Integrated circuit equipment data optimization monitoring system and method based on big data

Abstract

The invention discloses an integrated circuit equipment data optimization monitoring system and method based on big data, and relates to the technical field of integrated circuit manufacturing. The method is used for solving the problems of insufficient multi-physical field monitoring, difficulty in abnormal traceability and lack of closed-loop control in plasma etching. A plasma sheath thickness inversion model and an etching selection ratio model are constructed by collecting radio frequency reflection phase, mass spectrum ion strength and wafer temperature data, and process parameter-plasma state dynamic coupling is established. Interference image distortion features and electron microscope size data are fused, micro-groove geometric parameters are analyzed, morphology instability risk indexes are generated, and nanoscale early warning is achieved. And designing a dual-channel fusion network, embedding a physical constraint attention mechanism, and generating an etching rate optimization instruction. On the basis of incremental learning, model parameters are updated online, a'monitoring-decision-feedback 'closed-loop system is formed, anomaly detection sensitivity and decision reliability are improved, and technical support is provided for intelligence of integrated circuit equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of integrated circuit manufacturing, in particular to an integrated circuit equipment data optimization monitoring system and method based on big data. BACKGROUND

[0002] With the continuous upgrading of chip performance demand of new generation information technology, integrated circuit manufacturing technology is constantly evolving towards higher precision and more complex structure. In advanced process technology, plasma etching as a key link to form nanoscale device structure, its process control precision directly affects the chip performance and yield. However, the dynamic characteristics of plasma state and the evolution of wafer surface topography in etching process have strong coupling relationship, and traditional monitoring methods are difficult to realize the coordinated analysis of multi-physical field, resulting in narrow process window and high defect rate. Under this background, it is urgent to develop new monitoring methods to break through the real-time perception and closed-loop control problem in complex process environment.

[0003] The current integrated circuit equipment monitoring technology has the following core problems, the existing method relies on historical data to establish fixed model, cannot adapt to the dynamic change of plasma density fluctuation, reaction cavity state drift, etc., leading to significant etching rate prediction deviation. Most systems only process single type data, cannot model the coupling of multi-source data such as radio frequency parameters, gas state, temperature distribution, etc., and it is difficult to reveal the root cause of process abnormality. The traditional deep learning model does not effectively fuse the time series physical field data and spatial topography features, resulting in insufficient micro defect detection sensitivity, and unable to realize early warning. After adjusting the process parameters, there is no real-time feedback verification mechanism, and the monitoring model cannot be optimized and iterated online, and the performance degrades seriously after long-term operation. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides an integrated circuit equipment data optimization monitoring system and method based on big data, which solves the problems in the above background art.

[0005] In order to achieve the above object, the present application is realized by the following technical scheme: a kind of integrated circuit equipment data optimization monitoring method based on big data, comprising the following steps: S1. the reflection phase time sequence signal of radio frequency matching network in plasma etching process, residual gas mass spectrum fragment ion intensity data and wafer surface temperature distribution data are collected, the dynamic thickness distribution of plasma sheath is inverted by extracting radio frequency phase fluctuation characteristics, and dynamic prediction model of etching selection ratio is established based on the correlation analysis of mass spectrum ion intensity ratio and temperature gradient, to generate the dynamic coupling relationship matrix of process parameters and plasma state;S2. based on dynamic coupling relationship matrix, the distortion mode spatial distribution characteristics of scattered field interference image are analyzed, the microgroove distortion geometric parameter of wafer edge area is identified by deformable convolution network, the mapping relationship of surface roughness mutation feature and etching rate prediction value is constructed in combination with the key dimension data of scanning electron microscope image, and morphology instability risk index is generated;S3. the dynamic thickness distribution of plasma sheath and morphology instability risk index are input into feature fusion network, and plasma impedance characteristics and microgroove distortion mode are processed respectively by time sequence analysis and spatial feature extraction, the time sequence data confidence and spatial feature weight are dynamically allocated by physical constraint attention mechanism, etching rate optimization adjustment amount is generated and process parameter correction instruction is triggered;S4. etching rate optimization adjustment amount is fed back to etching equipment controller in real time, and the radio frequency reflection phase data after process correction and updated interference image features are synchronously collected, and dynamic coupling relationship matrix and feature fusion network parameters are updated.

[0006] Further, the specific process of collecting the reflection phase time sequence signal of radio frequency matching network in plasma etching process, residual gas mass spectrum fragment ion intensity data and wafer surface temperature distribution data, and inverting the dynamic thickness distribution of plasma sheath by extracting radio frequency phase fluctuation characteristics is as follows: the radio frequency reflection phase time sequence signal is subjected to time-frequency analysis, the phase fluctuation amplitude and frequency offset in the characteristic frequency band are extracted, and the signal is divided into multiple process cycle segments by sliding time window;A dynamic inversion model is established based on the plasma sheath transmission equation, the phase fluctuation amplitude is mapped to the instantaneous change amount of sheath thickness, and the thickness gradient graph of wafer radial distribution is generated by spatial interpolation algorithm;The thickness gradient graphs of adjacent process cycles are subjected to difference calculation, the plasma density abnormal fluctuation area is identified, the thickness measurement error caused by thermal drift is corrected in combination with wafer table temperature distribution data, the corrected dynamic thickness distribution of sheath is cross-validated with radio frequency matching network impedance data, and abnormal inversion results beyond physical law are removed.

[0007] Furthermore, the construction logic of the dynamic prediction model for etching selectivity based on the correlation analysis between mass spectrometry ion intensity ratio and temperature gradient is as follows: The mass spectrometry fragment ion intensity data is normalized, the concentration ratio of key reactive groups is calculated, random noise interference is eliminated through moving average filtering, local hotspot regions in the temperature gradient distribution on the wafer surface are extracted, and a nonlinear correlation function between temperature gradient and etching selectivity is established. The concentration ratio of reactive groups and temperature gradient are input into a dual-channel neural network, and the influence weights of both on etching selectivity are dynamically allocated through an attention mechanism, outputting the dynamic prediction value under the current process window. The prediction model is corrected in real time using online etching rate monitoring data. When the prediction deviation continuously exceeds the tolerance range, the mass spectrometer calibration and temperature sensor re-inspection process is triggered.

[0008] Furthermore, the specific process for generating the dynamic coupling relationship matrix between process parameters and plasma state is as follows: A three-dimensional coupling relationship matrix is ​​constructed, where the X-axis represents RF power parameters, including reflected power and impedance matching; the Y-axis represents plasma state parameters, including sheath thickness and electron density; and the Z-axis represents process output parameters, including etching rate and selectivity. The inverted dynamic sheath thickness distribution and the predicted etching selectivity are mapped to the corresponding dimensions of the matrix, and key coupling factors are extracted using a tensor decomposition algorithm. Based on process stability constraints, the coupling relationship matrix is ​​sparsified, weakly correlated parameter combinations are eliminated, and strongly correlated feature channels are retained. The effectiveness of the matrix is ​​verified using real-time acquired process yield data. When the statistical process control data for key wafer dimensions exceeds the control limits, the coupling factor weight coefficients are recalculated.

[0009] Furthermore, the spatial distribution characteristics of distortion patterns in the scattered field interferometric image are analyzed, and the specific process of identifying the geometric parameters of microgroove distortion in the wafer edge region using a deformable convolutional network is as follows: The original interferometric image is preprocessed, and noise interference is eliminated by phase unwrapping algorithm to extract gradient information of the height distribution on the wafer surface; the input layer of the deformable convolutional network is constructed, and the gradient information is mapped to the geometric deformation field of the feature map. The sampling position of the convolution kernel is adaptively adjusted to match the curvature of the microgroove edge. The global morphological contour and local distortion components are separated in the deep features of the network, and multi-scale distortion patterns are captured by spatial pyramid pooling; the depth, tilt angle, and radius of curvature of the microgroove are output, and the spatial location distribution of the distortion region is determined by combining the wafer coordinate system.

[0010] Furthermore, by combining key dimensional data from scanning electron microscope (SEM) images, a mapping relationship is constructed between surface roughness abrupt change features and etching rate predictions. The specific process for generating a morphology instability risk index is as follows: Sub-pixel edge detection is performed on the SEM images to extract the linewidth, sidewall angle, and bottom roughness parameters of key structures. The SEM measurement data and the microgroove geometric parameters obtained from the interferometric image are spatially aligned, and coordinate system offset errors are eliminated through affine transformation. A roughness abrupt change feature vector is constructed, including gradient change rate, local curvature extreme point density, and abnormal spike height distribution. The feature vector is input into a random forest regression model, and the morphology instability risk index is calculated by combining it with the current etching rate prediction.

