Method, device, electronic equipment, medium and product for controlling semiconductor process

Abstract

The present application relates to the technical field of semiconductor process, and discloses a kind of control method, device, electronic equipment, medium and product of semiconductor process, method includes: obtaining current multi-dimensional detection data of current measurement period;Current process feature vector is extracted from current multi-dimensional detection data;State deviation amount under current process parameter is determined based on current process feature vector and hierarchical mixed model, hierarchical mixed model includes mechanism model and compensation model, mechanism model is used to characterize the corresponding relationship of process parameter and process feature vector;Compensation model is used to correct mechanism model, and the output is state deviation amount;Based on state deviation amount, determine the drift index in the execution process of semiconductor process;If drift index is greater than preset drift index, based on state deviation amount, adjust current process parameter, make the execution state of semiconductor process return target execution state.The present application can realize the accurate closed-loop control of process, improve product yield.

Description

Technical Field

[0001] This invention relates to the field of semiconductor process technology, and more specifically to a semiconductor process control method, apparatus, electronic device, medium, and product. Background Technology

[0002] As the integration level and manufacturing capacity of semiconductor structures increase, semiconductor manufacturing processes are becoming increasingly complex. The control of process parameters during semiconductor manufacturing is directly related to whether the yield of semiconductor devices can be improved.

[0003] In semiconductor manufacturing processes, such as etching, depth / thickness data is acquired through methods like Laser Etch Profilometry (LEP) to reflect the process state. Then, process parameters are adjusted based on process mechanism models (such as plasma dynamics models) to bring the process state closer to the target state. However, one-dimensional or limited physical quantities are insufficient to accurately reflect the process state. Furthermore, actual processes involve complex physicochemical reactions, equipment state drift, and chamber aging. Process mechanism models struggle to precisely describe the relationship between process parameters and process state, resulting in low accuracy of the final process parameters and necessitating frequent calibration. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, medium, and product for controlling semiconductor processes, in order to solve the problem of low accuracy of process parameters output by process mechanism models, which requires frequent calibration.

[0005] In a first aspect, the present invention provides a method for controlling a semiconductor process, the method comprising: acquiring current multidimensional detection data of the current measurement cycle when fabricating a semiconductor structure using a semiconductor process, wherein the current multidimensional detection data includes three-dimensional topographic data of the semiconductor structure and reflectivity data reflecting the film thickness; extracting a current process feature vector from the current multidimensional detection data, wherein the current process feature vector is used to reflect the current execution state of the semiconductor process; determining a state deviation amount under the current process parameters based on the current process feature vector and a hierarchical hybrid model, wherein the hierarchical hybrid model includes a mechanistic model and a compensation model, the mechanistic model is used to characterize the correspondence between process parameters and process feature vector, the input is the current process parameters, and the output is a predicted process feature vector under the current process parameters; the compensation model is used to correct the mechanistic model, the input is a deviation vector between the predicted process feature vector and the current process feature vector and the current process parameters, and the output is a state deviation amount; determining a drift index during the execution of the semiconductor process based on the state deviation amount; if the drift index is greater than a preset drift index, adjusting the current process parameters based on the state deviation amount to obtain the next process parameters, so that the execution state of the semiconductor process returns to the target execution state.

[0006] In one optional implementation, adjusting the current process parameters based on the state deviation includes: obtaining an initial adjustment amount from an optimal parameter lookup table based on the state deviation, wherein the optimal parameter lookup table includes multiple correspondences between different state deviations and process parameter adjustment amounts; optimizing the target adjustment amount in the neighborhood of the initial adjustment amount using gradient descent based on a cost function and constraints of the process parameters, wherein the cost function is used to characterize the deviation between the next process feature vector and the target process feature vector corresponding to the target execution state, the next process feature vector is the predicted process feature vector corresponding to the adjusted current process parameters, and the prediction model is determined based on a hierarchical hybrid model; and adjusting the current process parameters based on the target adjustment amount to obtain the next process parameters.

[0007] In one optional implementation, based on the constraints of the cost function and process parameters, gradient descent is used to optimize within the neighborhood of the initial adjustment amount to obtain the target adjustment amount. This includes: if the number of optimizations using gradient descent is less than or equal to a preset number, and the deviation value corresponding to the cost function is less than or equal to a first preset threshold, then the adjustment amount corresponding to the target deviation value is determined as the target adjustment amount, where the target deviation value is a deviation value less than or equal to the first preset threshold; if the number of optimizations is greater than a preset number, and the deviation value corresponding to the cost function is greater than the first preset threshold, then the adjustment amount corresponding to the minimum deviation value is determined as the target adjustment amount.

[0008] In an optional implementation, the method further includes: if an initial adjustment amount is not obtained from the optimal parameter lookup table, then the current process parameter is determined as the next process parameter; if an initial adjustment amount is not obtained from the optimal parameter lookup table, or if the number of optimizations performed using the gradient descent method is less than or equal to a preset number, and the deviation value corresponding to the cost function is greater than a first preset threshold, then based on a first multi-objective optimization algorithm, the process parameter adjustment amount corresponding to the state deviation amount is determined; and based on the process parameter adjustment amount corresponding to the state deviation amount, the optimal parameter lookup table is updated.

[0009] In one alternative implementation, the cost function includes a penalty, which is characterized by the difference between the previous process parameter in the previous measurement cycle and the current process parameter.

[0010] In an optional implementation, before obtaining the initial adjustment amount from the optimal parameter lookup table based on the state deviation, the method further includes: determining the optimal parameter combination under different execution states based on a second multi-objective optimization algorithm and historical process parameters; and determining and storing the optimal parameter lookup table based on the optimal parameter combination under different execution states.

[0011] In one optional implementation, acquiring current multidimensional detection data for the current measurement period includes: acquiring three-dimensional topographic data for the current measurement period from a dual-wavelength phase-shifting structured light projection device; and acquiring reflectivity data reflecting the film thickness for the current measurement period using multi-wavelength reflection spectroscopy.

[0012] In one alternative implementation, the projection wavelength emitted by the dual-wavelength phase-shift structured light projection device differs from the characteristic emission spectral lines in the semiconductor process.

[0013] In one alternative implementation, the dual-wavelength phase-shift structured light projection device emits a projection wavelength to detect three-dimensional topography data when the radio frequency power switch in the semiconductor process equipment is off or during the period of minimum plasma afterglow in the semiconductor process equipment.

[0014] In one optional implementation, the current process feature vector includes the current defect feature, and the method further includes: fitting and determining the rate of change of defect density over time based on the current defect feature and multiple historical defect features prior to the current measurement cycle; if the slope of the rate of change of defect density over time is positive and greater than a second preset threshold in multiple consecutive measurement cycles, then a first alarm signal is output.

[0015] In one optional implementation, the method further includes: inputting the current defect characteristics into a lifetime prediction model, predicting the remaining lifetime of the semiconductor process equipment based on the lifetime prediction model; and outputting a second alarm signal if the remaining lifetime is less than a preset lifetime.

[0016] In one alternative implementation, the semiconductor process is an etching process, and the current process feature vector includes at least one of depth features, sidewall features, bottom features, film thickness features, and defect features.

