Overlay data analysis and prediction system based on feedback feedforward simulation

Abstract

The invention provides an Overlay data analysis and prediction system based on feedback and feedforward simulation, and relates to the field of semiconductor alignment optimization and intelligent control, the system comprises a main body system, the main body system comprises a data acquisition module, a data processing and storage module, an analysis simulation module and a prediction and correction module; the data acquisition module is arranged on a sensor at a key position of photoetching equipment, comprises 0.1 nm resolution optical encoders deployed on an exposure unit, a workpiece table and a lens group, a 64-channel MEMS vibration sensor and an infrared thermal imaging unit, and is used for acquiring Overlay data of a production batch in real time. And a deep residual network algorithm is adopted, so that key features can be accurately extracted from massive photoetching data. Through deep analysis of the features, potential rules hidden behind the data are mined, so that analysis of Overlay is not limited to the surface any more, potential process problems can be found in advance, and powerful support is provided for preventive maintenance and process optimization.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor alignment optimization and intelligent control, and specifically to an Overlay data analysis and prediction system based on feedback-feedforward simulation. Background Art

[0002] In the field of semiconductor manufacturing, lithography technology, as a key link in chip manufacturing, its accuracy directly affects the performance and production yield of chips. Overlay (overlay accuracy) is an important indicator to measure the accuracy of the lithography process. Precise analysis and prediction of it are crucial for optimizing the lithography process and improving the quality of chip manufacturing.

[0003] Currently, in terms of Overlay data processing and lithography process optimization, the following technical means are mainly adopted:

[0004] Basic data collection and simple analysis: By installing basic sensors on lithography equipment, some key data during the lithography process are collected, such as position information and exposure time. Simple statistical analysis is performed on these data.

[0005] Traditional model prediction: Some classical physical models or empirical models are used to predict Overlay based on the parameters of lithography equipment and process conditions.

[0006] Single algorithm processing: Some enterprises adopt a single algorithm, such as a simple regression algorithm or a neural network algorithm, to analyze and predict Overlay data.

[0007] Manual experience adjustment: When problems occur in the lithography process, mainly rely on the experience of operators to adjust lithography parameters to improve Overlay.

[0008] Although the existing technical means can monitor and adjust the lithography process to a certain extent, there are still many deficiencies:

[0009] Insufficient data mining depth: The existing basic data collection and simple analysis methods cannot extract key features and potential laws from a large amount of lithography data, resulting in the analysis of Overlay staying on the surface and making it difficult to discover potential process problems in advance.

[0010] Limited prediction accuracy: Due to the simplification of the lithography process and insufficient consideration of complex factors in traditional models, as well as the limitations of single algorithms in processing complex data, the accuracy of Overlay prediction is difficult to meet the requirements of high-precision chip manufacturing, easily leading to lithography deviation and affecting the performance and production yield of chips.

[0011] Lack of real-time feedback and adaptive adjustment: Most of the existing methods lack a real-time feedback mechanism and cannot dynamically adjust the process according to the real-time data during the lithography process. At the same time, it is difficult to achieve adaptive adjustment in the face of changes in process parameters and fluctuations in equipment status, reducing the stability and reliability of the lithography process.

[0012] Relying on manual experience: The method of adjusting lithography parameters based on manual experience lacks scientific basis and systematicness. The adjustment effects of different operators vary greatly, and it is difficult to quickly respond to changes in the production process, which is not conducive to improving production efficiency and product quality.

[0013] Therefore, a system for analyzing and predicting Overlay data based on feedback-feedforward simulation is needed to solve the above problems. Summary of the Invention

[0014] Technical problems to be solved

[0015] In view of the deficiencies of the prior art, the present invention provides a system for analyzing and predicting Overlay data based on feedback-feedforward simulation, which solves the problems in the above background technology.

[0016] Technical solution

[0017] To achieve the above objectives, the present invention is realized through the following technical solutions: A system for analyzing and predicting Overlay data based on feedback / feedforward simulation, including a main system, which contains a data acquisition module, a data processing and storage module, an analysis and simulation module, and a prediction and correction module;

[0018] The data acquisition module is a sensor set at key positions of the lithography equipment, including an optical encoder with a resolution of 0.1 nm, a 64-channel MEMS vibration sensor, and an infrared thermal imaging unit deployed in the exposure unit, the workpiece stage, and the lens group, which can collect Overlay data of the production batch in real time;

[0019] The exposure unit, the workpiece stage, and the lens group here are all main components equipped in the existing lithography equipment.

[0020] The data processing and storage module adopts a distributed storage architecture based on the IPFS protocol, performs joint noise reduction of wavelet transform-independent component analysis on the collected data, and stores it in the blockchain database through the SM4 national secret encryption unit after removing noise interference;

[0021] The analysis and simulation module uses a two-way coupling system of a physically driven model and a data-driven model of a deep residual network; dynamically allocates model weights of 0.3-0.7 through an attention mechanism to simulate the production process under different parameters;

[0022] The prediction and correction module predicts the Overlay deviation based on the simulation results and adjusts the lithography parameters through the forward compensation and reverse correction functions. The bidirectional coupling system of the physical drive model and the deep residual network data drive model is used to dynamically allocate the model weight of 0.3-0.7 through the attention mechanism to simulate the production process under different parameters.