[0011] Furthermore, the specific process of inputting the plasma sheath dynamic thickness distribution and morphological instability risk index into the feature fusion network, and processing the plasma impedance characteristics and microgroove distortion modes through temporal analysis and spatial feature extraction is as follows: A dual-branch feature encoder is designed: the temporal branch performs multi-scale wavelet decomposition on the plasma sheath thickness sequence to extract transient impact response components and steady-state oscillation components; the spatial branch models the microgroove distortion mode using a graph convolutional network, constructing a topological adjacency matrix to describe the spatial correlation of the distortion region; the temporal and spatial features are mapped to a unified latent space through cross-modal projection, and adversarial training strategies are used to eliminate distribution differences between modes. A feature similarity matrix is ​​calculated in the latent space, and high-confidence cross-modal correlation feature pairs are selected.

[0012] Furthermore, the specific process of dynamically allocating the confidence level of time-series data and the weight of spatial features through a physical constraint attention mechanism to generate an etching rate optimization adjustment amount and trigger process parameter correction commands is as follows: Input the transient impact response amplitude of the time-series branch and the energy density of the distortion region of the spatial branch, and convert the plasma sheath equation into an inequality constraint to limit the allocation range of attention weights; calculate the time-series confidence level based on the signal-to-noise ratio of RF impedance data and the periodicity of steady-state oscillation components, and calculate the spatial confidence level based on the matching consistency between the micro-trench distortion mode and the scanning electron microscope measurement data; use the time-series confidence level and the spatial confidence level as dynamic weights to perform weighted fusion of multimodal features to generate an etching rate optimization adjustment amount. When the adjustment amount exceeds the process safety window, the RF power matching parameter adjustment command and the reactive gas flow rate ratio adjustment command are triggered sequentially.

[0013] Furthermore, the specific process of synchronously acquiring RF reflection phase data after process correction and updated interferometric image features, and updating the dynamic coupling relationship matrix and feature fusion network parameters is as follows: Store the RF phase data, interferometric images and corresponding etching rate adjustments before and after process correction, and establish a data association index based on timestamps and process batch numbers; Map new samples to the missing dimensions of the original three-dimensional matrix using a streaming tensor decomposition algorithm, control the growth of the matrix rank by truncating low singular values, freeze the pre-trained weights of the spatial branches of the feature fusion network, perform online fine-tuning only on the temporal branches, and introduce an elastic weight solidification algorithm to constrain the update magnitude of important parameters.

[0014] A big data-based integrated circuit equipment data optimization and monitoring system includes the following modules: a multi-source data dynamic modeling module, a morphology instability quantification module, a cross-modal optimization decision-making module, and a closed-loop self-updating module. The multi-source data dynamic modeling module collects the reflection phase timing signal of the RF matching network, residual gas mass spectrometry fragment ion intensity data, and wafer surface temperature distribution data during the plasma etching process. It extracts RF phase fluctuation characteristics to invert the dynamic thickness distribution of the plasma sheath and establishes a dynamic prediction model for etching selectivity based on the correlation analysis of mass spectrometry ion intensity ratio and temperature gradient, generating a dynamic coupling relationship matrix between process parameters and plasma state. The morphology instability quantification module analyzes the spatial distribution characteristics of distortion modes in the scattered field interferometric image based on the dynamic coupling relationship matrix and identifies wafer edge regions through a deformable convolutional network. The microgroove distortion geometry parameters of the domain are combined with key dimensional data from scanning electron microscope images to construct a mapping relationship between surface roughness abrupt changes and etching rate predictions, generating a morphological instability risk index. The cross-modal optimization decision module is used to input the dynamic thickness distribution of the plasma sheath and the morphological instability risk index into the feature fusion network. It processes the plasma impedance characteristics and microgroove distortion modes through temporal analysis and spatial feature extraction, respectively. It dynamically allocates the confidence level of temporal data and the weight of spatial features through a physical constraint attention mechanism to generate etching rate optimization adjustment and trigger process parameter correction commands. The closed-loop self-updating module is used to feed back the etching rate optimization adjustment to the etching equipment controller in real time, synchronously collect the RF reflection phase data after process correction and the updated interference image features, and update the dynamic coupling relationship matrix and feature fusion network parameters.

[0015] The present invention has the following beneficial effects: (1) A data optimization monitoring method for integrated circuit equipment based on big data significantly improves the accuracy and reliability of process status monitoring through dynamic sensing of multi-source heterogeneous data and multi-physics field coupling modeling. Based on the inversion of the dynamic thickness distribution of the plasma sheath by the radio frequency phase fluctuation characteristics, and combined with mass spectrometry and temperature gradient data, a dynamic prediction model of etching selectivity is established, which can respond to changes in the process environment in real time and solve the problem of inaccurate monitoring caused by chamber deposition and gas ratio fluctuations in traditional static models. By linking process parameters, plasma state and micro-morphology data through dynamic coupling relationship matrix, the mapping relationship between the distortion mode of the scattered field interference image and the key dimensions of the scanning electron microscope is analyzed, realizing the causal tracing of cross-scale process anomalies and breaking through the limitations of single data dimension analysis. A morphology instability risk index is constructed to quantify the nonlinear correlation between the surface roughness mutation characteristics and etching rate, which can trigger early warning in the early stage of micro-defect formation and avoid batch scrapping of wafers.

[0016] (2) A big data-based integrated circuit equipment data optimization monitoring system, relying on physical constraint feature fusion and closed-loop feedback mechanism, realizes intelligent process optimization decision-making and system self-evolution capability. A time-space dual-channel feature fusion network is designed, and plasma transport equation constraints are embedded in the attention mechanism to ensure that the etching rate optimization adjustment conforms to the physical laws of the process and avoids the risk of overfitting of the pure data-driven model. By real-time feedback of multi-source data after process correction, combined with incremental learning algorithm to dynamically update the coupled model and network parameters, it continuously adapts to long-term changes such as equipment aging and process drift, and maintains the high accuracy and stability of the monitoring system. A lightweight model architecture and streaming computing framework are adopted to support multimodal data parsing and real-time control command generation at the equipment edge, meeting the stringent requirements of integrated circuit manufacturing for low latency and high concurrency.

[0017] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0018] Figure 1 This is a flowchart of a data optimization and monitoring method for integrated circuit equipment based on big data, according to the present invention.

[0019] Figure 2 This is a flowchart of a big data-based integrated circuit equipment data optimization and monitoring system according to the present invention. Detailed Implementation

[0020] This application provides a big data-based integrated circuit equipment data optimization monitoring system and method. By constructing a cross-scale correlation model of plasma state, process parameters, and microstructure, and combining a feature fusion algorithm based on physical constraints with a closed-loop feedback mechanism, it solves the core problems of insufficient dynamic response capability and poor multimodal data coordination in traditional monitoring technologies, and provides real-time optimization decision support for high-precision etching processes.

[0021] The overall concept of the solution in this application embodiment is as follows: Dynamic sensing of multi-source heterogeneous data: Simultaneously acquire radio frequency reflection phase timing signals, residual gas mass spectrometry data, wafer surface temperature distribution and scattering field interference images in the plasma etching process, and extract multi-dimensional process state characterization quantities such as radio frequency phase fluctuation characteristics, plasma sheath thickness gradient distribution and micro-trench geometric distortion parameters through time-frequency analysis, spatial interpolation and other algorithms.

[0022] Multiphysics cross-scale coupling modeling: Establish a dynamic correlation model between process parameters (RF power, gas flow rate), plasma state (sheath thickness, electron density) and microstructure (surface roughness, critical dimensions), extract cross-scale coupling factors through tensor decomposition algorithm, and reveal the transmission law of process anomalies in multiphysics.

[0023] Intelligent decision-making under physical constraints: Design a temporal-spatial dual-channel feature fusion network to analyze the dynamic characteristics of plasma impedance in the temporal branch and identify the morphology distortion mode in the spatial branch. Dynamically allocate feature weights through an attention mechanism embedded in the plasma transport equation to generate etching rate optimization instructions that conform to physical laws.

[0024] Closed-loop feedback-driven model iteration: Optimization instructions are fed back to the etching equipment controller in real time, multi-source data after process adjustment are collected synchronously, and the parameters of the dynamic coupling model and feature fusion network are updated online through incremental learning algorithms to achieve continuous self-optimization of the monitoring system.

