Device binning parameter extraction method, computer device and storage medium

Abstract

The application discloses a device binning parameter extraction method, a computer device and a storage medium, and relates to the technical field of semiconductors. The device binning parameter extraction method comprises the following steps: obtaining a first geometric parameter of a first device, inputting the first geometric parameter into a pre-trained first fusion model, and determining an observation vector corresponding to each physical parameter in the first device through the first fusion model; constructing a coefficient matrix corresponding to each physical parameter according to the corresponding relationship between the binning parameters and the geometric dimensions of the first device; constructing a linear equation group of each physical parameter based on the observation vector and the coefficient matrix; and solving each linear equation group to obtain binning parameters corresponding to each physical parameter. The technical scheme aims to improve the extraction accuracy of the binning parameters of semiconductor devices.

Description

Technical Field

[0001] This application relates to the field of semiconductor technology, and in particular to a method for extracting device binning parameters, a computer device, and a computer-readable storage medium. Background Technology

[0002] As semiconductor processes shrink to the nanometer scale, the emergence of various new devices has led to a dramatic increase in the complexity of compact models such as BSIM-CMG (Berkeley Short-Channel IGFET Model-Common Multi-Gate, an advanced compact model for simulating metal-oxide-semiconductor field-effect transistors). This results in a significantly increased number of device parameters and a markedly longer parameter extraction cycle. Faced with the challenge of high-dimensional parameter extraction at advanced process nodes, the industry's main automated extraction scheme relies on purely data-driven inverse neural network parameter extraction strategies. However, pure neural network models are prone to overfitting, and the generated electrical characteristic parameters may violate physical laws, causing simulator errors. Furthermore, this approach is highly dependent on the amount of data; its extrapolation prediction ability is extremely poor within bias voltage or temperature ranges not covered by the training data, failing to guarantee the physical correctness of the model. Summary of the Invention

[0003] The main objective of this application is to provide a method for extracting device binning parameters, a computer device, and a computer-readable storage medium, with the aim of improving the accuracy of extracting semiconductor device binning parameters.

[0004] To achieve the above objectives, this application provides a method for extracting device binning parameters, the method comprising: The first geometric parameters of the first device are obtained, and the first geometric parameters are input into the pre-trained first fusion model. The observation vectors corresponding to each physical parameter in the first device are predicted and determined by the first fusion model. Based on the correspondence between the boxing parameters and geometric dimensions of the first device, a coefficient matrix corresponding to each physical parameter is constructed. Based on the observation vector and the coefficient matrix, a system of linear equations is constructed for each physical parameter; Solve the linear equations to obtain the binning parameters corresponding to each physical parameter.

[0005] In one embodiment, the step of determining the observation vectors corresponding to each physical parameter in the first device through the prediction of the first fusion model includes: The first fusion model is used to sample the geometric space corresponding to the first geometric parameter to generate multiple sets of physical parameter values, wherein the number of sets of physical parameter values ​​is equal to the preset number of sampling points; For each physical parameter in the first device, an observation vector corresponding to each physical parameter is generated based on the values ​​of each group of physical parameters.

[0006] In one embodiment, the correspondence between the binning parameters and the geometric dimensions is a functional relationship, and the step of constructing the coefficient matrix corresponding to each physical parameter based on the correspondence between the binning parameters and the geometric dimensions of the first device includes: The functional relationship between the boxing parameters and geometric dimensions is converted into matrix form to obtain the coefficient matrix corresponding to each physical parameter; The dimension of the coefficient matrix is ​​determined by the number of groups of physical parameter values ​​and the number of binning parameters.

[0007] In one embodiment, the step of constructing a system of linear equations for each physical parameter based on the observation vector and the coefficient matrix includes: Initialize a one-dimensional parameter vector matrix, wherein the parameter vector matrix is ​​composed of unknowns corresponding to multiple bin parameters; For each physical parameter, a system of linear equations is generated based on the coefficient matrix, the parameter vector matrix, and the observation vector, wherein the system of linear equations represents the product of the coefficient matrix and the parameter vector matrix as equal to the observation vector.

[0008] In one embodiment, after the step of obtaining the first geometric parameters of the first device, the method further includes: The number of partitions is initialized to 2. Based on the preset multiple sampling points and the number of partitions, the geometric space corresponding to the first geometric parameter is divided into two sub-regions. For each sub-region, the observation vector corresponding to each physical parameter in the first device is determined by prediction using the first fusion model; For each sub-region, a coefficient matrix corresponding to each physical parameter is constructed based on the correspondence between the binning parameters and geometric dimensions of the first device; For each sampling point, the local binning parameters corresponding to each sub-region are determined based on the observation vector and coefficient matrix corresponding to each physical parameter. Substitute the two sets of local binning parameters corresponding to each sampling point into the preset simulator to obtain the first simulation result; Based on the first simulation results corresponding to each sampling point and the original test results of the first device, calculate the fitting error corresponding to each sampling point. The local binning parameters corresponding to the sampling point with the smallest fitting error are determined as the binning parameters of the first device.

[0009] In one embodiment, after the step of solving each of the linear equations to obtain the binning parameters corresponding to each of the physical parameters, the method further includes: The binning parameters corresponding to each of the physical parameters of the first device are input into a preset simulator for simulation to obtain the second simulation result of the first device. Based on the second simulation result of the first device and the original test result of the first device, calculate the root mean square error of the second simulation result; If the root mean square error of the second simulation result is greater than or equal to the first error threshold, then return to the execution step: divide the geometric space corresponding to the first geometric parameter into two sub-regions according to the preset multiple sampling points and number of partitions.