[0023] Preferably, in the analysis and simulation module, the feedback algorithm establishes a thermal-mechanical coupling model based on the Navier-Stokes equations, and performs dynamic compensation according to the difference between the actual data and the expected data in the production process. The feedforward algorithm extracts high-order features through a deep residual network, and predicts and simulates the production process according to pre-set parameters and models. An incremental learning strategy is introduced to trigger model retraining every 50 batches of data processed, and dynamically optimize the algorithm parameters.

[0024] Preferably, in the prediction and correction module, the forward compensation reduces the Overlay deviation by adjusting the exposure energy and alignment offset parameters of the lithography equipment, and the reverse correction uses the XGBoost intelligent decision tree model to automatically select the correction strategy according to the historical process scene types, and reversely correct the over-adjustment in the correction process, and is equipped with a dynamic time warping algorithm to achieve spatiotemporal synchronization of multi-sensor data to ensure the accuracy of lithography alignment.

[0025] Preferably, the data acquisition module uses a fiber Bragg grating strain measurement unit to collect overlay data in real time; it is also equipped with an adaptive acquisition frequency adjustment function, automatically switches the sampling rate within the range of 50Hz-10kHz according to the vibration amplitude of the equipment, and integrates a temperature compensation circuit to eliminate the influence of ±0.5℃ temperature drift on the sensor accuracy.

[0026] Preferably, the data processing and storage module adopts a data sharding storage architecture to store process data in blocks according to timestamps; uses the LZMA compression algorithm to reduce storage space, and implements blockchain traceability management of key data through a Merkle tree version tracing mechanism and attribute-based encrypted access control.

[0027] Preferably, the main system also includes a digital twin visualization module, which can display the analysis results in the form of intuitive graphics and charts, including real-time trend charts and comparative analysis charts, so that operators can quickly understand the production status. Through the CFD thermal field distribution cloud map, the ±0.1°C temperature change is dynamically displayed, and the AR-assisted calibration interface is used to be compatible with the Hololens2 device to achieve virtual-real superposition alignment, and a three-level deviation warning system is established, setting yellow (>3nm), orange (>5nm), and red (>8nm) alarm thresholds.

[0028] Preferably, the main system has a multi-device collaboration function, which is realized through a quantum secure communication unit. The NTRU lattice-based encryption is integrated to ensure data transmission between lithography devices. The channel security is detected in real time through the BB84 protocol monitoring module, and the 256-bit quantum key is automatically refreshed every hour.

[0029] Preferably, a fault diagnosis module is introduced into the main system. By analyzing the anomalies in the collected data, the faults of the lithography equipment can be predicted in advance. The faults include a process knowledge graph engine: a fault entity relationship network is constructed using the BERT-BiLSTM-CRF model, historical maintenance solutions are matched through a case-based reasoning mechanism, and the adaptive update of fault features is realized using a dynamic node embedding algorithm.

[0030] Preferably, the main system has an online learning function, which is specifically manifested as follows: a model fusion unit is equipped to automatically adjust the weights of the physical-data model according to the prediction error. In an emergency condition, a 10kHz high-frequency acquisition mode is started to update the compensation parameters in real time; incremental retraining is triggered every 50 batches of data.

[0031] Beneficial effects

[0032] The present invention provides an Overlay data analysis and prediction system based on feedback-feedforward simulation, which has the following beneficial effects:

[0033] 1. This system adopts the wavelet transform-independent component analysis joint denoising technology and the deep residual network algorithm, and can accurately extract key features from a large amount of lithography data. Through the in-depth analysis of these features, the potential laws hidden behind the data are mined, so that the analysis of Overlay is no longer limited to the surface, and potential process problems can be discovered in advance, providing strong support for preventive maintenance and process optimization. In the data acquisition module, rich data is obtained through a variety of high-precision sensors, and then processed by the complex algorithms of the data processing and storage module, enabling the subsequent analysis and simulation module to deeply mine based on high-quality data and timely detect the subtle factors that may affect the lithography quality.

[0034] 2. This system uses a two-way coupling system of a physics-driven model and a deep residual network data-driven model, and dynamically allocates model weights through an attention mechanism. This innovative model architecture fully considers various complex factors in the lithography process, overcomes the simplification drawbacks of traditional models and the limitations of single algorithms. Thus, the accuracy of Overlay prediction is significantly improved, the lithography deviation is effectively reduced, and the performance and production yield of the chip are guaranteed. In the analysis and simulation module, through continuous model training and parameter optimization, the system's Overlay prediction under different process parameters and equipment states is more accurate, providing reliable prediction results for the lithography process.

[0035] 3. Based on the feedback-feedforward simulation mechanism, the system of the present invention has real-time feedback ability and can dynamically adjust the process according to the real-time data during the lithography process. When process parameters change or equipment status fluctuates, the system can respond quickly, automatically adjust relevant parameters, and achieve adaptive adjustment. In the prediction and correction module, the Overlay deviation is predicted based on the real-time collected data, and the lithography parameters are adjusted in a timely manner through the forward compensation and reverse correction functions, ensuring that the lithography process is always in a stable and reliable state, greatly enhancing the stability and reliability of the lithography process.