[0025] Please see Figure 1This invention provides a technical solution: a data optimization and monitoring method for integrated circuit equipment based on big data, comprising the following steps: S1. Acquiring the reflection phase timing signal of the radio frequency matching network, residual gas mass spectrometry fragment ion intensity data, and wafer surface temperature distribution data during the plasma etching process; extracting radio frequency phase fluctuation characteristics to invert the dynamic thickness distribution of the plasma sheath; and establishing a dynamic prediction model for etching selectivity based on the correlation analysis of mass spectrometry ion intensity ratio and temperature gradient, generating a dynamic coupling relationship matrix between process parameters and plasma state; S2. Based on the dynamic coupling relationship matrix, analyzing the spatial distribution characteristics of distortion modes in the scattered field interference image; and identifying the micro-trench distortion geometric parameters of the wafer edge region through a deformable convolutional network. By combining key dimension data from scanning electron microscope images, a mapping relationship between surface roughness abrupt changes and etching rate predictions is constructed to generate a morphological instability risk index. S3. The dynamic thickness distribution of the plasma sheath and the morphological instability risk index are input into a feature fusion network. Plasma impedance characteristics and microgroove distortion modes are processed through temporal analysis and spatial feature extraction, respectively. A physical constraint attention mechanism dynamically allocates the confidence level of temporal data and the weights of spatial features, generating an etching rate optimization adjustment amount and triggering process parameter correction commands. S4. The etching rate optimization adjustment amount is fed back to the etching equipment controller in real time. Simultaneously, the RF reflection phase data after process correction and the updated interferometric image features are acquired, updating the dynamic coupling matrix and feature fusion network parameters.

[0026] In this implementation scheme, step S1: By acquiring and modeling multi-dimensional sensor data, a dynamic correlation model between process parameters and plasma physical state is constructed to provide a physical field reference benchmark for subsequent morphology prediction and control optimization. Data acquisition: Radio frequency reflection phase timing signal: used to characterize the impedance matching between radio frequency power and plasma; its phase fluctuation reflects plasma density and its time-varying characteristics. Residual gas mass spectrometry ion intensity: by monitoring the intensity changes of specific ions in the cavity, the active concentration of chemical groups in the etching reaction is quantified. Wafer surface temperature distribution: the two-dimensional temperature field of the wafer during the etching process is acquired to determine the risk of local etching non-uniformity caused by thermal inhomogeneity. Feature extraction and modeling: the phase fluctuation characteristics of the radio frequency signal are extracted using time-frequency analysis techniques such as wavelet transform, and its dynamic thickness distribution is inverted by combining the plasma sheath theoretical model. A nonlinear correlation model is established based on the mass spectrometry ion intensity ratio and temperature gradient to construct a dynamic prediction mechanism for etching selectivity. By summarizing multiple physical field variables such as radio frequency, plasma, and temperature, a three-dimensional coupling relationship tensor is constructed to describe its synergistic effect with etching process output parameters (rate, selectivity). Plasma sheath: The charge barrier region between the plasma and the solid boundary; its thickness affects the acceleration path and energy distribution of ions. Etching selectivity: The etching rate ratio between the target material and the mask or substrate; a core indicator for precision control of the etching process. Step S2: Multimodal correlation between plasma state and wafer surface morphology changes enables early identification and risk level quantification of nanoscale morphology defects. Interference image analysis: Surface height perturbation data is extracted using scattered field interference images; spatial distortion patterns of interference fringes are analyzed to determine structural anomalies such as microgrooves and sidewall fluctuations. A deformable convolutional neural network (DCN) is introduced to automatically adjust the sampling position of the convolution kernel according to the deformation region, thereby extracting the distortion geometric parameters of microgrooves at the wafer edge with high precision. Cross-modal alignment and roughness modeling: Key dimensional parameters such as linewidth, aspect ratio, and sidewall angle measured in scanning electron microscope images are spatially aligned with interferogram data to achieve cross-domain data fusion. The abrupt changes in local roughness characteristics of the wafer surface (such as gradient extreme density) are quantified, and a regression mapping is established with the etching rate prediction deviation. Risk Index Generation: Based on a random forest model, morphological features and etching deviation indicators are integrated to output a morphological instability risk score, supporting subsequent multi-level early warning control (low / medium / high risk classification). Deformable Convolutional Network (DCN): A neural network structure with adaptive spatial awareness, suitable for complex distortion feature extraction tasks. Morphological Instability: Refers to microstructural anomalies occurring during wafer etching, such as edge collapse and irregular sidewall corners, often caused by plasma inhomogeneity. Step S3: By integrating temporal physical data and spatial morphological features, combined with plasma physical constraints, etching process correction instructions conforming to manufacturing rules are generated to achieve intelligent control optimization.Dual-channel feature processing: Temporal path: Multi-scale wavelet decomposition is used to analyze the sheath thickness sequence, extracting high-frequency perturbations and low-frequency steady-state components to reflect the dynamic response characteristics of plasma. Spatial path: A spatial correlation model between local distortion regions on the wafer is constructed using a graph convolutional network (GCN) to extract the distortion propagation trend. Physically constrained attention mechanism: Radio frequency signal-to-noise ratio and image sharpness are introduced as feature confidence indicators, and the fusion ratio of temporal and spatial features is dynamically adjusted through an attention mechanism. The plasma sheath equation is embedded to constrain the distribution of attention weights, ensuring that the optimization suggestions have physical consistency. Optimization instruction generation: First-level instruction: Quickly adjust the parameters of the radio frequency matching network (such as impedance matching tolerance) to stabilize the plasma. Second-level instruction: Refine the adjustment of the flow rate ratio of the reactive gas components to achieve selectivity precision control. Attention mechanism: Feature weighting strategy used in deep learning to focus on key information. Plasma sheath equation: Describes the basic physical laws of the relationship between voltage, charge density, and distance during the acceleration of sheath ions. Step S4: Based on a real-time feedback mechanism, the model is subjected to online learning and adaptive optimization to improve the robustness and long-term adaptability of the system. Feedback Data Acquisition and Storage: Samples such as RF phase sequences, interferometric images, and correction amounts before and after each etching correction are stored in an incremental learning sample library, organized by batch and time sequence. Dynamic Coupling Update: Streaming tensor decomposition is performed on the original 3D coupling matrix to absorb new data. Low-rank constraints are used to suppress measurement noise and maintain matrix structure stability. Model Structure Iteration: An elastic weight solidification algorithm is used to calculate the importance of model parameters, penalizing historical key parameters to prevent "catastrophic forgetting" of the model. Spatial feature extraction branches are frozen, with only temporal channel parameters fine-tuned to ensure the model can quickly adapt to new process states while retaining its morphology recognition capabilities. Streaming Tensor Decomposition: An incrementally updatable tensor dimensionality reduction algorithm suitable for real-time high-dimensional data structure updates. Elastic Weight Solidification: An incremental learning strategy that uses importance-weighted restrictions to limit model parameter updates and prevent the forgetting of old knowledge.

[0027] Specifically, the process of collecting the reflection phase timing signal of the RF matching network, residual gas mass spectrometry fragment ion intensity data, and wafer surface temperature distribution data during the plasma etching process, and extracting RF phase fluctuation characteristics to invert the dynamic thickness distribution of the plasma sheath is as follows: Time-frequency analysis is performed on the RF reflection phase timing signal to extract the phase fluctuation amplitude and frequency offset within the characteristic frequency band, which is then divided into multiple process cycle segments using a sliding time window; a dynamic inversion model is constructed based on the plasma sheath transport equation, mapping the phase fluctuation amplitude to the instantaneous change in sheath thickness, and a thickness gradient map of the wafer radial distribution is generated using a spatial interpolation algorithm; differential calculation is performed on the thickness gradient maps of adjacent process cycles to identify regions of abnormal plasma density fluctuations, and thickness measurement errors caused by thermal drift are corrected by combining wafer stage temperature distribution data; the corrected dynamic thickness distribution of the sheath is cross-validated with the RF matching network impedance data, and abnormal inversion results that exceed physical laws are eliminated.

