A hybrid constraint optimization method and system for regression model for semiconductor measurement
By constructing an enhanced objective function with constraints in regression analysis and using augmented Lagrangian multiplication method and frame projection method, the problem of unconstrained independent variables in semiconductor measurement is solved, and the physical significance of the results is achieved and accuracy improve the user experience and algorithm performance of measurement equipment.
Patent Information
- Application Number
- CN202411175450.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-08-26
AI Technical Summary
The existing regression analysis methods fail to effectively consider the constraints of independent variables in semiconductor measurements, resulting in the output results that may be meaningless or violate physical constraints, affecting subsequent calculations and data visualization.
The hybrid constraint optimization method of regression model is adopted to construct an enhanced objective function with constraints, combined with the augmented Lagrangian multiplication method and the frame projection method, the search points and the update constraint parameters are dynamically adjusted to ensure that the results meet the physical constraint conditions.
It effectively avoids the occurrence of numerical values without physical significance, improves the compatibility and generalization capabilities of the algorithm, ensures that the output results are within the acceptable range of users, and improves the user experience of semiconductor measurement equipment.
Smart Images

Figure CN118690341B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of semiconductor measurement technology, and in particular to a regression model hybrid constraint optimization method and system for semiconductor measurement. Background Art
[0002] Optical Critical Dimension (OCD) measurement technology and thin film thickness (ThinFilm) measurement technology are widely used in the field of semiconductor chip manufacturing measurement, and are of great significance to the quality monitoring, process optimization, cost control, and yield improvement of semiconductor products. The mathematical principle of measurement technology is to obtain the actual physical morphology and critical dimensions of semiconductor chips through a series of complex algorithms, that is, the combination of direct and inverse problems, without destroying the sample to be tested.
[0003] Regression analysis is a common inverse problem solving method in the field of semiconductor measurement. Its main idea is to obtain measured data from the measurement equipment, and infer the actual physical size of the semiconductor device by solving the appropriate regression analysis mathematical model. Common regression algorithms include the steepest descent method (or gradient descent method), Gauss-Newton method (Gauss-Newton or GN), Levenberg-Marquardt method (Levenberg-Marquardt or LM), and quasi-Newton methods such as BFGS. Among them, the LM method can be regarded as a combination of the steepest descent method and the GN method: when the distance to the optimal solution is far, the LM method tends to the steepest descent method and can quickly determine the descent direction; when gradually approaching the optimal solution, the LM method tends to the GN method and can quickly converge to the optimal solution.
[0004] However, the common regression algorithms mentioned above do not explicitly consider the constraints of the independent variables and are unconstrained optimization algorithms. In actual application scenarios, the dimensional parameters of semiconductor samples are often subject to strict physical constraints. For example, the width and thickness of each layer of material cannot be negative, the sum of the thickness of multiple layers of material cannot exceed a certain threshold, and each parameter to be measured must not exceed its corresponding upper and lower limits. If the various constraints on the independent variables are not properly handled, there may be a situation where "mathematically reasonable but physically meaningless" occurs, such as negative variables, exceeding the preset range, and the coupling relationship between variables cannot be satisfied. This will cause the output results to be meaningless or violate the user-defined constraints, thereby affecting subsequent optical information calculations, data visualization, physical model drawing and other operations. Summary of the invention
[0005] In view of some or all of the problems in the prior art, the present invention provides a regression model hybrid constraint optimization method for semiconductor measurement, which comprises the following steps:
[0006] Input initial information and constraints to construct an enhanced objective function with constraints ; wherein the initial information includes the actual spectral data measured by the device and the physical meaning of the parameters of each degree of freedom, the parameter values, the upper and lower boundaries of the parameters, the coupling relationship between the parameters and the parameter sensitivity, and the physical meaning of the parameters includes the width or thickness or angle of the sample material to be measured; wherein the constraints include the inequality constraints g(x) of the upper and lower boundaries of the parameters and the equality constraints h(x) between the parameters, and the inequality constraints g(x) and the equality constraints h(x) are expressed by the following formula:
[0007]
[0008] in, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints;
[0009] Calculate the constraint regression model to get the regression result point vector , and dynamically adjust the search point during the calculation process;
[0010] According to the constraints, the regression result point vector Perform box projection to get the projection point vector ;
[0011] Compute constraint convergence indices When the constraint convergence index meets the convergence condition, the final result is output ;as well as
[0012] The final result Converted into a morphological image of the semiconductor sample to be tested.
[0013] Furthermore, based on the augmented Lagrange multiplier method, the enhanced objective function with constraints is constructed as follows:
[0014]
[0015]
[0016]
[0017] in, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints, μ, λ, σ are all constraint parameters;
[0018] It is expressed by the following formula:
[0019]
[0020] in, represents the calculated theoretical spectrum, Indicates the spectrum measured by the device, represents the L2 norm, represents the first-order partial derivative of the calculated theoretical spectrum with respect to the perturbation amount, It represents the first-order partial derivative of the measured spectrum of the device with respect to the disturbance quantity, represents the coefficient of the first-order partial derivative.
