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the CMP process optimization method based on multiphysics coupling as described above.

[0021] To achieve the above and other related objectives, a fourth aspect of this application provides a computer program product comprising computer program code that, when executed on a computer, causes the computer to implement the CMP process optimization method based on multiphysics coupling as described above.

[0022] As described above, this application has the following beneficial effects:

[0023] (1) Improve modeling accuracy and simulation reliability: The target chip layout is divided into grids through a three-level grid system to extract multi-scale feature quantization data of each micro-grid unit. Based on the CMP multi-physics field fully coupled mathematical model, physical simulation operation is performed, which can fully characterize the complex evolution law of the three-dimensional morphology of the wafer surface under advanced process, improve the accuracy and consistency of simulation results, and overcome the problems of simplified assumptions and large prediction deviations in traditional models.

[0024] (2) Achieve multi-physics coupling: Incorporate multi-physics effects such as contact mechanics, chemical reaction, and thermodynamics into a unified modeling framework, and reveal the intrinsic relationship between process parameters and flattening effect through fully coupled physical simulation, providing scientific and systematic theoretical support for process optimization.

[0025] (3) Improve prediction accuracy and quantify uncertainty: The simulation result data is compensated and corrected by a machine learning intelligent compensation network, the final prediction result is output, and the confidence interval is given at the same time. This can intuitively represent the credibility of the prediction, quantify the reliability of the prediction, and provide risk basis for subsequent process decisions.

[0026] (4) Possesses real-time adaptive and closed-loop optimization capabilities: Based on a preset rolling time-domain optimization strategy, a preset process knowledge graph, and a multi-objective optimization algorithm, the CMP target process parameter control sequence is solved online. It can dynamically match the current process state and risk level by combining the final prediction results and confidence intervals of the micro-grid units, thereby achieving real-time, accurate, and adaptive control of process parameters. The CMP target process parameter control sequence is sent to the CMP equipment to perform polishing operations. Real-time data of actual equipment operation and real measurement data of each micro-grid unit are collected. The prediction error is calculated and the error source is traced by combining the preset process knowledge graph to locate the cause of process deviation. Using the prediction error and causal analysis results, the parameters of the predictive control model are updated through incremental learning or parameter calibration, and the rule weights of the preset process knowledge graph are optimized simultaneously, realizing a complete closed loop of prediction-optimization-execution.

[0027] (5) Integrating physical mechanisms and data-driven approaches: By combining the prior mechanisms of physical simulation with data-driven machine learning, the value of massive process data can be fully explored. Physical simulation provides underlying mechanism constraints, while intelligent compensation networks compensate and correct the simulation results, forming complementary advantages and improving the generalization ability of the model, which can adapt to the polishing requirements of different processes and different types of wafers. Attached Figure Description

[0028] Figure 1 The diagram shown is a flowchart of a CMP process optimization method based on multiphysics coupling in one embodiment of this application.

[0029] Figure 2 The diagram shows a flowchart illustrating the process of generating simulation result data for each micro-mesh unit in one embodiment of this application.

[0030] Figure 3 The diagram shown is a flowchart illustrating the distribution prediction operation performed in one embodiment of this application.

[0031] Figure 4 The diagram shown is a schematic representation of the structure of a preset machine learning intelligent compensation network in one embodiment of this application.

[0032] Figure 5 The diagram shown is a flowchart illustrating the parameter adjustment operation performed in one embodiment of this application.

[0033] Figure 6 The diagram shown is a schematic block diagram of a CMP process optimization system based on multiphysics coupling in one embodiment of this application. Detailed Implementation

[0034] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0035] In the embodiments of this application, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" do not necessarily imply that they are different.

[0036] It should be noted that, in the embodiments of this application, the words "exemplary" or "for example" indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0037] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0038] Before providing a further detailed description of the present invention, the nouns and terms used in the embodiments of the present invention are explained, and the nouns and terms used in the embodiments of the present invention are subject to the following interpretations:

[0039] <1> CMP (Chemical Mechanical Polishing), also known as chemical mechanical planarization, is a core process in semiconductor manufacturing that achieves wafer surface planarization through the synergistic effect of chemical etching and mechanical polishing. It is widely used in critical steps such as multilayer wiring and dielectric layer planarization in integrated circuit chips.

[0040] <2> DOE (Design of Experiments): In semiconductor manufacturing (such as CMP process optimization), it is a systematic and efficient method for planning and executing experiments. The core objective is to accurately identify which process parameters (such as pressure, rotation speed, slurry concentration, etc.) have the greatest impact on the final results (such as thickness uniformity, defect rate) with the fewest number of experiments, and to establish a quantitative relationship between parameters and results, thereby finding the optimal process window.

[0041] To facilitate understanding of the embodiments of this application, firstly, in conjunction with Figure 1 Detailed explanation. Figure 1 This document illustrates a flowchart of a CMP process optimization method based on multiphysics coupling, as described in an embodiment of the present invention. The CMP process optimization method based on multiphysics coupling in this embodiment mainly includes the following steps:

[0042] S101: Based on a pre-built three-level grid system, perform grid division operation on the target chip layout to obtain multiple micro-grid units, and extract multi-scale feature quantization data of each micro-grid unit.

[0043] In this embodiment, the target chip layout is input into a three-level mesh system to perform nested mesh partitioning at macro, meso, and micro levels. The three-level mesh system uses the entire area of ​​the target chip layout as the analysis range, partitioning it according to a preset first size (e.g., 50×50μm) to obtain multiple macro-mesh units. Within each macro-mesh unit, it is further partitioned according to a preset second size (e.g., 5×5μm) to obtain multiple meso-mesh units. Within each meso-mesh unit, it is further partitioned according to a preset third size (e.g., 0.5×0.5μm) to obtain multiple micro-mesh units. The macro-scale captures wafer-level uniformity trends, the meso-scale analyzes standard cell-level process effects, and the micro-scale resolves nanometer-level edge details. This hierarchical processing mechanism can capture long-range density gradient effects and identify localized pressure concentration areas, providing comprehensive and accurate input features for subsequent process simulations. The system has adaptive adjustment capabilities, automatically optimizing mesh size and feature weights according to different process nodes.