[0010] In one embodiment, after the step of determining the local binning parameters corresponding to the sampling point with the smallest fitting error as the binning parameters of the first device, the method further includes: The local binning parameters corresponding to the sampling point with the smallest fitting error are determined as the binning parameters of the first device and input into a preset simulator for simulation to obtain the third simulation result of the first device. Based on the third simulation result of the first device and the original test result of the first device, calculate the root mean square error of the third simulation result; If the root mean square error of the third simulation result is greater than or equal to the second error threshold, then the number of sub-regions is increased. Based on the preset number of sampling points and partitions, the geometric space corresponding to the first geometric parameter is divided into multiple sub-regions, and the execution steps are returned: for each sub-region, the observation vector corresponding to each physical parameter in the first device is predicted and determined by the first fusion model until the root mean square error of the third simulation result is less than the second error threshold.

[0011] In one embodiment, the device binning parameter extraction method further includes: Initialize the second fusion model, wherein the second fusion model includes at least a micro neural network and a physical equation module; Obtain the second geometric parameters and bias voltage of the second device; The second geometric parameter is input into the micro neural network, and the bias voltage is input into the physical equation module; The second geometric parameter is mapped to multiple sets of physical parameter values ​​through the micro neural network; By combining the bias voltage with each set of physical parameter values ​​through the physical equation module, the first current data or the first capacitance data is output. Based on the error between the first current data or the first capacitance data and the second current data or the second capacitance data in the original test results of the second device, the network parameters of the micro neural network are iteratively updated until a preset stopping condition is met, thus obtaining the first fusion model.

[0012] In addition, this application also provides a computer device, which includes at least: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the device binning parameter extraction method applied to the computer device as described above.

[0013] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the device binning parameter extraction method described above.

[0014] In addition, to achieve the above objectives, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the device binning parameter extraction method described above.

[0015] This application provides a method for extracting device binning parameters. The method includes: firstly, obtaining the first geometric parameters of a first device; inputting the first geometric parameters into a pre-trained first fusion model; predicting and determining the observation vectors corresponding to each physical parameter in the first device using the first fusion model; then, constructing a coefficient matrix corresponding to each physical parameter based on the correspondence between the binning parameters and geometric dimensions of the first device; then, constructing a system of linear equations for each physical parameter based on the observation vectors and the coefficient matrix; and finally, solving the system of linear equations to obtain the binning parameters corresponding to each physical parameter. Compared to traditional parameter extraction schemes, the pre-trained first fusion model in this application outputs the observation vectors corresponding to intermediate physical parameters instead of directly outputting electrical characteristic parameters, enabling the model to satisfy physical laws and avoid overfitting. Furthermore, the method combines the correspondence between binning parameters and geometric dimensions to construct a coefficient matrix and a system of linear equations to solve for the binning parameters. Moreover, the output of the first fusion model is the observation vectors corresponding to the physical parameters, which helps to reduce dependence on training data by utilizing prior physical knowledge during training, ensuring the training efficiency and model performance of the first fusion model. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the device compartment parameter extraction method in the embodiments of this application; Figure 2 This is a schematic diagram comparing the third simulation result with the original test result in the embodiments of this application; Figure 3 This is a schematic diagram of the training process of the first fusion model in the embodiments of this application; Figure 4 This is a schematic diagram of a feasible model training, device binning parameter extraction and verification process in an embodiment of this application; Figure 5 This is a schematic diagram of the hardware operating environment of the device involved in the device compartment parameter extraction method in this application embodiment.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a brief explanation of the technical terms involved in the technical solution of this application will be given first.

[0023] Fusion Model: Refers to a hybrid modeling architecture that integrates a physical equation framework with a micro neural network to capture device characteristics.

[0024] Binning Parameters: In the BSIM standard model, these are scaling factors used to describe how model parameters vary with geometric dimensions (L, W), such as LPARAM, WPARAM, PPARAM, etc.

[0025] Global Model: A unified set of parameters that covers the entire range of geometric dimensions allowed by the process design rules.

[0026] DTCO: Design Technology Co-Optimization.

[0027] As semiconductor processes shrink to the nanometer scale, the emergence of new devices such as FinFET (Fin Field-Effect Transistor) and GAAFET (Gate-All-Around FET) has led to a sharp increase in the complexity of compact models for devices like BSIM-CMG, significantly increasing the number of device parameters and lengthening the parameter extraction cycle. Faced with the challenge of high-dimensional parameter extraction at advanced process nodes, the industry mainly adopts two technical solutions: one is the traditional "local-then-global" manual extraction scheme. To cover the full-size range of the process design kit (PDK), this scheme first requires manually dividing the geometric space into multiple regions (Bin). Then, for typical dimensions within each Bin (e.g., L=16nm, W=100nm), the model parameters are repeatedly adjusted manually based on expert experience to fit the measured IV (current-voltage) and CV (capacitance-voltage) data, thereby extracting multiple sets of local model parameters. Finally, during circuit simulation, the actual parameters are calculated by interpolating the local parameters. This process is highly dependent on manual intervention, and the iteration cycle typically lasts for several weeks. The second approach is a purely data-driven inverse neural network parameter extraction scheme. This method abandons the constraints of traditional physical equations and constructs an inverse mapping network from electrical characteristics to model parameters. Specifically, it uses the device's test data (IV / CV) as input after characterization and the model parameters to be extracted as output for end-to-end training. It then uses the nonlinear fitting capability of deep neural networks to directly infer the binning parameters of the device.