[0021] Further, the constraint regression model is calculated to obtain the regression result point vector include:
[0022] The computational constraint regression model is to solve the minimum problem min(F), where F represents the enhanced objective function with constraints. ;
[0023] Solve the minimum problem min(F) and get the regression result, that is, the result point vector ;
[0024] The constrained regression model is calculated using iterative calculations, starting from the initial point vector Start searching, repeatedly update the search point to calculate the theoretical spectrum, compare the theoretical spectrum with the measured spectrum, and get the fit of the spectrum.
[0025] Furthermore, dynamically adjusting the search point during the calculation process includes:
[0026] Use the following formula to dynamically adjust the search point and correct the parameters with negative values:
[0027]
[0028] in, is to adjust the weight, represents the search point for the kth regression iteration calculation, represents the initial point vector, and i represents the parameter of the i-th degree of freedom.
[0029] Further, according to the constraint conditions, the regression result point vector Perform box projection to get the projection point vector include:
[0030] The resulting point vector Projecting to the feasible domain, the expression of the projection point vector is:
[0031]
[0032] in, represents the resulting point vector, Represents the projection point vector obtained after projection. represents the upper boundary of the feasible region, represents the lower boundary of the feasible region, i represents the parameter of the i-th degree of freedom;
[0033] Euclidean distance between point vectors before and after projection and the minimum Euclidean distance for,
[0034]
[0035]
[0036] Among them, N represents the number of degree of freedom parameters, and V represents the number of degree of freedom parameters that exceed the feasible region.
[0037] Furthermore, the constraint convergence index is calculated When the constraint convergence index meets the convergence condition, the final result is output include:
[0038] Constrained Convergence Index The expression of is,
[0039]
[0040] in, represents the constraint convergence index calculated by the kth regression iteration, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints, λ and σ are constraint parameters, represents the projection point vector obtained after the projection is performed, and k represents the kth regression iteration calculation;
[0041] When , the constraint convergence index satisfies the convergence condition and the final result is output;
[0042] in, Indicates the set threshold value. ;
[0043] The final result is calculated In the formula ,Will As the final result .
[0044] Further, When the constraint convergence index does not meet the convergence condition, the constraint parameters are updated, the constraint regression model is calculated again to obtain the regression result, and the search point is dynamically adjusted during the calculation process. According to the constraint condition, the regression result is subjected to frame projection to obtain the projection point vector, until the constraint convergence index meets the convergence condition, and the final result is output;
[0045] The method to update the constraint parameters is,
[0046] When ;
[0047] When ;
[0048] in, represents the constraint convergence index calculated by the kth regression iteration, represents the constraint convergence index calculated by the k-1th regression iteration; represents the constraint parameters for the k+1th regression iteration calculation, Represents the constraint parameters during the k-th regression iteration calculation; Indicates the set value, ; represents the constraint parameter update coefficient, ;
[0049] use Update the constraint parameters, the update formula is,
[0050]
[0051]
[0052] in, represents the constraint parameters for the k+1th regression iteration calculation, represents the constraint parameters for the kth regression iteration calculation, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints, Represents the projection point vector obtained after projection.
[0053] The present invention also provides a regression model hybrid constraint optimization system for semiconductor measurement, the system comprising the following modules:
[0054] The enhanced objective function building module is configured to input initial information and constraints and build an enhanced objective function with constraints ; wherein the initial information includes the actual spectral data measured by the device or the physical meaning of the parameters of each degree of freedom or the parameter value or the upper and lower boundaries of the parameters or the coupling relationship between the parameters or the parameter sensitivity, and the physical meaning of the parameters includes the width or thickness or angle of the sample material to be measured; wherein the constraint conditions include the inequality constraint conditions g(x) of the upper and lower boundaries of the parameters or the equality constraint conditions h(x) between the parameters, and the inequality constraint conditions g(x) and the equality constraint conditions h(x) are expressed by the following formula:
[0055]
[0056] in, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints;
[0057] The regression result calculation module is configured to calculate the constraint regression model to obtain the regression result point vector , and dynamically adjust the search point during the calculation process;
[0058] The frame projection module is configured to project the regression result point vector according to the constraint conditions. Perform box projection to get the projection point vector ;as well as
[0059] The final result output module is configured to calculate the constraint convergence index When the constraint convergence index meets the convergence condition, the final result is output .
[0060] The present invention also provides a computer system, comprising:
[0061] a processor configured to execute machine-readable instructions;
[0062] A graphics card with an artificial intelligence chip configured to train a regression model hybrid constraint optimization method for semiconductor metrology; and
[0063] The memory is configured to store machine-readable instructions, which, when executed by a processor and / or a graphics card, perform the steps of the regression model hybrid constraint optimization method for semiconductor measurement.
