Remote plasma assisted semiconductor etch surface cleaning method

Abstract

The application relates to the technical field of semiconductor surface cleaning, and discloses a remote plasma assisted semiconductor etching surface cleaning method. The method is configured with a remote plasma generating device and sets an initial parameter combination; then, semiconductor wafer surface state data is collected, a surface state change matrix is constructed after multidimensional preprocessing; the matrix is input into a multimodal feature integration model for feature integration, and an abnormal feature matrix of plasma action is constructed accordingly; subsequently, a dynamic optimization algorithm is used to iteratively optimize the initial parameter combination, and the device is set to run parameters according to the optimized parameters; then, a three-dimensional surface cleaning state model is constructed based on wafer surface impedance distribution data, and a mapping relationship between model parameters and surface contaminant components is established; finally, the device running parameters are dynamically adjusted according to the mapping relationship. Through multidimensional data processing, feature integration, parameter optimization and dynamic adjustment, the method improves the precision and adaptability of semiconductor etching surface cleaning.

Description

Technical Field

[0001] This invention relates to the field of semiconductor surface cleaning technology, specifically a remote plasma-assisted semiconductor etching surface cleaning method. Background Technology

[0002] In semiconductor manufacturing, etching is a crucial step in forming intricate circuit structures, and the surface cleanliness after etching directly affects the performance and reliability of devices. During etching, various contaminants, including polymer residues, metallic impurities, and particulate contaminants, can easily remain on the wafer surface. If these contaminants are not effectively removed, they can lead to problems such as poor contact, short circuits, or leakage in subsequent processes, severely reducing device yield.

[0003] Traditional semiconductor etching surface cleaning methods primarily rely on wet cleaning and close-range plasma cleaning technologies. While wet cleaning can remove some soluble contaminants, its effectiveness in removing stubborn polymer residues and microparticles formed after etching is limited. Furthermore, it easily introduces new chemical impurities and generates large amounts of waste liquid, which is inconsistent with the development trend of green manufacturing. Close-range plasma cleaning technology achieves cleaning through the direct interaction of plasma with the wafer surface. However, the close proximity of the plasma source to the wafer allows high-energy particles to easily cause physical damage to the wafer surface, disrupting the already formed delicate circuit structure. This damage has a more significant impact on device performance, especially in advanced processes.

[0004] Existing plasma cleaning technologies often rely on empirical values ​​or fixed procedures for parameter settings, lacking dynamic response to the real-time condition of the wafer surface. The composition and distribution of contaminants on wafer surfaces vary across different batches and regions. Plasma cleaning with fixed parameters struggles to adapt to complex and variable surface conditions, frequently resulting in over- or under-cleaning. Furthermore, traditional methods for monitoring wafer surface conditions often employ single-dimensional data acquisition, such as obtaining surface images solely through optical detection. This fails to comprehensively reflect the composition, thickness, and distribution characteristics of contaminants, leading to low accuracy in identifying anomalies during the cleaning process and hindering timely adjustments to the cleaning strategy.

[0005] As semiconductor devices evolve towards higher integration and smaller feature sizes, higher demands are placed on the precision and stability of post-etching surface cleaning. The shortcomings of existing technologies in cleaning efficiency, surface damage control, parameter adaptability, and comprehensive condition monitoring have become bottlenecks restricting the development of advanced semiconductor manufacturing processes, necessitating a precise, adaptive, and low-damage surface cleaning method. Summary of the Invention

[0006] The purpose of this invention is to provide a remote plasma-assisted semiconductor etching surface cleaning method to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a remote plasma-assisted semiconductor etching surface cleaning method, the method comprising:

[0008] Configure the remote plasma generator and set the initial parameter combination;

[0009] Collect semiconductor wafer surface state data and perform multi-dimensional preprocessing; construct a surface state change matrix based on the wafer surface state data.

[0010] The surface state change matrix is ​​input into the multimodal feature integration model for feature integration, and the plasma interaction anomaly feature matrix is ​​constructed based on the output of the multimodal feature integration model.

[0011] A dynamic optimization algorithm is used to iteratively optimize the initial parameter combination, and the operating parameters of the remote plasma generator are set according to the optimized parameter combination.

[0012] A three-dimensional surface cleanliness state model is constructed based on wafer surface impedance distribution data, and the mapping relationship between the parameters of the three-dimensional surface cleanliness state model and the composition of surface contaminants is established.

[0013] The operating parameters of the plasma generator are dynamically adjusted based on the mapping relationship.

[0014] Preferably, the acquisition of semiconductor wafer surface state data and the multi-dimensional preprocessing include:

[0015] Wafer surface morphology data, surface composition spectral data, and surface impedance distribution data are acquired using multi-channel sensors.

[0016] The collected surface state data are formatted and standardized. Surface state data with missing or noisy data are repaired and normalized. Key surface state features are extracted from the normalized data and combined with the key features to form a preprocessed dataset.

[0017] Preferably, constructing the surface state change matrix based on wafer surface state data includes:

[0018] A dynamic scanning window is set based on the sampling frequency and spatial distribution characteristics of the surface state data. The preprocessed dataset is divided into multiple spatiotemporal subsets according to the dynamic scanning window. A reference spatiotemporal subset is selected as the reference for state changes. The state deviation between each spatiotemporal subset and the reference spatiotemporal subset is analyzed. A surface state dynamic correction model is established based on the state deviation analysis results. The correction parameters of each spatiotemporal subset are calculated by applying the surface state dynamic correction model. The corrected spatiotemporal subsets are merged to generate a surface state change matrix.