[0028] However, traditional manual extraction methods are inefficient and suffer from geometric discontinuities: advanced node models such as BSIM-CMG contain over 100 key parameters with strong coupling between them. The "local-global" extraction process, reliant on expert experience, struggles to handle high-dimensional parameter spaces, resulting in extraction cycles lasting several weeks, which fails to meet the rapid iteration requirements of DTCO. Furthermore, the step-by-step process of local extraction and global fitting leads to parameter jumps (discontinuities) and non-differentiability at the boundaries of different bins, affecting the convergence of circuit simulations. On the other hand, purely data-driven approaches lack physical constraints and have poor extrapolation capabilities: pure neural network models lack a physical equation framework, making them highly susceptible to overfitting. The generated parameters may violate physical laws (e.g., exhibiting negative mobility or failing to satisfy the Zero-in, Zero-out characteristic), causing simulator errors. Simultaneously, this approach is highly dependent on the amount of data; its extrapolation prediction ability is extremely poor within bias voltage or temperature ranges not covered by the training data, failing to guarantee the physical correctness of the model. Furthermore, high-dimensional geometric scaling parameters are difficult to solve automatically: existing automated extraction tools struggle to directly handle complex geometric dependency formulas in BSIM models (such as the joint solution of LPARAM (word parameters), WPARAM (long integer parameters), etc.). Traditional regression methods often get stuck in local optima when faced with global fitting tasks involving multiple sizes and multiple parameters, making it difficult to obtain high-precision model parameters that describe the full-size characteristics of the device in one go.

[0029] To overcome the shortcomings of the aforementioned binning parameter extraction strategies, this application provides a method for extracting device binning parameters, referring to... Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the device binning parameter extraction method of this application. The device binning parameter extraction method includes: Step S10: Obtain the first geometric parameters of the first device, input the first geometric parameters into the pre-trained first fusion model, and predict and determine the observation vectors corresponding to each physical parameter in the first device through the first fusion model; The first device is the semiconductor device from which binning parameters need to be extracted. The first geometric parameter of the first device refers to the relevant parameters of the first device in the geometric dimension, such as channel length, channel width, and number of fins. The first fusion model is a pre-trained fusion model (which may include neural networks and physical equations) capable of predicting device parameters under arbitrary geometric dimensions (i.e., implicitly mastering the global scaling rules). After receiving the input first geometric parameter, the first fusion model can scan and generate intermediate physical parameter values ​​and the observation vectors corresponding to the intermediate physical parameter values ​​from the geometric space corresponding to the first geometric parameter, reflecting the physical characteristics of the semiconductor device.

[0030] The physical parameters may include one or more of the following: basic electrical parameters (e.g., threshold voltage, flat-band voltage), mobility (e.g., low-field mobility), conductivity parameters (e.g., carrier saturation velocity), geometry and size effect parameters (e.g., channel length modulation coefficient, short-channel effect parameters, narrow-channel effect parameters), and capacitance and parasitic parameters (e.g., gate oxide thickness, source / drain parasitic resistance, junction capacitance parameters).

[0031] Step S20: Based on the correspondence between the binning parameters and geometric dimensions of the first device, construct the coefficient matrix corresponding to each physical parameter. The correspondence between the binning parameters and geometric dimensions of the first device can be determined based on relevant common knowledge in the field. For example, the correspondence between the binning parameters and geometric dimensions can be obtained according to the Scaling rules defined in the BSIM model manual (such as the BSIM-CMG general formula), and then converted into a coefficient matrix form.

[0032] Step S30: Based on the observation vector and coefficient matrix, construct a system of linear equations for each physical parameter; It is understandable that there is a certain correlation between the observation vector and the coefficient matrix of the physical parameters corresponding to the same first device, and the correlation between the two is closely related to the binning parameters. Specifically, the product of the coefficient matrix and the parameter vector matrix is ​​equal to the observation vector. Based on this correlation, a system of linear equations for each physical parameter can be established.

[0033] Step S40: Solve each system of linear equations to obtain the binning parameters corresponding to each physical parameter.

[0034] After obtaining the linear equations for each physical parameter, various existing solution methods can be used, such as least squares method and singular value decomposition. Finally, all binning parameters (i.e., binning parameters) corresponding to each physical parameter are calculated.

[0035] It should be noted that when constructing the observation vector and coefficient matrix for the entire geometric space of the first device, the calculated binning parameters are global parameters. When a single global parameter cannot cover the entire size, a single global model is insufficient to describe the nonlinear information of the geometric space. Therefore, the geometric space can be divided into bins, and the binning parameters corresponding to each bin can be calculated to improve the accuracy of the binning parameters.

[0036] Furthermore, in a feasible embodiment, step S10, which involves predicting and determining the observation vectors corresponding to each physical parameter in the first device using the first fusion model, may include: Step S11: Spatial sampling is performed on the geometric space corresponding to the first geometric parameter through the first fusion model to generate multiple sets of physical parameter values, wherein the number of sets of physical parameter values ​​is equal to the preset number of sampling points; Step S12: For each physical parameter in the first device, generate an observation vector corresponding to each physical parameter based on the values ​​of each group of physical parameters.

[0037] The purpose of this application embodiment is to automatically extract the matrix modeling of Binning parameters based on the first fusion model, Fusion Mode. The goal is to explicitly transform the implicitly learned rules of the Fusion Model into Binning parameters specified by the BSIM standard. Before extracting the Binning parameters, geometric space sampling is required.