[0036] 4. The system of the present invention realizes scientific adjustment of lithography parameters through intelligent algorithms and automated processes. It no longer relies on the personal experience of operators, avoiding the problem of inconsistent adjustment effects caused by differences in the experience of different operators. At the same time, the system can quickly respond to changes in the production process, make accurate parameter adjustments in a timely manner, effectively improve production efficiency and product quality. The fault diagnosis module predicts lithography equipment failures in advance through abnormal analysis of the collected data, and the online learning function module automatically adjusts the model weights according to the prediction errors, all of which make the entire lithography production process more scientific, efficient, and reduce the interference of human factors. Description of the Drawings

[0037] Figure 1 It is the main system framework flow chart of the present invention;

[0038] Figure 2 It is the data acquisition simulation diagram of the present invention;

[0039] Figure 3 It is the real-time trend simulation diagram of the present invention;

[0040] Figure 4 It is the comparative analysis simulation diagram of the present invention;

[0041] Figure 5 It is the CFD thermal field distribution cloud diagram of the present invention;

[0042] Figure 6 It is the AR-assisted calibration interface simulation diagram of the present invention;

[0043] Figure 7 It is the three-level deviation warning system simulation diagram of the present invention.

[0044] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. Specific Embodiment 1:

[0046] As shown Figures 1-7 in the figure, the Overlay data analysis and prediction system based on feedback / feedforward simulation includes a main system, which contains a data acquisition module, a data processing and storage module, an analysis and simulation module, and a prediction and correction module;

[0047] The data acquisition module is a sensor set at key positions of the lithography equipment, including an optical encoder with a resolution of 0.1 nm, a 64-channel MEMS vibration sensor, and an infrared thermal imaging unit deployed in the exposure unit, the worktable, and the lens group, which can collect Overlay data of production batches in real time;

[0048] The data processing and storage module adopts a distributed storage architecture based on the IPFS protocol, performs joint denoising of wavelet transform - independent component analysis on the collected data, and stores it in the blockchain database through the SM4 national secret encryption unit after removing noise interference;

[0049] The analysis and simulation module uses a bidirectional coupling system of a physically driven model and a data-driven model of a deep residual network; dynamically assigns model weights of 0.3 - 0.7 through an attention mechanism to simulate the production process under different parameters

[0050] The prediction and correction module predicts the Overlay deviation according to the simulation results, and adjusts the lithography parameters through forward compensation and reverse correction functions. It uses a bidirectional coupling system of a physically driven model and a data-driven model of a deep residual network, and dynamically assigns model weights of 0.3 - 0.7 through an attention mechanism to simulate the production process under different parameters.

[0051] In the analysis and simulation module, the feedback algorithm establishes a thermal-mechanical coupling model based on the Navier-Stokes equation, and performs dynamic compensation according to the difference between the actual data and the expected data in the production process. The feedforward algorithm extracts high-order features through a deep residual network, and pre-judges and simulates the production process based on preset parameters and models; an incremental learning strategy is introduced, and the model is retrained every 50 batches of data are processed to dynamically optimize the algorithm parameters.

[0052] In the prediction and correction module, forward compensation reduces the Overlay deviation by adjusting the exposure energy and alignment offset parameters of the lithography equipment. Reverse correction uses an XGBoost intelligent decision tree model to automatically select a correction strategy according to the types of historical process scenarios, and performs reverse correction on the over-adjustment during the correction process. At the same time, a dynamic time warping algorithm is equipped to achieve spatio-temporal synchronization of multi-sensor data. Ensure the accuracy of lithography alignment.

[0053] The data acquisition module uses fiber Bragg grating strain measurement units to collect Overlay data in real time. It is also equipped with an adaptive acquisition frequency adjustment function, which automatically switches the sampling rate within the range of 50Hz - 10kHz according to the vibration amplitude of the equipment, and integrates a temperature compensation circuit to eliminate the influence of temperature drift of ±0.5°C on the sensor accuracy.

[0054] The data processing and storage module adopts a data sharding storage architecture, stores process data in blocks according to timestamps, uses the LZMA compression algorithm to reduce storage space, and realizes blockchain traceability management of key data through the Merkle tree version tracing mechanism and attribute-based encrypted access control.

[0055] The main system also includes a digital twin visualization module, which can display the analysis results in the form of intuitive graphs and charts, including real-time trend charts and comparative analysis charts, to facilitate operators to quickly understand the production status. Through the CFD thermal field distribution cloud map, the temperature change of ±0.1°C is dynamically displayed. Using the AR-assisted calibration interface, it is compatible with the Hololens2 device to achieve virtual-real superposition alignment, and a three-level deviation warning system is established, with yellow (>3nm), orange (>5nm), and red (>8nm) alarm thresholds set.

[0056] The main system has a multi-device collaboration function, which is realized through a quantum secure communication unit. It integrates NTRU lattice-based encryption to ensure data transmission between lithography devices, and the BB84 protocol monitoring module is used to detect the channel security in real time. At the same time, the 256-bit quantum key is automatically refreshed every hour.

[0057] A fault diagnosis module is introduced into the main system. By analyzing the anomalies in the collected data, it can predict the faults of the lithography equipment in advance. The faults include a process knowledge graph engine: using the BERT-BiLSTM-CRF model to construct a fault entity relationship network, matching historical maintenance solutions through a case reasoning mechanism, and using a dynamic node embedding algorithm to achieve adaptive update of fault features.

[0058] The main system has an online learning function, which is specifically manifested as: equipped with a model fusion unit to automatically adjust the weights of the physical-data model according to the prediction error, starting the 10kHz high-frequency acquisition mode under emergency conditions to update the compensation parameters in real time; triggering incremental retraining every 50 batches of data.