[0028] In this implementation scheme, the time-series signal decomposition and phase fluctuation feature extraction are described as follows: The reflected phase (Phase of Reflected Power) signal φ(t) acquired by the RF matching network is subjected to time-frequency joint analysis to extract signal features reflecting the dynamic changes of the plasma. Processing flow: Continuous wavelet transform (CWT) is used to perform multi-scale decomposition of φ(t) to capture high-frequency disturbance features within a local time window; the instantaneous phase fluctuation amplitude and frequency drift within the characteristic frequency band (e.g., 10kHz~200kHz) are extracted; a sliding time window (the sliding window width is set to one cycle of the etching cavity, 1s) is used to slice the signal, generating multiple time segment segments, and constructing a time-series feature vector. ;in: Phase fluctuation amplitude within the characteristic frequency band Frequency center offset Phase derivative (rate of change). Inverse Modeling of Sheath Thickness: Physical Background: Plasma sheath thickness is a function of radio frequency reflection phase fluctuations. Based on the plasma boundary layer transport model, the inverse equation is established: Parameter definition: : The change in sheath thickness at wafer radius position r within the j-th time window; λ: Plasma wavelength, typically the wavelength of the incident radio frequency signal, such as 22.1 μm at 13.56 MHz. : Reflection phase fluctuation value. This formula indicates that, under the condition that the plasma sheath is approximately a dielectric reflective film, the reflection phase fluctuation is approximately linearly related to the change in sheath thickness. Explanation of radial thickness distribution map construction and spatial interpolation: Based on the inversion obtained... Discrete points are used to reconstruct the sheath thickness map of the entire wafer region using bivariate radial basis function (RBF) values. ,in Using wafer plane coordinates; constructing a continuous distribution image for subsequent thermal drift correction and differential analysis. Thickness differential and plasma density fluctuation region identification instructions: Perform pixel-by-pixel differential operations on the sheath thickness map between consecutive process cycles: The threshold method is used to identify regions with abnormal thickness variations (such as those exceeding 3σ of the average variation), characterizing regions of unstable plasma density. Temperature distribution data assists in thermal drift correction. Wafer surface temperature distribution is explained. This can affect the spatial uniformity of the sheath thickness, so a thermal drift model is introduced for correction. ;in: Graph showing the change in sheath thickness after temperature correction Temperature gradient diagram Thermal expansion related compensation coefficient. Inversion verification and physical consistency rejection explanation: [The remaining text appears to be incomplete and requires further context.] Impedance parameters within the same period (such as load impedance) Compare the results and calculate the physical consistency score. : When the score is below a threshold (e.g., 0.7), it is considered an anomaly and data is removed. Parameter definition: Theoretical impedance calculated based on the sheath model Tolerance factor for impedance difference.

[0029] Specifically, the construction logic of the dynamic prediction model for etching selectivity based on the correlation analysis between mass spectrometry ion intensity ratio and temperature gradient is as follows: Mass spectrometry fragment ion intensity data are normalized, the concentration ratio of key reactive groups is calculated, random noise interference is eliminated through moving average filtering, local hotspot regions in the temperature gradient distribution on the wafer surface are extracted, and a nonlinear correlation function between temperature gradient and etching selectivity is established. The concentration ratio of reactive groups and temperature gradient are input into a dual-channel neural network, and the influence weights of both on etching selectivity are dynamically allocated through an attention mechanism, outputting the dynamic prediction value under the current process window. The prediction model is corrected in real time using online etching rate monitoring data. When the prediction deviation continuously exceeds the tolerance range, the mass spectrometer calibration and temperature sensor re-inspection process is triggered.

[0030] In this implementation plan, the techniques for normalizing mass spectrometry ion intensity and calculating the concentration ratio of key reactive groups are described as follows: Time-series data of mass spectrometry fragment ion intensities under multiple mass-charge ratios are collected and expressed as follows: ,in Indicates the first Representative ions, such as , , To reduce system-scale inconsistencies among process variables, data normalization is necessary. Parameter explanation: :No. Ions at time The normalized relative intensity (dimensionless). Total number of ion species monitored. Raw ionic strength data. Extracting the concentration ratio of key reactive groups representing etching selectivity. For example, with and Ratio construction: This ratio reflects the changing trends of etching rate and selectivity, and is an important input feature for subsequent modeling. Explanation of moving average filtering for eliminating high-frequency noise: The original concentration ratio signal... It may be affected by discharge noise, pump speed fluctuations, etc., so a sliding window is used for smoothing, as shown below: Parameter explanation: The concentration ratio of reactive groups after smoothing. Window width, in units of sampling points. Technical description for extracting temperature gradient hotspot region features: Acquiring two-dimensional temperature field on the wafer surface. Local heating regions can be identified by calculating gradient magnitude: Set threshold Extract regions with significant temperature gradients (hot spots): Then, the average temperature gradient of the hotspot region is calculated: Parameter explanation: Temperature gradient magnitude. : A collection of pixels in a local hotspot region. : Number of hotspot pixels. The average hotspot gradient characteristics of the wafer at the current moment. Technical explanation of constructing a nonlinear correlation model between temperature and concentration ratio: The smoothed concentration ratio... With average temperature gradient By combining the input features, a nonlinear prediction function is constructed: ; : is a non-linear mapping function that can be fitted by a machine learning model (neural network). At any moment The etching selection ratio is higher than the predicted value. Dual-channel neural network + attention mechanism modeling structure description: A structured neural network is used for modeling: dual-channel input + attention fusion output. The model structure is as follows: Channel 1 input: Corresponding channel output 2. Channel 2 Input: Corresponding channel output Use learnable attention weights The contribution of allocating two channels to the final selection ratio: Attention factor The definition is: Parameter description: : These represent the attention scores generated by the intermediate layers of Channel 1 and Channel 2, respectively (which can be obtained through MLP+Softmax). The model outputs the predicted etching selectivity. Real-time correction mechanism and error feedback logic: The system has an embedded etching rate monitoring module to obtain the actual etching selectivity. Calculate the prediction error: When continuous The error in a single process cycle exceeds the set tolerance threshold. : Then the following maintenance operations are triggered: Mass spectrometer automatic calibration program: re-verify the linear response coefficient of the mass spectrometer signal; Temperature sensor drift re-inspection process: initiate redundant thermocouple comparison and standard point correction; Neural network structure fine-tuning or weight update: perform local retraining or momentum optimizer sliding update.

[0031] Specifically, the process of generating the dynamic coupling relationship matrix between process parameters and plasma state is as follows: A three-dimensional coupling relationship matrix is ​​constructed, where the X-axis represents RF power parameters, including reflected power and impedance matching; the Y-axis represents plasma state parameters, including sheath thickness and electron density; and the Z-axis represents process output parameters, including etching rate and selectivity. The inverted dynamic sheath thickness distribution and the predicted etching selectivity are mapped to the corresponding dimensions of the matrix, and key coupling factors are extracted using a tensor decomposition algorithm. Based on process stability constraints, the coupling relationship matrix is ​​sparsified, weakly correlated parameter combinations are eliminated, and strongly correlated feature channels are retained. The effectiveness of the matrix is ​​verified using real-time acquired process yield data. When the statistical process control data for key wafer dimensions exceeds the control limits, the weight coefficients of the coupling factors are recalculated.

[0032] In this implementation plan, the objective of constructing a three-dimensional process coupling relationship matrix is: to construct a three-dimensional matrix. This indicates the dynamic coupling strength relationship between the process input plasma state output parameters. : Indicates the radio frequency power dimension, parameters include: reflected power Impedance matching : Represents the plasma state dimension, parameters include: sheath thickness electron density : Indicates the process output dimension, parameters include: etching rate Etching selectivity Define matrix: ;in This is a nonlinear, multi-parameter interaction function, actually obtained through data modeling and fitting, representing the dynamic correlation strength among the three factors. The data is mapped to a matrix dimension, and the data obtained from the preceding inversion and prediction process includes: the dynamic thickness distribution of the sheath. Polar coordinates averaged to Dimensions; Etching selectivity ratio prediction Normalized mapping to Dimensions; RF matching network parameters : Standardized mapping to Dimensions. Each time window Tensor elements are generated. Accumulation is performed to form a time-series tensor set. Tensor decomposition extracts key coupling factors from the above three-dimensional tensor. Perform higher-order decomposition to extract the most representative coupling factor subspace. CP decomposition is employed. Parameter description: Coupling factor The weight of the module (reflecting the degree of strong coupling). : Radio frequency parameter factor matrix, Sheath and plasma state factor matrix The process output factor matrix. This decomposition operation extracts the L most influential core coupling patterns from complex 3D process interactions, serving as feature channels for dynamic process control. To avoid excessive coupling or noise introduced by interfering dimensions, a sparse regularization term is further introduced to optimize the weights, incorporating process stability constraints and sparsity processing. Parameter description: : Indicates tensor reconstruction error, The sparsity coefficient controls the number of coupling channels; a larger coefficient results in greater sparsity. Only principal factors with stable, high-weight coupling relationships between RF and state parameters and process output are retained, while weakly correlated edge channels are removed to improve the physical reliability of the model. The process feedback mechanism and coupling factor weight updates are implemented through real-time monitoring of process yield indicators, such as critical dimensions. like This refers to the SPC alarm; therefore, it applies to the tensor set within the current time window. Re-execute the decomposition; adjust And renormalized to: It enables adaptive dynamic adjustment of the importance weights of coupling modes.