[0017] Secondly, the present invention provides a control device for a semiconductor process, the device comprising: an acquisition module, configured to acquire current multidimensional detection data of the current measurement cycle when fabricating a semiconductor structure using a semiconductor process, wherein the current multidimensional detection data includes three-dimensional morphology data of the semiconductor structure and reflectivity data reflecting the film thickness; an extraction module, configured to extract a current process feature vector from the current multidimensional detection data, wherein the current process feature vector reflects the current execution state of the semiconductor process; and a deviation determination module, configured to determine the state deviation amount under the current process parameters based on the current process feature vector and a hierarchical hybrid model, wherein the hierarchical hybrid model includes a mechanism. The system comprises a model and a compensation model. The mechanistic model characterizes the correspondence between process parameters and process feature vectors, taking the current process parameters as input and outputting the predicted process feature vector under the current process parameters. The compensation model corrects the mechanistic model, taking the deviation vector between the predicted and current process feature vectors and the current process parameters as input, and outputting the state deviation. The drift determination module determines the drift index during semiconductor process execution based on the state deviation. The parameter adjustment module adjusts the current process parameters based on the state deviation if the drift index exceeds a preset drift index, obtaining the next process parameters to bring the semiconductor process execution state back to the target execution state.

[0018] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the control method of the semiconductor process described in the first aspect or any corresponding embodiment thereof.

[0019] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to perform the control method of the semiconductor process described in the first aspect or any corresponding embodiment thereof.

[0020] Fifthly, the present invention provides a computer program product, including computer instructions for causing a computer to execute a control method for a semiconductor process according to the first aspect or any corresponding embodiment thereof.

[0021] The semiconductor process control method, apparatus, electronic device, medium, and product provided by this invention have at least the following effects: This invention extracts process feature vectors from multidimensional detection data to reflect the process state. Compared with single-dimensional information, it can more accurately reflect the real-time processing state of semiconductor structures. At the same time, it uses a mechanistic model to predict the predicted process feature vector under the current process parameters. Then, based on the deviation vector between the predicted process feature vector and the extracted current process feature vector, a compensation model is used to determine the state deviation under the current process parameters. Compared with the method of directly determining the state deviation based on the mechanistic model, this invention further considers the influence of local effects that are difficult for the mechanistic model to describe accurately. The accuracy of the obtained state deviation is higher, so that process parameters can be more accurately controlled during process drift, so that the process state returns to the target trajectory.

[0022] This invention can improve the yield of semiconductor structure fabrication, improve intra-batch uniformity by more than 30%, and reduce critical dimension deviation by 50%; save costs by reducing offline measurement and rework costs by about 40% and improving equipment utilization by 20%; enhance process portability by shortening the matching time of the same formula between different machines by 60% through adaptive calibration of the layered mixing model.

[0023] Furthermore, when adjusting process parameters, an optimal parameter lookup table created through offline global search is first used to determine the initial adjustment amount. Then, the gradient descent algorithm is used for online local optimization of the initial adjustment amount to obtain the target adjustment amount that enables the process to reach the target execution state. This can significantly reduce optimization time and improve real-time control response speed. The closed-loop time from detection to parameter adjustment in this invention is less than 5 seconds, which can effectively prevent defects from occurring.

[0024] Furthermore, a penalty is introduced into the cost function used in the optimization process, which can achieve a balance between control accuracy and process stability.

[0025] Furthermore, after adjusting the process parameters, this invention also monitors the defect process by using the second derivative of the defect trend, which can capture early signals of accelerated defect deterioration and intervene before the defect deteriorates irreversibly, reducing the probability of major failures and improving process stability and product yield.

[0026] Furthermore, this invention monitors the remaining lifespan of semiconductor process equipment based on a lifespan prediction model, enabling the early replacement of worn-out devices and avoiding the impact of unplanned downtime on the production line.

[0027] Experimental verification showed that, after adopting the semiconductor process control method provided by this invention, the standard deviation of etching depth decreased from ±7.5nm (uncontrolled) to ±2.1nm, and uniformity improved by 72%. The defect rate did not accelerate in the later stages of the process, and predictive maintenance provided an early warning of chamber contamination 10 days in advance, avoiding sudden downtime. The average batch yield increased from 86.3% to 95.8%, an improvement of 9.5 percentage points. Attached Figure Description

[0028] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0029] Figure 1 This is a schematic flowchart of a semiconductor process control method according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the determination of state deviation based on a hierarchical hybrid model according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating another semiconductor process control method according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the adaptive control decision module adjusting process parameters according to an embodiment of the present invention; Figure 5 This is a schematic flowchart of another semiconductor process control method according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the training and prediction processes of the Bi-LSTM model according to an embodiment of the present invention; Figure 7 This is a schematic flowchart of another semiconductor process control method according to an embodiment of the present invention; Figure 8 This is a comparative schematic diagram of the etching depth control effect according to an embodiment of the present invention; Figure 9 This is a comparative schematic diagram of the etching uniformity within a batch according to an embodiment of the present invention; Figure 10 This is a schematic diagram illustrating the trend of defect rate over time according to an embodiment of the present invention; Figure 11 This is a schematic diagram of the remaining lifetime prediction curve based on the defect rate according to an embodiment of the present invention; Figure 12 This is a schematic diagram comparing batch yield before and after control according to an embodiment of the present invention; Figure 13 This is a structural block diagram of a semiconductor process control device according to an embodiment of the present invention; Figure 14 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0032] This invention provides a method, apparatus, electronic device, dielectric, and product for controlling semiconductor processes. By simultaneously acquiring multi-dimensional data such as three-dimensional topography and thin film thickness, and through domain knowledge-driven fusion analysis and intelligent optimization, process parameters are dynamically adjusted to achieve precise closed-loop control of the process, thereby improving product yield and equipment utilization. Furthermore, the multi-dimensional data can also include defect distribution; by analyzing the defect distribution, predictive maintenance can be achieved, avoiding unexpected downtime.

[0033] According to an embodiment of the present invention, a method for controlling a semiconductor process is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0034] This embodiment provides a method for controlling a semiconductor process, which can be used in electronic devices such as tablet computers, computers, or servers. Figure 1 This is a schematic flowchart of a semiconductor process control method according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S101: When fabricating a semiconductor structure using semiconductor technology, acquire the current multidimensional detection data for the current measurement cycle.

[0035] Specifically, semiconductor processes can be etching processes, deposition processes, photolithography processes, or other processes required to fabricate semiconductor structures. Current multidimensional detection data includes three-dimensional morphological data of semiconductor structures and reflectivity data reflecting film thickness.

[0036] During semiconductor manufacturing processes, integrated optical inspection systems acquire data such as the three-dimensional morphology and film thickness of the semiconductor structure surface according to measurement cycles. The measurement cycle can be 5 seconds or 10 seconds, etc. Electronic devices obtain the current multidimensional detection data of the current measurement cycle from the integrated optical inspection system. The current multidimensional detection data is the multidimensional detection data acquired in the measurement cycle corresponding to the current moment (i.e., the current measurement cycle).

[0037] Step S102: Extract the current process feature vector from the current multidimensional detection data.

[0038] The current process feature vector reflects the current execution state of the semiconductor process. It is the process feature vector corresponding to the current measurement cycle and reflects the execution state (process state) of the semiconductor process in the current measurement cycle. The execution state refers to the state of the processed surface of the semiconductor structure after processing the semiconductor structure using the current process parameters. Process parameters may include at least one of RF power, gas flow rate, chamber pressure, and temperature.