[0059] The system realizes precise control of lithography Overlay through a closed-loop control of multi-source perception, intelligent analysis, and dynamic correction. High-precision sensors (0.1nm optical encoder, 64-channel MEMS vibration array, infrared thermal imager) deployed at key stations of the lithography machine collect exposure position, equipment vibration, and temperature field data in real time. The adaptive sampling frequency dynamically switches between 50Hz - 10kHz according to the working conditions, and temperature compensation is embedded to eliminate environmental interference. After the collected data is processed by wavelet-ICA joint noise reduction, it is stored in the blockchain database through the IPFS distributed architecture, and Merkle tree tracing and SM4 encryption are used to ensure data security. The analysis module integrates a physical model (Navier-Stokes thermo-coupling equation) and a data model (improved ResNet network), and dynamically allocates the weights of the two (0.3 - 0.7) through an attention mechanism. The model parameters are updated by incremental learning every 50 batches of data processed. The prediction and correction module combines fuzzy PID forward compensation (adjusting the exposure dose by ±0.5mJ / cm 2 、alignment offset by ±2nm) and XGBoost backward correction (decision tree matching 20 process scenarios), and the six-degree-of-freedom nano platform performs closed-loop calibration with an accuracy of ≤1.5nm. The digital twin interface visualizes the CFD thermal field, AR-assisted alignment (integrated with Hololens2), and three-level early warning (yellow / orange / red). The quantum communication unit (NTRU encryption + BB84 protocol) ensures the security of multi-device data collaboration. The fault diagnosis engine dynamically associates historical cases through a knowledge graph, and the online learning system continuously optimizes the model weights and acquisition strategy, forming a full-process autonomous optimization of "perception - analysis - decision - verification". Specific Embodiment 2:

[0061] As Figures 1-7 shown, the key algorithms mentioned in Embodiment 1 are analyzed in detail below, including their core mathematical formulas and explanations:

[0062] Wavelet transform - independent component analysis joint noise reduction:

[0063] Wavelet transform:

[0064] The formula for discrete wavelet transform (DWT) is:

[0065]

[0066] W f (m,n) is the wavelet coefficient, which is the coefficient value of the original signal f(k) after wavelet transform at scale m and translation n, representing the characteristic information of the original signal at different scales and positions.

[0067] f(k) is the original signal, that is, the Overlay data collected from the lithography equipment sensor, and k represents discrete time or space sampling points.

[0068] is the conjugate of the wavelet basis function, ψ m,n (k) is the wavelet basis function, which is the function after being scaled by scale m and translated by n. m is the scale parameter that controls the scaling degree of the wavelet function. A larger m corresponds to low-frequency components, and a smaller m corresponds to high-frequency components; n is the translation parameter that controls the position of the wavelet function in time or space.

[0069] Wavelet transform is a time-frequency analysis method. It decomposes the original signal f(k) into different frequency components by performing an inner product operation (i.e., m,n ) with wavelet basis functions ψ ) at different scales and translations, and observes the signal at different scales and positions, thereby realizing the time-frequency analysis of the signal.

[0070] In the data processing and storage module, it is used to decompose the collected overlav data. Since the frequency characteristics of noise and useful signals are different, wavelet transform can highlight the details and trends in the signal, distinguish noise and useful signals at different scales, prepare for subsequent independent component analysis, help remove noise interference, improve data quality, and make the data for subsequent storage and analysis more accurate and reliable.

[0071] Independent Component Analysis (ICA):

[0072] Assume that the observed signal x = [x1, x2,..., x n ) T is obtained by linearly mixing the independent components s = [s1, s2,..., s n , that is, x = As, where A is the mixing matrix. The goal of ICA is to find a demixing matrix W such that y = Wx is as close as possible to the independent components s. Usually, it is solved by maximizing non-Gaussianity, using negentropy as the metric. The negentropy formula is: T J(y) = H(y

[0073] ) - H(y) gauss )

[0074] where x is the observed signal vector, which is the mixed signal collected from the lithography equipment and preliminarily processed. x i represents the i-th observed signal among them;

[0075] s is the independent component vector, which is the mutually independent original signal components that we hope to separate from the mixed signal. s i represents the i-th independent component among them;

[0076] A is the mixing matrix, which describes how the independent components are linearly mixed to form the observed signal. The element a in the matrix ijIndicates the contribution degree of the j-th independent component to the i-th observed signal;

[0077] W is the demixing matrix. By finding an appropriate W, a linear transformation is performed on the observed signal x (y = Wx) such that the transformed signal y is as close as possible to the independent components s;

[0078] J(y) is the negentropy, which is used to measure the difference degree between the signal y and the Gaussian distribution. The larger the negentropy, the more the signal y deviates from the Gaussian distribution;

[0079] H(y gauss ) is the entropy of the Gaussian distribution having the same variance as y. Entropy is a measure of the uncertainty of a random variable, and the entropy of the Gaussian distribution is a fixed value, which is only related to the variance;

[0080] H(y) is the entropy of the signal y, which reflects the uncertainty or information content of the signal y.

[0081] ICA assumes that the observed signals are linearly mixed by mutually independent components, and the demixing matrix W is found by maximizing the non-Gaussianity (measured by negentropy here). Since the independent components are usually non-Gaussian distributed, and the Gaussian distribution is the most "random" distribution, when the negentropy of the separated signals is the largest, it means that they are least like the Gaussian distribution and are thus closest to the independent components.

[0082] In the data processing and storage module, in combination with wavelet transform, the noise components in the data after wavelet transform are further removed. The demixing matrix is found through ICA to separate the noise and useful signals in the mixed signals, obtaining the true useful signals, improving the reliability of the data, facilitating subsequent storage and analysis, and ensuring that the stored data truly reflects the working state of the lithography equipment.