[0019] Preferably, the step of inputting the surface state change matrix into the multimodal feature integration model for feature integration includes:

[0020] A temporal feature extraction module is constructed to extract time-series features from the surface state change matrix, and a spatial feature extraction module is constructed to capture the spatial distribution features from the surface state change matrix. The features output by the temporal feature extraction module and the spatial feature extraction module are dynamically weighted and fused, and the fused multimodal feature dataset is output as the result of the multimodal feature integration model.

[0021] Preferably, constructing the plasma interaction anomaly feature matrix based on the output of the multimodal feature integration model includes:

[0022] An anomaly feature association matrix is ​​established based on multi-dimensional features in a multimodal feature dataset. A variable threshold detection mechanism is used to analyze the anomaly feature association matrix, identify the anomalous action nodes in the anomaly feature association matrix, and output the information of all anomalous action nodes.

[0023] Preferably, the iterative optimization of the initial parameter combination using a dynamic optimization algorithm includes:

[0024] Define the objective function for plasma parameter optimization, initialize the parameters of the particle swarm optimization algorithm, execute the iterative search process of the particle swarm optimization algorithm, and measure the particle state and update the optimization parameter set after each iteration converges.

[0025] Preferably, the construction of the three-dimensional surface cleanliness state model based on wafer surface impedance distribution data includes:

[0026] Set the wafer surface spatial boundary parameters and mesh generation parameters, construct a three-dimensional surface mesh structure based on the wafer surface spatial boundary, and assign impedance characteristic values ​​to each node in the three-dimensional surface mesh structure;

[0027] Based on the three-dimensional surface mesh structure, the spatial topological connection relationship of the wafer surface is identified; the nodes in the three-dimensional surface mesh structure are divided into regions according to the spatial distance threshold; the statistical distribution characteristics of the node impedance characteristic values ​​are calculated in each segmented region; the region impedance characteristic vector is generated according to the statistical distribution characteristics; and the region impedance characteristic vector is associated with and stored in relation to the spatial topological connection relationship.

[0028] Preferably, establishing the mapping relationship between the parameters of the three-dimensional surface cleanliness state model and the components of surface contaminants includes:

[0029] The correlation between the impedance eigenvalues ​​of mesh nodes and the composition of contaminants in a three-dimensional surface cleanliness model is analyzed, and a database of impedance eigenvalue-contaminant composition mapping relationships is constructed.

[0030] Preferably, the dynamic adjustment of the operating parameters of the plasma generator according to the mapping relationship includes:

[0031] The system monitors changes in the impedance characteristic value of the wafer surface in real time, queries the database of impedance characteristic value-contaminant composition mapping to obtain information on changes in contaminant composition, generates plasma parameter adjustment instructions based on the information on changes in contaminant composition, and sends the plasma parameter adjustment instructions to the remote plasma generator.

[0032] Preferably, configuring the remote plasma generator and setting the initial parameter combination includes:

[0033] Select the remote plasma source type and set the initial power parameters, set the initial reaction gas type and mixing ratio parameters, determine the initial processing chamber pressure parameters, and set the initial distance parameters between the wafer surface and the plasma source.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] This remote plasma-assisted semiconductor etching surface cleaning method provides a more efficient and adaptable solution for post-etching surface cleaning through a multi-stage collaborative design. By configuring a remote plasma generator, the plasma undergoes sufficient diffusion and neutralization before reaching the wafer surface, reducing the direct impact of high-energy particles on the wafer surface and lowering the risk of physical damage. This method is particularly suitable for advanced processes where surface integrity requirements are extremely high.

[0036] By collecting surface state data from semiconductor wafers and performing multi-dimensional preprocessing, the characteristics of surface contaminants can be captured from multiple perspectives, including their distribution density, morphological features, and electrical properties. The constructed surface state change matrix provides a comprehensive and accurate data foundation for subsequent feature analysis. This multi-dimensional data processing approach avoids information biases that may result from a single data source, making the judgment of surface state closer to reality.

[0037] By inputting the surface state change matrix into a multimodal feature integration model, the advantages of different data types can be combined to achieve accurate extraction of anomalous features during plasma interaction. By constructing a plasma interaction anomaly feature matrix, potential problems such as localized overreaction or underreaction during the cleaning process can be identified in a timely manner, providing clear guidance for subsequent parameter adjustments.

[0038] A dynamic optimization algorithm is employed to iteratively optimize the initial parameter combination, eliminating reliance on fixed parameters or empirical values. This process can dynamically adjust the operating parameters of the plasma generator according to the actual state of the wafer surface, ensuring that key indicators such as plasma energy and density match the requirements for contaminant removal, thus improving the adaptability of the cleaning process.

[0039] A three-dimensional surface cleaning state model constructed based on wafer surface impedance distribution data can intuitively and comprehensively reflect the state changes of the wafer surface during the cleaning process. Establishing a mapping relationship between model parameters and surface contaminant composition provides a basis for real-time judgment of contaminant type and residual state, making the dynamic adjustment of plasma generator operating parameters more targeted and helping to reduce unnecessary energy consumption while ensuring cleaning effectiveness. Attached Figure Description

[0040] Figure 1 This is a timing diagram of the remote plasma-assisted semiconductor etching surface cleaning method described in this invention;

[0041] Figure 2 A flowchart for multi-dimensional preprocessing of semiconductor wafer surface state data;

[0042] Figure 3 A flowchart for integrating multimodal features of the surface state change matrix;

[0043] Figure 4 This is a flowchart for constructing a three-dimensional surface cleanliness state model based on wafer surface impedance distribution data. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Please see Figure 1 This invention provides a remote plasma-assisted semiconductor etching surface cleaning method, the method comprising:

[0046] The process involves configuring a remote plasma generator and setting initial parameter combinations, including selecting the plasma source type, power parameters, reactant gas types and mixing ratios, processing chamber pressure, and the distance between the wafer surface and the plasma source. Semiconductor wafer surface state data is collected and preprocessed in multiple dimensions. Based on this data, a surface state change matrix is ​​constructed, including information such as surface morphology, spectral composition, and impedance distribution. This surface state change matrix is ​​input into a multimodal feature integration model for feature integration. This model extracts and fuses temporal and spatial features to generate a fused feature dataset. Based on the output of the multimodal feature integration model, a plasma interaction anomaly feature matrix is ​​constructed to identify anomalous interaction nodes. A dynamic optimization algorithm iteratively optimizes the initial parameter combinations, defining an optimization objective function and executing a search process. After updating the parameter set, the operating parameters of the remote plasma generator are set. A three-dimensional surface cleanliness state model is constructed based on wafer surface impedance distribution data. A spatial grid structure is established, impedance characteristic values ​​are assigned, and spatial topological relationships are analyzed. A mapping relationship between the parameters of the three-dimensional surface cleanliness state model and the surface contaminant composition is established, and a database is constructed by analyzing the correlation. Based on the mapping relationship, data changes are monitored in real time, and parameter adjustment instructions are generated by querying the database to dynamically adjust the operating parameters of the plasma generator.

[0047] Example 1: See Figure 2 The acquisition of semiconductor wafer surface condition data is accomplished collaboratively by multiple sensing devices based on different principles. A high-precision non-contact laser scanner is responsible for acquiring wafer surface topography data. This device moves above the wafer along a preset path, emitting a 650nm laser beam, and generating three-dimensional point cloud coordinates after receiving the reflected signal, achieving a spatial resolution at the micrometer level. Surface composition spectral data is acquired using a Fourier transform infrared spectrometer. Its probes are arranged in a ring array 50mm above the wafer, scanning an 8cm² area every 10 seconds, covering a wavelength range of 2500-16000nm. The generated spectral data is recorded as a two-dimensional matrix showing the absorption peak intensity distribution. Surface impedance distribution data is acquired through a 64-channel capacitance sensor array, arranged in a 5mm×5mm grid on the bottom of the wafer carrier. Each sensor measures the local capacitance value at a frequency of 100Hz and converts it into an impedance value. The data acquisition process of the three sensors is synchronized in time, achieving millisecond-level alignment through a hardware clock signal.

[0048] The collected raw data first underwent format standardization. The 3D point cloud coordinates of the surface topography data were transformed to an XYZ Cartesian coordinate system, with the wafer center defined as the origin. The data was resampled to a grid density of 0.1mm × 0.1mm, forming an M × N × 3 matrix structure (M and N correspond to the number of grids along the wafer diameter). The surface composition spectral data were reordered according to the wavenumber scale, within the range of 400-4000 cm⁻¹. -1400 data points are taken at equal intervals within the range, with each spatial location corresponding to a 400-dimensional vector. Impedance data are mapped to spatial grid points based on sensor coordinates, and uncovered areas are filled using bilinear interpolation to form a matrix of the same dimension as the topographic data.

[0049] Missing value repair is performed on data in a uniform format. If five or more consecutive acquisition points show zero impedance, the sensor is considered faulty, and a weighted average algorithm based on neighboring regions is applied to fill the missing data points: valid measurements within a 1mm radius of the fault point are taken, and a weighted average is calculated based on the reciprocal of the distance. Anomaly handling for spectral data employs a sliding window mechanism, performing high-pass filtering on the 10 spectral sequences before and after each spatial point to remove low-frequency noise in the 0-0.05Hz band while retaining characteristic absorption peak information. Outlier correction for topographic data is achieved through statistical analysis, calculating the mean and standard deviation of the height in a 10μm × 10μm local area; points exceeding three times the standard deviation are replaced with the neighborhood average.

[0050] The repaired data enters the normalization stage. The Z-axis height value of the morphology is normalized using the min-max normalization method, with the lowest point of the entire wafer as 0 and the highest point as 1, and linearly scaled. Spectral intensity data is processed independently for each wavelength, and the maximum and minimum values ​​of each column of data are calculated separately and then mapped to the [0,1] interval. The impedance value is relatively dimensionalized according to the wafer substrate impedance reference value (standard impedance of silicon substrate is 500Ω): the actual measured value is divided by the reference value and then logarithmically transformed, and the result is mapped to the range of -1 to 1.

[0051] Key feature values ​​were extracted based on normalization. Surface roughness (Sa parameter) within a 10μm × 10μm local window was calculated from morphological data, with a window sliding step of 1μm, yielding one roughness feature for each grid point. Characteristic peak positions and peak height ratio parameters were extracted from spectral data: [The text abruptly shifts to a different topic] ...in SiO2 characteristic peak (1060 cm⁻¹) -1 Take ±20cm at this location. -1 The position of the maximum value within the window is used as the frequency shift; the CH absorption peak (2920 cm⁻¹) is used as the value. -1 ) and Si-O peak (1100 cm⁻¹) -1 The intensity ratio of impedance data is used as an indicator of organic pollution. Impedance data feature extraction includes: calculating the standard deviation of each grid point from its eight neighboring points as an indicator of local homogeneity; per 1 mm 2 The peak-to-valley difference in impedance values ​​within a region characterizes cleanliness fluctuations. Eigenvalues ​​and normalized data are combined according to spatial location indexes. Each spatial point contains six key feature fields and original data reference values, all stored together in a distributed time-series database.