[0033] Specifically, the process of analyzing the spatial distribution characteristics of distortion patterns in the scattered field interferometric image and identifying the geometric parameters of microgroove distortion in the wafer edge region using a deformable convolutional network is as follows: The original interferometric image is preprocessed, noise interference is eliminated using a phase unwrapping algorithm, and gradient information of the wafer surface height distribution is extracted; the input layer of the deformable convolutional network is constructed, mapping the gradient information to the geometric deformation field of the feature map; the sampling position of the convolutional kernel is adaptively adjusted to match the curvature of the microgroove edge; the global topography contour and local distortion components are separated in the deep features of the network; multi-scale distortion patterns are captured using spatial pyramid pooling; the depth, tilt angle, and radius of curvature of the microgroove are output, and the spatial location distribution of the distortion region is determined using the wafer coordinate system.

[0034] In this implementation scheme, the interferometric image preprocessing and phase unwrapping reflect the optical phase difference of the surface scattering field, which often exhibits a 2π periodic jump problem. To obtain a continuous and physically solvable surface height distribution function... Phase unwrapping processing must be performed first. The original phase map exhibits modulo 2π folding. Continuous phase distribution after unwrapping; Image pixel space coordinates; Interferogram sequence number or imaging channel index. Noise errors caused by jumps are eliminated through phase gradient fitting and path tracking algorithms. The height gradient is calculated as follows: Where: λ: interferometric wavelength; n: effective refractive index of the medium; ∇: two-dimensional gradient operator. The input features for constructing a deformable convolutional network incorporate the aforementioned height gradient information. As input image, a geometric deformation field is introduced. The convolution offset parameter is input into the deformable convolutional network. The sampling position of the classic convolutional kernel is fixed at the center of the window. Nearby grid The DCN structure, on the other hand, introduces a spatial offset. : Where: x(⋅): input feature map (i.e., gradient map); y^{}(⋅): convolution output feature map; Convolution kernel weights The local deformation-aware offset is learned by the network. A shape component decoupling module is introduced in the intermediate layer of the network to separate local distortion from global shape features, thus separating the global contour. Such as macroscopic smoothness, edge contour; local distortion. Examples include microgrooving, indentation, tilting, and undulation. Implementation: Let the feature tensor be... Two branches are generated using a channel attention mechanism: The structure features at different frequency responses are captured separately. This decoupling structure helps prevent local distortions from being overwhelmed by the global background. Spatial pyramid pooling enables multi-scale distortion recognition. For distortion patterns at different scales (e.g., trench widths varying from micrometers to tens of micrometers), a spatial pyramid pooling module (SPP) is introduced: This approach extracts regional responses through multi-scale windows to capture the hierarchical spatial features of trench distortion, enhancing the network's robustness to trenches of varying scales. It outputs micro-trench geometric parameters and performs spatial localization. The final output layer performs regression prediction for each detection region (ROI): trench depth. trench inclination angle radius of curvature Parameter explanation: Trench number, Estimated by the absolute amplitude of the phase difference. Estimate along the normal gradient direction, It is obtained from the surface curvature formula. Combined with a wafer coordinate calibration system. All The numbered grooves are mapped to the global spatial region to construct a spatial distribution map.

[0035] Specifically, the process of constructing a mapping relationship between surface roughness abrupt change features and etching rate predictions by combining key dimensional data from scanning electron microscope (SEM) images to generate a morphology instability risk index is as follows: Sub-pixel edge detection is performed on the SEM images to extract the linewidth, sidewall angle, and bottom roughness parameters of key structures. The SEM measurement data and the microgroove geometric parameters obtained from the interferometric image are spatially aligned, and coordinate system offset errors are eliminated through affine transformation. A roughness abrupt change feature vector is constructed, including gradient change rate, local curvature extreme point density, and abnormal spike height distribution. The feature vector is input into a random forest regression model, and the morphology instability risk index is calculated by combining the current etching rate prediction.

[0036] In this implementation scheme, the scanning electron microscope (SEM) image processing and geometric feature extraction steps are described as follows: Subpixel-level edge detection is performed on the SEM image, and key size parameters are extracted, including line width. Sidewall angle Bottom roughness ,in: : The structure number identified in the wafer surface image; : Refers to the actual width of the lithographic lines or aperture array; The angle between the sidewall and the base; The root mean square roughness of the bottom morphology. An edge detection algorithm based on Canny+Spline fitting is used, combined with sub-pixel grayscale interpolation to improve geometric boundary detection accuracy to the sub-pixel level (<5nm). Roughness parameters are extracted from the contour lines of the cross-section. The step of aligning the microgroove parameters with the interferometric image is explained as follows: The structural coordinates in the SEM image are compared with the microgroove parameters obtained from the interferometric image analysis in the previous process. Alignment is performed. The formula is as follows: Let the SEM image coordinate system be... The coordinate system of the interferometric image is The correspondence is expressed by a two-dimensional affine transformation as follows: ;in: : Affine transformation matrix; : Coordinate offset vector; solved by multi-point registration (≥4 feature points) and This transformation allows feature data from heterogeneous images to be processed within the same coordinate system. The steps for constructing the roughness abrupt change feature vector are as follows: Based on the spatially aligned roughness data, extract the surface distortion feature vector. This is used to characterize the degree of abrupt changes in the local surface morphology of a wafer. Eigenvector composition: Parameter explanation: : The rate of change of bottom roughness along the structural direction; : Density of local curvature extrema; The average or maximum height of abnormal spikes is used to identify over-etching or localized micro-explosion phenomena. These dimensions constitute the input criteria for microstructure surface morphology instability. The steps for morphology risk assessment combined with etching rate predictions are as follows: The roughness feature vector and the obtained etching rate predictions are jointly input into a machine learning model (such as a random forest regressor) to calculate the morphology instability risk index. Model structure: Let the regression function be... Then we have: ;in: This indicates the risk index of topographic instability at this location; the larger the value, the closer it is to the instability boundary. A regression model was constructed using a training sample set (historical SEM data + process feedback tags). The model can achieve nonlinear feature cross-modeling using a 10-tree depth RandomForest regressor. The results are applied to closed-loop logic, when... When the threshold value is exceeded (e.g., 0.75), the following operation is triggered: Mark the structural point. The wafer region in question is a hotspot for topographic instability; corresponding process window parameters are fine-tuned (e.g., reducing RF power or gas flow); if multiple structural regions are triggered consecutively, system-level deviation correction and equipment inspection are performed.

[0037] Specifically, the process of inputting the dynamic thickness distribution of the plasma sheath and the morphological instability risk index into a feature fusion network, and processing the plasma impedance characteristics and microgroove distortion modes through temporal analysis and spatial feature extraction is as follows: A dual-branch feature encoder is designed: the temporal branch performs multi-scale wavelet decomposition on the plasma sheath thickness sequence to extract transient impact response components and steady-state oscillation components; the spatial branch models the microgroove distortion mode using a graph convolutional network, constructing a topological adjacency matrix to describe the spatial correlation of the distortion region; the temporal and spatial features are mapped to a unified latent space through cross-modal projection, and adversarial training strategies are used to eliminate distribution differences between modes. A feature similarity matrix is ​​calculated in the latent space, and high-confidence cross-modal correlation feature pairs are selected.

[0038] In this implementation scheme, a dual-branch feature encoder design is used. Temporal branching and wavelet decomposition are employed to process the sheath dynamics: the plasma sheath thickness is essentially a non-stationary signal that fluctuates over time, containing both transient process disturbances and stable oscillations. Multi-scale wavelet decomposition divides the original thickness sequence into several frequency bands: the high-frequency part reflects transient impact responses, such as discharge instability and plasma reconstruction; the low-frequency part represents steady-state oscillation trends, corresponding to the stable operating mode under the process setpoint. This processing preserves the temporal resolution of plasma dynamic behavior, which is beneficial for identifying nonlinear disturbances that may cause morphological instability. Spatial branching and graph convolution modeling are used to model the microgroove distortion region: microgroove distortion exhibits significant spatial clustering and structural correlation. The distortion region is treated as a spatial graph node, and a topological adjacency matrix is ​​constructed to describe the adjacency relationships between structural units (e.g., proximity, similar curvature, similar roughness). Graph convolutional networks (GCNs) are used to aggregate and encode the embedded features of these nodes, forming a high-order spatial abstraction of the microstructural instability behavior. This method can identify the propagation path of the influence of local distortion on the surrounding structure, improving the model's ability to perceive the overall structural instability of the wafer. Cross-modal feature projection and unified representation space construction: Due to the inherent differences in data form and distribution between temporal data (sheath thickness) and spatial data (distortion patterns), they need to be mapped to a unified representation space for easy subsequent fusion. A cross-modal projection network is used to map the feature encodings of the two modalities to a unified latent space vector domain. This process ensures geometric consistency between features through a shared structural layer. An adversarial training strategy is introduced, where a discriminator is used during training to force the embedding distributions of the two modalities to converge, thereby eliminating distributional differences (scale, variance) between modalities. Cross-modal similarity analysis and high-confidence feature pair selection: In the unified latent space, the similarity between the features of the two modalities is calculated (using cosine similarity) to construct a feature similarity matrix, which is used to discover potential cross-modal association structures. Furthermore, a confidence evaluation mechanism (clustering of feature similarity distributions) is used to select high-confidence feature pairs. These feature pairs represent a significant coupling relationship between sheath perturbation signals and specific distortion patterns, and can ultimately serve as a key criterion for determining the risk of morphological instability.