[0064] Compared with the prior art, the present invention has the following advantages:
[0065] 1. The hybrid constraint optimization method of the regression model for semiconductor measurement provided by the present invention improves the traditional regression analysis method, transforms the unconstrained optimization into the constrained optimization, and overcomes the disadvantage that the unconstrained regression model is prone to numerical values without physical meaning.
[0066] 2. The hybrid constraint optimization method for the regression model for semiconductor measurement provided by the present invention comprehensively utilizes the augmented Lagrange multiplier method, the dynamic adjustment method and the frame projection method to perform constraint optimization on the regression process and results from multiple angles. The constraint effect is better than using any of the above methods alone, which effectively improves the compatibility and generalization ability of the algorithm and improves the user experience of semiconductor measurement equipment.
[0067] 3. The hybrid constraint optimization method of the regression model for semiconductor measurement provided by the present invention utilizes sensitivity analysis prior calculation to selectively perform parameter constraints, which can achieve the constraint effect while avoiding excessive interference of constraints in the regression process.
[0068] 4. The hybrid constraint optimization method of the regression model for semiconductor measurement provided by the present invention optimizes the common method of using spectral fitting as the objective function, adds the spectral first-order derivative fitting term, and improves the performance of the regression algorithm. BRIEF DESCRIPTION OF THE DRAWINGS
[0069] To further illustrate the above and other advantages and features of various embodiments of the present invention, a more specific description of various embodiments of the present invention will be presented with reference to the accompanying drawings. It will be understood that these drawings only depict typical embodiments of the present invention and are therefore not to be considered as limiting the scope thereof. In the accompanying drawings, for clarity, identical or corresponding parts will be represented by identical or similar reference numerals.
[0070] Figure 1 A schematic diagram showing a flow chart of a hybrid constraint optimization method for a regression model for semiconductor measurement according to an embodiment of the present invention; and
[0071] Figure 2 A schematic diagram of a regression model hybrid constraint optimization system for semiconductor measurement according to an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0072] In the following description, the present invention is described with reference to various embodiments. However, those skilled in the art will recognize that various embodiments may be implemented without one or more specific details or with other replacement and / or additional methods or components. In other cases, well-known structures or operations are not shown or described in detail to avoid obscuring the inventive point of the present invention. Similarly, for the purpose of explanation, specific numbers and configurations are set forth to provide a comprehensive understanding of embodiments of the present invention. However, the present invention is not limited to these specific details.
[0073] In this specification, reference to "one embodiment" or "the embodiment" means that a particular feature, structure, or characteristic described in conjunction with the embodiment is included in at least one embodiment of the present invention. The phrase "in one embodiment" appearing in various places in this specification does not necessarily all refer to the same embodiment.
[0074] It should be noted that the embodiments of the present invention describe the method steps in a specific order, but this is only for the purpose of illustrating the specific embodiment, rather than limiting the order of the steps. On the contrary, in different embodiments of the present invention, the order of the steps can be adjusted according to actual needs.
[0075] In the present invention, each module of the system according to the present invention can be implemented using software, hardware, firmware or a combination thereof. When the module is implemented using software, the function of the module can be implemented through a computer program flow, for example, the module can be implemented by a code segment (such as a code segment of a language such as C, C++, etc.) stored in a storage device (such as a hard disk, memory, etc.), wherein when the code segment is executed by a processor, the corresponding function of the module can be implemented. When the module is implemented using hardware, the function of the module can be implemented by setting a corresponding hardware structure, for example, the function of the module can be implemented by hardware programming a programmable device such as a field programmable gate array (FPGA), or by designing an application-specific integrated circuit (ASIC) including electronic devices such as multiple transistors, resistors and capacitors to implement the function of the module. When the module is implemented using firmware, the function of the module can be written into a read-only memory such as an EPROM or EEPROM of the device in the form of a program code, and when the program code is executed by a processor, the corresponding function of the module can be implemented. In addition, some functions of the module may need to be implemented by separate hardware or by collaboration with the hardware, for example, the detection function is implemented by corresponding sensors (such as proximity sensors, acceleration sensors, gyroscopes, etc.), the signal transmission function is implemented by corresponding communication devices (such as Bluetooth devices, infrared communication devices, baseband communication devices, Wi-Fi communication devices, etc.), the output function is implemented by corresponding output devices (such as displays, speakers, etc.), and so on.
[0076] The hybrid constraint optimization method for the regression model for semiconductor measurement provided by the present invention adds a correction term to the objective function of the traditional regression model according to the actual constraint conditions that conform to the physical meaning, transforms the unconstrained optimization into the constrained optimization, constructs the constrained optimization mathematical model and solves its minimum value, and combines the dynamic adjustment method and the frame constraint projection method to enhance the constraint effect. The hybrid constraint optimization method for the regression model for semiconductor measurement provided by the present invention can ensure that the intermediate process and the final output result all meet the constraint conditions. The hybrid constraint optimization method for the regression model for semiconductor measurement provided by the present invention can be compatible with a wider range of initial points. Even if convergence cannot be achieved when the initial point is very poor, it can ensure that the output result is constrained within a range acceptable to the user, thereby improving the generalization ability and versatility of the algorithm.