[0052] The state change matrix is ​​constructed using a dynamic window mechanism. The time dimension window size is adaptively adjusted based on data volatility: when the feature standard deviation continuously exceeds 0.2, the window shrinks to 10 seconds; when the volatility is below 0.05, it expands to 60 seconds. The spatial dimension uses 30mm×30mm blocks as the basic unit, with each unit containing 900 spatial point data points. The mean of the first 30 seconds of complete acquisition is selected as the reference dataset. State deviation calculation uses vector space analysis: each spatiotemporal subset is treated as a high-dimensional vector, and the angle between it and the reference vector is used as the overall deviation. Simultaneously, the cosine similarity of each feature dimension is calculated as a local deviation index. After deviation analysis, a quadratic surface fitting model is constructed as a dynamic correction algorithm: using the reference data and real-time data as input, the third-order polynomial parameters are solved using the nonlinear least squares method, and the output is a transformation matrix containing translation, rotation, and scaling parameters. Each spatiotemporal subset is projected onto the reference space after processing by the transformation matrix. The corrected data is reorganized using a quadtree index structure: the wafer surface is divided into four quadrants, and each quadrant node stores the mean and variance statistics of the features of that region, forming a tree-like data structure as the storage form of the surface state change matrix. The time dimension data is compressed into keyframes per minute, and the relative change codes are stored between adjacent keyframes. The final constructed matrix has a dual-dimensional structure of spatial quadtree index and time series.

[0053] This embodiment realizes the end-to-end processing of wafer state data, from acquisition to state matrix construction, forming a traceable and analyzable basic data structure. Multidimensional data from different sources establishes a precise spatial location mapping relationship through strict spatial alignment and time synchronization mechanisms. The correlation between the original data and feature data is preserved during data transformation, providing a complete information foundation for subsequent analysis. The adaptive construction mechanism of the state change matrix effectively balances the requirements of data granularity and processing efficiency, enabling the system to respond to monitoring requirements under different operating conditions.

[0054] Example 2: See Figure 3The multimodal feature integration model employs a parallel dual-channel architecture to receive the surface state change matrix data output from Example 1. The temporal feature extraction module is deployed on a long short-term memory neural network with 128 hidden units. The input data is expanded into a continuous sampling sequence along the time axis. Each time step input contains a single frame of spatial point cloud data, with dimensions corresponding to 900 spatial point features of a 30mm × 30mm block on the wafer surface. The network learns the evolution pattern of the wafer state through a gating mechanism. The forget gate calculates the retention ratio of information based on the previous state and the current input, the input gate adjusts the degree of new information fusion, and the output gate controls the state propagation of the hidden layer. The hidden layer outputs a 32-dimensional temporal feature vector, where the first 16 dimensions represent short-term fluctuation trends, and the last 16 dimensions capture periodic change patterns. The spatial feature extraction module constructs a three-layer convolutional neural network architecture: the first layer uses eight 5×5 convolutional kernels to extract features from the spatial dimension of the surface state change matrix, with a stride of 1. Each convolutional kernel traverses the wafer plane to generate a feature map. The second layer deploys four 3×3 convolutional kernels to capture region-related features by increasing the receptive field. The final layer uses a max-pooling layer to compress the data dimension to 1 / 4 of its original size. The activation function for the convolutional layers is ReLU nonlinear transformation, and the output of each layer is batch normalized. The final spatial feature tensor contains 48 feature channels, each corresponding to a specific spatial pattern response.

[0055] The feature fusion layer employs a dual-input dynamic weighting mechanism. Temporal feature vectors are expanded to the same dimension as spatial features via a fully connected layer, mapping to a 576-dimensional vector (48 channels × 12 × 12 grid). The spatial feature tensor is expanded into a sequence of feature vectors. The weighted fusion unit calculates the contribution factors of the two types of features: static weight components are calculated by real-time analysis of feature variance; a base weight increase of 0.6 is obtained when the temporal feature variance exceeds a preset threshold of 0.25; dynamic weight components are calculated based on feature mutual information, using a sliding window to statistically analyze the joint distribution characteristics of the two feature sets at 256 time points. The final fusion weight ω calculation formula includes both static and dynamic components, with a sum constraint of 1. The fusion process is performed at the feature element level: the fusion feature value corresponding to each spatial point is calculated as temporal feature value × ω + spatial feature value × (1 - ω). The fusion result is stored in a two-dimensional array, with rows corresponding to the 576 feature dimensions and columns indexing the time series points. Array element values ​​are normalized feature values ​​in the interval [-1, 1].

[0056] The plasma interaction anomaly feature matrix was constructed by extracting features from the fused dataset. Six core features were selected: maximum fluctuation amplitude in the time dimension (feature dimension 8), local entropy value in the spatial dimension (feature dimension 23), gradient change of spatiotemporal coupling (feature dimension 112), peak value of cross-correlation (feature dimension 256), anomaly event density (feature dimension 301), and pattern stability (feature dimension 488). These six features constituted the initial feature space, and the position of each data point in this space was projected onto three-dimensional coordinates through dimensionality reduction. The anomaly feature correlation matrix adopted a 12×12 symmetric matrix structure, with matrix elements representing the interaction strength between features, calculated using the cosine of the angle between feature vectors. The diagonal elements of the matrix were set as the autocorrelation index of each feature, and the position value of the off-diagonal element (i,j) was calculated as the product of the covariance of feature i and feature j within the sliding time window divided by the feature standard deviation.

[0057] A variable threshold detection mechanism establishes a three-level dynamic judgment criterion. The primary detection scans the off-diagonal elements of the correlation matrix; when an element value exceeds the range [0.85, 1.15], an anomaly marker is triggered. The secondary detection analyzes the distribution of matrix eigenvalues: calculating the roots of the matrix's eigenpolynomial; if a eigenvalue with a modulus greater than 3 exists, a significant anomaly pattern is determined. The tertiary detection spatially locates the anomaly correlation patterns: using deconvolution operations, the anomaly features are mapped back to their positions on the wafer surface, achieving a localization resolution of 1 mm². Anomaly node identification employs a density clustering algorithm: based on feature spatial points, a radius threshold of 0.35 is set; when the number of points in the neighborhood exceeds 25, a core node is formed; edge nodes are grouped into the same anomaly cluster when the distance to the core node is less than 0.45. After the center of each anomaly cluster is determined, its original spatial coordinates are traced and mapped to a specific location on the wafer surface through coordinate transformation.