[0039] The current process feature vector extracted from the current multidimensional data is not a general image statistic, but a key physical indicator used by semiconductor process engineers to diagnose the process status. When the semiconductor process is an etching process, the current process feature vector is taken as a chip cell or test pattern and includes at least one of the following: depth features, sidewall features, bottom features, film thickness features, and defect features.

[0040] Depth features may include average etching depth D and depth uniformity. And at least one of the pattern density effect indices, the average etching depth D refers to the average depth of all etching trenches in the measurement area, used to reflect the etching progress; depth uniformity The standard deviation of the depth of all etched trenches within the measurement area is used to reflect the in-plane uniformity of the etching rate; the pattern density effect index is the ratio of the depth of densely measured areas to that of isolated measured areas, used to quantify the loading effect (RIE-1ag).

[0041] Sidewall features may include sidewall angles Sidewall roughness And at least one of the sidewall protective layer thickness evaluation indicators, sidewall angle The sidewall roughness is obtained by fitting the straight line of the etched trench sidewall profile and is used to reflect the anisotropy of the etching. The sidewall roughness refers to the root mean square roughness of the etched trench sidewall profile and is used to reflect the quality of the passivation layer. The sidewall protective layer thickness evaluation index is obtained by reconstructing the etched trench sidewall profile by multi-angle structured light and indirectly estimating the passivation layer deposition status.

[0042] Bottom features may include at least one of bottom flatness (PV) and microgroove budding index. PV refers to the peak-to-valley value of the bottom height of the etched trench, reflecting the flatness of the bottom surface. The microgroove budding index is the deviation of the local depth from the average depth within a region 0.5 μm from the sidewall, and tracks the growth rate of this deviation over time. Microgrooves can be located at the bottom of the etched trench, with a depth of 3%-10% (e.g., 20 nm-200 nm) of the total depth of the etched trench.

[0043] Film thickness characteristics can include the mean residual film thickness T and film thickness uniformity. At least one of the following, the average remaining film thickness T refers to the remaining thickness of the etching stop layer or mask layer, and the film thickness uniformity. It refers to the standard deviation of film thickness at multiple test points within the measurement area, used to reflect the in-plane variation of etching selectivity.

[0044] Defect characteristics may include defect density. Maximum defect size Defect growth rate At least one of the following: defect density and defect spatial distribution entropy. This refers to the number of defects per square centimeter; the maximum defect size. This refers to the maximum defect size; the defect growth rate. It refers to the rate of change of defect density between consecutive batches, used to assess the rate of contamination in the chamber; the spatial distribution entropy of defects refers to the degree of spatial aggregation of defects. High entropy (greater than or equal to the preset distribution entropy) indicates uniform distribution, while low entropy (less than the preset distribution entropy) indicates local aggregation. Different defect types correspond to different failure mechanisms.

[0045] Step S103: Based on the current process feature vector and the hierarchical hybrid model, determine the state deviation under the current process parameters.

[0046] The hierarchical hybrid model includes a mechanistic model and a compensation model. The mechanistic model is used to characterize the correspondence between process parameters and process feature vectors. The input is the current process parameters, and the output is the predicted process feature vector under the current process parameters. The compensation model is used to correct the mechanistic model. The input is the deviation vector between the predicted process feature vector and the current process feature vector, as well as the current process parameters. The output is the state deviation.

[0047] Specifically, this invention uses a primary-secondary hierarchical hybrid modeling approach to establish a quantitative relationship between process feature vectors and process parameters, rather than simple residual compensation.

[0048] The primary model is the mechanistic model, which is used to capture the dominant trend of process parameters and ensure the extrapolation capability of the mechanistic model under unseen operating conditions. The secondary model is the compensation model, which is used to compensate for local effects that the mechanistic model cannot accurately describe (or does not consider). Local effects can refer to pattern density effects, chamber wall deposition state, window transmittance changes, and equipment individual differences, etc.

[0049] The compensation model does not directly correct the predicted process feature vector output by the mechanism model. Instead, it corrects the key adjustable parameters in the mechanism model (such as the reaction probability coefficient and ion flux efficiency coefficient), giving the mechanism model adaptive calibration capability. This mechanism avoids treating the mechanism model as a complete black box and maintains the physical interpretability of the mechanism model.

[0050] The process feature vector contains features of different dimensions, and the mechanistic model can be a collection of sub-models corresponding to multiple features of different dimensions. In other words, the main model is not a single equation, but a family of multi-output models that takes process parameters as input and outputs process feature vectors such as depth, uniformity, and sidewall angle to represent the process state.

[0051] Taking semiconductor etching as an example, where the process feature vector includes depth, sidewall angle, and uniformity, the mechanism model can be constructed by using the time integral of the etching rate to obtain the depth, using the rate difference at different locations (radial, azimuth angle) to obtain the uniformity, and using the ion incident angle and sidewall reaction probability to obtain the sidewall angle. The specific construction process is as follows: Step a1: Based on plasma physics and surface reaction kinetics, a simplified mechanism equation for the etching rate is constructed.

[0052] For example, for silicon etching (assuming ion-assisted etching is dominant), the etching rate ER can be expressed as shown in formula (1):

[0053] In the formula, Represents the proportionality coefficient. Indicates ion flux, and radio frequency power. and chamber pressure Related, This represents the energy-dependent yield of ions. This represents the surface coverage of the reactants and is related to the gas flow rate F. Based on experimental data, the above formula (1) can be simplified to formula (2):

[0054] In the formula, The power index is used to reflect the trend of plasma density as power increases. This represents the pressure index, used to reflect the relationship between ion flux and pressure. The apparent activation energy is used to reflect the effect of temperature on the reaction rate. This represents the ideal gas constant, typically 8.314 J / (mol·K). This indicates the temperature of the semiconductor structure.

[0055] It should be noted that, It is used to reflect the combined effect of all factors other than explicit parameters (power, pressure, gas, temperature) on the etching rate, including chamber geometry (electrode spacing, gas inlet distribution), material coefficients (probability of silicon's reaction to different free radicals) and equipment aging status (electrode surface roughness, polymer deposition thickness), etc. Not an arbitrary constant, during model calibration, It was obtained by fitting experimental data; in real-time control... Dynamic updates are identified online using recursive least squares (RLS) to reflect the drift in the chamber state.

[0056] Step a2: Based on the simplified mechanism equation of etching rate, establish a depth integral model and use the depth integral model as a sub-model corresponding to the depth feature.

[0057] Specifically, using a recursive formula, the depth integral model can be represented as shown in formula (3):

[0058] In the formula, This indicates the predicted etching depth for the next measurement cycle under the current process parameters. This indicates the etching depth in the current measurement cycle. This indicates the etching rate within the current measurement cycle. Indicates the time interval of the measurement period. Measured.

[0059] Step a3: Based on the simplified mechanism equation of etching rate, establish a uniformity model and use the uniformity model as a sub-model corresponding to the uniformity characteristics.

[0060] Specifically, the semiconductor structure (wafer) is divided into three regions (central region, edge region, and transition region between the central and edge regions), and the etching rate correction factor for each region is assigned. The correction factor for the central region is... f center The revision factor for the transition region is f middle The revision factor for the edge region is f edge Due to the micro-load effect, f center> f edge Then the region i etching rate As shown in formula (4):

[0061] In the formula, Indicates the region i The revision factor, This represents the etching rate without considering the regional effects. The uniformity (standard deviation) is calculated from the difference in the rates of each region.