[0044] In this embodiment, the multi-scale feature quantization data includes, but is not limited to, geometric complexity index (GCI), process sensitivity index (PSI), and cross-scale density ratio (CSDR). The geometric complexity index (GCI), process sensitivity index (PSI), and cross-scale density ratio (CSDR) of each micro-grid unit are calculated respectively, thereby achieving comprehensive capture of chip layout features.

[0045] In this embodiment, the calculation method for the geometric complexity index (GCI) of the microgrid cells includes:

[0046] Formula (1)

[0047] Wherein, GCI represents the geometric complexity index data of the microgrid cell; Indicates the first weighting coefficient; This represents the composite linewidth value of the microgrid cells after normalization. This represents the second weighting coefficient; This represents the edge spacing of the microgrid cells after normalization. PAR represents the third weighting coefficient; PAR represents the perimeter density ratio of the normalized microgrid cells.

[0048] In this embodiment, the linewidth values ​​of all patterns within a microgrid unit are statistically analyzed, and a weighted average is calculated to obtain the comprehensive linewidth value of the microgrid unit. The spacing between adjacent patterns within a microgrid unit is statistically analyzed, and the minimum spacing value is taken as the edge spacing value of that microgrid unit. The total perimeter of the patterns within a microgrid unit and the area of ​​the microgrid unit are statistically analyzed, and the total perimeter of the patterns is divided by the area of ​​the microgrid unit to obtain the perimeter density ratio value of the microgrid unit. The comprehensive linewidth value, edge spacing value, and perimeter density ratio value of the microgrid unit are normalized to map to a uniform numerical range of 0 to 1.

[0049] In this embodiment, the cross-scale density ratio data (CSDR) of the micro-grid cells is calculated as follows:

[0050] Formula (II)

[0051] Wherein, CSDR represents the cross-scale density ratio data of the micro-grid cell; ρ_micro represents the local pattern density of the micro-grid cell; and ρ_macro represents the average pattern density of the macro-grid cell to which the micro-grid cell belongs.

[0052] In this embodiment, the calculation method for the comprehensive score of the process sensitivity of the micro-grid unit includes:

[0053] Formula (III)

[0054] in, This represents the overall score indicating the process sensitivity of the micro-grid unit. This represents the feature weight coefficient corresponding to the i-th layout feature of the microgrid cell; This represents the i-th layout feature of a microgrid cell; represents the interaction term coefficient; GCI represents the geometric complexity index data of microgrid cells; CSDR represents the cross-scale density ratio data of microgrid cells.

[0055] In this embodiment, a historical CMP process dataset is collected, which includes: historical map feature data corresponding to multiple historical microgrid cells, and multiple historical process result data. The historical map feature data includes: historical geometric complexity index data, historical grid location information data, and historical layer information data. The historical process result data includes: erosion depth data, dish depth data, and thickness loss data. A random forest regression algorithm is used, with the historical map feature data as the model input and the historical process sensitivity comprehensive score constructed based on the historical process result data as the model output, to train a regression model. Feature importance analysis is performed on the regression model to obtain the feature weight coefficients corresponding to different map features. , ,......, }

[0056] In this embodiment, based on the geometric complexity index (GCI), cross-scale density ratio (CSDR), grid location information, layer information, and other layout features of the current microgrid cell, the comprehensive process sensitivity score of the current microgrid cell is calculated using formula (II). The comprehensive process sensitivity score of the current microgrid cell is then normalized and mapped to a uniform numerical range of 0 to 1 to obtain the process sensitivity index (PSI).

[0057] In this embodiment, the random forest regression algorithm is a regression prediction algorithm based on the idea of ​​ensemble learning. Its core is to construct multiple decision tree regressors and output results through an ensemble strategy: multiple sample subsets are randomly extracted from the historical CMP process dataset, and an independent decision tree regression model is trained for each subset; each decision tree regression model learns the mapping relationship between "historical layout feature data" and "comprehensive process sensitivity score" based on randomly selected partial layout features, and finally outputs the mean of all decision tree regression results as the predicted value. By performing feature importance analysis on the regression model, that is, by replacing feature values ​​and calculating the difference in prediction accuracy, the contribution of each layout feature to the comprehensive process sensitivity score is quantified, thereby obtaining the feature weight coefficients corresponding to different layout features, providing a basis for the subsequent calculation of process sensitivity index (PSI) data.

[0058] In this embodiment, a comprehensive capture of chip layout features is achieved by constructing a three-level grid system encompassing macroscopic, mesoscopic, and microscopic dimensions. By calculating the ratio of the local layout pattern density of a micro-grid cell to the average layout pattern density of the macroscopic grid cell to which the micro-grid cell belongs, the cross-scale density ratio (CSDR) data of the micro-grid cell is obtained. This allows for the identification of isolated structures and regions of abrupt density changes, establishing a bridge between microscopic structure and macroscopic performance. Simultaneously, the geometric complexity index (GCI) data of the micro-grid cell is calculated, comprehensively evaluating multi-dimensional geometric parameters such as linewidth, spacing, and perimeter density ratio, as well as calculating the process sensitivity index (PSI), thereby establishing a complete feature quantification index system.

[0059] S102: Based on the CMP multiphysics fully coupled mathematical model, and according to the multi-scale characteristic quantization data of each microgrid unit, perform physical simulation operations to generate simulation result data for each microgrid unit.

[0060] In this embodiment, as Figure 2 The diagram illustrates the flowchart for generating simulation result data for each microgrid unit in an embodiment of the present invention. The method of generating simulation result data for each microgrid unit by performing a physical simulation operation based on the CMP multiphysics fully coupled mathematical model and according to the multi-scale characteristic quantization data of each microgrid unit includes:

[0061] S1021: Based on the preset contact mechanics model and the multi-scale feature quantization data of each microgrid unit, the mechanical removal rate component of each microgrid unit is calculated.

[0062] In this embodiment, the preset contact mechanics model includes:

[0063] Formula (IV)

[0064] in, This represents the mechanical removal rate component; represents the Preston coefficient; P represents the contact pressure distribution. This represents the pressure gradient correction factor; Indicates the pressure gradient; Indicates relative sliding speed; Indicates the edge enhancement factor; Indicates the density suppression coefficient; This indicates the local metal density.