[0077] The technical solutions in the embodiments of the present invention are described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.
[0078] Figure 1 FIG. 1 is a flow chart of a hybrid constraint optimization method for semiconductor measurement using a regression model according to an embodiment of the present invention. Figure 1 As shown, the regression model hybrid constraint optimization method for semiconductor measurement includes the following steps:
[0079] First, the initial information and constraints are input to construct an enhanced objective function with constraints.
[0080] The initial information includes the actual spectral data measured by the device, the physical meaning of the parameters of each degree of freedom, the parameter value, the upper and lower boundaries of the parameters, the coupling relationship between the parameters, the sensitivity of the parameters, etc. Among these information, the actual spectral data is obtained by the measurement component in the device, the physical meaning of the parameters generally includes the width, thickness, angle, etc. of each layer of the material of the sample to be measured, the parameter value, the upper and lower boundaries of the parameters, etc. are defined by the user input at the software interface, and the coupling relationship between the parameters can also be determined in the form of input equations.
[0081] Parameter sensitivity analysis belongs to prior information. After determining the floating parameters, make small perturbations of equal proportion to each parameter, and calculate the change of the theoretical spectrum caused by the small perturbation. The degree of change can be characterized by the first-order derivative of the fit between the theoretical spectrum and the measured spectrum with respect to the perturbation amount. The greater the degree of change, the higher the sensitivity of the corresponding parameter, and the greater the constraint strength of the parameter should be; and the parameters with lower sensitivity can retain unconstrained optimization, thereby controlling the number of constraints, reducing the computational complexity, and weakening the constraint interference of the search process.
[0082] The upper and lower bounds of the parameters for each degree of freedom are essentially inequality constraints, denoted by g(x). If the upper and lower bounds of the parameters are linear constraints, then this inequality constraint is a box constraint. The equation expressions between the parameters are equality constraints, denoted by h(x). Therefore, various constraint conditions obtained externally can be regarded as a mixed constraint containing inequality constraints and equality constraints. The subspace restricted by this mixed constraint in the multi-dimensional parameter space is the feasible region for regression calculation. For example, in a certain semiconductor device OCD measurement scenario, through initial information and prior calculation, the morphology of a layer of silicon nitride material SiN in the device to be measured is determined to be trapezoidal. The degrees of freedom to be measured, initial values, sensitivities, and constraint conditions of this layer of material are as follows: bottom width b = 100 nm (the range of the bottom width is 80 nm < b < 120 nm), thickness t = 25 nm (the range of the thickness is 20 nm < t < 30 nm), bottom side angle a = 45° (the range of the bottom side angle is 0° < a < 90°), and top width u = 40 nm (with low sensitivity and no constraint), then these parameter boundaries are the inequality group g(x) or box constraints; and since this layer of material is defined as trapezoidal, there are equality constraints between these parameters, that is, u + 2[t / tan(a)] = b; more complex geometric models may have more equality relationships, and these relationships form the equality constraint equation system h(x).
[0083] The core objective of the traditional regression model is to solve the global minimum of the objective function. Specifically in the field of semiconductor metrology, the objective function is generally defined as the L2 norm fitting degree between the theoretical spectrum Scal calculated from the forward problem and the measured spectrum Sreal of the device , which is expressed by the following formula
[0084]
[0085] where represents the L2 norm
[0086] The optimization problem of the traditional regression model is to solve min .
[0087] To improve the accuracy of the regression algorithm and approximate the measured spectrum as much as possible, based on the traditional objective function, this invention combines the prior information obtained from the sensitivity analysis in this step and adds the first-order partial derivative value as part of the fitting signal to construct an enhanced objective function , which is expressed by the following formula
[0088]
[0089] where represents the calculated theoretical spectrum represents the measured spectrum of the device represents the L2 norm, represents the first-order partial derivative of the calculated theoretical spectrum with respect to the perturbation amount, It represents the first-order partial derivative of the measured spectrum of the device with respect to the disturbance; Represents the coefficient of the first-order partial derivative, in order to avoid the new term being too large and covering the original spectral term, Generally set to a smaller value, such as <0.1, the actual value can be tuned by testing.
[0090] Combining the above hybrid constraints with the regression model of the enhanced objective function, assuming that there are l equality constraints and m inequality constraints, the unconstrained optimization problem min[ ] is transformed into a constrained optimization problem, which can be expressed mathematically as:
[0091]
[0092] Based on the augmented Lagrange multiplier method, the objective function is enhanced. Add correction terms or penalty terms on the basis of , and construct the enhanced objective function with constraints as follows:
[0093]
[0094]
[0095]
[0096] in, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints, μ, λ, and σ are constraint parameters.
[0097] The values of the constraint parameters can be adjusted based on a certain initial parameter set according to actual tests. The principle of value selection is to ensure that the penalty term has an effective constraint on the original function, but the constraint strength cannot be too large to excessively interfere with the search process. In one embodiment of the present invention, the initial values of the constraint parameters can be set to μ=0.1, λ=0.1, σ=2, and then adjusted according to actual tests.