[0058] The final anomaly node information is output in a structured record format: the location field records the wafer coordinates (X, Y) and height layer (Z); the timestamp field is accurate to the millisecond level; the anomaly type field is classified into six encoding forms according to the feature combination pattern; and the feature deviation field records the percentage offset of the six-dimensional feature values ​​from the standard values. Each anomaly record is accompanied by an association matrix slice data, stored as a 64-bit floating-point array. The data processing is implemented through a pipelined architecture: the temporal feature extractor is deployed on the FPGA chip for real-time processing, spatial feature extraction runs on the GPU computing cluster, and the fusion layer and anomaly detection module complete the calculation in the CPU main memory. Data blocks are transmitted between modules via a high-speed PCIe bus. The anomaly node information is transmitted to the control center via industrial Ethernet, with messages encapsulated in JSON format and the communication protocol following the Modbus TCP specification.

[0059] Example 3: See Figure 4The dynamic optimization algorithm employs an improved particle swarm optimization framework and designs a dedicated objective function for the plasma parameter combination optimization problem. This objective function... Defined as:

[0060]

[0061] in: This represents a parameter vector consisting of power value, gas ratio, and chamber pressure. The plasma uniformity index is calculated using the coefficient of variation of wafer surface impedance measurements. This represents a cleaning efficiency parameter, based on the measurement of pollutant removal rate per unit time. This is an energy cost function, calculated by combining power consumption and gas usage. Weighting coefficients. , , Pre-set according to process requirements, to meet The constraints.

[0062] The algorithm initialization phase sets up 60 particles, with each particle's position vector containing five optimization variables: radio frequency power (range 50-500W), argon flow rate (10-100 sccm), oxygen content (5-30%), fluorocarbon concentration (1-15%), and chamber pressure (0.1-10 Torr). The particle velocity vector is initialized with random values, limited to ±20% of the search interval for each variable. Cognitive coefficient. Set to 1.8, social coefficient Set to 1.6, inertia weight The value decreases linearly from 0.9 to 0.4. The iterative process employs an asynchronous update strategy: the global optimal solution is updated after every 10 evaluations, and a boundary check is performed after each position update, with parameters exceeding the search space handled through reflection.

[0063] Particle state assessment was achieved through an experimental measurement system. Upon particle position update, the control system automatically configured the corresponding parameter combination of the plasma generator, initiated a 15-second stabilization process, and then began data acquisition. Wafer surface impedance data was acquired using a 64-point array sensor, with a sampling frequency of 1 kHz and a duration of 10 seconds. Uniformity... Calculated as the ratio of the standard deviation to the mean of the impedance values ​​at 100 spatial points. Cleaning efficiency. The energy consumption was determined by comparing the rate of change in the carbon peak area in the X-ray photoelectron spectra before and after treatment. The cumulative consumption of power meters and gas mass flow meters is recorded in real time. The evaluation results are transmitted to the optimization algorithm master node and stored in the historical database for convergence analysis.

[0064] Convergence is determined using a composite conditional mechanism. The primary convergence metric is the population fitness variance; convergence occurs when the variance is less than a certain value for 20 consecutive generations. Convergence is triggered at certain times. Auxiliary conditions include: the global optimum has not improved for 50 consecutive generations, or the average particle movement distance is less than 0.1% of the search space diameter. Upon reaching convergence, the algorithm outputs a non-dominated solution set containing 15 Pareto optimal parameter combinations. The top three combinations are selected as recommended configurations after being sorted by objective function weights.

[0065] The 3D surface cleanliness model was constructed based on impedance measurement data. The wafer space boundary was defined as a circular region with a diameter of 200 mm. An adaptive triangulation algorithm was used for mesh generation: the initial mesh side length was 5 mm, and in regions with an impedance gradient greater than 10 Ω / mm, the mesh was automatically refined to a resolution of 1 mm. Each mesh node was associated with an impedance measurement value, and missing data was filled in using radial basis function interpolation. The kernel function was selected as a cubic spline.

[0066]

[0067] in: This represents the Euclidean distance between the interpolation point and the sample point. Spatial topological relationships are stored using a half-edge data structure, recording the vertex index, adjacent edge information, and normal vector direction of each triangle. Region segmentation employs a spectral clustering algorithm based on impedance values: calculating the node impedance similarity matrix, performing k-means clustering on the first three eigenvectors after eigenvalue decomposition, forming 5-8 feature regions.

[0068] Regional characteristic statistics include four core indicators: mean impedance reflects overall cleanliness; coefficient of variation characterizes regional uniformity; skewness coefficient indicates the distribution tendency of outliers; and autocorrelation length describes spatial continuity. These indicators constitute a four-dimensional feature vector, which, along with the regional spatial coordinates, is stored in a graph database. The database employs an attribute graph model, where nodes represent feature regions, edges store the relationships between adjacent regions, and node attributes include timestamps, process parameters, and feature vector values. The query interface supports spatiotemporal range filtering, enabling rapid retrieval of the evolution history of regional characteristics under specific treatment conditions.

[0069] The model update mechanism is designed as an incremental learning framework. After each plasma processing cycle, newly acquired impedance data triggers local mesh reconstruction: only the mesh of regions with changes exceeding 5% is updated, and the topological relationships and feature vectors of the affected regions are recalculated. Historical data retention employs a sliding window management strategy, keeping the most recent 50 processing records online, while earlier data is compressed and archived. The visualization subsystem projects the 3D model onto a 2D plane, uses gradient colors to map the impedance value distribution, and overlays characteristic region boundaries and outlier markers, with a refresh rate synchronized with the data processing cycle.