[0065] It should be noted that the Preston coefficient The Preston coefficient, measured through nanoindentation experiments, represents the material's fundamental response to polishing pressure, reflecting the ease with which the material itself is removed during chemical mechanical polishing. The contact pressure distribution P is obtained through contact mechanics simulation. Pressure gradient correction factor. Calibrated using cross-scale density ratio data (CSDR), its magnitude shows a positive linear correlation with the cross-scale density ratio data (CSDR). Pressure gradient Reflects the density abrupt change effect. The relative sliding speed v is a device setting. Edge enhancement factor. Correlation with the perimeter-density ratio in the Geometric Complexity Index (GCI) data. Density suppression coefficient. The local metallic density is calibrated using the average pattern density ρ_macro of the macroscopic grid cells to which the microscopic grid cell belongs. This refers to the proportion of the area of ​​metallic materials (such as copper, aluminum, and other wiring layers) to the total area of ​​the microgrid within a microgrid unit.

[0066] S1022: Based on the preset chemical corrosion model and the multi-scale feature quantization data of each micro-grid unit, the chemical removal rate component of each micro-grid unit is calculated.

[0067] In this embodiment, the preset chemical corrosion model includes:

[0068] Formula (5)

[0069] in, This represents the chemical removal rate component; Represents the rate constant of a chemical reaction; Indicates the concentration of active material in the grinding slurry; e represents the natural constant. R represents activation energy; T represents gas constant; T represents interface temperature. This represents the geometric complexity correction function.

[0070] It should be noted that the chemical reaction rate constant The concentration of active material in the grinding slurry was measured through electrochemical experiments. Activation energy is a process setpoint. These are inherent material properties. The gas constant R is a physical constant. The interface temperature T is obtained from thermodynamic simulations. Geometric complexity correction function. It is obtained by relying on the Geometric Complexity Index (GCI) data.

[0071] S1023: Based on the preset thermodynamic model and the multi-scale characteristic quantization data of each microgrid unit, the heat effect removal rate component of each microgrid unit is calculated.

[0072] In this embodiment, the preset thermodynamic model includes:

[0073] Formula (VI)

[0074] in, This represents the heat removal rate component; Indicates the coefficient of thermal expansion; This indicates a localized temperature rise; Indicates the effective contact time; This indicates process sensitivity correction.

[0075] It should be noted that the coefficient of thermal expansion Measured by thermomechanical analysis. Local temperature rise. Obtained from thermodynamic simulation. Effective contact time. This refers to the effective contact time between a specific micro-grid cell on the wafer surface and the polishing pad during the CMP process. It is determined by the relative speed of the polishing head and the actual contact area between the two. (Process sensitivity correction) It is obtained by relying on process sensitivity index (PSI) data.

[0076] S1024: Based on the mechanical removal rate component, chemical removal rate component, and thermal effect removal rate component of each microgrid unit, and based on the CMP multiphysics fully coupled mathematical model, the total material removal rate of each microgrid unit is calculated.

[0077] In this embodiment, the CMP multiphysics fully coupled mathematical model includes:

[0078] Formula (VII)

[0079] in, Indicates the total material removal rate; Indicates the contribution weight of mechanical grinding; Indicates the contribution weight of chemical corrosion; Indicates the weight of the contribution of the thermal effect; This represents the mechanical removal rate component; This represents the chemical removal rate component; This represents the heat removal rate component.

[0080] It should be noted that the contribution weight of mechanical grinding Calibrated using DOE experiments, the range is 0.4-0.7. Chemical corrosion contribution weight. Calibrated via DOE experiments, ranging from 0.2 to 0.5. Weight of thermal effect contribution. Calibrated by DOE experiments, range 0.1-0.3.

[0081] In this embodiment, a pressure gradient term and a linewidth sensitivity term (i.e., a pressure gradient correction factor) are introduced into the CMP multiphysics fully coupled mathematical model. By incorporating edge enhancement factors, a material removal rate model that more closely reflects the actual physical process is established. Through multi-field coupling calculations, the interaction between abrasive pad deformation, abrasive fluid flow, surface chemical reactions, and frictional heat effects can be accurately simulated, improving the accuracy of thickness prediction.

[0082] S1025: Perform a physical simulation operation based on the total material removal rate of each of the micro-grid units to generate simulation result data for each of the micro-grid units.

[0083] In this embodiment, a physical simulation operation is performed based on the total material removal rate of each of the micro-grid units. Using the divided micro-grid units as the basic simulation units, the total material removal rate of each micro-grid unit is taken as input, and combined with a preset CMP polishing time threshold, the real-time material removal depth of each micro-grid unit during the polishing process is calculated. Based on the real-time material removal depth of each micro-grid unit during the polishing process, simulation result data for each micro-grid unit is generated. The simulation result data includes: simulated thickness distribution data, simulated erosion depth data, simulated dish depth data, simulated thickness loss data, etc.

[0084] S103: Based on the simulation result data of each microgrid unit and a preset machine learning intelligent compensation network, perform a distribution prediction operation to generate the final prediction result data of each microgrid unit and the confidence interval corresponding to each final prediction result data.

[0085] In this embodiment, as Figure 3 The diagram illustrates a flowchart of the distribution prediction operation in an embodiment of the present invention. Figure 4 The diagram illustrates the structure of a pre-defined machine learning intelligent compensation network in an embodiment of the present invention. The pre-defined machine learning intelligent compensation network includes a feature embedding layer, an attention mechanism layer, a multi-scale feature fusion layer, and a prediction quantization layer. The method for generating the final prediction result data for each microgrid unit and the corresponding confidence interval for each final prediction result data, based on the simulation result data of each microgrid unit and the pre-defined machine learning intelligent compensation network, includes:

[0086] S1031: The feature embedding layer is used to perform spatial alignment and integration operations based on the simulation result data of the micro-grid unit and the multi-scale feature quantization data of the micro-grid unit, so as to generate simulation feature vector data of the micro-grid unit and layout multi-scale feature vector data.

[0087] In this embodiment, the feature embedding layer performs spatial alignment and integration operations on the simulation result data of the microgrid cells and the multi-scale feature quantization data of the microgrid cells to convert them into dense numerical vectors that can be uniformly processed by the deep neural network, and use them as the joint input of the network.

[0088] S1032: The attention mechanism layer is used to perform evaluation and focusing operations based on the simulation feature vector data and layout multi-scale feature vector data of the micro-grid unit to generate physical simulation key feature data and multi-scale process-sensitive feature data of the micro-grid unit.