[0098] This step obtains initial information from the outside, defines the constraints of each degree of freedom parameter through sensitivity analysis, introduces the processed mixed constraints into the traditional unconstrained optimization model, and obtains the enhanced objective function with constraints. , which is provided to the subsequent calculation steps.
[0099] Next, the constraint regression model is calculated to obtain the regression result, and the search point is dynamically adjusted during the calculation process.
[0100] The computational constraint regression model is to solve the minimum problem min(F), where F represents the enhanced objective function with constraints. .
[0101] Solve the minimum problem min(F) and get the regression result, that is, the result point vector .
[0102] The regression algorithm for solving min(F) can be selected from the common steepest descent method, Gauss-Newton method, Levenberg-Marquardt method (LM), quasi-Newton method, etc. Among them, the steepest descent method is relatively simple to implement and has a small amount of calculation, but it converges slowly and lacks accuracy and robustness; the Newton method makes up for the above shortcomings, but it needs to calculate the second-order Hessian matrix, which increases the amount of calculation; the LM method inherits the advantages of the Newton method, replaces the second-order Hessian matrix with approximate calculation, and overcomes the non-positive definite problem by adding a damping term on its basis. Considering the speed and accuracy requirements, the LM method is mostly used in the semiconductor field to solve inverse problems; under some special needs, other regression algorithms can also be used.
[0103] The constrained regression model is calculated using iterative calculations, starting from the initial point vector Start searching, repeatedly update the search point and calculate the positive problem, that is, calculate the theoretical spectrum from the physical size of the current point, compare the theoretical spectrum with the measured spectrum, and get the fit of the spectrum. The solution to the positive problem can be achieved by rigorous coupled wave analysis, finite element method, finite difference time domain method or neural network, etc.; then, by comparing the theoretical spectrum with the measured spectrum provided by the above steps, the fit between the spectra is obtained, and constraints are introduced so that under the constraints, the spectral fit is gradually improved, and the variables of each degree of freedom gradually approach the actual physical size of the sample to be tested, and the calculation results can better restore the sample morphology.
[0104] In the iterative calculation process, from a mathematical perspective, although the regression model with constraints can guide the search direction to stay within the feasible domain as much as possible, there will still be cases where the intermediate quantities have negative values. This is mathematically reasonable, but lacks actual physical meaning. Parameters such as material thickness and width cannot be negative. For example, in a certain iterative calculation step, a layer of SiN material of the semiconductor device to be tested is not constrained and has a negative value u=-10nm due to the low sensitivity of the top width u. At this time, the algorithm can still perform subsequent calculations normally, but for the intermediate quantity of this iteration, the -10nm width obviously has no physical meaning and will mislead the users of the measuring equipment. Therefore, in addition to the external constraints introduced in the above steps, this step continuously checks the intermediate quantities of the parameters in the process of solving the regression model with constraints. If a negative value appears, the degree of freedom where the negative value appears is corrected using a dynamic adjustment method. The search point is dynamically adjusted using the following formula:
[0105]
[0106] in, is to adjust the weight, represents the search point for the kth regression iteration calculation, represents the initial point vector, i represents the parameter of the i-th degree of freedom. is dynamically adjusted if If the value is negative, the dynamic adjustment method is used to search for a set of weights. =0.75, =0.25, then after the iterative calculation step where a negative value appears, the top width u can be adjusted to u=2.5nm according to the above weights, and then the adjusted non-negative u value can be used to perform subsequent calculations.
[0107] This step combines the dynamic adjustment method to solve the regression model with constraints and obtain the result point vector .
[0108] Next, according to the constraint conditions, a frame projection is performed on the regression result to obtain a projection point vector.
[0109] Compared with the traditional unconstrained optimization problem, the constrained regression model can effectively guide the search point to remain within the feasible domain, but the constraint ability of the penalty term is limited and cannot completely prevent the violation of the constraint conditions. In other words, the result point vector obtained above There is still a possibility that the parameters are out of bounds. Therefore, in the upper and lower boundary constraints of the above steps, the linear inequality constraints can be used as box constraints to check the result point vector obtained above. Whether there are parameters that cross the boundary. Assuming that there are V degrees of freedom parameters among the N degrees of freedom parameters that cross the boundary, the parameters of these V degrees of freedom are projected to the boundary closest to them according to the box constraint conditions. According to the constraint conditions, the regression result is subjected to box projection, and the projection point vector obtained includes:
[0110] The resulting point vector Projecting to the feasible domain, the expression of the projection point vector is:
[0111]
[0112] in, represents the resulting point vector, Represents the projection point vector obtained after projection. represents the upper boundary of the feasible region, represents the lower boundary of the feasible region, i represents the parameter of the i-th degree of freedom;
[0113] Euclidean distance between point vectors before and after projection and the minimum Euclidean distance for,
[0114]
[0115]
[0116] Among them, N represents the number of degree of freedom parameters, and V represents the number of degree of freedom parameters that exceed the feasible region.