[0083] Feedback algorithm - thermo-mechanical coupling model based on the Navier-Stokes equation:

[0084] Navier-Stokes equation (incompressible fluid):

[0085] Continuity equation:

[0086] is the Hamiltonian operator It is a vector differential operator used for performing derivative operations on functions in the spatial direction.

[0087] u = (u, v, w): velocity vector, where u, v, and w are the velocity components in the x, y, and z directions respectively, describing the motion velocity of the fluid in space. This equation represents the mass conservation of the fluid, that is, the mass of the fluid flowing into a certain microelement within a unit time is equal to the mass of the fluid flowing out of the microelement, meaning that there is no source or sink of mass in the flow field.

[0088] Momentum equation:

[0089] where ρ is the fluid density, which is the mass of the fluid per unit volume and reflects the density of the fluid, t is the time, used to describe the change of fluid motion over time, is the partial derivative of the velocity vector u with respect to time t, representing the rate of change of velocity over time, is the convection term, which describes the momentum transfer caused by the macroscopic motion (velocity u) of the fluid, is a vector operator, representing the directional derivative along the velocity direction, is the pressure gradient, p is the pressure, represents the rate of change of pressure in space, and the force generated by it drives the fluid motion. μ is the dynamic viscosity, reflecting the viscous characteristics inside the fluid, that is, the ability of the fluid to resist deformation, is the Laplace operator acting on the velocity vector It describes the second-order spatial change of velocity and is related to the viscous force. f is the external force, such as the external forces of gravity and electromagnetic force acting on the fluid.

[0090] The continuity equation ensures the conservation of mass during fluid flow and is the embodiment of the law of conservation of mass in fluid mechanics. The momentum equation describes that the rate of change of the momentum of a fluid element is equal to the sum of various forces acting on the element, including the pressure gradient force viscous force and external force (f), as well as the momentum change caused by the fluid's own motion is the application of Newton's second law in fluid mechanics.

[0091] In the analysis and simulation module, it is used to establish a thermal-mechanical coupling model. During the operation of a lithography device, problems related to the flow of fluid (such as cooling medium), heat transfer, and stress distribution are involved. Through the Navier-Stokes equation, the flow situation of the fluid can be calculated based on the temperature and pressure data during the production process, and then the stress distribution can be analyzed. Dynamic compensation is carried out according to the difference between the actual data and the expected data, making the simulation results more in line with the actual production situation, improving the accuracy of the simulation, better predicting the performance of the lithography device under different working conditions, and providing a basis for optimizing the production process.

[0092] Feedforward algorithm - Deep Residual Network:

[0093] Basic unit of Deep Residual Network (ResNet)

[0094] Assume the input is x, after a series of convolution operations (denoted as F(x)), the output is:

[0095] y = F(x) + x

[0096] Where x is the input data, which in this solution is the data of the lithography equipment production process after preprocessing (such as data acquisition and preliminary noise reduction), containing information about the equipment status, process parameters, etc. F(x) is the residual function, which is usually composed of multiple convolutional layers and activation functions. It represents the operation of feature extraction and transformation of the input x, and extracts high-order features from the input data. y is the output data, which is the result after residual connection. It not only contains the features extracted by the convolution operation (F(x)), but also retains the information of the original input data x.

[0097] Traditional neural networks are prone to gradient vanishing or gradient exploding problems when the number of layers is increased, making training difficult. ResNet introduces residual connections), allowing the network to learn the residual F(x) instead of directly learning the complex input-to-output mapping relationship. In this way, gradients can be more easily transferred during the back-propagation process, thereby more effectively training deep networks.

[0098] In the analysis and simulation module, it is used to extract high-order features from the production process data. By training the deep residual network with a large amount of historical data, it can learn the complex feature patterns in the data. According to the pre-set parameters and models, the production process is predicted and simulated, and the learned features are used to predict the future production process, improve the accuracy of the prediction of the future production process, discover possible problems in advance, and provide support for production decisions.

[0099] Prediction and Correction Module - Forward Compensation:

[0100] Forward compensation parameter adjustment:

[0101] Assuming that the exposure energy of the original lithography equipment is E0 and the alignment offset is (x0, y0), according to the predicted Overlay deviation Δo, the adjusted exposure energy is:

[0102] E=E0+k1Δo

[0103] The adjusted alignment offset is:

[0104] (x,y)=(x0+k2Δo,y0+k3Δo)

[0105] Among them, E0 is the exposure energy of the original lithography equipment, which is an important parameter in the lithography process and affects the exposure degree of the photoresist and the quality of pattern transfer. (x0, y0) is the original alignment offset, that is, the initial deviation of the lithography equipment during the alignment process. x0 and y0 respectively represent the offsets in the x and y directions; Δo is the predicted Overlay deviation, which is the deviation value between the lithography pattern predicted by the analysis simulation module and the expected pattern. E is the adjusted exposure energy. According to the predicted Overlay deviation Δo, the original exposure energy E0 is adjusted according to the coefficient k1 to compensate for possible deviations. (x, y) is the adjusted alignment offset, and in the x and y directions respectively, according to the predicted Overlay deviation Δo, the original alignment offset (x0, y0) is adjusted according to the coefficients k2 and k3. k1, k2, and k3 are adjustment coefficients, and these coefficients are determined according to experiments or experience, and they determine the amplitude of adjusting the exposure energy and alignment offset according to the Overlay deviation.