[0089] In this embodiment, the attention mechanism layer includes a physical feature attention head and a process feature attention head. The physical feature attention head automatically analyzes the simulation feature vector data of the micro-grid cells, focusing on areas with high uncertainty or potential distortion in the simulation (e.g., boundaries, complex structures) to generate key physical simulation feature data for the micro-grid cells. The process feature attention head automatically analyzes the layout multi-scale feature vector data of the micro-grid cells, focusing on design-sensitive areas that have the greatest impact on the final process results (e.g., isolated micro-shapes) to generate multi-scale process-sensitive feature data for the micro-grid cells. This allows the network to learn "where to look," concentrating computational resources on the critical parts that most need compensation, improving the model's efficiency and interpretability. Both the key physical simulation feature data and the multi-scale process-sensitive feature data are attention-weighted data.

[0090] S1033: The multi-scale feature fusion layer is used to perform feature fusion operations based on the physical simulation key feature data and multi-scale process-sensitive feature data of the micro-grid unit to generate multi-scale fused feature data of the micro-grid unit.

[0091] In this embodiment, the multi-scale feature fusion layer performs feature fusion operations on the physical simulation key feature data with attention weights and the multi-scale process-sensitive feature data to fully capture the multi-scale coupling law of CMP process, so as to generate multi-scale fused feature data of micro-mesh unit, which integrates information such as simulation key area and layout process-sensitive area.

[0092] S1034: The prediction quantization layer is used to perform a distribution prediction operation based on the multi-scale fusion feature data of the microgrid unit to generate the final prediction result data of the microgrid unit and the confidence interval corresponding to the final prediction result data.

[0093] In this embodiment, the prediction quantization layer includes a temporal memory unit layer, an uncertainty estimation layer, and a Bayesian output layer; the distribution prediction operation includes a temporal correction operation, a prediction quantization operation, and a result correction operation; wherein, the prediction quantization layer is used to perform a distribution prediction operation based on the multi-scale fusion feature data of the microgrid units to generate the final prediction result data of the microgrid units and the confidence interval corresponding to the final prediction result data in the following ways:

[0094] (1) The temporal memory unit layer is used to perform temporal correction operations based on the multi-scale fusion feature data of the microgrid unit and historical process state data to generate temporal dependency compensation feature data of the microgrid unit.

[0095] (2) The uncertainty estimation layer is used to compensate feature data according to the temporal dependency of the microgrid unit and perform prediction quantization operation to generate prediction uncertainty feature data of the microgrid unit.

[0096] (3) The Bayesian output layer is used to perform result correction operation based on the prediction uncertainty feature data of the microgrid unit to generate the final prediction result of the microgrid unit and determine the confidence interval corresponding to the final prediction result.

[0097] In this embodiment, the temporal memory unit layer can adopt a Long Short-Term Memory (LSTM) network structure. Its function is to remember and associate historical process states, enabling the network to understand time dependencies and thus accurately compensate for errors with cumulative effects or time-varying characteristics. Historical process state data includes prediction results from previous steps (such as the predicted material removal depth at the previous moment) and equipment parameters (such as the time series of polishing pressure, polishing disc rotation speed, and slurry flow rate). Specifically, the temporal memory unit layer selectively forgets and updates historical information based on the multi-scale fusion feature data of microgrid units and historical process state data through a gating mechanism. It selectively discards historical noise data that has no significant impact on the current prediction, retains the temporal correlation information that has a key impact on the current prediction, and performs temporal correction operations. Finally, it outputs temporal dependency compensation feature data of microgrid units that integrates spatial and temporal dynamic features, solving the technical problem that traditional models cannot capture the cumulative effects of CMP processes over time.

[0098] In this embodiment, the uncertainty estimation layer is used to quantify the degree of uncertainty in the model's prediction of the process results of each micro-grid unit. For regions with sparse data, complex features, or large known deviations in physical simulation, it will provide a larger uncertainty quantification result, providing a basis for subsequent prediction result correction.

[0099] In this embodiment, the confidence interval is used to characterize the credibility of the final prediction result, i.e., the probability that the actual measured value falls within the interval. The width of the confidence interval reflects the reliability of the model prediction and the risk level of the corresponding process region: When the confidence interval is narrow (e.g., 350±2nm), it indicates that the model has a high degree of confidence in the prediction result of the current micro-grid cell. This type of region usually corresponds to a conventional layout region with simple design rules, mature physical models, and abundant historical data, and has a low process risk level. In the process optimization process, performance targets can be pursued more aggressively. When the confidence interval is wide (e.g., 350±15nm), it indicates that the model has a low degree of confidence in the prediction result of the current micro-grid cell, and can be regarded as a high-risk warning for the process. This type of region usually corresponds to a region with extremely complex layout structure (e.g., regions with high geometric complexity exponents and high cross-scale density ratios), significant multi-physics coupling effects, and potential inherent simplification or failure of the physical simulation model. In the process control and decision-making process, it is necessary to increase redundancy margins and strengthen verification.

[0100] In this embodiment, the Bayesian output layer performs result correction operations based on the prediction uncertainty characteristic data of the micro-grid cells, outputting the final prediction result of the micro-grid cells in the form of a probability distribution (such as mean and variance), and simultaneously outputting the confidence interval corresponding to the final prediction result. The Bayesian output layer does not overturn the simulation result data of the micro-grid cells, but rather learns the systematic error pattern between the physical simulation and the actual measured values ​​based on historical data, predicts a "deviation compensation value," performs data-driven fine-tuning of the simulation result data, and outputs the final prediction result and its corresponding confidence interval in the form of a probability distribution, quantifying the prediction reliability and providing a risk basis for subsequent process decisions.

[0101] In this embodiment, the final prediction result data includes the thickness distribution data after micro-grid cell correction, the erosion depth data after correction, the dish-shaped depth data after correction, and the thickness loss data after correction.

[0102] It is worth noting that the pre-defined machine learning intelligent compensation network achieves a complementary advantage between physical mechanisms and data-driven approaches. Employing a hierarchical attention mechanism—namely, physical feature attention heads and process feature attention heads—it adaptively focuses on key influencing factors, autonomously identifying critical process parameters and sensitive design regions, thus enhancing the model's interpretability. A temporal memory unit layer captures instantaneous and cumulative effects, and an uncertainty estimation layer is introduced to quantify the uncertainty of the model's predictions for process results at each microgrid unit, providing direct evidence for process risk assessment. The Bayesian output layer outputs the final prediction results for each microgrid unit in the form of a probability distribution (such as mean and variance), along with the corresponding confidence interval, providing decision confidence support for intelligent manufacturing.