[0117] It should be noted that in practical applications, numerical instability problems are likely to occur at the boundaries, which may lead to large errors in calculation results, incorrect judgments, or situations where calculations cannot be performed. Therefore, based on the parameter boundary conditions provided in the above steps, the boundaries of the parameters are shrunk by a certain proportion to form more stringent linear constraints, that is, a sub-feasible region contained within the feasible region is formed. Then, the projection operator is executed according to the multi-dimensional space frame constraints of this sub-feasible region to avoid the above-mentioned numerical instability situations. Shrinking the boundaries is to enhance the stability at the original boundaries and cannot be set to a large value; if the shrinking ratio is too large, the parameter search space will rapidly decrease, which will greatly interfere with the search process of regression analysis and may even exclude the global optimal solution from the feasible region, reducing the algorithm performance. In an embodiment of the present invention, the shrinking ratio can be <10%. For example, in a certain layer of SiN material in the above steps, if the shrinking ratio is set to 10%, the ranges of each parameter become: bottom width 82nm < b < 118nm, thickness 20.5nm < t < 29.5nm, bottom side angle 4.5° < a < 85.5°; the result point vector , in this vector, the thickness t = 30nm > 29.5nm, b = 75nm < 80nm < 82nm, that is, the thickness t satisfies the original constraint conditions but does not satisfy the shrunk constraint conditions, and the bottom width b does not satisfy both before and after shrinking; it is necessary to project these two parameters so that = 29.5, = 82, and substitute and for t and b to obtain the projected point vector that satisfies the shrinking constraint conditions.
[0118] In this step, through the above-mentioned frame constraint projection method, the search points that are difficult to constrain are forcibly restricted within the feasible region, forming a complementary relationship with the method of introducing penalty terms in the above steps, enhancing the constraint strength, and improving the effect of constraint optimization.
[0119] Finally, calculate the constraint convergence index. When the constraint convergence index meets the convergence conditions, the final result is output.
[0120] The constraint convergence index is calculated from the constraint conditions and constraint parameters , and the expression of the constraint convergence index is,
[0121]
[0122] where, represents the constraint convergence index obtained from the k-th regression iteration calculation, represents the i-th equality constraint condition, represents the number of equality constraint conditions, represents the jth inequality constraint, represents the number of inequality constraints, λ and σ are constraint parameters, represents the projection point vector obtained after the projection is performed, and k represents the kth regression iteration calculation.
[0123] By comparing the constraint convergence index β with the set threshold , determine whether the convergence condition is met.
[0124] When , the constraint convergence index satisfies the convergence condition and the final result is output.
[0125] in, Indicates the set threshold value. .
[0126] When the constraint convergence index does not meet the convergence condition, the constraint parameters are updated, the constraint regression model is calculated again to obtain the regression result, and the search point is dynamically adjusted during the calculation process. According to the constraint condition, the regression result is subjected to frame projection to obtain the projection point vector, until the constraint convergence index meets the convergence condition, and the final result is output.
[0127] The method to update the constraint parameters is,
[0128] When ;
[0129] When ;
[0130] in, represents the constraint convergence index calculated by the kth regression iteration, represents the constraint convergence index calculated by the k-1th regression iteration; represents the constraint parameters for the k+1th regression iteration calculation, Represents the constraint parameters during the k-th regression iteration calculation; Indicates the set value, ; represents the constraint parameter update coefficient, ;
[0131] use Update the constraint parameters, the update formula is,
[0132]
[0133]
[0134] in, represents the constraint parameters for the k+1th regression iteration calculation, represents the constraint parameters for the kth regression iteration calculation, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints, Represents the projection point vector obtained after projection.
[0135] when When satisfied, the current is the approximate minimum value that satisfies the constraints, and the calculation is terminated and As the final result Transferred to the subsequent components of the measurement equipment software, according to Through data post-processing, visualization of calculated spectra, and CAD drawing of physical models, the final regression calculation results that meet all constraints are converted into intuitive morphological images of the semiconductor samples to be tested and presented to the user.
[0136] For example, for a layer of SiN material mentioned above, the threshold value is set =1e-5, after k regression iterations, we get the point vector , where b=82nm, t=28nm, a=75°, u=67nm, The parameters satisfy the trapezoidal equation u+2[t / tan(a)]=b, and all parameters remain within the contraction range or on the boundary, then all parameters must be within the original range, and the calculation is =3e-6< , obviously On the basis of satisfying all constraints and meeting the convergence conditions, then That is ,Will By providing it to subsequent software components, the actual morphology of the trapezoidal SiN material layer can be reproduced.