[0106] According to the predicted Overlay deviation, the exposure energy and alignment offset are adjusted in a certain proportion. When an Overlay deviation is predicted, the exposure effect of the photoresist can be changed by adjusting the exposure energy, and the lithography pattern can be aligned more accurately by adjusting the alignment offset, thereby reducing the deviation.

[0107] In the prediction and correction module, by adjusting the parameters of the lithography equipment, the predicted Overlay deviation is directly compensated. During the lithography process, the Overlay deviation will affect the manufacturing accuracy and performance of the chip. Through positive compensation, the accuracy of lithography alignment can be improved, the product quality problems caused by the deviation can be reduced, and the yield of chip manufacturing can be improved.

[0108] Prediction and correction module - Reverse correction - XGBoost intelligent decision tree model:

[0109] XGBoost objective function:

[0110]

[0111] Among them, Obj(θ) is the objective function, which is used to measure the performance of the model. θ is the model parameter, including the structure of the decision tree and the node splitting conditions. is the loss function part, n is the number of samples, y i is the true value of the i-th sample. In this solution, it may be the actual Overlay deviation or the true information such as the lithography process scenario category. is the predicted value of the model for the i-th sample. represents the predicted value and the true value y iThe differences between them. Common loss functions include mean squared error (for regression problems) and cross entropy (for classification problems). By minimizing this part of the loss function, the predicted values of the model can be made closer to the true values. is the regularization term, K is the number of trees, and f k is the k-th tree, and Ω(f k ) is used to measure the complexity of the k-th tree, including factors such as the depth of the tree and the number of leaf nodes. The role of the regularization term is to prevent the model from overfitting, enabling the model to have better generalization ability and avoiding the model performing well on the training data but poorly on the test data or in actual applications.

[0112] XGBoost trains the model by minimizing the objective function. The loss function measures the prediction error of the model, prompting the model to continuously adjust the parameters to improve the prediction accuracy; the regularization term limits the complexity of the model, preventing the model from overfitting the training data and enabling the model to better adapt to different actual situations.

[0113] In the prediction and correction module, the correction strategy is automatically selected according to the types of historical process scenarios, and the over-adjustment during the correction process is reversely corrected. The XGBoost model can learn the features and patterns in the historical process scenario data. When encountering a new process scenario, it can automatically select an appropriate correction strategy based on the learned knowledge. At the same time, after the lithography parameters are adjusted forward, there may be over-adjustment. XGBoost can reversely correct the over-adjustment according to the historical data and the current situation, and use its powerful classification and regression capabilities to improve the accuracy and effectiveness of the correction strategy selection, and further improve the accuracy of lithography alignment. Specific Embodiment Three:

[0115] As Figures 1-7 shown, the following is a description of the specific application logic steps of each module and algorithm in the Overlay data analysis and prediction system based on feedback-feedforward simulation:

[0116] Data acquisition module:

[0117] First, the Overlay data of the production batch is collected in real time through sensors set at key positions of the lithography equipment. The sampling rate is dynamically adjusted according to the vibration amplitude of the equipment. When the vibration amplitude of the equipment is small, a lower sampling rate (50Hz) is used to save resources; when the vibration amplitude is large, it is switched to a higher sampling rate (10kHz) to ensure that the data can accurately reflect the equipment status.

[0118] Hardware composition and description:

[0119] 0.1nm Resolution Optical Encoder: Deployed in the exposure unit, workpiece stage, and lens group, it is used to accurately measure the position information of these key components. With a resolution of up to 0.1nm, it can provide high-precision position basic data for Overlay data.

[0120] 64-channel MEMS Vibration Sensor: It monitors the vibration of the device in real time. The 64-channel design can comprehensively sense vibrations in different directions and provide vibration amplitude data for adaptive acquisition frequency adjustment.

[0121] Infrared Thermal Imaging Unit: It is used to monitor the temperature distribution of key parts of the device and helps analyze the impact of the device's thermal state on Overlay.

[0122] Fiber Bragg Grating Strain Measurement Unit: It acquires Overlay data in real time. Utilizing the characteristics of optical fibers, which are sensitive to temperature and strain changes, it can accurately obtain relevant data.

[0123] Temperature Compensation Circuit: Integrated in the sensor, it eliminates the influence of a temperature drift of ±0.5°C on the sensor's accuracy and ensures the accuracy of sensor data in different temperature environments.

[0124] Data Processing and Storage Module:

[0125] First, perform joint noise reduction of wavelet transform - independent component analysis on the collected data to remove noise interference. Then, encrypt the data through the SM4 national cryptography encryption unit. Next, adopt a distributed storage architecture based on the IPFS protocol or a data sharding storage architecture to store the process data in blocks according to the timestamp, and use the LZMA compression algorithm to reduce the storage space. Finally, through the Merkle tree version tracing mechanism and attribute-based encryption access control, realize the blockchain traceability management of key data.

[0126] Hardware Composition and Description:

[0127] SM4 National Cryptography Encryption Unit: A hardware encryption module that encrypts data using the SM4 encryption algorithm to ensure the security of data during storage and transmission.

[0128] Blockchain Database Storage Hardware: Used to build a distributed storage architecture, and realize the distributed storage of data based on the IPFS protocol, improving the reliability and scalability of data storage.