[0039] Specifically, the process of dynamically allocating the confidence level of time-series data and the weight of spatial features through a physical constraint attention mechanism to generate an etching rate optimization adjustment amount and trigger process parameter correction commands is as follows: The transient impact response amplitude of the time-series branch and the energy density of the distortion region of the spatial branch are input, and the plasma sheath equation is converted into an inequality constraint to limit the allocation range of attention weights; the time-series confidence level is calculated based on the signal-to-noise ratio of the RF impedance data and the periodicity of the steady-state oscillation component, while the spatial confidence level is calculated based on the matching consistency between the micro-trench distortion mode and the scanning electron microscope measurement data; the time-series confidence level and the spatial confidence level are used as dynamic weights to weight and fuse the multimodal features to generate an etching rate optimization adjustment amount. When the adjustment amount exceeds the process safety window, the RF power matching parameter adjustment command and the reactive gas flow rate ratio adjustment command are triggered sequentially.

[0040] In this implementation scheme, the input feature construction and preprocessing, and the time-series branch input, extract the transient impact response amplitude of the plasma sheath thickness, denoted as... This represents the short-term plasma perturbation energy level over a timescale and exhibits high time-varying characteristics. Spatial branch input: The distortion intensity of the microgroove region is measured by energy and defined as the distortion energy density. The calculation is based on the product of the extreme value distribution of curvature of the microstructure, the abrupt change in roughness, and the distortion amplitude after area normalization. II. Introducing an Inequality Attention Mechanism with Physical Constraints: To ensure the physical rationality of the attention weights, simplified transport equation constraints of the plasma sheath are introduced, constructing an inequality constraint of the following form: ; Attention weights assigned to time-series branches (the first branch) (Perturbation-like patterns). Attention weights for spatial branch assignment (the first) (Deformed appearance). The upper limit of the maximum acceptable energy disturbance response derived from the sheath steady-state transport model corresponds to the first... A typical discharge steady-state solution. This constraint ensures that the total weights will not exceed the physical stability limit under the material processing window, preventing the model from generating unrealizable feedback commands. Confidence modeling and weight allocation, time-series confidence calculation: based on the signal-to-noise ratio of RF reflection signals ( Periodicity measures of steady-state oscillation components (such as wavelet dominant frequency) (Stability factors), jointly constructing a confidence score: ;in: These are the empirical weighting coefficients; in the denominator... 1. Variance representing frequency fluctuations; the smaller the variance, the more stable the structure. 2. Spatial confidence calculation: A structural matching score is obtained by comparing the morphological consistency between the distorted structure and key dimensional data (such as linewidth and sidewall angles) extracted from the scanning electron microscope image. Spatial confidence is expressed as: ;in: Standard deviation of the inclination angle of the micro-grooves in the distorted region; Spatial confidence weighting coefficient. IV. Multimodal feature weighted fusion and optimization adjustment: The fusion process uses dynamic confidence weights to weight the features of the two modalities. ;in: , This fusion feature The input is fed into a nonlinear regression module (such as an attention regressor or an LSTM regressor) to generate an optimized etching rate adjustment. The triggering mechanism for process correction commands defines the upper and lower limits of the process safety window within the system. When the adjustment exceeds the tolerance range: Sequential triggers: RF power matching parameter adjustment: Prioritize adjusting RF reflection power and matching network impedance to optimize coupling efficiency; Reaction gas flow rate ratio adjustment: Under confirmed plasma stability conditions, adjust the F / C ratio or add / remove dilution gas to optimize the etching reaction rate.

[0041] Specifically, the process of synchronously acquiring RF reflection phase data after process correction and updated interferometric image features, and updating the dynamic coupling relationship matrix and feature fusion network parameters is as follows: Store the RF phase data, interferometric images and corresponding etching rate adjustments before and after process correction, and establish a data association index based on timestamps and process batch numbers; Map new samples to the missing dimensions of the original three-dimensional matrix using a streaming tensor decomposition algorithm, control the growth of the matrix rank by truncating low singular values, freeze the pre-trained weights of the spatial branches of the feature fusion network, perform online fine-tuning only on the temporal branches, and introduce an elastic weight solidification algorithm to constrain the update magnitude of important parameters.

[0042] In this implementation plan, an incremental indexing system for process correction data is constructed, and the collected content includes: RF reflection phase data before and after process correction. ; Feature tensors of interferometric images before and after correction ; Corresponding etching rate adjustment amount Data index construction method: using a unified timestamp. and process batch number Perform data association and establish mapping relationships: This indexing mechanism ensures the traceability and reliability of all new samples in the dynamic model, providing a data foundation for subsequent incremental updates of coupling relationships. Online completion of the dynamic coupling relationship matrix: To avoid frequent reconstruction of the coupling model, streaming tensor decomposition employs a streaming tensor completion method, which applies new data to the existing three-dimensional tensor of the coupling relationship. Online expansion is performed. Tensor interpolation dimension: If a new sample corresponds to a missing region in the original matrix (e.g., a specific reflection phase or interferogram pattern that has not appeared before), it is mapped to a new sub-tensor slice. Rank growth control mechanism: To control model complexity, a truncated singular value decomposition (TruncatedSVD) strategy is introduced, setting an upper limit on the rank. Only retain the principal singular value components. This is to eliminate the influence of weakly perturbed samples. The formal expression is: ;in ; Tensor outer product; : Orthogonal basis corresponding to each modal component. The local fine-tuning mechanism of the feature fusion network is based on the fact that most process perturbations have a delayed effect on spatial morphology and structure stability; therefore: Freezing spatial branch weights: Weights of spatial feature extraction branches (convolutional layers, morphology attention modules). Set a freeze flag to exclude it from current updates. Online fine-tuning of timing branches: Perform small-step online fine-tuning on timing branches handling RF phase changes; the optimizer only updates the weights. An elastic weight fixation algorithm is introduced to prevent the generalization ability of the original weights from being destroyed during online updates; a regularization term for updating important parameters is introduced into the loss function. ; The optimal value learned in the previous task; : No. Fisher information estimates for each parameter, representing its importance in the original task; : Weight constraint strength adjustment factor. Advantages of incremental update process: Timeliness: No need to retrain the entire model, only fine-tuning of key parameters; Robustness: Through low-rank control and weight solidification, overfitting to new data or catastrophic forgetting is prevented; Physical consistency: The dimension and physical structure of the coupling matrix remain unchanged, improving parameter interpretability and traceability.

[0043] A big data-based integrated circuit equipment data optimization and monitoring system includes the following modules: a multi-source data dynamic modeling module, a morphology instability quantification module, a cross-modal optimization decision-making module, and a closed-loop self-updating module. The multi-source data dynamic modeling module collects the reflection phase timing signal of the RF matching network, residual gas mass spectrometry fragment ion intensity data, and wafer surface temperature distribution data during the plasma etching process. It extracts RF phase fluctuation characteristics to invert the dynamic thickness distribution of the plasma sheath and establishes a dynamic prediction model for etching selectivity based on the correlation analysis of mass spectrometry ion intensity ratio and temperature gradient, generating a dynamic coupling relationship matrix between process parameters and plasma state. The morphology instability quantification module analyzes the spatial distribution characteristics of distortion modes in the scattered field interferometric image based on the dynamic coupling relationship matrix and identifies wafer edge regions through a deformable convolutional network. The microgroove distortion geometry parameters of the domain are combined with key dimensional data from scanning electron microscope images to construct a mapping relationship between surface roughness abrupt changes and etching rate predictions, generating a morphological instability risk index. The cross-modal optimization decision module is used to input the dynamic thickness distribution of the plasma sheath and the morphological instability risk index into the feature fusion network. It processes the plasma impedance characteristics and microgroove distortion modes through temporal analysis and spatial feature extraction, respectively. It dynamically allocates the confidence level of temporal data and the weight of spatial features through a physical constraint attention mechanism to generate etching rate optimization adjustment and trigger process parameter correction commands. The closed-loop self-updating module is used to feed back the etching rate optimization adjustment to the etching equipment controller in real time, synchronously collect the RF reflection phase data after process correction and the updated interference image features, and update the dynamic coupling relationship matrix and feature fusion network parameters.