[0038] Specifically, during the geometric space sampling process, a corresponding geometric space is pre-constructed using the first geometric parameters. Then, a pre-trained Fusion Model is used to scan within the geometric space, generating N (N being the number of sampling points) sets of intermediate physical parameter values ​​under different (L, W) combinations. Let P be a physical parameter to be extracted (e.g., ...). Then, construct the corresponding observation vector B, which can be represented as: ; Here, "geometric space" refers to the multi-dimensional vector space constituted by the physical dimensional parameters of the device as defined by semiconductor process design rules. The specific definition is as follows: Spatial dimension: Includes, but is not limited to, key geometric variables such as channel length (L), channel width (W), number of fins (NFIN, for FinFET devices) or number of nanosheets (for GAA devices). It should be noted that the sampling range of the geometric space can cover everything from the smallest dimension allowed by the process (…). , ) to maximum size ( , A complete continuous interval. For example, a trained Fusion Model is used as a high-precision interpolator to perform high-density gridded scanning within the aforementioned geometric space. For instance, virtual test points are generated in steps of 5 nm within the L = [14 nm, 100 nm] interval. This sampling method overcomes the limitations of discrete and sparse geometric dimensions in actual chip fabrication testing, generating a sufficient number of samples (N sets of data) to meet the mathematical requirements of subsequently constructing a system of linear equations, ensuring the stability of the subsequent solution to the linear equations.

[0039] Furthermore, in a feasible embodiment, the correspondence between the binning parameters and the geometric dimensions is a functional relationship, and the step S20 of constructing the coefficient matrix corresponding to each physical parameter based on the correspondence between the binning parameters and the geometric dimensions of the first device may include: Step S21: Convert the functional relationship between the boxing parameters and geometric dimensions into matrix form to obtain the coefficient matrix corresponding to each physical parameter; The dimension of the coefficient matrix is ​​determined by the number of groups of physical parameter values ​​and the number of binning parameters.

[0040] For example, the functional relationship between the binning parameter P and the geometric dimensions can be expressed as: P= + / + / + / ( ); Wherein, P (Target Parameter) refers to the actual value of the target physical parameter to be extracted under a specific geometric size (e.g., the mobility corresponding to a specific L, W). ); (Reference Parameter) refers to the basic reference value, which usually represents the parameter value of a long-channel, wide-width device. This refers to the effective channel length and effective channel width of the device. (LPARAM) is a sensitivity coefficient that reflects the change of parameters with the reciprocal of the channel length (1 / L), and is used to characterize the short-channel effect. (WPARAM) is a sensitivity coefficient that reflects the change of parameters as a function of the inverse of the channel width (1 / W), and is used to characterize the narrow width effect. (PPARAM): The cross-term coefficient reflecting the variation of the parameter with the inverse of the area 1 / (L*W), used to characterize the coupling effect between length and width. The above... , , , These are the BinningParameters that are ultimately written into the SPICE (a simulation circuit simulator) model card, which are the objects to be solved by the system of equations in this application embodiment.

[0041] Furthermore, the process of converting the above functional relationship into matrix A is equivalent to extracting the coefficients from the above function. The resulting coefficient matrix A can be expressed as: ; The coefficient matrix A has a dimension of N×M, where N is the number of sampling points and M is the number of Binning parameters to be calculated.

[0042] In one feasible embodiment, step S30, which constructs a system of linear equations for each physical parameter based on the observation vector and the coefficient matrix, may include: Step S31: Initialize a one-dimensional parameter vector matrix, wherein the parameter vector matrix is ​​composed of unknowns corresponding to multiple bin parameters; Step S32: For each physical parameter, generate a system of linear equations based on the coefficient matrix, parameter vector matrix, and observation vector. The system of linear equations is used to represent that the product of the coefficient matrix and the parameter vector matrix is ​​equal to the observation vector.

[0043] Initializing a one-dimensional parameter vector matrix is ​​equivalent to defining the unknowns of the binning parameters to be computed. For example, the unknowns of the binning parameters may include the following parameters: , , as well as (corresponding to the previous text) , , , The above unknowns can be grouped into a one-dimensional parameter vector matrix X, which can be expressed as: ; It should be noted that the above method for constructing the parameter vector matrix can be extended to binning parameters that include other higher-order terms such as NPARAM. NPARAM refers to the parameter inversely proportional to the number of fins / nanofalves for FinFET / GAA devices (1 / The coefficient of change.

[0044] Furthermore, based on the previously determined coefficient matrix A and observation vector B, a linear system of equations AX=B can be constructed.

[0045] For the above system of linear equations, in general, since the value of N is usually much larger than the value of M, this is an overdetermined system of equations. Solving it using the least squares method or singular value decomposition (SVD) yields the following: .

[0046] In another feasible embodiment, after the step of obtaining the first geometric parameters of the first device, the method may further include: Step A10: Initialize the number of partitions to 2. Based on the preset multiple sampling points and number of partitions, divide the geometric space corresponding to the first geometric parameter into two sub-regions. Step A20: For each sub-region, the observation vector corresponding to each physical parameter in the first device is determined by prediction using the first fusion model. Step A30: For each sub-region, construct the coefficient matrix corresponding to each physical parameter based on the correspondence between the binning parameters and geometric dimensions of the first device; Step A40: For each sampling point, determine the local binning parameters corresponding to each sub-region based on the observation vector and coefficient matrix corresponding to each physical parameter. It should be noted that steps A10 to A40 and steps S10 to S40 are parallel technical solutions.