[0137] The hybrid constraint optimization method for regression model for semiconductor measurement provided by the present invention converts the optimization problem of solving the global minimum value without constraints into the problem of searching for an approximate optimal solution that satisfies the constraints on the basis of the traditional regression model. By adding correction terms or penalty terms to the objective function, a regression model with constraints is constructed, and the minimum value of the constraint regression model is iteratively searched; in the iterative calculation process, the intermediate quantity of the measured parameter is continuously checked and modified by the dynamic adjustment method to prevent the occurrence of numerical values without physical meaning; the regression result is further corrected by the frame constraint projection method, and it is determined whether to update the constraint parameters and continue searching according to the convergence conditions until the conditions are met, the calculation is terminated and the final result is output. The hybrid constraint optimization method for regression model for semiconductor measurement provided by the present invention comprehensively utilizes the augmented Lagrange multiplier method, the dynamic adjustment method and the frame projection method, and performs constraint optimization on the regression process and results from multiple angles. The constraint effect is better than that of using any of the above methods alone, effectively improving the compatibility and generalization ability of the algorithm, and improving the user experience of semiconductor measurement equipment.
[0138] Figure 2 FIG. 1 is a schematic diagram of a regression model hybrid constraint optimization system for semiconductor measurement according to an embodiment of the present invention. Figure 2 As shown, the system includes the following modules:
[0139] The enhanced objective function building module is configured to input initial information and constraints and build an enhanced objective function with constraints ; wherein the initial information includes the actual spectral data measured by the device or the physical meaning of the parameters of each degree of freedom or the parameter value or the upper and lower boundaries of the parameters or the coupling relationship between the parameters or the parameter sensitivity, and the physical meaning of the parameters includes the width or thickness or angle of the sample material to be measured; wherein the constraint conditions include the inequality constraint conditions g(x) of the upper and lower boundaries of the parameters or the equality constraint conditions h(x) between the parameters, and the inequality constraint conditions g(x) and the equality constraint conditions h(x) are expressed by the following formula:
[0140]
[0141] in, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints;
[0142] The regression result calculation module is configured to calculate the constraint regression model to obtain the regression result point vector , and dynamically adjust the search point during the calculation process;
[0143] The frame projection module is configured to project the regression result point vector according to the constraint conditions. Perform box projection to get the projection point vector ;as well as
[0144] The final result output module is configured to calculate the constraint convergence index When the constraint convergence index meets the convergence condition, the final result is output .
[0145] In one embodiment of the present invention, the present invention further provides a computer system, which includes a processor, a graphics card and a memory, wherein the memory is configured to store machine-readable instructions, the graphics card is configured to train a regression model hybrid constraint optimization method for semiconductor measurement, and the processor is configured to execute the machine-readable instructions. When the processor and / or the graphics card executes the machine-readable instructions, the following processing steps are implemented: inputting initial information and constraint conditions, constructing an enhanced objective function containing constraint conditions ; Calculate the constraint regression model to get the regression result point vector , and dynamically adjust the search point during the calculation process; according to the constraints, the regression result point vector Perform box projection to get the projection point vector Compute constraint convergence indices When the constraint convergence index meets the convergence condition, the final result is output ; and the final result Converted into a morphological image of the semiconductor sample to be tested.
[0146] The graphics card may preferably have a GPU computing power higher than 5.0. Since the amount of data to be trained is large, providing a graphics card configuration can significantly increase the training speed.
[0147] The memory includes: a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and other media that can store machine-readable instructions.
[0148] It can be understood that in addition to the memory and processor mentioned above, the above-mentioned computer system also includes other software and hardware components not listed in this specification. The specific components can be determined according to the model of specific data processing equipment in different application scenarios, and this specification will not list them one by one in detail.
[0149] Although various embodiments of the present invention are described above, it should be understood that they are presented as examples only and not as limitations. It is obvious to those skilled in the relevant art that various combinations, modifications and changes can be made thereto without departing from the spirit and scope of the present invention. Therefore, the breadth and scope of the present invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should only be defined according to the technical solutions of the present invention and their equivalents.
Claims
1. A hybrid constraint optimization method for regression model for semiconductor measurement, characterized in that: The steps include: Input initial information and constraints to construct an enhanced objective function with constraints ; wherein the initial information includes the actual spectral data measured by the device and the physical meaning of the parameters of each degree of freedom, the parameter values, the upper and lower boundaries of the parameters, the coupling relationship between the parameters and the parameter sensitivity, and the physical meaning of the parameters includes the width or thickness or angle of the sample material to be measured; wherein the constraints include the inequality constraints g(x) of the upper and lower boundaries of the parameters and the equality constraints h(x) between the parameters, and the inequality constraints g(x) and the equality constraints h(x) are expressed by the following formula: in, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints; Based on the augmented Lagrange multiplier method, the enhanced objective function with constraints is constructed as follows: in, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints, μ, λ, σ are all constraint parameters; It is expressed by the following formula: in, represents the calculated theoretical spectrum, Indicates the spectrum measured by the device, represents the L2 norm, represents the first-order partial derivative of the calculated theoretical spectrum with respect to the perturbation amount, It represents the first-order partial derivative of the measured spectrum of the device with respect to the disturbance quantity, represents the coefficient of the first-order partial derivative; Calculate the constraint regression model to get the regression result point vector , and dynamically adjust the search point during the calculation process; According to the constraints, the regression result point vector Perform box projection to get the projection point vector ; Compute constraint convergence indices When the constraint convergence index meets the convergence condition, the final result is output ;as well as The final result Converted into a morphological image of the semiconductor sample to be tested.