[0103] S104: Based on the final prediction result data of each of the microgrid cells and the confidence interval corresponding to each of the final prediction result data, perform parameter adjustment operations to generate the CMP target process parameter control sequence.

[0104] In this embodiment, as Figure 5 The diagram illustrates a flowchart of the parameter adjustment operation in an embodiment of the present invention. The method for generating a CMP target process parameter control sequence by performing parameter adjustment operations based on the final prediction result data of each microgrid unit and the corresponding confidence interval of each final prediction result data includes:

[0105] S1041: Based on the final prediction result data of each of the micro-grid units and the confidence interval corresponding to each of the final prediction result data, and based on the preset rolling time-domain optimization strategy, a predictive control model is constructed to generate an initial process parameter control sequence.

[0106] In this embodiment, the preset rolling time-domain optimization strategy refers to preset two time windows: a prediction time domain (N steps) and a control time domain (M steps, M≤N). Within each control cycle, the prediction time domain covers the process for several future control cycles, while the control time domain serves as the time span for the parameters to be optimized. For example, the process results for the next 10 control cycles (N=10) are predicted, but only the control parameters for the next 3 cycles (M=3) are optimized, thereby constructing a predictive control model. This predictive control model, based on the time windows of the prediction and control time domains, uses the final prediction result data of each microgrid cell and the confidence interval corresponding to each final prediction result data, combined with the actual operating status data of the current CMP equipment (including current partition polishing pressure, polishing disc speed, etc.), to achieve dynamic prediction of future process results and quantification of prediction uncertainty, and outputs the initial process parameter control sequence for several future control cycles. For example, the predictive control model is based on a preset rolling time-domain optimization strategy and uses a CMP multiphysics fully coupled mathematical model to predict the process results under different time windows at the current stage. For example, it predicts the process results for the next 10 seconds, and optimizes the process parameters only for the results of the next 3 seconds.

[0107] S1042: Based on the preset process knowledge graph, perform causal association analysis on the final prediction result data of each microgrid unit and the confidence interval corresponding to each final prediction result data to identify multiple high-risk areas and determine the process parameter optimization variables and process parameter optimization weights corresponding to each high-risk area.

[0108] In this embodiment, the preset process knowledge graph mainly describes the following types of relationships:

[0109] (1) Physical field coupling relationship: describes the interaction mechanism between mechanical, chemical, fluid and thermal fields.

[0110] (2) Parameter-effect relationship: Describe how controllable parameters (pressure, speed, etc.) affect the process results (removal rate, uniformity, etc.).

[0111] (3) Feature-sensitivity relationship: describes the sensitivity of layout geometric features (density, line width, etc.) to specific process defects.

[0112] (4) Time-series degradation relationship: describes the aging pattern of consumables (grinding pads, dressers) over time.

[0113] (5) Constraint Dependency: Describes the mutual constraints between equipment physical limits and process windows.

[0114] A pre-defined process knowledge graph tells the system, "When A occurs, B usually occurs because mechanism C is at work." It provides physical constraints and a reasoning framework for subsequent data-driven models, ensuring that the learning direction aligns with fundamental process principles and avoiding erroneous conclusions that "conform to data but not to physics." For example:

[0115] [High CSDR region] — (leading to) —> [Local pressure overshoot]

[0116] ↑ ↓

[0117] [Grinding pad deformation]

[0118] ↓

[0119] [Narrow linewidth structure] — (exacerbated) —> [Abnormally high local removal rate]

[0120] ↑ ↓

[0121] [Distribution / Copper Removal Selectivity Imbalance]

[0122] ↓

[0123] [Abrasive pad aging] —(weakening) —> [Abrasive delivery efficiency] —(impact) —> [Dish defect formation].

[0124] In this embodiment, based on a preset process knowledge graph, causal correlation analysis is performed on the final prediction results of each microgrid unit and the corresponding confidence intervals. The deviations of the final prediction results of each microgrid unit from the preset process standard threshold are compared. Combined with the prediction uncertainty quantified by the confidence interval, microgrid units whose prediction results exceed the standard threshold and whose confidence intervals fluctuate significantly (wide confidence intervals) are identified as multiple high-risk regions. The causes of errors in these high-risk regions are analyzed to determine the process parameter optimization variables and process parameter optimization weights corresponding to each high-risk region. For example, the high uncertainty in high-risk region A, analyzed using the preset process knowledge graph, is mainly caused by uneven local pressure and has a weak relationship with rotational speed. Therefore, "zone pressure" rather than "spindle speed" is used as the process parameter optimization variable for region A, and it is given a higher process parameter optimization weight.

[0125] S1043: Based on the process parameter optimization variables and process parameter optimization weights corresponding to each of the high-risk areas, the initial optimal process parameter control sequence is calibrated, and based on a multi-objective optimization algorithm, the calibrated initial optimal process parameter control sequence is optimized to generate the CMP target process parameter control sequence.

[0126] In this embodiment, the initial optimal process parameter control sequence is calibrated based on the process parameter optimization variables and process parameter optimization weights corresponding to each high-risk area. During the calibration process, the setting range of the process parameter optimization variables corresponding to the high-risk area is adjusted. At the same time, the control intensity of the process parameter optimization variables corresponding to the high-risk area is strengthened by combining the process parameter optimization weights, while the control of parameters with less impact is weakened, so as to ensure that the calibrated initial process parameter control sequence can specifically solve the process hidden dangers in the high-risk area.

[0127] In this embodiment, the multi-objective optimization algorithm combines multiple process objectives (such as maximizing uniformity, minimizing defect rate, and minimizing material cost) to transform these objectives into a mathematical problem, constructing a total cost function. Taking the final predicted result data as the thickness distribution data corrected by the micro-grid cells as an example, the constructed total cost function includes:

[0128] Formula (8)

[0129] Where J represents the total cost function; This represents the fourth weighting coefficient; This represents the variance of the uniformity of the thickness distribution data; This represents the fifth weighting coefficient; This represents the probability that the predicted thickness will violate the risk threshold. It is calculated from the confidence interval. The greater the uncertainty, the higher the probability of violating the risk constraint. This represents the sixth weighting coefficient; This indicates the energy consumption for pressure adjustment.