[0129] Analysis and Simulation Module:

[0130] A bidirectional coupling system that combines a physics-driven model and a data-driven deep residual network model dynamically allocates model weights of 0.3 - 0.7 through an attention mechanism to simulate the production process under different parameters. The feedback algorithm establishes a thermal-mechanical coupling model based on the Navier-Stokes equation and performs dynamic compensation according to the difference between the actual data and the expected data during the production process. The feedforward algorithm extracts high-order features through a deep residual network and pre-judges and simulates the production process based on preset parameters and models. The model is retrained every 50 batches of data to dynamically optimize the algorithm parameters.

[0131] Hardware composition and description: It mainly relies on computer hardware resources, such as high-performance CPUs and GPUs, to run complex models and algorithms for large-scale data calculation and analysis.

[0132] Prediction and correction module:

[0133] Predict the Overlay deviation based on the simulation results of the analysis and simulation module. For positive compensation, adjust the exposure energy and alignment offset parameters of the lithography equipment to reduce the Overlay deviation. For reverse correction, use the XGBoost intelligent decision tree model to automatically select the correction strategy according to the types of historical process scenarios, and perform reverse correction on the over-adjustment during the correction process. At the same time, a dynamic time warping algorithm is equipped to achieve spatio-temporal synchronization of multi-sensor data to ensure the accuracy of lithography alignment.

[0134] Hardware composition and description: It also relies on computer hardware resources to run prediction and correction algorithms and may need to interact with the control hardware of the lithography equipment to adjust the lithography parameters.

[0135] Digital twin visualization module:

[0136] Display the analysis results in the form of intuitive graphs and charts, including real-time trend graphs and comparative analyses. Through the CFD thermal field distribution cloud map, dynamically display temperature changes of ±0.1°C. Use the AR-assisted calibration interface to be compatible with the Hololens2 device to achieve virtual-real superposition alignment, and establish a three-level deviation warning system with yellow (>3nm), orange (>5nm), and red (>8nm) alarm thresholds to facilitate operators to quickly understand the production status.

[0137] Hardware composition and description:

[0138] Display device: Such as a high-resolution display screen, used to display various graphs, charts, and cloud maps.

[0139] AR device (Hololens2): Achieve virtual-real superposition alignment and assist operators in calibration work.

[0140] Multi-device collaboration function:

[0141] Multi-device collaboration is achieved through a quantum-secure communication unit. NTRU lattice-based encryption is integrated to ensure data transmission between lithography devices. The BB84 protocol monitoring module is used to detect the channel security in real time. Meanwhile, the 256-bit quantum key is automatically refreshed every hour to ensure the security and stability of data transmission between multiple devices.

[0142] Hardware composition and description:

[0143] Quantum-secure communication unit: It includes the hardware of quantum key distribution devices to achieve secure communication based on quantum technology.

[0144] NTRU lattice-based encryption hardware: It is used to encrypt the transmitted data to ensure the security of data transmission.

[0145] Fault diagnosis module: By analyzing the anomalies in the collected data, it can predict the faults of lithography devices in advance. The BERT-BiLSTM-CRF model is used to construct a fault entity relationship network. The historical maintenance solutions are matched through a case-based reasoning mechanism, and the adaptive update of fault features is realized by using a dynamic node embedding algorithm.

[0146] Hardware composition and description: It depends on computer hardware resources for data processing and model operations, and is used to run fault diagnosis algorithms and store historical fault data and maintenance solutions.

[0147] Online learning function:

[0148] It is equipped with a model fusion unit to automatically adjust the weights of the physics-data model according to the prediction error. In case of emergency conditions, a 10kHz high-frequency acquisition mode is started to update the compensation parameters in real time. Incremental retraining is triggered every 50 batches of data to continuously optimize the model performance.

[0149] Hardware composition and description: It mainly depends on computer hardware resources and is used to run model fusion algorithms and data acquisition mode switching control programs. Specific embodiment four:

[0151] As Figures 1-7 shown, the following is the explanation in sequence according to the generated simulation diagram:

[0152] Real-time trend graph:

[0153] This graph shows the changing trends of the actual Overlay data (blue solid line) and the expected Overlay data (red dashed line) over time. The X-axis represents time, and the Y-axis represents the value of the Overlay data. By observing the two curves, operators can intuitively see how the difference between the actual Overlay data and the expected value evolves over time during the actual production process. If the actual data curve deviates from the expected data curve, it may mean that there are abnormalities in the production process and further analysis and adjustment are required.

[0154] Comparison Analysis Chart:

[0155] This chart presents the difference between the actual Overlay data and the expected Overlay data in the form of a bar chart. The X-axis represents time, and the Y-axis represents the numerical value of the Overlay data difference. Each bar represents the difference between the actual data and the expected data at a specific time point. Bars in the positive half of the Y-axis indicate that the actual data is greater than the expected data, while bars in the negative half indicate that the actual data is less than the expected data. Through this comparison, it is clear at which time points the differences are significant, thus helping to analyze the fluctuations and problems in the production process.

[0156] CFD Thermal Field Distribution Cloud Chart:

[0157] This chart simulates the temperature distribution in a two-dimensional area, which is the CFD (Computational Fluid Dynamics) thermal field distribution. The X-axis and Y-axis represent the coordinate positions of the simulation area respectively. Different colors in the chart represent different temperature values. Through the color change, it is possible to visually see the temperature distribution in this area, such as whether there are areas with higher or lower temperatures and the trend of temperature change. The color bar is used to correspond to the colors and specific temperature values, helping to accurately understand the distribution characteristics of the thermal field.