[0062] Step a4: Based on the ion incident angle distribution and anisotropic etching ratio model, the sidewall angle is determined by the process parameters.

[0063] Specifically, the sub-model corresponding to the sidewall angle can be shown in formula (5):

[0064] In the formula, This indicates the sidewall angle of the etched trench after the etching process is performed according to the current process parameters. Indicates the maximum sidewall angle. Indicates the pressure influence coefficient. This indicates the actual working pressure of the etching reaction chamber. Indicates the reference chamber pressure. This represents the passivation ratio influence coefficient. , This indicates the flow rate of the passivating gas. This indicates the flow rate of the etching gas.

[0065] For example, the present invention can use a gradient boosting tree (XGBoost) as a compensation model. During training, the input of the compensation model is the current process parameters and the historical feature sequence. The initial output is the deviation vector between the process feature vector predicted by the mechanism model and the current process feature vector. Based on the deviation vector, the adjustable model parameters in the mechanism model are corrected to obtain the corrected deviation vector. After the mechanism model and the compensation model converge, the output of the compensation model is the predicted deviation vector (i.e., the state deviation).

[0066] XGBoost can handle nonlinearity and feature interactions, and provides a ranking of feature importance, making it easier for process engineers to understand. The historical feature sequence refers to the process feature vectors (historical feature vectors) obtained from multiple measurement cycles prior to the current measurement cycle, as well as the first-order difference (rate of change) and second-order difference (acceleration) of the historical feature vectors.

[0067] Historical feature sequences are used to enable compensation models (such as XGBoost) to learn the dynamic evolution of the process, rather than just the current static state. For example, if the etching rate has decreased for three consecutive measurement cycles, even if the current rate is still within the threshold, it indicates possible chamber contamination, and the positive second derivative of the sidewall angle indicates that anisotropic loss is occurring.

[0068] Furthermore, the input to the compensation model can also include the identifier of the semiconductor process equipment, which refers to the unique number or ID of the etching machine. Even when running the same recipe, different machines will have different process performance due to differences in hardware tolerances, aging levels, and maintenance history. By adding the identifier of the semiconductor process equipment to the model input when predicting process feature vectors, the model can automatically adjust the output bias or slope for different machines by learning the equipment's identifier, capturing the individual differences of the equipment, and improving the accuracy of the model's predictions.

[0069] For example, suppose there are two etching machines, A and B. Machine A has a newer chamber, while machine B's chamber has been used for two years without cleaning. Under the same process parameters, the actual etching rate of machine B is 5% lower than that of machine A. If the model is not given a device identifier, it will misjudge this as process drift and adjust the parameters incorrectly. However, by giving the device identifier, the model can learn that for machine B, the rate prediction should be compensated by an additional 5%, thus avoiding erroneous actions.

[0070] Step S104: Determine the drift index during the semiconductor process based on the state deviation.

[0071] Specifically, if the state deviation is less than or equal to a preset deviation, there is no drift in the semiconductor process, and the drift index is 0. If the state deviation is greater than the preset deviation, there is drift in the semiconductor process, and the drift index is proportional to the state deviation; the larger the state deviation, the larger the drift index. For example, a depth deviation exceeding 5nm can be marked as drift, and a sidewall angle deviation exceeding 0.5° can also be marked as drift.

[0072] Step S105: If the drift index is greater than the preset drift index, then adjust the current process parameters based on the state deviation amount to obtain the next process parameters, so that the execution state of the semiconductor process returns to the target execution state.

[0073] Specifically, when the drift index is greater than the preset drift index, the current process parameters are optimized to make the state deviation less than the preset deviation. The optimized current process parameters are then used as the next process parameters, and the semiconductor process is executed to fabricate the semiconductor structure. If the state deviation is less than the preset deviation, it indicates that the execution state of the semiconductor process meets expectations, and the process state can return to the target trajectory. When the drift index is less than or equal to the preset drift index, it indicates that the current process state is still on the target trajectory. The next measurement cycle is then used as the current measurement cycle, and steps S101 to S105 are re-executed.

[0074] For example, the state deviation includes multiple sub-deviations in different dimensions. In this case, the drift index can include a sub-drift index corresponding to the sub-deviation in each dimension. If any sub-drift index is greater than the corresponding preset drift index, the current process parameters are adjusted based on the state deviation to obtain the next process parameters.

[0075] The semiconductor process control method provided in this embodiment extracts process feature vectors from multi-dimensional detection data to reflect the process state. Compared with single-dimensional information, it can more accurately reflect the real-time processing state of the semiconductor structure. At the same time, it uses a mechanistic model to predict the predicted process feature vector under the current process parameters. Then, based on the deviation vector between the predicted process feature vector and the extracted current process feature vector, a compensation model is used to determine the state deviation under the current process parameters. Compared with the method of directly determining the state deviation based on the mechanistic model, it further considers the influence of local effects that are difficult for the mechanistic model to describe accurately. The accuracy of the obtained state deviation is higher, so that the process parameters can be more accurately controlled when the process drifts, so that the process state returns to the target trajectory.

[0076] Taking semiconductor etching process as an example, combined with the attached Figure 2 The semiconductor process control method provided by this invention will be described in detail.

[0077] During the process, an integrated optical inspection system is used to acquire the morphological features of the semiconductor structure's surface being processed according to a measurement cycle. Each time the electronic equipment acquires a new set of multidimensional inspection data (raw measurement data), such as... Figure 2 As shown, the multidimensional detection data is first preprocessed, such as denoising, background correction, and pixel-level registration. Then, feature vectors with clear physical meaning are extracted from the preprocessed multidimensional detection data. These process feature vectors can include depth features, sidewall features, bottom features, film thickness features, and defect features.

[0078] Then, a feature vector for the current process is constructed based on the extracted features. , This indicates the average etching depth under the current process parameters. This indicates the depth uniformity under the current process parameters. This indicates the sidewall angle under the current process parameters. This represents the average remaining film thickness under the current process parameters. This indicates the defect density under the current process parameters.

[0079] Predicting the current process parameters using mechanistic models The expected features (predicted process feature vector) are as follows. And calculate the deviation vector. Then, the deviation vector and the current process parameters are used. The XGBoost residual model is input for compensation to obtain the corrected bias estimate (state bias). The process drift index is determined. If the process drift index is greater than the preset drift index, the adaptive control decision module is triggered. Based on the state deviation, the current process parameters are adjusted to obtain the next process parameters so that the process state returns to the target trajectory. If the process drift index is less than or equal to the preset drift index, the process returns to the data preprocessing step and continues monitoring.

[0080] This embodiment provides another method for controlling semiconductor processes, which can be used in electronic devices such as tablet computers, computers, or servers. Figure 3 This is a flowchart illustrating another semiconductor process control method according to an embodiment of the present invention, such as... Figure 3 As shown, the process includes the following steps: Step S301: When fabricating a semiconductor structure using semiconductor technology, acquire the current multidimensional detection data for the current measurement cycle.

[0081] Specifically, the integrated optical detection system includes a dual-wavelength phase-shifting structured light projection device and an optical device that outputs multi-wavelength reflection spectra. Step S301 may include: Step b1: Obtain the three-dimensional topographic data of the current measurement cycle from the dual-wavelength phase-shifting structured light projection device.

[0082] Specifically, by using a dual-wavelength phase-shifting structured light projection device and combining it with a multi-frequency heterodyne phase calculation algorithm, the surface height map (three-dimensional topography data) of the semiconductor structure is reconstructed, with a horizontal resolution of 1 μm and a height resolution of 1 nm.