[0130] In this embodiment, with the goal of minimizing the total cost function, an optimization solution operation is performed on the calibrated initial optimal process parameter control sequence. During this solution process, different combinations of process parameters, including partition pressure, table speed, and polishing fluid flow rate, are simulated within the future time window corresponding to the prediction time domain. The predictive control model is used to predict the process results corresponding to each parameter combination and follows multiple constraints. The hard constraints are the physical limits of the equipment (such as maximum pressure) and the process safety boundary, while the soft constraints are the dynamic safety margin based on the prediction confidence. The upper and lower limits of the thickness target are automatically tightened for high-risk areas to reserve sufficient safety buffer. For example, the thickness constraint for high-risk areas (wide confidence interval) is automatically tightened from [target ± 10 nm] to [target ± 5 nm] to create a dynamic safety margin. Subsequently, a multi-objective optimization algorithm (such as NSGA-II or interior point method) searches within this constraint space to minimize the total cost function J, and finally generates the CMP target process parameter control sequence, for example: {at time t+1: pressure in region 1 = 30.5 kPa, pressure in region 2 = 29.8 kPa; at time t+2: pressure in region 1 = 30.6 kPa, ...}.

[0131] In this embodiment, NSGA-II is an evolutionary algorithm for multi-objective optimization. Its principle is to simulate natural selection and evolution by iteratively searching the solution space to generate a set of Pareto optimal solutions. First, it performs a "non-dominated sort" on all candidate solutions, dividing non-dominated solutions into different frontier levels and prioritizing solutions with higher levels. Then, through "crowding calculation," it selects solutions with a more even distribution within the same frontier level to ensure the diversity of the solution set. Finally, it generates the next generation population through selection, crossover, and mutation operations, iterating repeatedly until convergence, thus obtaining a set of solutions that achieve optimal trade-offs among multiple objectives, rather than a single "optimal solution."

[0132] In this embodiment, the interior-point method is an efficient numerical algorithm for solving constrained single-objective optimization problems. Its principle is to introduce a "barrier function," integrating the constraints into the objective function and transforming the constrained optimization problem into a series of unconstrained subproblems. Starting from an initial point within the feasible region, it iteratively approaches the optimal solution, remaining within the feasible region throughout the iteration process to avoid directly touching the constraint boundaries. As the iteration progresses, the influence of the barrier function gradually weakens, and the solution gradually approaches the constraint boundaries, eventually converging to the optimal solution that satisfies all constraints.

[0133] In this embodiment, the method further includes: controlling the CMP target process parameter sequence according to the CMP target process parameter sequence, and acquiring the actual operating data of the CMP equipment and the real measurement data of each microgrid unit in real time, so as to perform parameter update operations on the predictive control model.

[0134] In this embodiment, the control command at time t+1 in the CMP target process parameter control sequence is extracted and sent to the CMP equipment to perform the CMP polishing operation. After execution at time t+1, the actual operating data of the equipment and the real measurement data of each micro-grid unit (including real thickness distribution, real erosion depth, real dish depth, real thickness loss, etc.) are obtained through the sensors (online thickness measurement sensors) on the equipment. Subsequently, the prediction error of each micro-grid unit is calculated, that is, the difference between the real measurement data of each micro-grid unit and the corresponding final prediction result data (e.g., real thickness - predicted thickness). Combining the preset process knowledge graph, a causal analysis operation is performed on the prediction error to analyze the root cause of the error. For example: "The prediction for region 1 is too thin, while the actual pressure is slightly lower than the set value. Based on the preset process knowledge graph, it is determined that the main reason is a slight hysteresis in the pressure control loop." Using this prediction error and the causal analysis result, the parameters of the predictive control model are fine-tuned through incremental learning or parameter calibration, and the rule weights of the preset process knowledge graph are optimized simultaneously to provide reliable support for the generation of the control sequence for the next control cycle, realizing closed-loop adaptive control.

[0135] In this embodiment, based on a preset rolling time-domain optimization strategy, a preset process knowledge graph, and a multi-objective optimization algorithm, the target process parameter control sequence for CMP is solved online. This allows for dynamic matching of the current process state and risk level by combining the final prediction results and confidence intervals of the micro-grid units, achieving real-time, accurate, and adaptive control of process parameters. During control execution, the online-solved CMP target process parameter control sequence is sent to the CMP equipment for polishing operations. Sensors collect real-time operating data of the equipment and actual measurement data of each micro-grid unit, calculating the prediction error and using the preset process knowledge graph to trace the source of the error and pinpoint the cause of process deviations. Utilizing this prediction error and causal analysis results, the predictive control model is updated through incremental learning or parameter calibration, and the rule weights of the preset process knowledge graph are simultaneously optimized. This provides reliable support for the generation of the control sequence for the next control cycle, achieving closed-loop adaptive control.

[0136] It is worth noting that the CMP process optimization method based on multiphysics coupling of the present invention has the following advantages:

[0137] (1) Improve modeling accuracy and simulation reliability: The target chip layout is divided into grids through a three-level grid system to extract multi-scale feature quantization data of each micro-grid unit. Based on the CMP multi-physics field fully coupled mathematical model, physical simulation operation is performed, which can fully characterize the complex evolution law of the three-dimensional morphology of the wafer surface under advanced process, improve the accuracy and consistency of simulation results, and overcome the problems of simplified assumptions and large prediction deviations in traditional models.

[0138] (2) Achieve multi-physics coupling: Incorporate multi-physics effects such as contact mechanics, chemical reaction, and thermodynamics into a unified modeling framework, and reveal the intrinsic relationship between process parameters and flattening effect through fully coupled physical simulation, providing scientific and systematic theoretical support for process optimization.

[0139] (3) Improve prediction accuracy and quantify uncertainty: The simulation result data is compensated and corrected by a machine learning intelligent compensation network, the final prediction result is output, and the confidence interval is given at the same time. This can intuitively represent the credibility of the prediction, quantify the reliability of the prediction, and provide risk basis for subsequent process decisions.