[0070] A closed-loop control mechanism is established for the collaborative workflow of parameter optimization and 3D modeling. Each new parameter combination generated in the particle swarm optimization iteration immediately triggers a 3D model update after experimental evaluation. Model analysis results are fed back to the optimization algorithm, dynamically adjusting the objective function weights: when large-area inhomogeneity is detected, the weights are increased. The weighting coefficient is adjusted; if areas are found to be incompletely cleaned, the weighting coefficient is increased. The proportion of [something unclear]. This real-time feedback mechanism enables the optimization process to adapt to the actual state changes of the wafer surface, avoiding getting trapped in local optima. All process data is recorded in a distributed time-series database, including raw measurements, feature parameters, optimization paths, and model snapshots, supporting subsequent offline analysis and algorithm improvement.

[0071] Example 4: The process of establishing the mapping relationship between the parameters of the three-dimensional surface cleanliness model and the components of contaminants is illustrated using silicon wafer surface treatment as an example. The model parameters are derived from the three-dimensional mesh node data constructed in Example 3. Each node contains three basic parameters: the real part of impedance, the imaginary part, and the temperature compensation value, as well as derived features such as mean, variance, and gradient extracted from regional statistics. The analysis of contaminant components was performed using a synchrotron X-ray fluorescence spectrometer. The measurement points were spatially aligned with the impedance mesh nodes, and the detected elements covered common contaminants such as carbon, oxygen, fluorine, copper, and iron.

[0072] The construction of the mapping database began with the data acquisition phase. Twelve batches of process verification wafers were selected, each batch containing five samples processed under the same conditions. Each wafer underwent impedance scanning and contaminant analysis immediately after plasma cleaning. Impedance measurements were performed at a constant temperature of 23°C, with a 100kHz sinusoidal excitation signal applied to the probe, and the amplitude and phase difference were recorded. Contaminant detection was performed in a vacuum chamber with an X-ray beam diameter of 50μm, and three spectra were acquired at each grid node and averaged. The raw data was preprocessed to form a structured record; the following is an example data table:

[0073]

[0074] The correlation analysis employed a hierarchical modeling approach. The first layer established a single-parameter linear regression model to analyze the correlation between various impedance parameters and pollutant concentrations. For example, the scatter plot of the real part of the impedance and carbon concentration showed a positive correlation, with the data points exhibiting a banded structure; the imaginary part showed a negative correlation with fluoride concentration. The second layer constructed a multi-parameter random forest model, with input including all impedance characteristic parameters and outputting the predicted concentrations of various pollutants. 80% of the sample data was used for model training, with the remaining 20% ​​used for validation. The depth of each decision tree was limited to 5 layers to prevent overfitting.

[0075] The database architecture is designed as a distributed storage system. The master node stores the complete mapping model, including the regression coefficient matrix and the parameter set of the random forest. Slave nodes are deployed on various testing devices to cache frequently used query results. Data updates employ a version control mechanism; each change in process parameters generates a new version of the mapping relationship, while historical records are retained for traceability and comparison. The query interface supports multi-condition combined searches; for example, it can query the distribution pattern of typical contaminants when the real part of the impedance is in the range of 120-130Ω, or specify the range of impedance characteristic values ​​when the copper content is below 20ppm.

[0076] An application example demonstrates a real-world scenario where the mapping relationship is used. When the online detection system detects a sudden increase in the real part of the impedance to 145Ω in a certain area, the database query returns the probability distribution of pollutants corresponding to this characteristic value: carbon pollution probability 78%, metal residue 15%, and others 7%. Based on this result, the control system adjusts the plasma process parameters, increasing the oxygen ratio by 2% to improve carbon removal efficiency. After treatment, a retest of the impedance value shows it has returned to the normal range, and X-ray sampling confirms that the carbon concentration has decreased to below the standard value.

[0077] The anomaly handling mechanism reflects the dynamic adaptability of the mapping relationship. A new type of photoresist residue was found in a batch of wafers, and the initial impedance characteristics deviated from the database records. The system automatically triggered a learning mode: it collected complete spectral data from 10 anomaly points, and laboratory analysis confirmed the addition of nitrogen element characteristic peaks. The database initiated an incremental update process, adding a nitrogen contamination category while maintaining the original mapping relationship and retraining the model parameters. The updated model can then identify this new type of contamination and accurately classify it in the next detection.

[0078] The data visualization subsystem provides an intuitive display of mapping relationships. A heatmap showing the correlation between impedance values ​​and contaminant concentrations is overlaid on a 3D wafer model, with different colors distinguishing contaminant types. Users can interactively select specific parameter ranges, and the system displays the corresponding contaminant distribution histogram in real time. The historical data comparison function supports side-by-side display of differences in mapping patterns under different process conditions, assisting engineers in analyzing the effects of parameter adjustments.

[0079] The quality control module utilizes mapping relationships to provide early warnings. When the online system detects that the impedance characteristics are starting to deviate from the normal range but have not yet exceeded the limit, it predicts the contaminant accumulation trend based on historical data. If the prediction indicates that copper contamination will exceed the threshold after three treatments, it suggests adjusting process parameters or replacing consumables in advance. This predictive maintenance mechanism significantly reduces wafer rework rates while avoiding surface damage caused by over-cleaning.

[0080] Database maintenance includes regular validation and optimization. Monthly cross-validation of stored mappings is performed, and model accuracy is evaluated using independent test datasets. When the prediction error for a particular pollutant is found to be continuously increasing, a data re-collection process is triggered. Storage optimization employs columnar compression technology, Delta encoding for impedance parameters, and dictionary compression for pollutant concentrations, reducing the database size by more than 60%. Access performance optimization is achieved by building feature value indexes, keeping common query response times within 200 milliseconds.