[0158] AR-assisted Calibration Interface Simulation:

[0159] This chart simply simulates the interface effect of AR (Augmented Reality)-assisted calibration. The background is a simulated virtual image, which may represent a certain scene or view in the lithography equipment. The red '+' mark in the chart simulates the actual alignment point, used to indicate the position that needs to be calibrated in the virtual image. Through this combination of virtual and real, in practical applications, virtual information can be superimposed on the real scene with the help of AR devices (such as Hololens2), helping operators to perform calibration operations more accurately.

[0160] Three-level Deviation Warning System:

[0161] This chart shows the working conditions of the three-level deviation warning system using a scatter plot. The X-axis represents time, and the Y-axis represents the Overlay data. The blue scatter points in the chart represent normal data points, that is, the points where the difference between the actual Overlay data and the expected data is within the normal range. The yellow, orange, and red scatter points correspond to different levels of warnings (yellow warning, orange warning, red warning) respectively. These points indicate that the difference between the actual data and the expected data exceeds the corresponding warning thresholds (yellow corresponds to > 3nm, orange corresponds to > 5nm, red corresponds to > 8nm). Through this chart, it is possible to clearly see at which time points different degrees of deviation warnings occur, so as to take timely measures for adjustment and processing.

[0162] It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising a reference structure" does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0163] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An Overlay data analysis and prediction system based on feedback / feedforward simulation, including a main system, characterized in that: The main system includes a data acquisition module, a data processing and storage module, an analysis and simulation module, and a prediction and correction module; The data acquisition module is a sensor set at a key position of the lithography equipment, including a resolution optical encoder, a vibration sensor and an infrared thermal imaging unit deployed on the exposure unit, the workpiece stage and the lens group, to collect the overlay data of the production batch in real time; The data processing and storage module adopts a distributed storage architecture based on the IPFS protocol to perform wavelet transform-independent component analysis joint noise reduction on the collected data, and after removing noise interference, stores it in the blockchain database through the SM4 national secret encryption unit; The analysis and simulation module uses a bidirectional coupling system of a physical driving model and a deep residual network data-driven model; The prediction and correction module predicts Overlay deviation according to simulation results and adjusts lithography parameters through forward compensation and reverse correction functions.

2. The Overlay data analysis and prediction system based on feedback / feedforward simulation according to claim 1, characterized in that: In the analysis and simulation module, the feedback algorithm establishes a thermal-mechanical coupling model based on the Navier-Stokes equations, and performs dynamic compensation according to the difference between the actual data and the expected data in the production process. The feedforward algorithm extracts high-order features through a deep residual network, and predicts and simulates the production process according to pre-set parameters and models. An incremental learning strategy is introduced to trigger model retraining every 50 batches of data processing, and dynamically optimize the algorithm parameters.

3. The overlay data analysis and prediction system based on feedback / feedforward simulation according to claim 2, characterized in that: In the prediction and correction module, forward compensation reduces Overlay deviation by adjusting the exposure energy and alignment offset parameters of the lithography equipment. The reverse correction uses the XGBoost intelligent decision tree model to automatically select the correction strategy according to the historical process scenario types, and reversely correct the excessive adjustments in the correction process. At the same time, it is equipped with a dynamic time warping algorithm to achieve spatiotemporal synchronization of multi-sensor data.

4. The overlay data analysis and prediction system based on feedback / feedforward simulation according to claim 3, wherein: The data acquisition module adopts a fiber Bragg grating strain measurement unit to collect overlay data in the application process in real time; and is also equipped with an adaptive acquisition frequency adjustment function.

5. The overlay data analysis and prediction system based on feedback / feedforward simulation according to claim 4, characterized in that: The data processing and storage module adopts a data sharding storage architecture to store process data in blocks according to timestamps; uses the LZMA compression algorithm to reduce storage space, and implements blockchain traceability management of key data through a Merkle tree version tracing mechanism and attribute-based encrypted access control.

6. The Overlay data analysis and prediction system based on feedback / feedforward simulation according to claim 1, characterized in that: The main system also includes a digital twin visualization module, which dynamically displays temperature changes through CFD thermal field distribution cloud maps, uses an AR-assisted calibration interface, and is compatible with Hololens2 devices to achieve virtual-real superposition alignment.

7. The overlay data analysis and prediction system based on feedback / feedforward simulation according to claim 1, characterized in that: The main system has a multi-device coordination function, which is realized through a quantum secure communication unit. It integrates NTRU lattice encryption to ensure data transmission between lithography devices, and detects channel security in real time through a BB84 protocol monitoring module. It also automatically refreshes the 256-bit quantum key every hour.

8. The Overlay data analysis and prediction system based on feedback / feedforward simulation according to claim 1, characterized in that: A fault diagnosis module is introduced into the main body system. By analyzing the anomalies of the collected data, faults of the lithography equipment can be predicted in advance. Among them, the faults include a process knowledge graph engine: a fault entity relationship network is constructed using the BERT-BiLSTM-CRF model, historical maintenance solutions are matched through a case reasoning mechanism, and an adaptive update of fault features is realized using a dynamic node embedding algorithm.

9. The overlay data analysis and prediction system based on feedback / feedforward simulation according to claim 1, wherein: The main body system has an online learning function, which is specifically manifested as follows: it is equipped with a model fusion unit to automatically adjust the weights of the physical-data model according to the prediction error, starts a 10kHz high-frequency acquisition mode under emergency conditions, and updates the compensation parameters in real time; incremental retraining is triggered every 50 batches of data.

Images