Wafer edge anti-fouling rinsing method and system

Abstract

The application discloses a wafer edge anti-fouling rinsing method and system, and belongs to the field of semiconductor wafer cleaning and process control. In order to solve the problem that residual liquid is prone to be retained and splash secondary pollution occurs in the edge area, the back surface outer ring and the clamping contact area of a high-speed rotating wafer, the application realizes the technical effects of reducing wafer edge residues, reducing splash pollution, and improving rinsing stability and cleanliness through multi-view synchronous sampling, abnormal segmentation, state fusion, digital twin prediction and constrained reinforcement learning closed-loop control.

Description

Technical Field

[0001] This invention relates to the field of semiconductor wafer cleaning and process control, and in particular to a method and system for preventing contamination at the edge of a wafer. Background Technology

[0002] As semiconductor manufacturing processes advance towards higher cleanliness and yield, the requirements for cleaning control of edge areas, the outer ring of the back surface, and the clamping contact area in wafer wet rinsing equipment are constantly increasing. In existing wafer rinsing processes, the wafer is usually in a high-speed rotating state, and the rinsing fluid is thrown out along the surface towards the edge under the action of centrifugal force. The top and bottom chamfers of the edges and the clamping and shielding areas are prone to insufficient coverage, fluid flow turbulence, and local residual fluid retention.

[0003] Meanwhile, the ejected rinsing fluid, after impacting the inner wall of the chamber, the liquid-blocking structure, or the clamping components, easily forms backsplash droplets that re-adhere to the wafer edge, back side, or clamping contact area, leading to problems such as particle residue, chemical residue, watermarks, and subsequent contamination diffusion. Existing solutions typically lack the ability to identify zones based on contamination status, backsplash risk, and residual liquid distribution, and also lack a closed-loop optimization mechanism that can link nozzle parameters, wafer rotation speed, and negative pressure drainage parameters.

[0004] Therefore, there is a need for a wafer edge anti-fouling rinsing method and system that can overcome the shortcomings of the existing technology. Summary of the Invention

[0005] One objective of this invention is to propose a wafer edge anti-fouling rinsing method and system. In view of the problem that residual liquid retention and backsplashing are prone to secondary contamination at the wafer edge, back outer ring and clamping contact area in the prior art, this invention proposes a technical solution based on multi-view synchronous sampling, few-sample anomaly segmentation, state fusion, digital twin prediction and constrained reinforcement learning control. This invention has the technical effect of reducing edge residue, mitigating backsplashing contamination and improving rinsing stability and cleanliness.

[0006] This invention provides a wafer edge anti-fouling rinsing method, comprising: S1, acquiring multi-view edge images synchronized with a high-speed rotating wafer, encoder angle, and flow rate, pressure, negative pressure, spindle current, and acoustic signals, and unfolding the wafer edge region and clamping contact area into a ring-shaped partition state map according to the encoder angle; S2, inputting the ring-shaped partition state map into a trained few-sample anomaly segmentation model to obtain a contamination mask and residual liquid film mask for each partition; S3, fusing the contamination mask, residual liquid film mask, and the flow rate, pressure, negative pressure, spindle current, and acoustic signals to obtain the residual liquid thickness, backsplash probability, and clamping shielding compensation coefficient for each partition; S4, rinsing the residual liquid... Thickness, splashing probability, clamping and shielding compensation coefficient, and candidate execution parameters are input into the digital twin model. Under preset process constraints, the residual liquid state and risk results of each partition at the next moment are predicted for each candidate execution parameter, and candidate execution parameters that meet the preset process constraints are selected. S5, The risk results corresponding to the selected candidate execution parameters and the preset process constraints are input into the constrained reinforcement learning controller. The controller selects and outputs action vectors including nozzle flow rate, spray angle, wafer rotation speed, edge auxiliary nozzle start / stop, and negative pressure drainage control values ​​of each partition from the selected candidate execution parameters, and controls the adjustable angle nozzle module and the segmented annular drainage negative pressure module to perform corresponding rinsing and recycling.

[0007] Optionally, S1 includes:

[0008] The edge top chamfer, edge vertex, bottom chamfer, back outer ring, and each clamping contact area are divided into multiple angular partitions according to polar coordinates. The edge images from at least two viewpoints are registered and expanded according to the encoder angle window corresponding to each angular partition to generate the annular partition state diagram.

