Dummy filling optimization method and system based on neural network CMP model

Abstract

The invention provides a dummy filling optimization method and system based on a neural network CMP model, and the method comprises the steps: dividing a chip layout into a plurality of windows, and taking each window as a basic filling unit for filling optimization; extracting a fillable area of each window according to a design rule; distributing a target height of each window based on a prediction height and a design rule of a neural network CMP model; and based on the fillable area and the target height, a multi-target genetic algorithm is utilized to generate filling parameters and optimize the filling parameters. The method is used for improving the surface flatness in the chemical mechanical polishing process and optimizing the filling layout design. According to the method, efficient and accurate multi-objective optimization is realized in combination with the multi-objective genetic algorithm, the prediction accuracy of height change is improved, and the comprehensive performance of the filling scheme is also remarkably optimized.

Description

Technical Field

[0001] The present invention relates to the field of integrated circuit manufacturing and computer science and technology, and in particular to a dummy element filling optimization method and system based on a neural network chemical mechanical polishing (CMP) model, and also to a corresponding computer terminal and a computer-readable storage medium. Background Art

[0002] Chemical Mechanical Polishing (CMP) is one of the key processes for achieving surface flatness in integrated circuit manufacturing, which directly affects the reliability and performance of multi-layer interconnect structures. In modern advanced processes, due to the unevenness of metal density in the layout, local height variations are prone to occur in the CMP process, which may lead to degradation of electrical performance, increased signal delays, and even manufacturing defects. As the technology node advances from 28nm to 7nm and even 3nm, the impact of these height variations on chip performance has become increasingly significant.

[0003] In order to reduce the height variation during CMP, dummy metal filling is usually inserted into the layout to balance the metal density. However, metal filling optimization is not just a simple geometric filling problem, but also involves the trade-off of multiple objectives such as height variation, capacitive coupling and filling amount. Existing metal filling methods are mainly divided into rule-based methods and model-based methods. Among them, rule-based filling methods rely on fixed design rules to allocate metal filling areas. These methods are simple and efficient, but lack flexibility, especially when facing complex layouts, it is difficult to achieve accurate optimization of multiple objectives. Model-based filling methods use physics-driven or data-driven models to predict height variations and other performance indicators during CMP. Although these methods have high optimization accuracy, they are usually computationally complex, especially when dealing with large-scale chip layouts, and it is difficult to meet the timeliness requirements in actual production. In addition, with the increase in process complexity, especially when facing nonlinear effects and complex geometric structures, the existing models have limited ability to describe height variations in CMP.

[0004] At present, no description or report of similar technology to the present invention has been found, and similar information at home and abroad has not been collected yet. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a dummy element filling optimization method and system based on a neural network chemical mechanical polishing (CMP) model, and also provides a corresponding computer terminal and computer-readable storage medium for layout optimization and flatness control in the chemical mechanical polishing process.

[0006] According to one aspect of the present invention, a dummy variable filling optimization method based on a neural network CMP model is provided, comprising:

[0007] Divide the chip layout into multiple windows, and use each window as a basic filling unit for filling optimization;

[0008] Extract the fillable area of ​​each window according to the design rules;

[0009] Assign a target height to each window based on the predicted height of the neural network CMP model and design rules;

[0010] Based on the fillable area and the target height, a filling parameter is generated and optimized using a multi-objective genetic algorithm.

[0011] According to another aspect of the present invention, a dummy element filling optimization system based on a neural network CMP model is provided, comprising:

[0012] Layout discretization module, which is used to divide the chip layout into multiple windows and use each window as a basic filling unit for filling optimization;

[0013] A fillable area extraction module, which extracts the fillable area of ​​each window according to the design rules;

[0014] A target height assignment module that assigns a target height to each window based on the predicted height of the neural network CMP model and design rules;

[0015] A filling optimization module generates and optimizes filling parameters based on the fillable area and the target height using a multi-objective genetic algorithm.

[0016] According to a third aspect of the present invention, there is provided a computer terminal comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, can be used to execute the method described above in the present invention, or to run the system described above in the present invention.

[0017] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can be used to execute the method described above in the present invention, or to run the system described above in the present invention.

[0018] Due to the adoption of the above technical solution, the present invention has at least one of the following beneficial effects compared with the prior art:

[0019] Improved accuracy: By introducing a neural network CMP model, the present invention can accurately predict the impact of layout density on surface height changes, which is significantly improved compared to the prediction accuracy of traditional physical models, especially in complex layout environments.

[0020] Enhanced optimization efficiency: The present invention combines a multi-objective genetic algorithm with parallel computing and a dynamic feedback mechanism to significantly reduce the time required for optimization. Compared with traditional optimization methods (such as single-objective optimization or rule-driven algorithms), the efficiency is improved by several times.

[0021] Global balance optimization: The present invention uses a multi-objective fitness function to comprehensively consider surface flatness, height variation, capacitive coupling and filling amount to achieve global balance optimization, effectively solving the problem of being unable to take into account multiple objective constraints in the prior art.

[0022] Wide applicability: The present invention supports multiple filling modes (such as rectangular, staggered, and distributed characteristic functions) and different process requirements (such as density restrictions and shape constraints), and can flexibly adapt to different node processes and complex layout environments.

[0023] Real-time adjustment capability: Through a dynamic feedback mechanism, the present invention can adjust the optimization strategy in real time according to layout and process changes, which significantly improves the adaptability and robustness of the system in the actual manufacturing environment.

[0024] Easy to expand and maintain: The modular system architecture of the present invention facilitates the addition or adjustment of functional modules, can quickly adapt to emerging process nodes and layout optimization requirements, and reduces maintenance and expansion costs.

[0025] The present invention combines a convolutional neural network to predict the height change of the layout during the polishing process, and globally optimizes the dummy filling strategy through a multi-objective genetic algorithm, taking into account the multiple constraints of flatness, capacitive coupling and filling amount, which not only improves the prediction accuracy of the height change, but also significantly optimizes the comprehensive performance of the filling scheme.

[0026] The present invention realizes accurate modeling of layout density and height through neural network, and generates optimal dummy element filling scheme using genetic algorithm, reduces height variation after filling, optimizes filling uniformity, and controls capacitive coupling effect, and is particularly suitable for high-precision integrated circuit manufacturing process optimization.