[0083] Step b2: Use multi-wavelength reflectance spectroscopy to obtain reflectance data of the film thickness for the current measurement period.

[0084] Specifically, reflectivity data reflecting the film thickness is obtained using multi-wavelength reflection spectroscopy, and then the remaining film thickness value of each pixel is fitted based on a thin-film interference model. In addition to the structured light wavelength, the multi-wavelength reflection spectroscopy also includes a broadband white light channel, and the film material can be silicon dioxide (SiO2) or silicon nitride (SiN), etc.

[0085] Furthermore, current multidimensional detection data can also include defect density maps. After acquiring height maps and reflectivity data (reflectance maps), defects (such as particles, scratches, dents, microgrooves, etc.) are identified from the height maps and reflectivity maps, and the coordinates, size and type of defects are recorded to generate defect density maps.

[0086] For example, the dual-wavelength phase-shift structured light projection device includes a projection module that uses a digital micromirror device (DMD) to project and emit a first projection wavelength and a second projection wavelength. The first projection wavelength is greater than the second projection wavelength. The first projection wavelength has better silicon penetration and is suitable for acquiring signals near the bottom of the etched trench. The second projection wavelength is used to acquire signals far from the bottom of the etched trench.

[0087] The projection wavelength emitted by the projection module differs from the characteristic emission lines in semiconductor processes. Characteristic emission lines in etching processes can be fluorine or chlorine atom lines, with wavelengths of 704 nm or 777 nm. A dual-wavelength combination of 650 nm (red light) and 850 nm (near-infrared) can be used for projection. 650 nm is used for high-precision measurement of shallow topography, while 850 nm has better silicon penetration and is suitable for signal acquisition at the bottom of deep trenches. Selecting the projection wavelength in the dark region of the plasma spectrum, avoiding the characteristic emission lines in typical etching processes, can suppress interference from plasma background light.

[0088] The dual-wavelength phase-shift structured light projection device also includes an imaging module, which is equipped with a narrowband filter (bandwidth ±10nm) corresponding to the projection wavelength to further suppress stray light. The imaging module uses a global shutter complementary metal-oxide-semiconductor (CMOS) camera and is equipped with an indium gallium arsenide (InGaAs) quantum dot enhanced sensor in the near-infrared channel to improve the signal-to-noise ratio at the bottom of the deep trench.

[0089] Optionally, when the RF power switch in the semiconductor process equipment is off or during the period of minimum plasma afterglow in the semiconductor process equipment (e.g., the last 10ms of each cycle), the dual-wavelength phase-shifting structured light projection device emits a projection wavelength to detect three-dimensional topographic data. Triggering the measurement when the plasma light intensity is weak can further reduce background light noise.

[0090] It should be noted that as the process progresses, polymer gradually deposits in the observation window of the chamber, causing a decrease in light transmittance. The integrated optical detection system monitors the intensity of reflected light at the reference point in real time. When the intensity of reflected light at the reference point is lower than the threshold, it automatically increases the exposure time or increases the intensity of the projected light to ensure stable measurement accuracy.

[0091] Step S302: Extract the current process feature vector from the current multidimensional detection data.

[0092] Please see details Figure 1 Step S102 of the illustrated embodiment will not be described again here.

[0093] Step S303: Based on the current process feature vector and the hierarchical hybrid model, determine the state deviation under the current process parameters.

[0094] Please see details Figure 1 Step S103 of the illustrated embodiment will not be described again here.

[0095] Step S304: Determine the drift index during the semiconductor process based on the state deviation.

[0096] Please see details Figure 1 Step S104 of the illustrated embodiment will not be described again here.

[0097] Step S305: If the drift index is greater than the preset drift index, adjust the current process parameters based on the state deviation to obtain the next process parameters, so that the execution state of the semiconductor process returns to the target trajectory.

[0098] Specifically, step S305 above may include: Step S3051: Based on the state deviation, obtain the initial adjustment amount from the optimal parameter lookup table.

[0099] The optimal parameter lookup table includes the correspondence between multiple different state deviations and process parameter adjustments.

[0100] Specifically, after determining the state deviation, the process parameter adjustment amounts corresponding to the state deviation in the optimal parameter lookup table are determined as the initial adjustment amounts. .

[0101] The optimal parameter lookup table (LUT) is created offline and pre-stored in the electronic device as key-value pairs. The key is the state deviation, and the value is the recommended process parameter adjustment. The LUT can cover more than 90% of normal operating conditions. The LUT is built before the initial adjustment is obtained from the optimal parameter lookup table based on the state deviation. The creation process of the optimal parameter lookup table includes: Step c1: Based on the second multi-objective optimization algorithm and historical process parameters, determine the optimal parameter combination under different execution states.

[0102] Step c2: Based on the optimal parameter combination under different execution states, determine and store the optimal parameter lookup table.

[0103] The second multi-objective optimization algorithm can be a multi-objective genetic algorithm (NSGA-II) or other multi-objective optimization algorithms.

[0104] Specifically, based on historical process parameters, a multi-objective optimization problem (such as target depth, uniformity, sidewall angle, etc.) is defined. Then, a multi-objective optimization algorithm is used to solve for Pareto optimal solution sets under different process states (different target depths, different pattern densities, different chamber usage times) to obtain the optimal combination of process parameters for each process state. Then, the state deviation is used as the key, and the process parameter adjustment amount obtained is used as the value and stored in the LUT to form an optimal parameter lookup table.

[0105] Step S3052: Based on the constraints of the cost function and process parameters, the gradient descent method is used to optimize within the neighborhood of the initial adjustment amount to obtain the target adjustment amount.

[0106] The cost function characterizes the deviation between the next process feature vector and the target process feature vector corresponding to the target execution state. The next process feature vector is the predicted process feature vector corresponding to the adjusted current process parameters, determined by the prediction model built based on the hierarchical hybrid model. The gradient in the gradient descent method is obtained by analytical differentiation or finite difference of the prediction model. Constraints include the allowable range of process parameters (power upper limit, gas flow range, pressure range, etc.) and the limit on the rate of change (e.g., power change rate ≤ 10 W / s).

[0107] In this embodiment, a hierarchical hybrid model is used as the prediction model, and the prediction model can be as shown in formula (6):

[0108] In the formula, This represents the predicted feature vector for the next process. This represents a hierarchical hybrid model.

[0109] Optionally, the cost function includes a penalty, which is characterized by the difference between the previous process parameter in the previous measurement cycle and the current process parameter.

[0110] Specifically, the cost function It can be shown in formula (7):

[0111] In the formula, The first element in the process feature vector represents the... i The weights of individual sub-features are as follows: for example, the weight of the sub-feature corresponding to depth is 0.5, the weight of the sub-feature corresponding to uniformity is 0.3, and the weight of the sub-feature corresponding to sidewall angle is 0.2. This represents the eigenvector of the predicted next process. i Individual characteristics, Represents the first element in the target process feature vector. i Individual characteristics, This represents the smoothness penalty coefficient (regularization parameter) for the control action. Indicates the change in process parameters The square of the Euclidean norm, , This indicates the previous process parameter.

[0112] This invention introduces a penalty term into the cost function. This is not only for smooth control, but also for the stability of the physical properties of the semiconductor process; frequent and drastic parameter changes can lead to plasma instability, introduce particulate contamination, and shorten the lifespan of the matcher. Therefore, this invention introduces a penalty term to achieve a balance between control precision and process stability.