2. The regression model hybrid constraint optimization method according to claim 1, characterized in that: Calculate the constraint regression model to get the regression result point vector include: The computational constraint regression model is to solve the minimum problem min(F), where F represents the enhanced objective function with constraints. ; Solve the minimum problem min(F) and get the regression result, that is, the result point vector ; The constrained regression model is calculated using iterative calculations, starting from the initial point vector Start searching, repeatedly update the search point to calculate the theoretical spectrum, compare the theoretical spectrum with the measured spectrum, and get the fit of the spectrum.
3. The regression model hybrid constraint optimization method according to claim 1, characterized in that: Dynamically adjusting the search point during the calculation process includes: Use the following formula to dynamically adjust the search point and correct the parameters with negative values: in, is to adjust the weight, represents the search point for the kth regression iteration calculation, represents the initial point vector, and i represents the parameter of the i-th degree of freedom.
4. The regression model hybrid constraint optimization method according to claim 1, characterized in that: According to the constraints, the regression result point vector Perform box projection to get the projection point vector include: The resulting point vector Projecting to the feasible domain, the expression of the projection point vector is: in, represents the resulting point vector, Represents the projection point vector obtained after projection. represents the upper boundary of the feasible region, represents the lower boundary of the feasible region, i represents the parameter of the i-th degree of freedom; Euclidean distance between point vectors before and after projection and the minimum Euclidean distance for, Among them, N represents the number of degree of freedom parameters, and V represents the number of degree of freedom parameters that exceed the feasible region.
5. The hybrid constraint optimization method for regression model according to claim 1, characterized in that: Compute constraint convergence indices When the constraint convergence index meets the convergence condition, the final result is output include: Constrained Convergence Index The expression of is, in, represents the constraint convergence index calculated by the kth regression iteration, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints, λ and σ are constraint parameters, represents the projection point vector obtained after the projection is performed, and k represents the kth regression iteration calculation; When , the constraint convergence index satisfies the convergence condition and the final result is output; in, Indicates the set threshold value. ; The final result is calculated In the formula ,Will As the final result .
6. The hybrid constraint optimization method for regression model according to claim 5, characterized in that: When the constraint convergence index does not meet the convergence condition, the constraint parameters are updated, the constraint regression model is calculated again to obtain the regression result, and the search point is dynamically adjusted during the calculation process. According to the constraint condition, the regression result is subjected to frame projection to obtain the projection point vector, until the constraint convergence index meets the convergence condition, and the final result is output; The method to update the constraint parameters is, When ; When ; in, represents the constraint convergence index calculated by the kth regression iteration, represents the constraint convergence index calculated by the k-1th regression iteration; represents the constraint parameters for the k+1th regression iteration calculation, Represents the constraint parameters during the k-th regression iteration calculation; Indicates the set value, ; represents the constraint parameter update coefficient, ; use Update the constraint parameters, the update formula is, in, represents the constraint parameters for the k+1th regression iteration calculation, represents the constraint parameters for the kth regression iteration calculation, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints, Represents the projection point vector obtained after projection.
7. A system for the hybrid constraint optimization method of a regression model as claimed in any one of claims 1 to 6, characterized in that: Includes the following modules: The enhanced objective function building module is configured to input initial information and constraints and build an enhanced objective function with constraints ; wherein the initial information includes the actual spectral data measured by the device or the physical meaning of the parameters of each degree of freedom or the parameter value or the upper and lower boundaries of the parameters or the coupling relationship between the parameters or the parameter sensitivity, and the physical meaning of the parameters includes the width or thickness or angle of the sample material to be measured; wherein the constraint conditions include the inequality constraint conditions g(x) of the upper and lower boundaries of the parameters or the equality constraint conditions h(x) between the parameters, and the inequality constraint conditions g(x) and the equality constraint conditions h(x) are expressed by the following formula: in, represents the i-th equality constraint, represents the number of equality constraints, represents the jth inequality constraint, represents the number of inequality constraints; The regression result calculation module is configured to calculate the constraint regression model to obtain the regression result point vector , and dynamically adjust the search point during the calculation process; The frame projection module is configured to project the regression result point vector according to the constraint conditions. Perform box projection to get the projection point vector ;as well as The final result output module is configured to calculate the constraint convergence index When the constraint convergence index meets the convergence condition, the final result is output .
8. A computer system, characterized in that: include: a processor configured to execute machine-readable instructions; A graphics card with an artificial intelligence chip configured to train a regression model hybrid constrained optimization method; as well as A memory configured to store machine-readable instructions, wherein the machine-readable instructions, when executed by a processor and / or a graphics card, perform the steps of the method according to any one of claims 1 to 6.