[0044] This implementation plan includes a multi-source data dynamic modeling module. This module comprehensively collects data from the following heterogeneous sensors: RF matching network reflected phase timing signal, reflecting the plasma-equipment matching state; residual gas mass spectrometry fragment ion intensity, revealing the concentration distribution of intermediate products in the etching reaction; and wafer surface temperature distribution data, reflecting the spatial differences in energy input conversion on the wafer surface. The RF phase fluctuation characteristics and plasma sheath thickness are physically inverted to form a timing field. A selectivity dynamic prediction model is established by combining the mass spectrometry ion intensity ratio (product / reactant) and local temperature gradient. The output results are uniformly integrated into a three-dimensional dynamic coupling relationship matrix to describe the deep relationship between process parameters, plasma state, and etching response. For the first time, plasma sheath modeling, ion composition behavior, and thermal field information are coupled for inverting the microscopic etching response. A structured data fusion tensor (a non-planar data structure) is introduced as the input data model, significantly improving the expressive ability of the correlation between parameters. The topography instability quantification module utilizes image analysis of scattered field interferometry guided by a coupling matrix to identify the spatial distribution characteristics of wafer surface distortion modes. A deformable convolutional network is employed to precisely align the micro-trenches' contour edges and extract geometric distortion parameters (such as depth, tilt angle, and radius of curvature). Spatially registration is performed with key dimensional information (CD, SW angle, bottom roughness) from the scanning electron microscope (SEM) image. A roughness abrupt change feature vector is constructed and combined with etching rate predictions to output a topography instability risk index. This module systematically establishes a quantification link from image interferometry to deep neural networks, SEM physical matching, and topography index output. The combination of deformable convolutional and pyramid spatial pooling mechanisms enables precise identification of multi-scale topography distortions. The module innovatively proposes a comprehensive index, the "topography instability risk index," to quantify the risk level of topography instability to etching consistency. Cross-modal optimization decision module: Dual-branch feature encoder architecture: Temporal branch: Multi-scale wavelet decomposition is used to process the dynamic sequence of sheath thickness, distinguishing between transient impacts and steady-state oscillations; Spatial branch: Graph convolutional network is used to model the spatial structure of the distorted morphology, constructing a topological adjacency matrix of the distorted region; The two types of modal features are projected to a unified latent space, and weights are adaptively allocated through a cross-modal attention mechanism; Physical constraint mechanisms (such as RF impedance response boundaries) are added to control the attention distribution, improving the physical rationality of the decision; The etching rate optimization adjustment is output, triggering correction commands for RF power and gas flow rate. A cross-modal feature fusion mechanism is proposed, simultaneously considering the plasma electric field state and morphological response, breaking through the limitations of traditional process monitoring modes; Physical constraints are applied to the attention mechanism, transforming the "black box feature allocation" behavior of deep models into a controllable parameter scheduling mechanism; A high-confidence cross-modal feature pair screening mechanism is introduced to improve model robustness and prevent abnormal samples from interfering with the decision.The closed-loop self-updating module: Real-time acquisition and archiving of system states (RF reflection phase, interferometric images, etc.) after optimization command execution; construction of an index system based on timestamps and process batch numbers to form structured pre- and post-correction data pairs; incremental updates of the dynamic coupling relationship matrix using a streaming tensor completion algorithm, and control of rank growth by truncating singular values; implementation of a local fine-tuning strategy for the feature fusion network: spatial branch freezing; introduction of an elastic weight solidification mechanism for temporal branches to prevent catastrophic forgetting; maintaining model stability and adaptability compatibility. The process correction behavior, process response, and model fine-tuning are tightly coupled into a closed loop, breaking through the limitation of traditional models that "only train and do not update"; innovatively introducing a streaming tensor decomposition algorithm for online dynamic relationship modeling, ensuring real-time performance while avoiding full retraining; and introducing a parameter importance constraint mechanism in the network training strategy, significantly enhancing the model's sustainable learning ability.

[0045] In summary, this application has at least the following effects: A data optimization monitoring system and method for integrated circuit equipment based on big data integrates multi-source data such as RF reflection phase, mass spectrometry ion intensity, and wafer temperature distribution. By establishing physical relationships between multi-dimensional parameters based on a dynamic coupling matrix, it effectively improves the characterization capability of plasma state changes and achieves accurate dynamic prediction of key process parameters such as etching selectivity and sheath thickness. Utilizing joint analysis of interferometric images and scanning electron microscopy data, combined with deformable convolutional networks and graph convolutional models, it can perform high-precision modeling of the distortion morphology of microtrench edges and quantitatively assess surface changes that may lead to defects through a morphology instability risk index, improving the timeliness and reliability of defect early warning. A temporal-spatial dual-branch feature fusion network is constructed, and a physically constrained attention mechanism is used to dynamically control the weights of multimodal features, effectively avoiding decision bias caused by inconsistent data distribution. This results in more accurate generation of etching rate optimization adjustments and automatic triggering of parameter correction operations. By introducing streaming tensor decomposition and elastic weight solidification strategies, the system possesses online update capabilities, enabling it to adapt to changes in new batches of data without losing historical learning effects. This maintains the long-term stability and robustness of the model's performance, supporting stable process operation in large-scale production. While optimizing and controlling equipment parameters, the entire system can track the evolution path of morphological defects and feed optimization suggestions back to the equipment controller in a closed-loop manner using intelligent commands. This significantly enhances the intelligent adaptive capabilities of the process, providing strong data support and a solid decision-making foundation for the mass production stability of advanced processes.

[0046] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0047] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0048] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0049] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0050] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0051] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A data optimization and monitoring method for integrated circuit equipment based on big data, characterized in that, Includes the following steps: S1. Collect the reflection phase timing signal of the radio frequency matching network, residual gas mass spectrometry fragment ion intensity data and wafer surface temperature distribution data in the plasma etching process. Extract the radio frequency phase fluctuation characteristics to invert the dynamic thickness distribution of the plasma sheath. Based on the correlation analysis of mass spectrometry ion intensity ratio and temperature gradient, establish a dynamic prediction model of etching selectivity and generate a dynamic coupling relationship matrix between process parameters and plasma state. S2. Based on the dynamic coupling relationship matrix, the spatial distribution characteristics of the distortion mode of the scattered field interference image are analyzed. The geometric parameters of the micro-trenches in the wafer edge region are identified by a deformable convolutional network. Combined with the key size data of the scanning electron microscope image, the mapping relationship between the surface roughness abrupt change characteristics and the etching rate prediction value is constructed to generate the morphological instability risk index. S3. Input the dynamic thickness distribution of the plasma sheath and the morphological instability risk index into the feature fusion network. Process the plasma impedance characteristics and micro-groove distortion mode through time series analysis and spatial feature extraction, respectively. Dynamically allocate the confidence of time series data and the weight of spatial features through the physical constraint attention mechanism to generate the etching rate optimization adjustment amount and trigger the process parameter correction command. S4. Feed back the etching rate optimization adjustment to the etching equipment controller in real time, and simultaneously collect the RF reflection phase data after process correction and the updated interference image features, and update the dynamic coupling relationship matrix and feature fusion network parameters.

2. The method for optimizing and monitoring integrated circuit equipment data based on big data according to claim 1, characterized in that: The specific process of acquiring the reflection phase timing signal of the RF matching network, residual gas mass spectrometry fragment ion intensity data, and wafer surface temperature distribution data during the plasma etching process, and extracting RF phase fluctuation characteristics to invert the dynamic thickness distribution of the plasma sheath is as follows: Time-frequency analysis is performed on the RF reflection phase timing signal to extract the phase fluctuation amplitude and frequency offset within the characteristic frequency band, and the signal is divided into multiple process cycle segments by a sliding time window. A dynamic inversion model is constructed based on the plasma sheath transport equation, which maps the phase fluctuation amplitude to the instantaneous change in sheath thickness. A thickness gradient map of the wafer radial distribution is generated by spatial interpolation algorithm. Differential calculations are performed on the thickness gradient maps of adjacent process cycles to identify regions with abnormal fluctuations in plasma density. The thickness measurement error caused by thermal drift is corrected by combining the wafer stage temperature distribution data. The corrected dynamic thickness distribution of the sheath is cross-validated with the impedance data of the RF matching network to eliminate abnormal inversion results that exceed physical laws.