[0081] The practical applications of this mapping relationship cover multiple stages of semiconductor manufacturing. In addition to plasma cleaning processes, it extends to scenarios such as surface inspection after chemical mechanical polishing and substrate condition assessment before thin film deposition. The mapping database establishes correlations between different processes, enabling cross-process contaminant propagation path analysis. When abnormal contamination is detected in a process, the impedance characteristics of upstream processes can be traced to quickly locate the contamination source.

[0082] Example 5: Dynamic adjustment of the plasma generator's operating parameters begins with real-time data acquisition. Wafer surface impedance distribution monitoring is performed using a 64-channel capacitive sensor array. The array refreshes global scan data every 10 milliseconds, with each measurement point covering a 3mm² area. Sensor signals are converted into a digital matrix by a 24-bit ADC and transmitted to the data processing unit via optical fiber. Impedance eigenvalue extraction includes three core parameters: fundamental frequency impedance magnitude, phase difference, and harmonic distortion. The calculation process uses a Fast Fourier Transform algorithm to process the raw waveform data. The eigenvalue spatial coordinates are precisely mapped to the wafer's physical location, with the coordinate system established as a planar grid centered on the wafer center.

[0083] The query operation of the impedance characteristic value-pollutant component mapping database adopts a distributed architecture. After receiving the real-time impedance data stream, the front-end server performs feature matching in the in-memory database. The matching condition is a range search in a multi-dimensional feature space: the input feature vector includes the current value, recent rate of change, and spatial gradient value, and Euclidean distance is calculated with the database records. When the matching distance is less than a preset threshold, a prediction result set is returned, containing the probability distribution of pollutant types and concentration change trajectories. The prediction results are encapsulated in a standard data structure, with pollutant classification coding using the ISO 14644-1 standard, and concentration data expressed in both percentage and ppm units.

[0084] The parameter adjustment command generation is based on a dual mechanism of rule engine and learning model. The central control unit is configured with a three-layer decision architecture: the basic layer is a preset rule table containing standard response plans for 20 typical pollution scenarios. For example, when the carbon pollution probability is detected to exceed 65%, rule C-03 is triggered, instructing an increase in the oxygen ratio by 3±0.5%; when an increase in copper residual concentration is detected, rule M-12 is activated, instructing a 15% increase in fluorocarbon gas flow. The middle layer is an adaptive learning module that records historical adjustment effects and optimizes rule parameters. When the same scenario occurs more than three times, the system automatically fine-tunes the response amplitude and delay time. The top layer has an expert intervention channel, supporting manual injection of specific command sets to deal with complex operating conditions.

[0085] Command transmission employs an Industrial Internet of Things (IIoT) protocol stack. Command encoding uses lightweight JSON format, including fields such as target device address, parameter modifications, execution time window, and checksum. Deterministic Ethernet technology is selected for network transmission, ensuring command delivery to the target device within 5 milliseconds through time-sensitive networking. The transmission path employs a dual-redundancy design: the primary channel is a fiber optic ring network, and the backup channel is a shielded twisted-pair network. An automatic switching mechanism activates after 50 microseconds of signal loss.

[0086] The parameter execution system of the plasma generator comprises multi-level control loops. RF power adjustment is achieved through a fully digital resonant converter, which changes the duty cycle of the PWM drive signal upon receiving commands, achieving a power setting resolution of 1W. Gas flow control utilizes a mass flow controller array, with each gas channel independently adjustable and a proportional valve opening control accuracy of 0.1%. Chamber pressure regulation is achieved by a stepper motor driving a throttle valve, with a position feedback encoder resolution of 0.01 degrees. The wafer distance adjustment mechanism is equipped with a linear motor-driven platform, achieving a repeatability accuracy of ±0.1mm. The status of all actuators is fed back to the central controller in real time, forming a closed-loop control verification mechanism.

[0087] The initial parameter configuration of the remote plasma generator establishes the equipment's baseline state. Upon system startup, a default configuration library is loaded, and a preset group is automatically selected based on the wafer size. For a 200mm silicon wafer, configuration group B2 is selected, which includes basic parameter settings: inductively coupled source type activated, RF power 200W, argon base flow rate 50sccm, oxygen percentage 15%, CF4 concentration 8%, operating pressure 1.2 Torr, and source-wafer distance 120mm. Configuration information is stored in non-volatile memory and supports offline retrieval.

[0088] The user interface enables visualized monitoring and intervention. The interface displays a 3D wafer impedance distribution heatmap, overlaid with predicted contaminant location markers. Parameter adjustment processes are shown through dynamic curves displaying historical trends, including a comparison of commanded and actual feedback values. The warning system has multi-level alarm thresholds; a yellow alert is triggered when the actual parameter deviates from the commanded value by more than 2%, and a red alarm is activated and the process is paused when it exceeds 5%. The data logging function saves a complete operation log, including timestamps, operator IDs, parameter change details, and execution result codes.

[0089] The system employs multiple safety protection mechanisms. Electrical parameters are set with upper and lower limit hard protection, limiting power output to an adjustable range of 50-500W; exceeding this range immediately cuts off the drive power. An interlock logic is implemented for the gas mixing ratio, automatically reducing the fluorocarbon gas flow when the oxygen concentration exceeds 30%. Physical limit switches are installed on moving mechanical parts to prevent exceeding safe travel limits. A temperature monitoring system detects the plasma source casing temperature, initiating forced cooling when it exceeds 80°C. All safety events trigger audible and visual alarms and generate fault code snapshots, stored in a separate safety log partition.