[0027] The present invention solves the trade-off problem among surface flatness, capacitive coupling and filling amount control by combining a neural network to predict the surface height change of the layout during the CMP process and a multi-objective genetic algorithm to optimize the dummy filling layout. It has the characteristics of high efficiency and precision and is suitable for the metal filling optimization needs in advanced integrated circuit manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0029] Figure 1 This is a workflow diagram of a dummy variable filling optimization method based on a neural network CMP model in a preferred embodiment of the present invention.

[0030] Figure 2 It is a schematic diagram of the component modules of a dummy filling optimization system based on a neural network CMP model in a preferred embodiment of the present invention.

[0031] Figure 3 The figure is a workflow diagram of a dummy variable filling optimization method based on a neural network CMP model in a specific application example of the present invention.

[0032] Figure 4 The present invention is a flowchart of an algorithm for performing filling optimization using a multi-objective genetic algorithm adopted in a specific application example of the present invention.

[0033] Figure 5 It is a schematic diagram of different filling modes used in a specific application example of the present invention.

[0034] Figure 6 This is a diagram showing the effect of dummy filling in a specific application example of the present invention. DETAILED DESCRIPTION

[0035] The following is a detailed description of the embodiments of the present invention: This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given. It should be pointed out that for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

[0036] An embodiment of the present invention provides a dummy element filling optimization method based on a neural network CMP (chemical mechanical polishing) model, which is used to improve the surface flatness in the chemical mechanical polishing process and optimize the filling layout design. The method combines a multi-objective genetic algorithm to achieve efficient and accurate multi-objective optimization, which not only improves the prediction accuracy of height changes, but also significantly optimizes the comprehensive performance of the filling scheme.

[0037] Specifically, Figure 1 As shown, the dummy variable filling optimization method based on the neural network CMP model provided in this embodiment may include the following operations:

[0038] S1, layout discretization: Divide the chip layout into multiple windows, and use each window as the basic filling unit for filling optimization;

[0039] S2, fillable area extraction: extract the fillable area of ​​each window according to the design rules;

[0040] S3, target height assignment: assigning the target height of each window based on the predicted height of the neural network CMP model and the design rules;

[0041] S4, multi-objective genetic algorithm filling: Based on the fillable area and target height, a multi-objective genetic algorithm is used to generate and optimize the filling parameters.

[0042] S1, layout discretization: Divide the chip layout into multiple windows, each window as the basic unit of filling optimization;

[0043] S2, fillable area extraction: extract the fillable area of ​​each window according to the design rules to ensure that the generated filling solution meets the manufacturing constraints;

[0044] S3, target height assignment: assigning the target height of each window based on the prediction results of the neural network CMP model and the design rules;

[0045] S4, multi-objective genetic algorithm filling: Optimize the filling layout based on a multi-objective genetic algorithm, taking into account surface flatness, capacitive coupling and filling amount control.

[0046] The technical solution provided by the above embodiment of the present invention is further described in detail below in conjunction with the preferred implementation manner.

[0047] In some preferred implementations, the above S1, dividing the chip layout into multiple windows, and using each window as a basic filling unit for filling optimization, may further include the following operations:

[0048] In the layout discretization step, the chip layout is divided into multiple windows of uniform size, each of which is used as a basic unit for filling optimization. The window division mechanism includes but is not limited to:

[0049] Fixed window division: The layout is divided into windows of fixed size. The window size is pre-set by the design rules or process requirements. This division is suitable for scenarios with regular layout and relatively uniform density distribution (i.e., the standard deviation of the metal line width and spacing in the layout is less than the set threshold a and / or the standard deviation of the metal density in the local area of ​​the chip is less than the set threshold b).

[0050] Floating window partitioning: Dynamically adjust the window size according to the metal density and design complexity of the local area. This partitioning is suitable for situations with large density changes or irregular layouts (i.e., the standard deviation of the metal line width and spacing in the layout is greater than the set threshold a and / or the standard deviation of the metal density in the local area of ​​the chip is greater than the set threshold b), which helps to improve the accuracy of local optimization.

[0051] Adaptive window partitioning: Dynamically generate window partitioning schemes based on historical optimization results to achieve a balance between global optimization goals and local characteristics (i.e., scenarios where the metal line width and spacing in the layout are greater than the set threshold c and the gradient of the metal density in the local area of ​​the chip is greater than the set threshold d).

[0052] In some preferred implementations, the above S2, extracting the fillable area of ​​each window according to the design rule, may further include the following operations:

[0053] Fillable area extraction means extracting the fillable area of ​​each window according to the design rules after the window division is completed to ensure that the generated filling solution meets the manufacturing constraints. The extraction methods include but are not limited to:

[0054] Minimum spacing-based extraction: Ensures that the distance between the fill area and the existing metal lines meets the minimum spacing requirements to avoid manufacturing rule violations.

[0055] Density-constrained extraction: Limit the metal density in the filling area to within the range specified by the design rules, and meet the requirements of local flatness and density uniformity by adjusting the filling area ratio in the window.

[0056] Shape-constrained extraction: The shape of the fill area (such as a rectangle or staggered shape) is restricted according to the requirements of a specific process (such as lithography or etching).

[0057] In some preferred implementations, the above S1 and S2 may further include the following operations:

[0058] The steps of layout discretization and fillable area extraction can be extended to the discretization and extraction of multi-layer layouts to meet the optimization requirements of multi-layer metal interconnect structures. The steps are as follows:

[0059] Independent extraction layer by layer: Window division and filling area extraction are performed on each metal layer separately;

[0060] Cross-layer joint extraction: Based on the coupling effect between multiple layers, the filling areas between different layers are jointly extracted.

[0061] In some preferred implementations, the above S3, based on the fillable area and the target height, generates and optimizes the filling parameters using a multi-objective genetic algorithm, and may further include the following operations:

[0062] The target height assignment step aims to combine the prediction results of the neural network CMP model with the design rules to assign a target height to each window in the layout, thereby providing a clear optimization target for subsequent fill optimization. The process includes the following specific steps:

[0063] Initial height extraction: The layout is predicted through the neural network CMP model to obtain the predicted initial surface height h of each window. initial .