[0009] Optionally, S2 includes:

[0010] The multi-scale features of the annular partition state map are extracted using a pre-trained visual feature extraction network. The multi-scale features are then input into a few-sample anomaly segmentation head for partition anomaly segmentation. Based on the segmentation results, the contamination mask and the residual liquid mask are output, wherein the contamination mask includes a particulate contamination mask and a drug residue mask.

[0011] Furthermore, the training process of the few-shot anomaly segmentation model includes: establishing a partitioned benchmark feature library based on normal rinsing images; using few-shot anomaly images labeled with particulate contamination, drug residue, and residual film to perform metric learning correction on the partitioned benchmark feature library; and using the corrected feature distribution to train the few-shot anomaly segmentation head.

[0012] Optionally, S3 includes:

[0013] The contamination mask includes a particulate contamination mask and a liquid residue mask: the particulate contamination mask, the liquid residue mask, the residual liquid film mask are time-aligned with the flow rate, pressure, negative pressure, spindle current and acoustic signals according to the partition number, and the residual liquid thickness, the backsplash probability and the clamping and shielding compensation coefficient are output respectively through the partition state fusion network, wherein the backsplash probability is determined according to the statistical frequency of backsplash droplets returning to the wafer surface.

[0014] Optionally, S4 includes:

[0015] The residual liquid thickness, the backsplash probability, the clamping and shielding compensation coefficient, and the candidate execution parameters consisting of nozzle flow rate, spray angle, wafer rotation speed, edge auxiliary nozzle start / stop, and negative pressure drainage control values ​​for each zone are input into the digital twin model. The predicted residual liquid thickness, backsplash landing point heat value, and water mark risk value for each zone at the next moment are calculated under the action of each candidate execution parameter. Candidate execution parameters that do not meet the wafer rotation speed upper limit, surface shear threshold, negative pressure capability threshold, clamping stability threshold, and liquid consumption threshold are eliminated. The retained candidate execution parameters are used as the source of candidate parameters for subsequent control steps.

[0016] Optionally, S5 includes:

[0017] The predicted value of residual liquid thickness at the next moment corresponding to the retained candidate execution parameters, the heat value of the splash landing point, and the watermark risk value are constructed into a state vector. The retained candidate execution parameters are constructed into an action set. The rotation speed deviation, surface shear deviation, negative pressure deviation, clamping stability deviation, and liquid consumption deviation are determined according to the wafer rotation speed, surface shear amount, negative pressure capacity occupancy value determined by the negative pressure drainage control value of each partition, clamping stability index, and the difference between the liquid consumption per unit time and the corresponding wafer rotation speed upper limit, surface shear threshold, negative pressure capacity threshold, clamping stability threshold, and liquid consumption threshold. The preset constraint budget is the upper limit set for the cumulative constraint cost corresponding to each of the above deviations.

[0018] The constrained reinforcement learning controller selects the action vector from the action set that ensures the cumulative constraint cost does not exceed the preset constraint budget, and controls the adjustable angle nozzle module and the segmented annular drainage negative pressure module according to the action vector, wherein the negative pressure drainage control value of each partition in the action vector is used to determine the negative pressure drainage setting value corresponding to each partition.

[0019] Furthermore, the contamination mask includes a particulate contamination mask and a drug residue mask. The few-sample anomaly segmentation model outputs the identification uncertainty of each partition while outputting the particulate contamination mask, the drug residue mask, and the residual film mask. The identification uncertainty is determined based on the segmentation probability entropy and the segmentation boundary confidence difference.

[0020] When the identification uncertainty is greater than the preset uncertainty threshold, a local re-inspection of the corresponding partition is triggered. The local re-inspection involves re-acquiring local multi-view images of the corresponding partition and recalculating the particulate contamination mask, the drug residue mask, the residual liquid film mask, and the identification uncertainty. The updated calculation results are written back to the annular partition state diagram. Based on the updated particulate contamination mask, the drug residue mask, and the residual liquid film mask, the residual liquid thickness, splash probability, and clamping and occlusion compensation coefficient of the corresponding partition are re-determined in the current control wheel.

[0021] The upper limit of nozzle flow rate, the upper limit of wafer rotation speed and the lower limit of the corresponding partition negative pressure drainage control value are switched to a preset conservative parameter set. The preset conservative parameter set is used to narrow the complete candidate execution parameter range of the current control wheel. The constrained reinforcement learning controller generates the complete action vector of the current control wheel within the narrowed candidate execution parameter range and limits the action vector component of the corresponding partition to the conservative action component selected from the conservative action set.