[0047] Steps S10 to S20 calculate the corresponding global binning parameters for the geometric space of the first device, while steps A10 to A40 divide the geometric space corresponding to the first geometric parameters into bins and calculate the binning parameters for each sub-region (the principle is the same as steps S10 to S40). This is because, generally, a single global model cannot cover the entire size, so binning is necessary. In this embodiment, the geometric space is divided into two sub-regions based on sampling points, and then the above matrix construction and solution process is repeated in each sub-region to automatically generate the local binning parameters corresponding to each sub-region.

[0048] It should be noted that when dividing the sub-regions in step A10, multiple different sampling points were used as boundary points to divide the region, resulting in multiple sets of sub-regions (each boundary point corresponds to one set, and each set contains two sub-regions).

[0049] For example, the multi-bin processing approach of this application embodiment can be based on an iterative boundary search algorithm for minimizing errors. Specific steps may include: Step A (candidate boundary generation): First, extract the unique geometric values ​​(such as all different L values) from all sampling points as candidate boundary points. Step B (Terrain Evaluation): Assuming we need to divide the dataset into k bins (e.g., k=2), we iterate through all possible combinations of splitting positions. Step C (Subdomain Regression): For each candidate boundary point, we divide the dataset into... and Two subsets are then used. Subsequently, coefficient matrices and observation vectors are independently constructed within each subset, and corresponding sets of phenomenon equations are constructed for each subset. Finally, the corresponding local Binning parameters are solved for each subset.

[0050] Step A50: Substitute the two sets of local binning parameters corresponding to each sampling point into the preset simulator to obtain the first simulation result; Step A60: Calculate the fitting error corresponding to each sampling point based on the first simulation result and the original test result of the first device corresponding to each sampling point. Step A70: Determine the local binning parameters corresponding to the sampling point with the smallest fitting error as the binning parameters of the first device.

[0051] After calculating the local binning parameters for each sub-region group, the optimal binning scheme is further screened. The screening criteria are based on simulation of the local binning parameters corresponding to the sub-region combination, comparing the simulation results with the original test results, calculating the fitting error (e.g., MAPE, Mean Absolute Percentage Error), and selecting the binning scheme corresponding to the sampling point with the smallest fitting error as the optimal binning scheme. The corresponding local binning parameters are then determined as the binning parameters of the first device.

[0052] The preset simulator can be a SPICE simulator. When the local binning parameters corresponding to each sampling point are input into the SPICE simulator, the simulator will output the first simulation result, which can be in the form of IV / CV data. The original test result is pre-prepared multi-size IV / CV test data. In calculating the fitting error, the measured current or capacitance under the same (L, W, Vd, Vg) conditions in the first simulation result can be compared with that in the original test result. Then, the current / capacitance in the first simulation result and the measured current / capacitance are substituted into the preset fitting error loss function to calculate the corresponding fitting error.

[0053] The simulation result determined by the local Binning parameter of the sampling point with the smallest fitting error is closest to the original test result, indicating that the Binning scheme is optimal and the corresponding local Binning parameter is also the most accurate. This achieves automatic optimal segmentation of geometric space and automatic optimization of Binning parameters.

[0054] In one feasible embodiment, after step S40 of solving each system of linear equations to obtain the binning parameters corresponding to each physical parameter, the method may further include: Step S50: Input the corresponding binning parameters of each physical parameter of the first device into the preset simulator for simulation to obtain the second simulation result of the first device; Step S60: Calculate the root mean square error of the second simulation result based on the second simulation result of the first device and the original test result of the first device; Step S70: If the root mean square error of the second simulation result is greater than or equal to the first error threshold, then return to step A10: Divide the geometric space corresponding to the first geometric parameter into two sub-regions according to the preset multiple sampling points and number of partitions.

[0055] As can be seen from the foregoing application embodiments, steps S10 to S40 provide a strategy for calculating global binning parameters. However, in some cases, global binning parameters are insufficient to describe the nonlinear information of geometric spaces of various sizes. In order to measure whether the global binning parameters are accurate, steps S50 to S70 provide a method for verifying the binning parameters corresponding to each of the physical parameters of the first device.

[0056] It should be noted that after obtaining the binning parameters corresponding to each physical parameter, the BSIM parameters can be written into the model card to generate the standard model card corresponding to the first device. Then, it can be run in the SPICE simulator to obtain the corresponding second simulation results (which may include electrical parameters such as voltage, current, and conductance). Then, the error of various electrical parameters under the same (L, W, Vd, Vg) conditions in the original test results and the second simulation results is compared and expressed in the form of root mean square error.

[0057] If the root mean square error (RMS) is less than the first error threshold (e.g., 5%), the accuracy of the currently obtained binning parameters is considered to meet the requirements; otherwise, a single global model is deemed insufficient to describe the nonlinear information of the geometric space. The system triggers a "binning" mechanism, automatically dividing the geometric space into two sub-regions (i.e., the method from steps A10 to A40), and returns to step A10. For each sub-region, matrix-based solutions are performed independently until the root mean square error of the simulation results corresponding to the local binning parameters of all sub-regions is less than the first error threshold.