[0064] Target height calculation: Combine the design rules and predicted height results to calculate the target height h of each window target .

[0065] In some preferred embodiments, the neural network CMP model construction refers to building a CMP model based on a convolutional neural network (CNN) to accurately predict the impact of metal density distribution on surface height changes in the layout, and provide height prediction support for filling optimization. The following steps are included: model input and feature extraction, network architecture design, objective function design, training and optimization, model verification, and model deployment.

[0066] Furthermore, in some preferred embodiments, the input features of the neural network CMP model include but are not limited to: layout image, layout parameters, process parameters, multi-layer coupling characteristics, spatial position encoding, and multi-input joint features.

[0067] in:

[0068] The layout image represents a two-dimensional image matrix of the chip layout, which is used to capture the spatial characteristics of the layout and can reflect the distribution shape and complexity of the metal lines.

[0069] Layout parameters include distribution characteristics such as metal line width, spacing, hierarchy, and fill area. These parameters are extracted by parsing the layout file and can more accurately describe the geometric characteristics of the layout.

[0070] Process parameters include process conditions involved in the chemical mechanical polishing process, such as polishing speed, pressure, slurry concentration, etc. These parameters provide external influencing factors of surface height variation and provide support for the generalization ability of the model.

[0071] Multilayer coupling characteristics, for multilayer metal layout, the coupling density characteristics between layers are added as input to optimize the flatness prediction of multilayer structures.

[0072] Spatial position encoding: Encode the layout window position to provide the model with the global position information of the window.

[0073] Multi-input joint features: The model supports the joint processing of multiple input data. It extracts the features of layout images, layout parameters, and process parameters through a multi-branch neural network structure, and fuses them in the fully connected layer to generate comprehensive layout prediction results.

[0074] In some preferred implementations, the above input features may further include: data standardization and feature extraction, the methods of which include but are not limited to:

[0075] Normalization: Map the input data to a fixed interval (such as [0,1]) to eliminate the scale differences between different data sources.

[0076] Principal component analysis: Dimensionality reduction of multi-dimensional input data, retaining the main information to reduce the computational complexity of the model;

[0077] Feature enhancement: Enhance the input data by calculating second-order or higher-order statistical features (such as mean, variance, skewness) to improve the model's ability to recognize complex layout characteristics.

[0078] In some preferred implementations, the above calculation of the target height of each window may further include the following operations:

[0079] The target height calculation method can flexibly adapt to different layout and process requirements, ensuring that the target height calculation result meets both design rules and actual manufacturing constraints, providing efficient support for filling optimization. The methods include but are not limited to:

[0080] Based on minimum density and rules: the target height is limited to the minimum height h of each layer min and empirical height h e The smaller value to meet density distribution and manufacturing requirements:

[0081] h target =min(h min ,h e )

[0082] Method based on multi-objective optimization: The target height is taken as part of the multi-objective optimization problem and calculated by dynamically adjusting the target weights:

[0083] h target = argmin(w d *|h pred -h target |+w c *C coupling +w f *F fill )

[0084] In the formula, is the adjusted target height, w d 、w c 、w f are the target weights of density, coupling capacitance and filling amount, respectively, and h pred is the predicted target height, C coupling is the coupling capacitor, F fill is the filling amount;

[0085] Method based on dynamic adjustment: The target height is dynamically adjusted based on real-time feedback from local optimization:

[0086]

[0087] Among them, α is the adjustment coefficient, and Δh is the height adjustment value fed back during the optimization process.

[0088] Combining design rules and empirical formulas, the target height can be directly calculated:

[0089] h target =h baseline +β*(ρ-ρ target )

[0090] Among them, h baseline is the target height reference value, β is the adjustment factor, ρ and ρ target They are the current area density and target density respectively.

[0091] In some preferred implementations, the above S4, multi-objective genetic algorithm optimization, may further include the following operations:

[0092] The population is evolved through a multi-objective genetic algorithm. Each individual in the population represents a possible filling solution. The characteristics of the individual include multiple parameters related to the filling, such as the filling geometry, position, density, direction, etc.

[0093] Among them, the individual is the smallest unit in the genetic algorithm, and each individual represents a complete population configuration, whose characteristics include but are not limited to the following parameters:

[0094] Fill Location: Fill is placed at a specific location (such as coordinates or area) on the chip.

[0095] Filled Shape: A filled geometric shape, such as a square, rectangle, or other custom shape.

[0096] Fill size: the width, length and other dimensions of the filling unit.

[0097] Fill direction: The direction or angle of the fill.

[0098] Fill spacing: the distance between different filling units.

[0099] A population is a collection of multiple individuals, each of which represents a different filling scheme. The size of the population directly affects the breadth of the search space and the diversity of solutions. During the evolution of the genetic algorithm, the population is continuously optimized through operations such as selection, crossover, and mutation to find the best filling configuration. The purpose of population evolution is to optimize the filling scheme so that the filling parameters can minimize the defects that may occur during the chemical mechanical polishing process while ensuring the chip manufacturing yield.

[0100] During the optimization process, the genetic algorithm evaluates each individual through multiple objective functions. The objective functions for evaluating the quality of individuals include but are not limited to the following categories:

[0101] Error minimization goal: This goal aims to reduce various errors caused by filling, including but not limited to:

[0102] Filling density deviation: Optimize the filling position and density to reduce the uneven filling density between areas.

[0103] Filling height deviation: Reduce the inconsistency of surface height after filling and ensure surface flatness.

[0104] Manufacturing cost minimization goal: This goal focuses on cost control of the filling process, including but not limited to:

[0105] Reduce the amount of filling: Reduce the use of filling materials through a reasonable filling plan.

[0106] Reduce filling time: Optimize the filling layout and process to reduce the time required for filling operations.

[0107] Process yield maximization goal: This goal aims to improve the process yield, including but not limited to:

[0108] Reduce defects after chemical mechanical polishing: Reduce defects during the CMP process caused by unreasonable filling design by optimizing the filling configuration.

[0109] Comprehensive goal: Taking into account multiple goals such as error, cost and yield, the genetic algorithm compromises multiple goals through strategies such as weighted sum method to find the best balance solution.