[0022] Furthermore, the state vector also includes a temporal residual liquid memory vector established for each angular partition. The temporal residual liquid memory vector is updated based on the predicted residual liquid thickness, backsplash landing heat value, and watermark risk value for 3 to 10 consecutive revolutions in the same angular partition. The temporal residual liquid memory vector and the clamping occlusion compensation coefficient together constitute an extended state vector and are input to the constrained reinforcement learning controller to determine the action vector.

[0023] On the other hand, the present invention also provides a wafer edge anti-fouling rinsing system, comprising:

[0024] The synchronous sampling module acquires multi-view edge images, encoder angles, flow rate, pressure, negative pressure, spindle current, and acoustic signals synchronized with the high-speed rotating wafer, and generates a ring-shaped partition state map. The anomaly segmentation module outputs the contamination mask and residual liquid film mask for each partition. The state assessment module outputs the residual liquid thickness, backsplash probability, and clamping obstruction compensation coefficient for each partition based on the contamination mask, the residual liquid film mask, and the flow rate, pressure, negative pressure, spindle current, and acoustic signals. The twin prediction and screening module predicts the residual liquid state and risk results corresponding to each candidate execution parameter based on a digital twin model, and screens candidate execution parameters that meet preset process constraints. The constrained control module inputs the screened candidate execution parameters and their corresponding prediction results into a constrained reinforcement learning controller, and controls the adjustable angle nozzle module and the segmented ring-shaped drainage negative pressure module to perform corresponding rinsing and recovery based on the solution results of the constrained reinforcement learning controller.

[0025] The beneficial effects of this invention are:

[0026] 1. By performing multi-view synchronous sampling of the edge region and clamping contact area of ​​the high-speed rotating wafer, and combining anomaly segmentation and state fusion processing, the distribution of contaminants and residual liquid status can be identified more accurately, thereby improving the targeted cleaning capability of the edge region.

[0027] 2. By using a digital twin model to predict the residual liquid state and risk results under candidate execution parameters, and screening candidate execution parameters under preset process constraints, the risk of secondary pollution from back splashing and over-rinsing can be reduced, and the stability of the rinsing process can be improved.

[0028] 3. By using a constrained reinforcement learning controller to link nozzle flow rate, spray angle, wafer rotation speed, and negative pressure drainage control value for closed-loop control, the system can balance cleaning effect, clamping stability, and liquid consumption requirements, thereby improving wafer rinsing cleanliness and process consistency. Attached Figure Description

[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0030] Figure 1 This is a flowchart of a method for preventing contamination at the edge of a wafer.

[0031] Figure 2 This is a flowchart of the constrained reinforcement learning control and conservative decision-making mechanism in step S5 of the present invention. Detailed Implementation

[0032] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0033] refer to Figures 1-2A wafer edge anti-fouling rinsing method includes: S1, acquiring multi-view edge images synchronized with a high-speed rotating wafer, encoder angle, and flow rate, pressure, negative pressure, spindle current, and acoustic signals, and unfolding the wafer edge region and clamping contact area into a ring-shaped partition state map according to the encoder angle; S2, inputting the ring-shaped partition state map into a trained few-sample anomaly segmentation model to obtain a contamination mask and residual liquid film mask for each partition; S3, fusing the contamination mask, residual liquid film mask, and the flow rate, pressure, negative pressure, spindle current, and acoustic signals to obtain the residual liquid thickness, splash probability, and clamping shielding compensation coefficient for each partition; S4, calculating the residual liquid thickness... The backsplash probability, clamping and shielding compensation coefficient, and candidate execution parameters are input into the digital twin model. Under the preset process constraints, the residual liquid state and risk results of each partition at the next moment are predicted for each candidate execution parameter, and candidate execution parameters that meet the preset process constraints are selected. S5. The risk results corresponding to the selected candidate execution parameters and the preset process constraints are input into the constrained reinforcement learning controller. The controller selects and outputs action vectors including nozzle flow rate, spray angle, wafer rotation speed, edge auxiliary nozzle start / stop, and negative pressure drainage control values ​​of each partition from the selected candidate execution parameters, and controls the adjustable angle nozzle module and the segmented annular drainage negative pressure module to perform corresponding rinsing and recycling.