[0058] For example, the second simulation result of the first device is compared with the original test result of the first device. Figure 2 As shown in the figure, the red circles correspond to the data of the original test results, and the blue continuous curve corresponds to the data of the second simulation results. Figure 2 The top left graph is used to characterize Ids-Vd (output characteristic curve). The horizontal axis represents Vd (drain voltage) and the vertical axis represents Ids (drain current). The unit is amperes (A), and the order of magnitude in the graph is 1e-5 A. Figure 2 The upper right graph is used to characterize the Ids-Vg (transfer characteristic curve, logarithmic scale), with Vg (gate voltage) on the horizontal axis and Ids (drain current) on the vertical axis, ranging from... A to A; Figure 2 The lower left figure is used to characterize the IGds-Vd (output conductance curve), with Vd (drain voltage) on the horizontal axis and Gds (output conductance) on the vertical axis. The unit is Siemens (s), and it is a logarithmic graph ranging from... S to S; Figure 2 The lower right figure is used to characterize the Gm-Vg (transconductance curve), with Vg (gate voltage) on the horizontal axis and Gm (transconductance) on the vertical axis, in Siemens (s). It is a linear graph ranging from 0 to 0.00015 S (i.e., 150 µS). As shown in the figure, the second simulation results for the first device are in excellent agreement with the original test results, indicating that the... Figure 2 The second simulation results shown have high accuracy, and the current binning parameters meet the accuracy requirements.

[0059] Furthermore, in another feasible embodiment, after step A70, which determines the local binning parameters corresponding to the sampling point with the smallest fitting error as the binning parameters of the first device, the method further includes: Step A71: Determine the local binning parameters corresponding to the sampling point with the smallest fitting error as the binning parameters of the first device, input them into the preset simulator for simulation, and obtain the third simulation result of the first device. Step A72: Calculate the root mean square error of the third simulation result based on the third simulation result of the first device and the original test result of the first device; Step A73: If the root mean square error of the third simulation result is greater than or equal to the second error threshold, then increase the number of sub-regions. Based on the preset number of sampling points and partitions, divide the geometric space corresponding to the first geometric parameter into multiple sub-regions, and return to execute step A20: For each sub-region, predict and determine the observation vector corresponding to each physical parameter in the first device through the first fusion model, until the root mean square error of the third simulation result is less than the second error threshold.

[0060] This application also provides a re-binding mechanism, which is an adaptive hierarchical refinement strategy designed to address the question of whether the current number of bins (e.g., 2) is sufficient to describe the nonlinear information of the device. As mentioned earlier, when the final error (RMS) calculated by the optimal 2-bin partitioning scheme is still higher than the second preset error threshold (e.g., 5%), it can be determined that the current model is underfitting, triggering the re-binding mechanism.

[0061] Specifically, the number of bins is increased (equivalent to increasing the number of sub-regions, one at a time, e.g., upgrading from 2-bin to 3-bin), and then the "multi-bin processing" process is re-invoked to find the optimal two dividing points in the geometric space (dividing the geometric space into three sub-regions). If the root mean square error of the third simulation result corresponding to the three sub-regions still does not meet the requirements, this process is repeated (e.g., from 1-bin to 2-bin, then to 3-bin...) until the root mean square error of the simulation result converges to below the second error threshold. This re-binning mechanism, which increases the number of bins from small to large, ensures that the model card generated by the local binning parameters meets the accuracy requirements without causing parameter redundancy due to excessive binning.

[0062] In another feasible embodiment, such as Figure 3 As shown, the device binning parameter extraction method may further include: Step B10: Initialize the second fusion model, wherein the second fusion model includes at least a micro neural network and a physical equation module; Step B20: Obtain the second geometric parameters and bias voltage of the second device; Step B30: Input the second geometric parameters into the micro neural network and input the bias voltage into the physical equation module; Step B40: The second geometric parameter is mapped to multiple sets of physical parameter values ​​using a micro neural network; Step B50: By combining the bias voltage with the values ​​of each set of physical parameters through the physical equation module, the first current data or the first capacitance data is output. Step B60: Based on the error between the first current data or the first capacitance data and the second current data or the second capacitance data in the original test results of the second device, the network parameters of the micro neural network are iteratively updated until the preset stopping condition is met, and the first fusion model is obtained.

[0063] This application provides a training method for a first fusion model. The second fusion model is a pre-initialized fusion model composed of a micro neural network and a physical equation module. The second geometric parameters and bias voltage of the second device are used as training data for the fusion model. It should be clarified that the training objective of the fusion model is to establish a model capable of implicitly learning the geometric dependencies of the devices through the neural network.

[0064] For the second fusion model, its input layer design includes using the second geometric parameters (channel length, channel width, number of fins, etc.) of the second device (the semiconductor device used to train the model) and the bias voltages (Vgs and Vds) as model inputs. It is important to note that the bias voltages are directly fed into the physical equations, while the micro-neural network only receives the geometric parameters. Furthermore, the micro-neural network employs a lightweight fully connected network to learn the nonlinear mapping from the geometric dimensions of semiconductor devices to intermediate physical parameters. It maps the input geometric parameters to corresponding intermediate physical parameter values ​​(i.e., a set of physical parameter values ​​corresponding to each physical parameter). The parameters output by the micro-neural network are then substituted into the core physical equations of the BSIM-CMG physical equation module (such as the long-channel equation and drift-diffusion equation) to calculate the first current or first capacitance data. The second current or capacitance data, obtained from pre-testing or simulation and reflecting the true multi-size IV / CV data of the second device, is then compared with the first current or first capacitance data. The error between the two is determined by calculating the root mean square error (RMSE). Based on the RMSE and the backpropagation algorithm, the model parameters of the micro-neural network (e.g., weights, learning rate, number of layers, etc.) are iteratively updated. The stopping condition can be that the error is below a preset error threshold (e.g., 5%), or that the RMSE converges or reaches a preset number of training rounds, or one or more of these conditions.