[0110] The filling parameters are the basic information input during the optimization process, including the basic characteristics of the filling (position, shape, size, spacing, etc.). The filling scheme is a specific filling layout generated based on the given filling parameters. The filling scheme defines the distribution, shape and amount of the filling. The filling result is the best filling scheme obtained based on the optimization algorithm. The filling parameters are the basis of the filling scheme and determine the basic characteristics of the filling. The filling scheme is generated by setting the filling parameters, and the filling result is based on the best filling scheme obtained by the optimization algorithm.

[0111] Specifically, the optimization process of the optimization method includes:

[0112] Population initialization: Generate an initial population with random filling parameters. The population is composed of multiple individuals, each of which represents a different filling parameter configuration.

[0113] Filling solution generation: Based on the fillable area, a corresponding initial filling solution is generated.

[0114] Filling result generation: Based on the filling parameters, the initial filling scheme and the target height, the corresponding filling result is generated, and the filling result includes: filling density, filling height, number of filling units and capacitive coupling.

[0115] Fitness evaluation: Based on the filling results, the fitness score of the filling parameters is calculated using the fitness function.

[0116] Genetic operation: Based on the fitness score, the next generation of filling parameters are generated through selection, crossover and mutation to improve the diversity of solutions and global exploration capabilities.

[0117] Termination condition: Stop the optimization when the number of iterations, fitness convergence and other conditions are met, and complete the optimization of the filling parameters.

[0118] In some preferred implementations, the above population initialization may further include the following operations:

[0119] Random initialization: Randomly generate padding parameters, including padding width, padding length, and padding spacing, and randomly distribute dummy padding within the layout.

[0120] Density-driven initialization: Based on the initial density of each window in the layout, a filling scheme is generated preferentially in the low-density area, and the filling parameters are dynamically adjusted according to the density gap (the difference from the target density).

[0121] Initialization based on layout characteristics: Generate populations based on key characteristics of the layout (such as wiring distribution and metal line density gradient) to avoid excessive filling in densely wired areas to reduce the impact of lateral capacitance.

[0122] Initialization based on rule templates: Use predefined templates (such as rectangles, staggered, distribution characteristic functions, etc.) to generate filling parameters, and the filling parameters (such as spacing, width) are randomly sampled within a reasonable range.

[0123] Mixed initialization: Combine the above methods to generate a population in a mixed ratio. For example, some individuals are randomly generated, and some are density-driven or template-based.

[0124] In some preferred implementations, the above fitness evaluation may further include the following operations:

[0125] Fitness evaluation needs to generate corresponding filling results according to the filling parameters of each individual, and calculate the fitness function value based on this to guide subsequent genetic operations. According to the filling parameters of each individual (such as filling width, length, spacing, etc.), a complete layout filling result is generated. The selection and generation process of the filling mode includes but is not limited to the following common methods:

[0126] Rectangular filling: Based on a regular rectangle, the cells are evenly distributed. Its parameters include filling width, filling length, cell spacing, etc. It is suitable for simple layouts that require fast and even distribution.

[0127] Staggered filling: Arrange the filling cells in a staggered pattern to reduce the parallel edges between the filling cells to optimize the capacitance. Its parameters include staggered spacing, row and column offset, etc., which is suitable for scenarios where lateral capacitance needs to be optimized.

[0128] Filling based on distribution characteristic function: Generates the position and size of filling cells according to a specific distribution characteristic function, which is suitable for filling control that requires high flexibility and precision.

[0129] Mixed filling: Combine rectangular filling with staggered filling, or combine different DCFs to optimize local areas of the layout.

[0130] Random filling: Generates filling units according to random distribution, which is suitable for exploring unknown layout characteristics or improving filling diversity.

[0131] In some preferred embodiments, the multi-objective genetic algorithm aims to optimize the filling scheme, and the core design of the fitness function includes but is not limited to the following optimization objectives: Overall Variation (OV), Fill Amount (FA) and Lateral Capacitance (LC). By constructing a multi-objective fitness function, these objectives are comprehensively balanced to guide the genetic algorithm to gradually converge to the global optimal solution. The general expression of the fitness function is:

[0132] FitnessScore=w1*S OV +w2*S FA +w3*S LC +w4*S Others

[0133] In the formula, FitnessScore represents the fitness score, S OV , S FA , S LC , S others They represent the fitness scores of height change, filling amount, lateral capacitance and other factors (density, density gradient, etc.), respectively. w1, w2, w3, and w4 are weight parameters used to dynamically adjust the importance of different goals.

[0134] In some preferred embodiments, the height-varying fitness score is designed to penalize the target height h target and predicted height h pred, to reduce the overall height variation of the layout. In addition to the commonly used quadratic function form, other forms of functions can also be set to adapt to different optimization requirements and constraints. Including but not limited to the following definitions and applicable scenarios of various function forms:

[0135] Quadratic function form: The quadratic function is the basic form of the highly variable score. It can significantly punish large deviations while having good resolution for small deviations. It is suitable for scenarios with relatively uniform deviation distribution and emphasizes smoothness.

[0136] S oV =C1+C2*(h target -h pred ) 2

[0137] In the formula, C1 and C2 are constant terms of the fraction.

[0138] Absolute value function form: The absolute value function is used to measure the deviation between the target value and the predicted value. It penalizes large deviations more significantly, but has a weaker penal effect on small deviations. It is suitable for complex layouts that require high flatness and have a large local adjustment range. It can effectively capture the global error after the filling parameter is adjusted.

[0139] S OV =C1+C2*|h target -h pred |

[0140] Exponential function form: The exponential function is used to significantly penalize large deviations while having a weak effect on small deviations. It is suitable for complex layouts that require a high degree of flatness and a large local adjustment range.

[0141]

[0142] Logarithmic function form: The logarithmic function slows down the penalty growth rate for large deviations. It is more suitable for scenarios that are sensitive to small deviations but have a certain tolerance for large deviations. It is suitable for scenarios that need to quickly adjust small deviations in the early stage of optimization and fine-tune large deviations in the later stage.

[0143] S OV =C1+C2*log(1+|h target -h pred |)

[0144] Piecewise linear function form: Different penalty strengths are defined in segments according to the deviation range, which is suitable for scenarios where stricter constraints need to be imposed on large deviation areas.