[0113] Step S3053: Adjust the current process parameters based on the target adjustment amount to obtain the next process parameters.

[0114] Specifically, the sum of the current process parameters and the target adjustment amount is determined as the next process parameter.

[0115] The semiconductor process control method provided in this embodiment uses an offline global search to create an optimal parameter lookup table to determine the initial adjustment amount, and then uses the gradient descent algorithm to locally optimize the initial adjustment amount online to obtain the target adjustment amount that enables the process state to reach the target execution state. This can significantly reduce the optimization time and improve the adjustment response speed.

[0116] In some alternative embodiments, step S3052 may include: Step d1: If the number of optimizations performed using gradient descent is less than or equal to the preset number of times, and the deviation value corresponding to the cost function is less than or equal to the first preset threshold, then the adjustment amount corresponding to the target deviation value is determined as the target adjustment amount.

[0117] The target deviation value is a deviation value that is less than or equal to the first preset threshold. The preset number of times can be 5 to ensure that the time consumed in a single optimization is small.

[0118] Step d2: If the number of optimizations is greater than the preset number of optimizations, and the deviation value corresponding to the cost function is greater than the first preset threshold, then the adjustment amount corresponding to the minimum deviation value is determined as the target adjustment amount.

[0119] Furthermore, if the initial adjustment amount is not obtained from the optimal parameter lookup table, the current process parameter is determined as the next process parameter; if the initial adjustment amount is not obtained from the optimal parameter lookup table, or if the number of optimizations using the gradient descent method is less than or equal to the preset number, and the deviation value corresponding to the cost function is greater than the first preset threshold, then based on the first multi-objective optimization algorithm, the process parameter adjustment amount corresponding to the state deviation amount is determined; based on the process parameter adjustment amount corresponding to the state deviation amount, the optimal parameter lookup table is updated.

[0120] The first multi-objective optimization algorithm can be particle swarm optimization (PSO) or other optimization algorithms.

[0121] Specifically, if the table lookup fails, or if the gradient descent method is used for local optimization and the error does not meet the requirements in a small number of iterations, the first multi-objective optimization algorithm is started in the background to determine the process parameter adjustment amount corresponding to the state deviation amount, and the optimal parameter lookup table is updated to shorten the time of the next process parameter adjustment. The current process parameter is determined as the next process parameter, or the suboptimal solution is used as the target adjustment amount to improve the response speed.

[0122] In an optional embodiment, the adaptive control decision module can adjust process parameters as follows: Figure 4 As shown, the process is divided into an offline stage and an online stage. The offline stage uses historical process data and a multi-objective genetic algorithm to create an optimal parameter lookup table. The online stage obtains the deviation vector... Next, the initial adjustment amount is obtained by looking up the optimal parameter lookup table. Then, the initial adjustment amount is locally optimized using the gradient descent algorithm. If the number of optimizations is less than or equal to 5, and the error determined by the cost function meets the requirements when the number of optimizations is less than or equal to 5, the optimized initial adjustment amount is directly output, and the parameters are sent through the communication standard protocol (SECS / GEM), and then the next measurement cycle begins. If the error does not meet the requirements when the number of optimizations is less than or equal to 5, the PSO global search is initiated to determine the target adjustment amount, and the optimal parameter lookup table is updated. In the current measurement cycle, the suboptimal solution is used or the original parameters are maintained.

[0123] This embodiment also provides another method for controlling semiconductor processes, which can be used in electronic devices such as tablet computers, computers, or servers. Figure 5 This is a flowchart illustrating another semiconductor process control method according to an embodiment of the present invention, such as... Figure 5 As shown, the process includes the following steps: Step S501: When fabricating a semiconductor structure using semiconductor technology, acquire the current multidimensional detection data for the current measurement cycle.

[0124] Please see details Figure 1 Step S101 of the illustrated embodiment or Figure 3 Step S301 of the illustrated embodiment will not be described again here.

[0125] Step S502: Extract the current process feature vector from the current multidimensional detection data.

[0126] In this embodiment, the current process feature vector includes the current defect features.

[0127] Step S503: Based on the current process feature vector and the hierarchical hybrid model, determine the state deviation under the current process parameters.

[0128] Please see details Figure 1 Step S103 of the illustrated embodiment will not be described again here.

[0129] Step S504: Determine the drift index during the semiconductor process based on the state deviation.

[0130] Please see details Figure 1 Step S104 of the illustrated embodiment will not be described again here.

[0131] Step S505: If the process drift index is greater than the preset drift index, adjust the current process parameters based on the state deviation amount to obtain the next process parameters, so that the execution state of the semiconductor process returns to the target execution state.

[0132] Please see details Figure 3 Step S305 of the illustrated embodiment will not be described again here.

[0133] Step S506: Based on the current defect characteristics and multiple historical defect characteristics prior to the current measurement cycle, fit and determine the rate of change of defect density over time.

[0134] Step S507: If the slope of the rate of change of defect density over time is positive and greater than the second preset threshold in multiple consecutive measurement cycles, then the first alarm signal is output.

[0135] Specifically, the contamination / aging process of a semiconductor chamber exhibits time-dependent and nonlinear acceleration characteristics. Initially, defect growth is slow (deposition on the chamber walls), but after reaching a critical point, it accelerates exponentially (particle shedding, window blurring). This invention uses a sliding window to linearly fit multiple defect features (current defect features and multiple historical defect features), calculates the rate of change of defect density over time, and monitors the rate of change (second derivative). When the second derivative remains positive and exceeds a second preset threshold, it indicates the entry into the acceleration phase, and a first alarm signal is output. This first alarm signal is used to alert personnel to the contamination / aging of the semiconductor chamber.

[0136] Step S508: Input the current defect characteristics into the lifetime prediction model, and predict the remaining lifetime of the semiconductor process equipment based on the lifetime prediction model.

[0137] Among them, the lifetime prediction model can be a bidirectional long short-term memory network (Bi-LSTM) model. The remaining lifetime of semiconductor process equipment refers to the minimum remaining lifetime of multiple wear-out components in the semiconductor process equipment. Wear-out components can be easily worn components such as electrostatic chucks, quartz windows, or O-rings.

[0138] Defect evolution depends not only on past trends but also on future goals (defects accelerate near maintenance). The bidirectional structure can capture contextual dependencies simultaneously and is more sensitive to inflection point prediction. Therefore, this invention uses a bidirectional long short-term memory network (Bi-LSTM) model as the lifetime prediction model, which can improve the accuracy of the predicted remaining lifetime.

[0139] Step S509: If the remaining lifespan is less than the preset lifespan, then output a second alarm signal.

[0140] The preset lifespan can be 5 days or 10 days, etc.

[0141] Specifically, when the predicted remaining lifespan is less than the preset lifespan, a second alarm signal is output through the work order system. This second alarm signal is used to prompt staff to replace worn parts and recommend maintenance periods (such as nighttime or weekends) to minimize the impact on production. When the predicted remaining lifespan is less than the preset lifespan, maintenance suggestions can also be generated in conjunction with the production plan, including maintenance time windows and recommended replacement parts.