3. The data optimization and monitoring method for integrated circuit equipment based on big data according to claim 2, characterized in that: The construction logic of the etching selectivity dynamic prediction model based on the correlation analysis between mass spectrometry ion intensity ratio and temperature gradient is as follows: The mass spectrometry fragment ion intensity data are normalized, the concentration ratio of key reactive groups is calculated, random noise interference is eliminated by moving average filtering, local hot spots in the temperature gradient distribution on the wafer surface are extracted, and a nonlinear correlation function between temperature gradient and etching selectivity is established. The reactive group concentration ratio and temperature gradient are input into a dual-channel neural network. The influence weights of the two on the etching selectivity are dynamically allocated through an attention mechanism, and the dynamic prediction value under the current process window is output. The prediction model is corrected in real time by monitoring online etching rate data. When the prediction deviation continues to exceed the tolerance range, the mass spectrometer calibration and temperature sensor re-inspection process is triggered.

4. The data optimization and monitoring method for integrated circuit equipment based on big data according to claim 3, characterized in that: The specific process for generating the dynamic coupling relationship matrix between process parameters and plasma state is as follows: A three-dimensional coupling matrix is ​​constructed, where the X-axis represents RF power parameters, including reflected power and impedance matching degree; the Y-axis represents plasma state parameters, including sheath thickness and electron density; and the Z-axis represents process output parameters, including etching rate and selectivity. The inverted dynamic sheath thickness distribution and the output etching selectivity prediction are mapped to the corresponding dimension of the matrix, and key coupling factors are extracted by tensor decomposition algorithm. Based on process stability constraints, the coupling relationship matrix is ​​sparsified to remove weakly correlated parameter combinations and retain strongly correlated feature channels. The effectiveness of the matrix is ​​verified by real-time acquired process yield data. When the statistical process control data of the wafer critical dimensions exceeds the control limit, the coupling factor weight coefficient is recalculated.

5. The data optimization and monitoring method for integrated circuit equipment based on big data according to claim 4, characterized in that: The specific process of analyzing the spatial distribution characteristics of distortion modes in the scattered field interferometric image and identifying the microgroove distortion geometric parameters in the wafer edge region using a deformable convolutional network is as follows: The original interferometric image is preprocessed, and noise interference is eliminated by phase unwrapping algorithm to extract gradient information of the height distribution on the wafer surface; The input layer of a deformable convolutional network is constructed, which maps gradient information to the geometric deformation field of the feature map. The sampling position of the convolutional kernel is adaptively adjusted to match the curvature of the microgroove edge. The global shape contour and local distortion components are separated in the deep features of the network. Multi-scale distortion patterns are captured by spatial pyramid pooling. Output the depth, tilt angle, and radius of curvature of the microtrestle, and use the wafer coordinate system to calibrate the spatial distribution of the distorted region.

6. The data optimization and monitoring method for integrated circuit equipment based on big data according to claim 5, characterized in that: The specific process of constructing a mapping relationship between surface roughness abrupt changes and etching rate predictions by combining key dimensional data from scanning electron microscope images, and generating a morphology instability risk index, is as follows: Subpixel edge detection is performed on scanning electron microscope (SEM) images to extract line width, sidewall angle, and bottom roughness parameters of key structures. The SEM measurement data is spatially aligned with the microgroove geometric parameters obtained from the interferometric image analysis, and coordinate system offset errors are eliminated through affine transformation. A roughness mutation feature vector is constructed, which includes the gradient change rate, the density of local curvature extreme points, and the distribution of abnormal spike height. The feature vector is input into a random forest regression model, and the morphological instability risk index is calculated by combining the current etching rate prediction value.

7. The data optimization and monitoring method for integrated circuit equipment based on big data according to claim 6, characterized in that: The specific process of inputting the dynamic thickness distribution of the plasma sheath and the morphological instability risk index into a feature fusion network, and processing the plasma impedance characteristics and microgroove distortion modes through time series analysis and spatial feature extraction is as follows: Design a dual-branch feature encoder: The temporal branch performs multi-scale wavelet decomposition on the plasma sheath thickness sequence to extract transient impulse response components and steady-state oscillation components; Spatial branching: Graph convolutional network modeling is used to model the microgroove distortion pattern, and a topological adjacency matrix is ​​constructed to describe the spatial correlation of the distortion region; By mapping temporal and spatial features to a unified latent space through cross-modal projection, adversarial training strategies are used to eliminate distribution differences between modalities. Feature similarity matrices are calculated in the latent space to select high-confidence cross-modal associated feature pairs.

8. The data optimization and monitoring method for integrated circuit equipment based on big data according to claim 7, characterized in that: The specific process of dynamically allocating the confidence level and spatial feature weights of time-series data through a physical constraint attention mechanism, generating etching rate optimization adjustment amounts and triggering process parameter correction commands is as follows: The transient impact response amplitude of the input time branch and the energy density of the distortion region of the input spatial branch are used to transform the plasma sheath equation into an inequality constraint to limit the range of attention weight allocation. The timing confidence is calculated based on the signal-to-noise ratio of RF impedance data and the periodicity of steady-state oscillation components, and the spatial confidence is calculated based on the matching consistency between the microgroove distortion mode and the scanning electron microscope measurement data. The temporal confidence and spatial confidence are used as dynamic weights to perform weighted fusion of multimodal features to generate an etching rate optimization adjustment amount. When the adjustment amount exceeds the process safety window, the RF power matching parameter adjustment command and the reaction gas flow rate ratio adjustment command are triggered in sequence.

9. The data optimization and monitoring method for integrated circuit equipment based on big data according to claim 8, characterized in that: The specific process of synchronously acquiring the modified RF reflection phase data and updated interferometric image features, and updating the dynamic coupling matrix and feature fusion network parameters is as follows: Radio frequency phase data, interferometric images, and corresponding etching rate adjustments before and after the memory process modification are used to establish a data association index based on timestamps and process batch numbers; The new samples are mapped to the missing dimensions of the original three-dimensional matrix by the streaming tensor decomposition algorithm. The growth of the matrix rank is controlled by truncating low singular values. The pre-trained weights of the spatial branches of the feature fusion network are frozen. Only the temporal branches are fine-tuned online. An elastic weight solidification algorithm is introduced to constrain the update magnitude of important parameters.

10. A big data-based integrated circuit equipment data optimization monitoring system, applied to the big data-based integrated circuit equipment data optimization monitoring method according to any one of claims 1-9, characterized in that, Includes the following steps: Multi-source data dynamic modeling module, morphological instability quantification module, cross-modal optimization decision module, closed-loop self-updating module; The multi-source data dynamic modeling module is used to collect the reflection phase timing signal of the radio frequency matching network, residual gas mass spectrometry fragment ion intensity data and wafer surface temperature distribution data in the plasma etching process. By extracting the radio frequency phase fluctuation characteristics, the dynamic thickness distribution of the plasma sheath is inverted, and an etching selectivity dynamic prediction model is established based on the correlation analysis between the mass spectrometry ion intensity ratio and the temperature gradient, generating a dynamic coupling relationship matrix between process parameters and plasma state. The morphological instability quantification module is used to analyze the spatial distribution characteristics of distortion modes in the scattered field interference image based on the dynamic coupling relationship matrix, identify the micro-groove distortion geometric parameters in the wafer edge region through a deformable convolutional network, and construct the mapping relationship between the surface roughness abrupt change characteristics and the etching rate prediction value by combining the key size data of the scanning electron microscope image, thereby generating a morphological instability risk index. The cross-modal optimization decision module is used to input the dynamic thickness distribution of the plasma sheath and the morphological instability risk index into the feature fusion network. It processes the plasma impedance characteristics and micro-groove distortion mode through time series analysis and spatial feature extraction, respectively. It dynamically allocates the confidence of time series data and the weight of spatial features through a physical constraint attention mechanism, generates the etching rate optimization adjustment amount and triggers the process parameter correction command. The closed-loop self-updating module is used to feed back the etching rate optimization adjustment to the etching equipment controller in real time, and simultaneously collect the RF reflection phase data after process correction and the updated interference image features, and update the dynamic coupling relationship matrix and feature fusion network parameters.

Images