[0145] In order to achieve more accurate fitness score calculation for height changes, in some preferred embodiments, the methods for predicting height based on the neural network CMP model include but are not limited to the following:

[0146] Method based on physical model: Use traditional physical formulas or numerical methods to establish a physical model of the CMP process and calculate the height distribution after filling. This method has certain theoretical explanations, but the calculation complexity is relatively high.

[0147] The neural network model-based method uses a machine learning model to learn the complex relationship between layout features and height distribution using training data. This method can quickly predict the height distribution in large-scale layouts with high accuracy and efficiency.

[0148] Hybrid model-based method: Combining the advantages of physical model and neural network model, further improving the prediction accuracy through hybrid prediction.

[0149] In some preferred embodiments, the filling amount fitness score is intended to reduce redundant filling, control the filling area, and avoid unnecessary process complexity and cost increase. In addition to the commonly used quadratic function form, other forms of functions can also be set according to different optimization requirements and constraints to improve flexibility and applicability. The following are the definitions and applicable scenarios of various function forms:

[0150] Quadratic function form: The infill fitness score is proportional to the square of the number of infill units. It is suitable for scenarios where excessive infill needs to be strongly punished, such as when the process has strict restrictions on infill density.

[0151] S FA =C1*(#Fill)

[0152] Where C1 refers to the constant parameter and #Fill refers to the number of fills.

[0153] Absolute value function form: Based on the filling amount and the target filling amount #Fill target It is suitable for scenarios where there is a target value requirement for the filling amount, such as the need to achieve precise filling ratio or uniformity.

[0154] S FA =C1*|#Fill-#Fill target |

[0155] Exponential function form: the number of filled cells exceeds the threshold #Fill threshold After that, the fitness score increases exponentially. It is suitable for scenarios where the amount of filling needs to be strongly restricted, such as in high-precision manufacturing processes, where excessive filling can cause serious performance degradation.

[0156]

[0157] Where k is a constant parameter that controls the growth rate.

[0158] Logarithmic function form: The fitness score grows with the logarithm of the number of infill units. It is suitable for scenarios with low sensitivity to infill amount, such as optimization objectives that allow infill amount fluctuations within a certain range but require smoothness.

[0159] S FA =C1*log(1+#Fill)

[0160] Piecewise linear function form: Set the piecewise penalty based on whether the filling amount is within the allowable range. It is suitable for scenarios where the filling amount needs to be controlled within the upper and lower limits, such as design rule constraints with strict density rules.

[0161] In some preferred embodiments, the fitness score of the lateral capacitor is intended to optimize the coupling effect between adjacent metal lines in the fill layout, reduce signal interference and improve electrical performance. The design methods of the fitness score include but are not limited to the following:

[0162] Quadratic function form:

[0163] S LC =C1+C2*(LC-LC target ) 2

[0164] Where LC refers to the lateral capacitance after filling, LC target C1 is the target capacitance value, which is set based on design requirements and process constraints. C1 and C2 are constant parameters.

[0165] Absolute value function form:

[0166] S LC =C1*|LC-LC target |

[0167] Exponential function form:

[0168]

[0169] Where k is a constant parameter that controls the growth rate.

[0170] Logarithmic function form:

[0171] S LC =C1*log(1+LC)

[0172] Piecewise linear function form: Set the piecewise penalty based on whether the filling amount is within the allowable range. It is suitable for scenarios where the filling amount needs to be controlled within the upper and lower limits, such as design rule constraints with strict density rules.

[0173] In some preferred embodiments, in order to accurately evaluate the lateral capacitance of the filled and original layouts, it is necessary to extract the capacitance. The extraction methods include but are not limited to the following:

[0174] Capacitance extraction based on analytical models uses physical formulas to directly calculate the lateral capacitance between metal lines. This method is suitable for regular layouts and has fast calculation speed, but there may be errors in complex layouts:

[0175]

[0176] In the formula, ∈ r is the dielectric constant, A is the overlapping area between metal lines, and d is the metal line spacing.

[0177] Capacitance extraction based on finite element analysis discretizes the layout through finite element analysis, and uses finite element solving tools (such as ANSYS and COMSOL) to accurately simulate the electric field distribution and capacitance values ​​between metal wires. This method has high accuracy and is suitable for complex layouts, but the calculation time is long.

[0178] Capacitance extraction based on circuit extraction tools uses the capacitance extraction module of commercial EDA tools (such as Cadence and Mentor Graphics) to automatically complete the analysis and extraction of layout capacitance, and can quickly extract lateral capacitance including filling cells.

[0179] Capacitance prediction based on neural network model directly predicts the lateral capacitance of the layout by training the neural network model. While ensuring high efficiency, it can handle complex layouts and has strong generalization ability.

[0180] Based on the same inventive concept, an embodiment of the present invention further provides a dummy element filling optimization system based on a neural network CMP model.

[0181] Specifically, Figure 2 As shown, the dummy element filling optimization system based on the neural network CMP model provided in this embodiment may include the following modules:

[0182] Layout discretization module, which is used to divide the chip layout into multiple windows and use each window as a basic filling unit for filling optimization;

[0183] A fillable area extraction module, which extracts the fillable area of ​​each window according to the design rules;

[0184] A target height assignment module that assigns a target height to each window based on the predicted height of the neural network CMP model and design rules;

[0185] The filling optimization module generates and optimizes the filling parameters based on the fillable area and the target height using a multi-objective genetic algorithm.

[0186] The dummy element filling optimization system based on the neural network CMP model provided in this embodiment aims to realize a comprehensive system for metal filling optimization, combining the neural network CMP model and the multi-objective optimization algorithm to realize the global optimization design of the filling layout, and significantly improve the surface flatness, capacitive coupling effect and filling density control performance.

[0187] The working contents of each functional module of the system provided by the above embodiment of the present invention are further described in detail below in conjunction with the preferred implementation manner.

[0188] The system provided by the above embodiment of the present invention mainly includes the following modules: a layout discretization module, a fillable area extraction module, a target height allocation module, and a filling optimization module. Among them:

[0189] The layout discretization module parses the input layout file (such as GDSII or OASIS format) and discretizes the layout into multiple basic unit windows. The fillable area extraction module extracts the fillable area to ensure that the filling design meets the manufacturing constraints.