[0142] In this embodiment, by monitoring the defect process using the second derivative of the defect trend, early signals of accelerated defect degradation can be captured. Intervention can be made before irreversible deterioration of defects, reducing the probability of major failures and improving process stability and product yield. Monitoring the remaining lifespan of semiconductor process equipment based on a lifetime prediction model allows for the early replacement of worn-out devices, avoiding the impact of unplanned downtime on the production line.

[0143] This embodiment describes the training and prediction processes of the Bi-LSTM model as follows: Figure 6As shown, to obtain accurate data for training the Bi-LSTM model, defect feature sequences (defect density, maximum defect size, defect type distribution, defect spatial distribution entropy) from the past N batches, as well as cumulative equipment operating time and cumulative process counts, are acquired. Then, the Bi-LSTM model is trained based on the collected data. The input feature is the defect feature sequence, and the output feature is the remaining useful life (RUL). The Bi-LSTM model's network structure includes a 12-dimensional input layer, two hidden layers, a bidirectional LSTM (64 units each), and an output layer. The output layer represents the mean and variance of the RUL, using negative log-likelihood loss to quantify uncertainty. The specific training strategy is as follows: when there is limited data on new machines, transfer learning is employed. The model is pre-trained using historical data from similar machines, and then the last layer is fine-tuned using a small amount of data from new machines to shorten the model's cold start time.

[0144] The trained Bi-LSTM model is configured in the electronic device to extract defect features from real-time defect flow data and input them into the trained Bi-LSTM model. The remaining lifetime is determined based on the RUL probability distribution output by the trained Bi-LSTM model. When the remaining lifetime is less than a threshold, a maintenance warning is triggered and a maintenance decision is generated.

[0145] For example, such as Figure 7 As shown, the semiconductor process control method provided by this invention can include four core steps: Step e1, real-time data acquisition: During the process, an integrated optical inspection system continuously acquires data on the three-dimensional morphology, film thickness, and defect distribution of the wafer surface. The integrated optical inspection system is specifically adapted for the semiconductor process environment to overcome interference from plasma emission, window contamination, and other factors.

[0146] Step e2, multi-parameter fusion analysis: The collected multi-dimensional data is spatiotemporally aligned, features with clear process physical significance are extracted, and a morphology-process parameter mapping relationship based on the "mechanism + data" hybrid model is established to diagnose process state drift in real time.

[0147] Step e3, adaptive control decision-making: Based on the fusion analysis results, a hybrid optimization strategy oriented towards real-time control is adopted to dynamically adjust process parameters (RF power, gas flow rate, chamber pressure, etc.) so that the process state can quickly return to the target trajectory.

[0148] Step e4, predictive maintenance: Based on defect growth trends and equipment health assessments, use time-series prediction models to provide early warnings of maintenance needs and reduce unplanned downtime.

[0149] Taking semiconductor etching as an example, the semiconductor etching control method provided by this invention is applied to a 12-inch wafer etching machine, with a target etching depth of 200nm and uniformity of <3%. After adopting the semiconductor etching control method provided by this invention, the etching depth of 50 consecutive batches is monitored and compared with the etching depth of 50 batches without using this method.

[0150] Figure 8 A comparative illustration of the etching depth control effect is provided, from... Figure 8 It can be seen that the uncontrolled batch has a large fluctuation in etching depth, with a standard deviation of 7.5nm; after control, the etching depth is stable at 200±2.1nm, and the uniformity is improved by 72%.

[0151] Figure 9 A comparative schematic diagram of etching uniformity within a batch is provided, from Figure 9 It can be seen that the average etching uniformity within the uncontrolled batch was 5.2nm, while that after control was reduced to 2nm, indicating a significant improvement in uniformity. Figure 10 A schematic diagram showing the trend of defect rate over time is provided, with the defect rate measured in defects per cm. 2 , Figure 11 A schematic diagram of the remaining lifetime prediction curve based on the defect rate is provided, from... Figure 10 and Figure 11 It can be seen that the defect rate increases at an accelerated rate over time, with the slope increasing after day 80. The LSTM predicts that the remaining lifetime will drop to 5 days on day 85, triggering a maintenance warning. In reality, the failure occurs on day 95, providing a warning 10 days in advance. Figure 12 A comparative diagram of batch yield before and after control is provided. The average yield of the first 25 batches was 86.3%, and the average yield of the last 25 batches was 95.8%, an improvement of 9.5 percentage points.

[0152] In summary, by adopting the semiconductor process control method provided by this invention, the standard deviation of etching depth was reduced from ±7.5 nm (uncontrolled) to ±2.1 nm, and uniformity was improved by 72%. The defect rate did not accelerate in the later stages of the process, and predictive maintenance provided an early warning of chamber contamination 10 days in advance, avoiding sudden downtime. The average batch yield increased from 86.3% to 95.8%, an improvement of 9.5 percentage points.

[0153] The semiconductor process control method provided by this invention improves real-time control response speed, with a closed-loop time of less than 5 seconds from detection to parameter adjustment, effectively preventing defects; significantly improves yield, improves batch uniformity by more than 30%, and reduces critical dimension deviation by 50%; saves costs, reducing offline measurement and rework costs by about 40%, increasing equipment utilization by 20%, and enabling predictive maintenance to avoid unplanned downtime and extend component life; enhances process portability, and through hybrid model adaptive calibration, shortens the matching time of the same formula between different machines by 60%.

[0154] This embodiment also provides a semiconductor process control device for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0155] This embodiment provides a control device for a semiconductor process, such as... Figure 13 As shown, it includes: The acquisition module 1301 is used to acquire the current multidimensional detection data of the current measurement cycle when a semiconductor structure is fabricated using a semiconductor process. The current multidimensional detection data includes the three-dimensional morphology data of the semiconductor structure and the reflectivity data reflecting the film thickness. The extraction module 1302 is used to extract the current process feature vector from the current multidimensional detection data, wherein the current process feature vector is used to reflect the current execution state of the semiconductor process; The deviation determination module 1303 is used to determine the state deviation amount under the current process parameters based on the current process feature vector and the hierarchical hybrid model. The hierarchical hybrid model includes a mechanism model and a compensation model. The mechanism model is used to characterize the correspondence between process parameters and process feature vectors. The input is the current process parameters, and the output is the predicted process feature vector under the current process parameters. The compensation model is used to correct the mechanism model. The input is the deviation vector between the predicted process feature vector and the current process feature vector and the current process parameters. The output is the state deviation amount. The drift determination module 1304 is used to determine the drift index during the semiconductor process based on the state deviation amount. The parameter adjustment module 1305 is used to adjust the current process parameters based on the state deviation amount if the drift index is greater than the preset drift index, so as to obtain the next process parameters and make the execution state of the semiconductor process return to the target execution state.

[0156] The semiconductor process control device provided in this embodiment of the invention can execute the semiconductor process control method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the various modules and units described above are the same as in the corresponding embodiments described above, and will not be repeated here.

[0157] Figure 14 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0158] like Figure 14 As shown, the electronic device may include a processor 1401 and a memory, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the functions defined in the semiconductor process control method provided in any of the above embodiments.

[0159] For example, the memory may include read-only memory 1402, random access memory 1403 and external memory 1404. The processor 1401, read-only memory 1402 and random access memory 1403 are communicatively connected to each other via bus 1405, and external memory 1404 is connected to bus 1405 via input / output multiplexing interface 1406.