[0065] Finally, after the preset stopping conditions are met, the second fusion model is set as the first fusion model (FusionModel), which has the ability to predict device parameters under arbitrary geometric dimensions (i.e., implicitly grasps the GlobalScaling law).

[0066] To facilitate understanding, and in conjunction with the aforementioned embodiments, a feasible process for extracting and verifying device compartmentation parameters will be further explained, such as... Figure 4As shown, firstly, the test data (IV / CV curves) of the semiconductor device are used as training data to train the fusion model. The fusion model consists of a micro neural network and a physical equation skeleton (i.e., physical equation module). The current / capacitance data output by the physical equation skeleton is combined with the training data to backpropagate and optimize the parameters of the micro neural network. After the model is trained, during the automated binning parameter extraction process, the geometric parameters of the device (represented as (I,W)) are extracted into the trained fusion model. The model performs geometric space sampling and constructs the observation matrix B (composed of observation vectors) and the coefficient matrix A. Then, the global parameters (referring to global binning parameters) are solved. The global binning parameters are substituted into the SPICE simulator for running. The simulation results are compared with the original test IV / CV curves to calculate the root mean square error (RMS). If the RMS is less than 5%, complete model cards are generated based on the obtained global binning parameters or local binning parameters. Otherwise, the binning is re-divided, and the process of solving the local binning parameters is returned.

[0067] The binning parameter extraction method in this application utilizes a matrix-based parameter extraction approach based on a fusion model. Compared to traditional physical model iteration schemes, it shortens the modeling cycle from weeks to minutes. Compared to purely data-driven neural network schemes, it leverages prior physical knowledge to address the overfitting problem, reducing the reliance on massive training data to a single set of standard IV / CV curves sufficient for full-scale model training. Furthermore, the parallel extraction strategy based on backpropagation can perform a one-time weight update of over a hundred key parameters of the BSIM model, significantly improving extraction efficiency to meet the rapid iteration requirements of DTCO. In addition, the above embodiments propose a global parameter solution method based on a re-binning strategy and matrix regression, transforming the geometric scaling rules into a system of linear equations and obtaining all global binning parameters through a single matrix operation, resulting in even higher efficiency.

[0068] It should be noted that this example is only for the purpose of assisting in understanding this application and does not constitute a limitation on the device compartment parameter extraction method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0069] This application also provides a computer device, which includes at least: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the device binning parameter extraction method in the above embodiments.

[0070] The following is for reference. Figure 5It shows a schematic diagram of the structure of a computer device suitable for implementing the embodiments of this application. Figure 5 The computer device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of this application.

[0071] like Figure 5 As shown, the computer device may include a processing unit 101 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 102 or a program loaded from storage device 103 into random access memory (RAM) 104. RAM 104 also stores various programs and data required for the operation of the computer device. The processing unit 101, ROM 102, and RAM 104 are interconnected via bus 105. Input / output (I / O) interface 106 is also connected to the bus. Typically, the following systems can be connected to I / O interface 106: input devices 107 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 108 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 103 including, for example, magnetic tapes, hard disks, etc.; and communication devices 109. Communication device 109 allows the computer device to exchange data via wireless or wired communication with other devices. Although the diagram shows computer equipment with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems may be implemented alternatively.

[0072] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 103, or installed from ROM 102. When the computer program is executed by processing device 101, it performs the functions defined in the methods of the embodiments of this application.

[0073] The computer device provided in this application, employing the device binning parameter extraction method described in the above embodiments, can improve the accuracy of semiconductor device binning parameter extraction. Compared with the prior art, the beneficial effects of the computer device provided in this application are the same as those of the device binning parameter extraction method provided in the above embodiments, and other technical feature values ​​in this computer device are the same as those disclosed in the method of the above embodiments, and will not be repeated here.

[0074] It should be understood that various parts of the embodiments of this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0075] The above description is merely a specific implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of the embodiments of this application. Therefore, the protection scope of the embodiments of this application should be determined by the protection scope of the above claims.

[0076] This application also provides a computer-readable storage medium storing a computer program that can run on a processor. The computer program is used to execute the device binning parameter extraction method in the above embodiments.

[0077] The computer-readable storage medium provided in this application embodiment may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0078] The aforementioned computer-readable storage medium may be included in a computer device or may exist independently and not assembled into a computer device.

[0079] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by a computer device, the computer device causes the computer device to: acquire the first geometric parameters of the first device; input the first geometric parameters into a pre-trained first fusion model; predict and determine the observation vectors corresponding to each physical parameter in the first device through the first fusion model; construct a coefficient matrix corresponding to each physical parameter based on the correspondence between the binning parameters and geometric dimensions of the first device; construct a system of linear equations for each physical parameter based on the observation vectors and the coefficient matrices; and solve each system of linear equations to obtain the binning parameters corresponding to each physical parameter.

[0080] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0081] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0082] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0083] The computer-readable storage medium provided in this application embodiment stores computer-readable program instructions for executing the above-described device binning parameter extraction method, which can improve the accuracy of semiconductor device binning parameter extraction. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application embodiment are the same as the beneficial effects of the device binning parameter extraction method provided in the above embodiments, and will not be repeated here.