[0140] (4) Possesses real-time adaptive and closed-loop optimization capabilities: Based on a preset rolling time-domain optimization strategy, a preset process knowledge graph, and a multi-objective optimization algorithm, the CMP target process parameter control sequence is solved online. It can dynamically match the current process state and risk level by combining the final prediction results and confidence intervals of the micro-grid units, thereby achieving real-time, accurate, and adaptive control of process parameters. The CMP target process parameter control sequence is sent to the CMP equipment to perform polishing operations. Real-time data of actual equipment operation and real measurement data of each micro-grid unit are collected. The prediction error is calculated and the error source is traced by combining the preset process knowledge graph to locate the cause of process deviation. Using the prediction error and causal analysis results, the parameters of the predictive control model are updated through incremental learning or parameter calibration, and the rule weights of the preset process knowledge graph are optimized simultaneously, realizing a complete closed loop of prediction-optimization-execution.

[0141] (5) Integrating physical mechanisms and data-driven approaches: By combining the prior mechanisms of physical simulation with data-driven machine learning, the value of massive process data can be fully explored. Physical simulation provides underlying mechanism constraints, while intelligent compensation networks compensate and correct the simulation results, forming complementary advantages and improving the generalization ability of the model, which can adapt to the polishing requirements of different processes and different types of wafers.

[0142] Furthermore, the CMP process optimization method based on multiphysics coupling in this application constructs a three-level grid system to extract multi-scale feature quantization data of each microgrid unit, establishes a reliable mechanistic model based on the fully coupled CMP multiphysics mathematical model, and utilizes a pre-set machine learning intelligent compensation network to capture complex nonlinear relationships, enabling real-time adaptive optimization of process parameters and achieving closed-loop process control. This architecture maintains the interpretability and extrapolation capability of the physical model while possessing the flexibility and adaptive characteristics of a data-driven model, providing a novel technical paradigm for semiconductor manufacturing process optimization.

[0143] Figure 6 This is a schematic block diagram of a CMP process optimization system based on multiphysics coupling provided in an embodiment of this application. Figure 6 As shown, the CMP process optimization system 600 based on multiphysics coupling includes:

[0144] The multi-scale feature extraction module 601 is used to perform grid division operation on the target chip layout based on a pre-constructed three-level grid system to obtain multiple micro-grid units and extract multi-scale feature quantization data of each micro-grid unit.

[0145] The multiphysics coupled simulation module 602 is used to perform physical simulation operations based on the CMP multiphysics fully coupled mathematical model and according to the multi-scale characteristic quantization data of each microgrid unit, so as to generate simulation result data of each microgrid unit.

[0146] The intelligent compensation module 603 is used to perform distribution prediction operations based on the simulation result data of each of the microgrid units and a preset machine learning intelligent compensation network, so as to generate the final prediction result data of each microgrid unit and the confidence interval corresponding to each final prediction result data.

[0147] The optimization control module 604 is used to perform parameter adjustment operations based on the final prediction result data of each of the microgrid cells and the confidence interval corresponding to each of the final prediction result data, so as to generate a CMP target process parameter control sequence.

[0148] It should be understood that the specific process of each module performing the above-mentioned steps has been described in detail in the above method embodiments, and will not be repeated here for the sake of brevity.

[0149] It should also be understood that the module division in the embodiments of this application is illustrative and only represents a logical functional division; in actual implementation, there may be other division methods. Furthermore, the functional modules in the various embodiments of this application can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0150] According to the method provided in the embodiments of this application, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the CMP process optimization method based on multiphysics coupling as described above.

[0151] According to the method provided in the embodiments of this application, this application also provides a computer program product, which includes computer program code. When the computer program code is run on a computer, the computer implements the CMP process optimization method based on multiphysics coupling as described above.

[0152] As used in this specification, the terms "component," "module," "system," etc., are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process and / or an execution thread, and components may be located on a single computer and / or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, and / or a network, such as the Internet interacting with other systems via signals).

[0153] Those skilled in the art will recognize that the various illustrative logical blocks and steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0154] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0155] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0156] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0157] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0158] In the above embodiments, the functions of each functional unit can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. A computer program product includes one or more computer instructions (programs). When the computer program instructions (programs) are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0159] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0160] In summary, this application provides a CMP process optimization method, system, medium, and product based on multi-physics coupling. It divides the target chip layout using a three-level mesh system and extracts multi-scale feature quantification data of micro-mesh units. Combined with a fully coupled CMP multi-physics mathematical model for physical simulation, it can fully characterize the complex evolution of the three-dimensional morphology of the wafer surface under advanced processes, improving the accuracy and consistency of simulation results and overcoming the problems of simplified assumptions and large prediction deviations in traditional models. Simultaneously, it incorporates contact mechanics, chemical reactions, thermodynamics, and other multi-physics effects into a unified modeling framework, achieving multi-physics coupling simulation and providing a scientific and systematic theoretical support for process optimization. A machine learning intelligent compensation network is used to compensate and correct the simulation results, outputting the final prediction results and corresponding confidence intervals, improving prediction accuracy and quantifying prediction reliability, providing a risk basis for process decisions. Based on a preset rolling time-domain optimization strategy, a preset process knowledge graph, and a multi-objective optimization algorithm, it solves the CMP target process parameter control sequence online, achieving real-time adaptive and closed-loop optimization capabilities. Therefore, this application effectively overcomes the various shortcomings of existing technologies and has high industrial application value.

[0161] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A CMP process optimization method based on multiphysics coupling, characterized in that, include: Based on a pre-built three-level grid system, a grid division operation is performed on the target chip layout to obtain multiple micro-grid units, and multi-scale feature quantization data of each micro-grid unit is extracted. Based on the CMP multiphysics fully coupled mathematical model, and according to the multi-scale characteristic quantization data of each microgrid unit, physical simulation operation is performed to generate simulation result data of each microgrid unit. Based on the simulation results of each microgrid unit, and based on a preset machine learning intelligent compensation network, a distribution prediction operation is performed to generate the final prediction results data of each microgrid unit and the confidence interval corresponding to each final prediction results data. Based on the final prediction results of each microgrid cell and the confidence interval corresponding to each final prediction result, a parameter adjustment operation is performed to generate a CMP target process parameter control sequence.