[0090] The control system's software architecture is modularly updatable. Core control algorithms are encapsulated as independent service components, supporting online upgrades without affecting hardware operation. User configuration strategies are exported as standard XML files, allowing process engineers to optimize and adjust schemes in an offline environment. The system maintenance mode supports full-parameter self-calibration, including sensor zero-point calibration, actuator stroke calibration, and control loop parameter tuning. The data interface reserves the OPCUA communication protocol for data integration with factory-level manufacturing execution systems.

[0091] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0092] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A remote plasma-assisted semiconductor etching surface cleaning method, characterized in that, Includes the following steps: Configure the remote plasma generator and set the initial parameter combination; Collect semiconductor wafer surface state data and perform multi-dimensional preprocessing; construct a surface state change matrix based on the wafer surface state data. The surface state change matrix is ​​input into the multimodal feature integration model for feature integration, and the plasma interaction anomaly feature matrix is ​​constructed based on the output of the multimodal feature integration model. A dynamic optimization algorithm is used to iteratively optimize the initial parameter combination, and the operating parameters of the remote plasma generator are set according to the optimized parameter combination. A three-dimensional surface cleanliness state model is constructed based on wafer surface impedance distribution data, and the mapping relationship between the parameters of the three-dimensional surface cleanliness state model and the composition of surface contaminants is established. The operating parameters of the plasma generator are dynamically adjusted according to the mapping relationship; The process of collecting semiconductor wafer surface state data and performing multi-dimensional preprocessing includes: Wafer surface morphology data, wafer surface composition spectrum data, and wafer surface impedance distribution data are acquired by multi-channel sensors. The acquired surface state data are formatted and standardized. Surface state data with missing or noisy data are repaired and normalized. Key surface state features are extracted from the normalized data and combined with the key features to form a preprocessed dataset. The step of inputting the surface state change matrix into the multimodal feature integration model for feature integration includes: A temporal feature extraction module is constructed to extract time-series features from the surface state change matrix, and a spatial feature extraction module is constructed to capture the spatial distribution features from the surface state change matrix. The features output by the temporal feature extraction module and the spatial feature extraction module are dynamically weighted and fused, and the fused multimodal feature dataset is output as the result of the multimodal feature integration model. The construction of the plasma interaction anomaly feature matrix based on the output of the multimodal feature integration model includes: An anomaly feature association matrix is ​​established based on multi-dimensional features in a multimodal feature dataset. A variable threshold detection mechanism is used to analyze the anomaly feature association matrix, identify the anomalous action nodes in the anomaly feature association matrix, and output the information of all anomalous action nodes.

2. The remote plasma-assisted semiconductor etching surface cleaning method according to claim 1, characterized in that, The construction of the surface state change matrix based on wafer surface state data includes: A dynamic scanning window is set based on the sampling frequency and spatial distribution characteristics of the surface state data. The preprocessed dataset is divided into multiple spatiotemporal subsets according to the dynamic scanning window. A reference spatiotemporal subset is selected as the reference for state changes. The state deviation between each spatiotemporal subset and the reference spatiotemporal subset is analyzed. A surface state dynamic correction model is established based on the state deviation analysis results. The correction parameters of each spatiotemporal subset are calculated by applying the surface state dynamic correction model. The corrected spatiotemporal subsets are merged to generate a surface state change matrix.

3. The remote plasma-assisted semiconductor etching surface cleaning method according to claim 2, characterized in that, The iterative optimization of the initial parameter combination using a dynamic optimization algorithm includes: Define the objective function for plasma parameter optimization, initialize the parameters of the particle swarm optimization algorithm, execute the iterative search process of the particle swarm optimization algorithm, and measure the particle state and update the optimization parameter set after each iteration converges.

4. The remote plasma-assisted semiconductor etching surface cleaning method according to claim 3, characterized in that, The construction of the three-dimensional surface cleanliness state model based on wafer surface impedance distribution data includes: Set the wafer surface spatial boundary parameters and mesh generation parameters, construct a three-dimensional surface mesh structure based on the wafer surface spatial boundary, and assign impedance characteristic values ​​to each node in the three-dimensional surface mesh structure; Based on the three-dimensional surface mesh structure, the spatial topological connection relationship of the wafer surface is identified; the nodes in the three-dimensional surface mesh structure are divided into regions according to the spatial distance threshold; the statistical distribution characteristics of the node impedance characteristic values ​​are calculated in each segmented region; the region impedance characteristic vector is generated according to the statistical distribution characteristics; and the region impedance characteristic vector is associated with and stored in relation to the spatial topological connection relationship.

5. The remote plasma-assisted semiconductor etching surface cleaning method according to claim 4, characterized in that, The mapping relationship between the parameters of the three-dimensional surface cleanliness model and the composition of surface contaminants includes: The correlation between the impedance eigenvalues ​​of mesh nodes and the composition of contaminants in a three-dimensional surface cleanliness model is analyzed, and a database of impedance eigenvalue-contaminant composition mapping relationships is constructed.

6. The remote plasma-assisted semiconductor etching surface cleaning method according to claim 5, characterized in that, The dynamic adjustment of the operating parameters of the plasma generator based on the mapping relationship includes: The system monitors changes in the impedance characteristic value of the wafer surface in real time, queries the database of impedance characteristic value-contaminant composition mapping to obtain information on changes in contaminant composition, generates plasma parameter adjustment instructions based on the information on changes in contaminant composition, and sends the plasma parameter adjustment instructions to the remote plasma generator.

7. The remote plasma-assisted semiconductor etching surface cleaning method according to claim 6, characterized in that, The configuration of the remote plasma generator and the setting of the initial parameter combination include: Select the remote plasma source type and set the initial power parameters, set the initial reaction gas type and mixing ratio parameters, determine the initial processing chamber pressure parameters, and set the initial distance parameters between the wafer surface and the plasma source.

Images