[0084] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the device binning parameter extraction method described above.

[0085] The computer program product provided in this application can improve the accuracy of extracting semiconductor device binning parameters. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the device binning parameter extraction method provided in the above embodiments, and will not be repeated here.

[0086] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent scope of this application.

Claims

1. A method for extracting device compartment parameters, characterized in that, The method for extracting device compartment parameters includes: The first geometric parameters of the first device are obtained, and the first geometric parameters are input into the pre-trained first fusion model. The first fusion model is used to perform spatial sampling on the geometric space corresponding to the first geometric parameters to generate multiple sets of physical parameter values. For each physical parameter in the first device, an observation vector corresponding to each physical parameter is generated based on each set of physical parameter values; The functional relationship between the boxing parameters and geometric dimensions is converted into matrix form to obtain the coefficient matrix corresponding to each physical parameter; Based on the observation vector and the coefficient matrix, a system of linear equations is constructed for each physical parameter; Solve the linear equations to obtain the binning parameters corresponding to each physical parameter.

2. The device compartment parameter extraction method as described in claim 1, characterized in that, The number of sets of physical parameter values ​​is equal to the preset number of sampling points.

3. The device compartment parameter extraction method as described in claim 1, characterized in that, The dimension of the coefficient matrix is ​​determined by the number of groups of physical parameter values ​​and the number of binning parameters.

4. The device compartment parameter extraction method as described in claim 1, characterized in that, The step of constructing a system of linear equations for each physical parameter based on the observation vector and the coefficient matrix includes: Initialize a one-dimensional parameter vector matrix, wherein the parameter vector matrix is ​​composed of unknowns corresponding to multiple bin parameters; For each physical parameter, a system of linear equations is generated based on the coefficient matrix, the parameter vector matrix, and the observation vector, wherein the system of linear equations represents the product of the coefficient matrix and the parameter vector matrix as equal to the observation vector.

5. The device compartment parameter extraction method as described in claim 1, characterized in that, After the step of obtaining the first geometric parameters of the first device, the method further includes: The number of partitions is initialized to 2. Based on the preset multiple sampling points and the number of partitions, the geometric space corresponding to the first geometric parameter is divided into two sub-regions. For each sub-region, the observation vector corresponding to each physical parameter in the first device is determined by prediction using the first fusion model; For each sub-region, a coefficient matrix corresponding to each physical parameter is constructed based on the correspondence between the binning parameters and geometric dimensions of the first device; For each sampling point, the local binning parameters corresponding to each sub-region are determined based on the observation vector and coefficient matrix corresponding to each physical parameter. Substitute the two sets of local binning parameters corresponding to each sampling point into the preset simulator to obtain the first simulation result; Based on the first simulation results corresponding to each sampling point and the original test results of the first device, calculate the fitting error corresponding to each sampling point. The local binning parameters corresponding to the sampling point with the smallest fitting error are determined as the binning parameters of the first device.

6. The device compartment parameter extraction method as described in claim 5, characterized in that, After the step of determining the local binning parameters corresponding to the sampling point with the smallest fitting error as the binning parameters of the first device, the method further includes: The binning parameters corresponding to each of the physical parameters of the first device are input into a preset simulator for simulation to obtain the second simulation result of the first device. Based on the second simulation result of the first device and the original test result of the first device, calculate the root mean square error of the second simulation result; If the root mean square error of the second simulation result is greater than or equal to the first error threshold, then return to the execution step: divide the geometric space corresponding to the first geometric parameter into two sub-regions according to the preset multiple sampling points and number of partitions.

7. The device compartment parameter extraction method as described in claim 5, characterized in that, After the step of determining the local binning parameters corresponding to the sampling point with the smallest fitting error as the binning parameters of the first device, the method further includes: The local binning parameters corresponding to the sampling point with the smallest fitting error are determined as the binning parameters of the first device and input into a preset simulator for simulation to obtain the third simulation result of the first device. Based on the third simulation result of the first device and the original test result of the first device, calculate the root mean square error of the third simulation result; If the root mean square error of the third simulation result is greater than or equal to the second error threshold, then the number of sub-regions is increased. Based on the preset number of sampling points and partitions, the geometric space corresponding to the first geometric parameter is divided into multiple sub-regions, and the execution steps are returned: for each sub-region, the observation vector corresponding to each physical parameter in the first device is determined by the first fusion model prediction until the root mean square error of the third simulation result is less than the second error threshold.

8. The method for extracting device compartment parameters as described in any one of claims 1 to 7, characterized in that, The method for extracting device binning parameters also includes: Initialize the second fusion model, wherein the second fusion model includes at least a micro neural network and a physical equation module; Obtain the second geometric parameters and bias voltage of the second device; The second geometric parameter is input into the micro neural network, and the bias voltage is input into the physical equation module; The second geometric parameter is mapped to multiple sets of physical parameter values ​​through the micro neural network; By combining the bias voltage with each set of physical parameter values ​​through the physical equation module, the first current data or the first capacitance data is output. Based on the error between the first current data or the first capacitance data and the second current data or the second capacitance data in the original test results of the second device, the network parameters of the micro neural network are iteratively updated until a preset stopping condition is met, thus obtaining the first fusion model.

9. A computer device, characterized in that, The computer device includes at least: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the device binning parameter extraction method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for implementing a device binning parameter extraction method, which is executed by a processor to implement the steps of the device binning parameter extraction method as described in any one of claims 1 to 8.

Images