[0160] The processor 1401 can perform various appropriate actions and processes based on the program stored in the read-only memory 1402 or the program loaded into the random access memory 1403 from the external memory 1404. The random access memory 1403 also stores various programs and data required for the operation of the electronic device. The processor 1401 can be a central processing unit, a graphics processing unit, etc., the read-only memory 1402 can be ROM, and the random access memory 1403 can be RAM.

[0161] Although Figure 14 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead. Figure 14 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0162] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the semiconductor process control method shown in the above embodiments is implemented.

[0163] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0164] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for controlling a semiconductor process, characterized in that, The method includes: When fabricating a semiconductor structure using semiconductor technology, the current multidimensional detection data of the current measurement cycle is acquired, wherein the current multidimensional detection data includes the three-dimensional morphology data of the semiconductor structure and the reflectivity data reflecting the film thickness; The current process feature vector is extracted from the current multidimensional detection data, wherein the current process feature vector is used to reflect the current execution state of the semiconductor process; Based on the current process feature vector and the hierarchical hybrid model, the state deviation under the current process parameters is determined. The hierarchical hybrid model includes a mechanistic model and a compensation model. The mechanistic model characterizes the correspondence between process parameters and process feature vectors, taking the current process parameters as input and outputting a predicted process feature vector under the current process parameters. The compensation model corrects the mechanistic model, taking the deviation vector between the predicted process feature vector and the current process feature vector, as well as the current process parameters, as input, and outputting the state deviation. Based on the state deviation, the drift index during the semiconductor process is determined; If the drift index is greater than the preset drift index, the current process parameters are adjusted based on the state deviation to obtain the next process parameters, so that the execution state of the semiconductor process returns to the target execution state.

2. The method according to claim 1, characterized in that, Adjusting the current process parameters based on the state deviation includes: Based on the state deviation, the initial adjustment amount is obtained from the optimal parameter lookup table, wherein the optimal parameter lookup table includes the correspondence between multiple different state deviations and process parameter adjustment amounts; Based on the constraints of the cost function and process parameters, the gradient descent method is used to optimize within the neighborhood of the initial adjustment amount to obtain the target adjustment amount. The cost function is used to characterize the deviation between the next process feature vector and the target process feature vector corresponding to the target execution state. The next process feature vector is the predicted process feature vector corresponding to the adjusted current process parameters, which is determined by the prediction model constructed based on the hierarchical hybrid model. The current process parameters are adjusted based on the target adjustment amount to obtain the next process parameters.

3. The method according to claim 2, characterized in that, The constraints based on the cost function and process parameters are optimized using gradient descent within the neighborhood of the initial adjustment amount to obtain the target adjustment amount, including: If the number of optimizations performed using gradient descent is less than or equal to a preset number, and the deviation value corresponding to the cost function is less than or equal to a first preset threshold, then the adjustment amount corresponding to the target deviation value is determined as the target adjustment amount, and the target deviation value is the deviation value that is less than or equal to the first preset threshold. If the number of optimizations exceeds the preset number, and the deviation value corresponding to the cost function is greater than the first preset threshold, then the adjustment amount corresponding to the minimum deviation value is determined as the target adjustment amount.

4. The method according to claim 2, characterized in that, The method further includes: If the initial adjustment amount is not obtained from the optimal parameter lookup table, the current process parameter is determined as the next process parameter. If the initial adjustment amount is not obtained from the optimal parameter lookup table, or if the number of optimizations performed using the gradient descent method is less than or equal to the preset number, and the deviation value corresponding to the cost function is greater than the first preset threshold, then the process parameter adjustment amount corresponding to the state deviation amount is determined based on the first multi-objective optimization algorithm. The optimal parameter lookup table is updated based on the process parameter adjustment amount corresponding to the state deviation.

5. The method according to any one of claims 2 to 4, characterized in that, The cost function includes a penalty, which is characterized by the difference between the previous process parameter in the previous measurement cycle and the current process parameter.

6. The method according to any one of claims 2 to 4, characterized in that, Before obtaining the initial adjustment amount from the optimal parameter lookup table based on the state deviation, the method further includes: Based on the second multi-objective optimization algorithm and historical process parameters, the optimal parameter combination under different execution states is determined; Based on the optimal parameter combination under different execution states, the optimal parameter lookup table is determined and stored.

7. The method according to any one of claims 1 to 4, characterized in that, The acquisition of current multidimensional detection data for the current measurement period includes: The three-dimensional topographic data of the current measurement cycle are obtained from a dual-wavelength phase-shifting structured light projection device; The reflectance data of the film thickness in the current measurement period are obtained by using multi-wavelength reflectance spectroscopy.

8. The method according to claim 7, characterized in that, The projection wavelength emitted by the dual-wavelength phase-shift structured light projection device is different from the characteristic emission spectrum in the semiconductor process.

9. The method according to claim 7, characterized in that, When the radio frequency power switch in the semiconductor process equipment is in the off state or during the period when the plasma afterglow in the semiconductor process equipment is minimal, the dual-wavelength phase-shift structured light projection device emits a projection wavelength to detect the three-dimensional topography data.

10. The method according to any one of claims 1 to 4, characterized in that, The current process feature vector includes current defect features, and the method further includes: Based on the current defect characteristics and multiple historical defect characteristics prior to the current measurement period, the rate of change of defect density over time is determined by fitting. If the slope of the rate of change of the defect density over time is positive and greater than the second preset threshold in multiple consecutive measurement cycles, then a first alarm signal is output.

11. The method according to claim 10, characterized in that, The method further includes: The current defect characteristics are input into the lifetime prediction model, and the remaining lifetime of the semiconductor process equipment is predicted based on the lifetime prediction model. If the remaining lifespan is less than the preset lifespan, a second alarm signal is output.

12. The method according to any one of claims 1 to 4, characterized in that, The semiconductor process is an etching process, and the current process feature vector includes at least one of depth features, sidewall features, bottom features, film thickness features, and defect features.

13. A control device for a semiconductor process, characterized in that, The device includes: The acquisition module is used to acquire the current multidimensional detection data of the current measurement cycle when a semiconductor structure is fabricated using a semiconductor process. The current multidimensional detection data includes the three-dimensional morphology data of the semiconductor structure and the reflectivity data reflecting the film thickness. An extraction module is used to extract a current process feature vector from the current multidimensional detection data, wherein the current process feature vector is used to reflect the current execution state of the semiconductor process; The deviation determination module is used to determine the state deviation amount under the current process parameters based on the current process feature vector and the hierarchical hybrid model. The hierarchical hybrid model includes a mechanistic model and a compensation model. The mechanistic model characterizes the correspondence between process parameters and process feature vectors, takes the current process parameters as input, and outputs a predicted process feature vector under the current process parameters. The compensation model corrects the mechanistic model, takes the deviation vector between the predicted process feature vector and the current process feature vector as input, and the current process parameters as output, and outputs the state deviation amount. A drift determination module is used to determine the drift index during the semiconductor process based on the state deviation amount. The parameter adjustment module is used to adjust the current process parameters based on the state deviation amount if the drift index is greater than the preset drift index, so as to obtain the next process parameters and make the execution state of the semiconductor process return to the target execution state.

14. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory storing computer instructions, and the processor executing the computer instructions to perform the control method of the semiconductor process according to any one of claims 1 to 12.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to perform the control method of the semiconductor process according to any one of claims 1 to 12.

16. A computer program product, characterized in that, It includes computer instructions for causing a computer to perform a control method for a semiconductor process according to any one of claims 1 to 12.

Images