2. The CMP process optimization method based on multiphysics coupling according to claim 1, characterized in that, The method of generating simulation result data for each microgrid unit by performing physical simulation operations based on the CMP multiphysics fully coupled mathematical model and according to the multi-scale characteristic quantization data of each microgrid unit includes: Based on the preset contact mechanics model and the multi-scale feature quantization data of each microgrid unit, the mechanical removal rate component of each microgrid unit is calculated. Based on the preset chemical corrosion model and the multi-scale feature quantization data of each micro-grid unit, the chemical removal rate component of each micro-grid unit is calculated. Based on a preset thermodynamic model and multi-scale characteristic quantization data of each microgrid unit, the heat effect removal rate component of each microgrid unit is calculated. Based on the mechanical removal rate component, chemical removal rate component, and thermal effect removal rate component of each microgrid unit, and based on the CMP multiphysics fully coupled mathematical model, the total material removal rate of each microgrid unit is calculated. Based on the total material removal rate of each of the microgrid cells, a physical simulation operation is performed to generate simulation result data for each of the microgrid cells.

3. The CMP process optimization method based on multiphysics coupling according to claim 1, characterized in that, The CMP multiphysics fully coupled mathematical model includes: ； in, Indicates the total material removal rate; Indicates the contribution weight of mechanical grinding; Indicates the contribution weight of chemical corrosion; Indicates the weight of the contribution of the thermal effect; This represents the mechanical removal rate component; This represents the chemical removal rate component; This represents the heat removal rate component.

4. The CMP process optimization method based on multiphysics coupling according to claim 1, characterized in that, The preset machine learning intelligent compensation network includes a feature embedding layer, an attention mechanism layer, a multi-scale feature fusion layer, and a prediction quantization layer; wherein, based on the simulation result data of each micro-grid unit and the preset machine learning intelligent compensation network, a distribution prediction operation is performed to generate the final prediction result data of each micro-grid unit and the confidence interval corresponding to each final prediction result data, including the following methods: The feature embedding layer is used to perform spatial alignment and integration operations based on the simulation result data of the micro-grid unit and the multi-scale feature quantization data of the micro-grid unit, so as to generate simulation feature vector data of the micro-grid unit and multi-scale feature vector data of the layout. The attention mechanism layer is used to perform evaluation and focusing operations based on the simulation feature vector data and layout multi-scale feature vector data of the micro-grid cells to generate physical simulation key feature data and multi-scale process-sensitive feature data of the micro-grid cells. The multi-scale feature fusion layer is used to perform feature fusion operations based on the physical simulation key feature data and multi-scale process-sensitive feature data of the micro-grid unit to generate multi-scale fused feature data of the micro-grid unit. The prediction quantization layer is used to perform a distribution prediction operation based on the multi-scale fusion feature data of the microgrid cells to generate the final prediction result data of the microgrid cells and the confidence interval corresponding to the final prediction result data.

5. The CMP process optimization method based on multiphysics coupling according to claim 4, characterized in that, The prediction quantization layer includes a temporal memory unit layer, an uncertainty estimation layer, and a Bayesian output layer; the distribution prediction operation includes a temporal correction operation, a prediction quantization operation, and a result correction operation; wherein, the prediction quantization layer is used to perform the distribution prediction operation based on the multi-scale fusion feature data of the microgrid units to generate the final prediction result data of the microgrid units and the confidence interval corresponding to the final prediction result data in the following ways: The temporal memory unit layer is used to perform temporal correction operations based on the multi-scale fusion feature data of the microgrid units and historical process state data to generate temporal dependency compensation feature data of the microgrid units. The uncertainty estimation layer is used to compensate feature data according to the temporal dependency of the microgrid cells and perform prediction quantization operations to generate prediction uncertainty feature data of the microgrid cells. The Bayesian output layer is used to perform result correction operations based on the prediction uncertainty feature data of the microgrid cells to generate the final prediction result of the microgrid cells and determine the confidence interval corresponding to the final prediction result.

6. The CMP process optimization method based on multiphysics coupling according to claim 1, characterized in that, The method for generating a CMP target process parameter control sequence by performing parameter adjustment operations based on the final prediction results of each microgrid unit and the corresponding confidence interval of each final prediction result includes: Based on the final prediction results of each microgrid unit and the confidence interval corresponding to each final prediction result, and based on a preset rolling time-domain optimization strategy, a predictive control model is constructed to generate an initial process parameter control sequence. Based on a preset process knowledge graph, causal association analysis is performed on the final prediction results data of each microgrid unit and the confidence intervals corresponding to each final prediction results data to identify multiple high-risk areas and determine the process parameter optimization variables and process parameter optimization weights corresponding to each high-risk area. Based on the process parameter optimization variables and process parameter optimization weights corresponding to each of the high-risk areas, the initial optimal process parameter control sequence is calibrated, and an optimization solution operation is performed on the calibrated initial optimal process parameter control sequence based on a multi-objective optimization algorithm to generate the CMP target process parameter control sequence.

7. The CMP process optimization method based on multiphysics coupling according to claim 6, characterized in that, Also includes: Based on the CMP target process parameter control sequence, and by acquiring the actual operating data of the CMP equipment and the real measurement data of each microgrid unit in real time, parameter update operations are performed on the predictive control model.

8. A CMP process optimization system based on multiphysics coupling, characterized in that, include: The multi-scale feature extraction module is used to perform grid division operations on the target chip layout based on a pre-built three-level grid system to obtain multiple micro-grid units and extract multi-scale feature quantization data of each micro-grid unit. The multiphysics coupling simulation module is used to perform physical simulation operations based on the CMP multiphysics fully coupled mathematical model and according to the multi-scale characteristic quantization data of each microgrid unit to generate simulation result data of each microgrid unit. The intelligent compensation module is used to perform distribution prediction operations based on the simulation result data of each microgrid unit and a preset machine learning intelligent compensation network to generate the final prediction result data of each microgrid unit and the confidence interval corresponding to each final prediction result data. The optimization control module is used to perform parameter adjustment operations based on the final prediction result data of each of the microgrid cells and the confidence interval corresponding to each of the final prediction result data, so as to generate the CMP target process parameter control sequence.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the CMP process optimization method based on multiphysics coupling as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes computer program code, which, when run on a computer, enables the computer to implement the CMP process optimization method based on multiphysics coupling as described in any one of claims 1 to 7.

Images