[0190] Its technical implementation includes but is not limited to:

[0191] 1. Fixed window division: Divide the layout into a regular grid based on design rules (such as window size, metal density restrictions).

[0192] 2. Dynamic window partitioning: Dynamically adjust the window size based on local metal density and layout complexity.

[0193] 3. Adaptive partitioning mechanism: adjust the window partitioning strategy based on real-time feedback to improve the optimization capability of complex areas.

[0194] 4. Hardware support: Multi-threaded processors are used to perform partitioning and extraction tasks in parallel, significantly improving the processing speed of large-scale layouts.

[0195] The target height allocation module uses the convolutional neural network (CNN) model to predict the impact of layout density and process parameters on surface height changes, providing a basis for target height allocation. Its technical implementation includes but is not limited to:

[0196] Input features: layout image, layout parameters, process parameters, cross-layer coupling characteristics and spatial position encoding, etc.

[0197] Network architecture: regression prediction model based on neural network, generation model based on neural network, etc.

[0198] Training and deployment: Use TensorFlow or PyTorch framework to train the model, and deploy the model on a high-performance graphics processor GPU (such as NVIDIA A100) or edge device to meet real-time prediction needs.

[0199] The target height assignment module assigns a target height to each window based on the neural network prediction results and design rules, ensuring the flatness and manufacturing constraints of the layout.

[0200] The filling optimization module uses a multi-objective genetic algorithm to optimize filling parameters (such as width, length, spacing) and generate a globally optimal filling layout. It includes:

[0201] Genetic Algorithm Design:

[0202] Population initialization: Combining random generation, density-driven and template-driven methods.

[0203] Genetic operations: Use mechanisms such as selection, crossover, and mutation to generate the next generation of solutions.

[0204] Fitness function: Combines height change, filling volume, and lateral capacitance to dynamically adjust the optimization target.

[0205] Its computing architecture supports:

[0206] Supports parallel optimization and fast genetic algorithm iteration through multi-core CPU or distributed GPU cluster.

[0207] Supports real-time optimization and readjusts target weights when process parameters change to achieve dynamic optimization.

[0208] Filling optimization module generates metal filling solutions that meet design rules based on the output of the optimization engine. Supports:

[0209] Multiple filling modes: supports multiple shapes (such as rectangular, staggered, and distributed characteristic function modes), and can optimize the filling shape and density according to different process requirements.

[0210] Multi-layer joint filling: Supports layer-by-layer filling: optimizes each metal layer separately, and supports cross-layer filling: combines multi-layer density and capacitive coupling relationship for overall optimization.

[0211] Output fill files that comply with industry standards such as GDSII.

[0212] The filling optimization module can also evaluate the performance of the filling results, including flatness, electrical performance and process compliance, and provide feedback on optimization suggestions. Support:

[0213] Height Variation Analysis: Calculate the overall standard deviation (OV) and maximum deviation (MV).

[0214] Lateral capacitance analysis: Use EDA tools or neural networks to predict lateral capacitance values.

[0215] Adjust optimization weights to dynamically optimize next-generation filling schemes.

[0216] Update the neural network model to improve prediction accuracy.

[0217] Furthermore, the system can also provide a data interface and storage module to provide data input, output and storage functions and support cross-platform operations.

[0218] Data input: Supports standard layout files (such as GDSII).

[0219] Data output: Generate optimized population files and provide reports.

[0220] Storage: Layout management based on cloud storage supports archiving of large-scale layouts and historical optimization solutions.

[0221] In summary, the system provided by the above-mentioned embodiment of the present invention realizes an efficient, accurate and flexible technical solution in metal filling optimization, which can significantly improve the flatness control and process reliability in integrated circuit manufacturing.

[0222] The technical solution provided by the above embodiment of the present invention is further described in detail below in conjunction with a specific application example.

[0223] Figure 3 This is a workflow diagram of the dummy variable filling optimization method based on the neural network CMP model involved in this specific application example.

[0224] like Figure 3 As shown, the dummy variable filling optimization method based on the neural network CMP model includes the following steps:

[0225] Step S100, layout discretization and fillable area extraction, further includes the following steps:

[0226] S101, layout discretization: Divide the entire layout into window areas of uniform size. The size of each window can be pre-set according to process requirements (such as 20μm×20μm).

[0227] S102, extracting fillable areas: extracting fillable areas that meet design rules in each window, specifically including:

[0228] S1021, shrink the metal line spacing according to the minimum spacing requirements to ensure compliance with design rules;

[0229] S1022, eliminating the filling area whose filling width or length is smaller than the minimum size;

[0230] S1023, merging the filling areas that meet the rules to generate a final fillable layout.

[0231] Step S200, target height allocation, further includes the following steps:

[0232] S201, initial height prediction: Use the neural network CMP model to predict the initial height h of each window initial Make predictions.

[0233] S202, target height calculation: Combine the design rules and the predicted height to calculate the target height h target Specific methods include but are not limited to:

[0234] S2021, based on minimum density and rules:

[0235] h target =min(h min ,h e )

[0236] S2022, based on multi-objective optimization method:

[0237] h target = argmin(w d *|h pred -h target |+w c *C coupling +w f *F fill )

[0238] S2023, based on a dynamic adjustment approach:

[0239]

[0240] Among them, α is the adjustment coefficient, and Δh is the height adjustment value fed back during the optimization process.

[0241] S2024, combined with design rules and empirical formulas:

[0242] h target =h baseline +β*(ρ-ρ target )

[0243] Among them, β is the adjustment factor, ρ and ρ target They are the current area density and target density respectively.

[0244] Step S300, filling optimization based on multi-objective genetic algorithm, such as Figure 4 As shown, further comprising the steps of:

[0245] S301, initializing the population, specifically including:

[0246] S3011, randomly initialize the initial population of the multi-objective genetic algorithm, each individual contains parameters such as filling width, length, spacing, etc.;

[0247] S3012: Generate a corresponding initial filling solution according to the filling area of ​​the layout.

[0248] In S301, the method of initializing the population includes:

[0249] S301.1, Uniform Random Initialization: Randomly generate padding width, length, and spacing within a predefined range;

[0250] S301.2, rule constraint initialization: Generate fill width, length and spacing in combination with density rules to ensure that each initial fill density complies with the design rules;

[0251] S301.3, Experience-driven initialization: Use the historical best population solution as the basis for the initial population.

[0252] S302, calculating the individual fitness value, specifically including:

[0253] S3021, based on the filling parameters of each individual, a corresponding filling result is generated, and the filling mode includes but is not limited to:

[0254] S30211, Rectangular Filling: Generate uniformly distributed rectangular cells in the filling area;

[0255] S30212, staggered filling: generating staggered filling units according to row and column offsets;

[0256] S30213, Distributed Characteristic Filling: Dynamically adjust the size and position of fill cells based on a specific distribution function (such as linear or nonlinear).

[0257] S3022, calculating the fitness function value according to the filling result, the fitness includes:

[0258] S30221, Density Adaptation: Measures whether the filling density meets the target density;

[0259] S30222, height fitness: score calculated based on the deviation between the height distribution predicted by the CMP model and the target height;

[0260] S30223, Filling Amount Fitness: Evaluate whether the filling amount meets the design constraints without overfilling;

[0261] S30223, Coupling Capacitor Fitness: Evaluate the filling scheme based on the optimization goal of the filling capacitor.

[0262] Figure 5 Figure 1 is a schematic diagram of different filling modes. Figure 5 As shown, in S302, the generation method of the filling pattern in calculating the individual fitness value includes:

[0263] S302.1, rectangular fill pattern: the fill cells are arranged in a regular rectangular grid;

[0264] S302.2, staggered filling mode: each row or column of filling cells is offset by a fixed distance to reduce lateral capacitance;

[0265] S302.3, Distribution characteristic filling mode: Generate the position and size of the filling unit based on a specific distribution function.

[0266] S303, Genetic operation generates a new population, specifically including:

[0267] S3031, generate the next generation population through selection, crossover and mutation operations:

[0268] S30311, selection: using elite retention, roulette wheel or tournament selection, giving priority to individuals with higher fitness;

[0269] S30312, Crossover: Perform single-point crossover or multi-point crossover on the parent individuals to generate new individuals;

[0270] S30313, mutation: Randomly adjust individual parameters or dynamically adjust parameter values ​​using local optimization feedback.

[0271] In S303, the specific implementation of the genetic operation includes:

[0272] S303.1, elite retention strategy: directly add the individuals with the highest fitness to the next generation;

[0273] S303.2, single-point crossover: randomly select the cutting point and exchange some parameters of the parent individuals;

[0274] S303.3, Random Mutation: Apply random perturbations to padding parameters, such as adjusting width, length, or spacing;

[0275] S303.4, Feedback-driven variation: Dynamically adjust the individual height parameters based on the deviation between the filling result and the target height.

[0276] S304, updating the optimization plan and determining the termination conditions, specifically including:

[0277] S3041, regenerate the filling result and calculate the fitness value according to the new population generated by the genetic operation;

[0278] S3042, determine whether the optimization termination condition is met (such as reaching the maximum number of generations or fitness convergence), otherwise return to S200 to continue iteration.

[0279] In S304, the determination of the optimization termination condition includes:

[0280] S304.1, fitness convergence: when the average fitness change of the population is lower than the preset threshold, the iteration is terminated;

[0281] S304.2, maximum number of generations limit: when the number of iterations reaches the preset maximum number of generations, the optimization is terminated;

[0282] S304.3, local optimal escape mechanism: If trapped in a local optimal solution, escape from the local optimal solution by increasing the mutation rate or introducing new individuals.

[0283] like Figure 6 As shown, this is the effect diagram after dummy variable filling in this specific application example.

[0284] The dummy element filling optimization method and system based on the neural network CMP model provided in the above embodiment of the present invention generates a dummy element filling scheme that meets the density rules, flatness requirements and capacitance control through the optimization design of the multi-objective genetic algorithm. Combining multiple filling modes and dynamically adaptive genetic operations, the present invention effectively reduces the height change after CMP, reduces the filling amount and optimizes the capacitance characteristics, significantly improving the efficiency and quality of filling optimization.

[0285] All matters not covered in the above embodiments of the present invention are well known in the art.

[0286] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art may make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A dummy variable filling optimization method based on a neural network CMP model, characterized in that: include: Divide the chip layout into multiple windows, and use each window as a basic filling unit for filling optimization; Extract the fillable area of ​​each window according to the design rules; Assign a target height to each window based on the predicted height of the neural network CMP model and design rules; Based on the fillable area and the target height, a filling parameter is generated and optimized using a multi-objective genetic algorithm.

2. The dummy variable filling optimization method based on the neural network CMP model according to claim 1 is characterized in that: The chip layout is divided into multiple windows, including any one or more of the following methods: For scenarios where the standard deviation of metal line width and spacing in the layout is less than a set threshold a and / or the standard deviation of metal density in a local area of ​​the chip is less than a set threshold b, the windows are divided into fixed-size windows according to a preset window size; For scenarios where the standard deviation of metal line width and spacing in the layout is greater than a set threshold a and / or the standard deviation of metal density in a local area of ​​the chip is greater than a set threshold b, the size of the partition window is dynamically adjusted according to the metal density or design complexity of the local area; For scenarios where the metal line width and spacing in the layout are greater than the set threshold c and the gradient of the metal density in the local area of ​​the chip is greater than the set threshold d, a window partitioning scheme is dynamically generated based on the local density and parameter optimization trend.

3. The dummy variable filling optimization method based on the neural network CMP model according to claim 1 is characterized in that: The extracting of the fillable area of ​​each window according to the design rules includes any one or more of the following methods: Based on the design rule of minimum spacing, the fillable area is extracted to ensure that the distance between the fill area and the existing metal line meets the minimum spacing design requirement; Based on the design rules of density constraints, the filling area ratio in the window is adjusted and the fillable area is extracted to limit the metal density in the filling area to within the specified design range; Based on the design rules of shape constraints, the fillable area is extracted to limit the shape of the fill area according to specific process requirements.

4. The dummy variable filling optimization method based on the neural network CMP model according to claim 1 is characterized in that: Through independent extraction layer by layer or cross-layer joint extraction, each metal layer is individually windowed and filled area extracted, or, based on the coupling effect between multiple layers, the filling areas between different layers are jointly extracted.

5. The dummy variable filling optimization method based on the neural network CMP model according to claim 1 is characterized in that: The prediction results and design rules based on the neural network CMP model are used to assign a target height to each window. The neural network CMP model is used to predict the initial surface height of each window; Combining the design rules with the predicted initial surface height, the target height for each window is calculated.

6. The dummy variable filling optimization method based on the neural network CMP model according to claim 5 is characterized in that: Also includes any one or more of the following: - The neural network CMP model adopts a convolutional neural network, whose input features include: layout image, layout parameters, process parameters, multi-layer coupling characteristics and / or spatial position encoding; - The calculation of the target height of each window by combining the design rules and the predicted initial surface height includes any one or more of the following methods: Based on the design rule of minimum density, the target height is limited to the minimum height h of each floor. min and empirical height h e A smaller value of to meet the density distribution requirements: h target =min(h min ,h e ) In the formula, h target is the target height; Based on the multi-objective optimization method, the target height is taken as part of the multi-objective optimization problem and calculated by dynamically adjusting the target weight: In the formula, is the adjusted target height, w d 、w c 、w f are the target weights of density, coupling capacitance and filling amount, respectively, and h pred is the predicted target height, C coupling is the coupling capacitor, F fill is the filling amount; Based on the dynamic adjustment method, the target height is dynamically adjusted according to the real-time feedback of local optimization: In the formula, α is the adjustment coefficient, and Δh is the height adjustment value fed back during the optimization process; The target height is calculated by density difference based on a combination of design rules and empirical formulas: h target =h baseline +β*(ρ-ρ target ) In the formula, h baseline is the target height reference value, β is the adjustment factor, ρ and ρ target They are the current area density and target density respectively.

7. The dummy variable filling optimization method based on the neural network CMP model according to claim 1 is characterized in that: The method of generating and optimizing filling parameters using a multi-objective genetic algorithm includes: First, randomly generate or initialize the padding parameters based on layout characteristics; Based on the fillable area, generating a corresponding initial filling solution; Based on the filling parameters, the initial filling scheme and the target height, a corresponding filling result is generated, wherein the filling result includes: filling density, filling height, number of filling units and capacitive coupling; Calculating the fitness value of the filling parameter according to the filling result; Based on the fitness value, generating next generation population parameters by selection, crossover and / or mutation; The optimization is stopped according to the number of iterations or the fitness convergence condition, and the optimization of the filling parameters is completed.

8. The dummy variable filling optimization method based on the neural network CMP model according to claim 7 is characterized in that: Also includes any one or more of the following: - The random generation or initialization of the filling parameters based on the layout characteristics may include any one or more of the following methods: Randomly generate filling parameters and randomly distribute dummy filling in the layout; Based on density drive, filling parameters are generated preferentially in low-density areas, and the filling parameters are dynamically adjusted according to the density gap; Generate filling parameters based on key features in the layout; Generate filling parameters based on a predefined rule template, where the filling parameters are randomly sampled within a set range; - Based on the fillable area, generating a corresponding initial filling solution, including: Select the fill mode of the fill unit according to the fillable area, including any one or more of the following: Rectangular fill mode, which arranges the fill cells in a regular grid; Staggered filling mode generates staggered filling cells according to row and column offsets; Distribution characteristic filling, generating filling unit position and size based on a specific distribution function; Mixed filling, combining rectangular filling and staggered filling to optimize the filling unit layout; Random filling, randomly generate filling units; -the step of calculating the fitness value of the filling parameter according to the filling result comprises: Calculate density fitness to measure whether the filling density meets the target density; Calculate the altitude change fitness and obtain the deviation between the target altitude and the predicted altitude; Calculate the filling amount fitness and control the number of filling units; Calculate lateral capacitance adaptability and optimize the coupling effect between adjacent metal lines in the layout; in: Fitness Score=w1*S OV +w2*S FA +w3*S LC +w4*S Others In the formula, Fitness Score represents the fitness score, S OV , S FA , S LC , S Others They represent the fitness scores of height change, filling amount, lateral capacitance and density respectively, and w1, w2, w3 and w4 are the corresponding weight parameters, which are used to dynamically adjust the importance of different goals; - a fitness function for calculating the fitness value, including: a quadratic function, an absolute value function, an exponential function, a logarithmic function and / or a piecewise linear function; -The randomly generated filling parameters include any one or more of the following methods: Uniform random initialization method randomly generates padding width, length and spacing within a predefined range; The rule-constrained initialization method combines density rules to generate fill width, length, and spacing to ensure that the initial fill density meets the design rules; Experience-driven initialization method, using the best historical filling scheme as the basis for the initial population; - Based on the fitness value, generating next generation population parameters by selection, crossover and / or mutation, including: Adopting the elite retention strategy, directly adding the individuals with the highest fitness to the next generation; Use single-point crossover, randomly select cutting points, and exchange some parameters of parent individuals; Using random mutation, random perturbations are imposed on the filling parameters; Adopt feedback-driven variation to dynamically adjust the individual height parameters according to the deviation between the filling result and the target height; -Stop the optimization according to the number of iterations or fitness convergence condition, including: When the average fitness change of the population is lower than the preset threshold, the iteration is terminated; When the number of iterations reaches the preset maximum number of generations, the optimization is terminated; When falling into a local optimal solution, the local optimal solution can be jumped out by increasing the mutation rate or introducing new individuals, and the iterative process can be re-executed.

9. A dummy filling optimization system based on a neural network CMP model, characterized in that: include: Layout discretization module, which is used to divide the chip layout into multiple windows and use each window as a basic filling unit for filling optimization; A fillable area extraction module, which extracts the fillable area of ​​each window according to the design rules; A target height assignment module that assigns a target height to each window based on the predicted height of the neural network CMP model and design rules; A filling optimization module generates and optimizes filling parameters based on the fillable area and the target height using a multi-objective genetic algorithm.

10. A computer terminal comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it can be used to perform the method described in any one of claims 1 to 8, or to run the system described in claim 9.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it can be used to perform the method described in any one of claims 1 to 8, or to run the system described in claim 9.

Images