[0034] In this specific embodiment, S1 includes:

[0035] Two edge cameras are fixedly installed on the equipment side to form at least two perspectives. The first camera faces the top chamfer and edge vertex area of ​​the wafer edge, and the second camera faces the bottom chamfer and back outer ring area. Both cameras use global shutter imaging and are triggered by the same hardware trigger source. The trigger source comes from the equal angle pulse of the spindle encoder to ensure that the image is synchronized with the angle of the high-speed rotating wafer.

[0036] The spindle encoder outputs per revolution A count of zero-index pulses is generated, and when the controller detects the zero-index pulse, the encoder angle at that moment is defined as the reference angle. In subsequent runs, the corresponding encoder angle is written to the edge image of each frame. ,in Converted from encoder count to The radian angle is recorded together with the frame image timestamp, flow signal, pressure signal, negative pressure signal, spindle current signal and acoustic signal to form synchronous sampling data within the same control wheel;

[0037] The wafer edge region and clamping contact area are partitioned angularly using polar coordinates, and the number of angular partitions is set to [number]. Each frame is assigned to a corresponding angular partition based on the encoder angle, and the partition number is determined by the formula:

[0038] ;

[0039] in Indicates the angular partition number and its value range is This indicates the encoder angle corresponding to that frame of the image. Indicates the reference angle defined by the zero-index pulse. Represents the angular width of each angular partition and satisfies This represents the modulo operation. This indicates the floor function;

[0040] The system divides the edge top chamfer, edge vertex, bottom chamfer, back outer ring, and each clamping contact area into five radial zones in polar coordinates and numbers them according to the same angular direction. Through, the normalized radius of the five types of zones in the radial direction with the wafer center as the origin. Expression, in which Corresponding to the outermost edge of the wafer, the five types of band regions correspond to pre-cured radius ranges and boundary values ​​are given in the device calibration file, thereby ensuring that each angular partition has a definite sampling position on the five types of band regions;

[0041] Registration of multi-view images is achieved using viewpoint mapping parameters obtained from offline calibration. During calibration, a calibration wafer with known circumferential boundaries is used, and images from two cameras are acquired in a static state. A least-squares fitting method is used to obtain a mapping lookup table from pixel coordinates to wafer polar coordinates for each camera. The mapping lookup table is formatted with the pixel coordinates of each camera. The mapping relationship is stored and the pixel resampling is completed by directly looking up the table at runtime, thereby uniformly mapping edge images from different viewpoints to the same... Coordinate system registration is achieved;

[0042] The unfolding process achieves consistent sampling using the encoder angle window corresponding to each angular partition. Set the encoder angle window to The controller selects synchronized frames from two cameras within this window and resamples them into partitioned patches of the same resolution according to five types of radial zones. Then, it performs brightness normalization on the two-view partitioned patches and fuses them according to zones, writing the results into the annular partitioned state diagram. The angular range of the clamping contact area is determined by the geometric mounting angle of the clamping mechanism and the encoder zero-position reference angle. The fixed conversion is marked as clamping and occlusion channels in the annular partition state diagram for direct use in subsequent steps. Finally, an annular partition state diagram with angular partitions as the main index, five types of radial zones as sub-indexes, and synchronously associated process signals is obtained for subsequent few-sample anomaly segmentation and state fusion.

[0043] In this specific embodiment, S2 includes:

[0044] The circular partition state diagram is denoted as... ,in Numbering from corner to section Organization and And each corner partition Includes the edge top chamfer, edge vertex, bottom chamfer, back outer ring, and clamping contact area corresponding to the strip area blocks and their channel annotation information;

[0045] The few-shot anomaly segmentation model is based on a pre-trained visual feature extraction network. with few sample anomaly segmentation head Composition, in which This is a combination of a four-level convolutional backbone network and a feature pyramid structure, with the output strides of the four levels of features being as follows: 32 pixels, and then horizontally connected to unify the number of feature pyramid channels. The parameter set is denoted as The pre-training process used self-supervised comparative learning of normal rinsing annular partition state images acquired by the same device under the same optical configuration. The pre-training data set consisted of 5000 annular partition state images, and the training epochs were set to 100 epochs, with only updates performed. To obtain multi-scale characterization of edge texture, liquid film reflection and clamping occlusion boundary stability;

[0046] Few-sample anomaly segmentation head The parameter set is denoted as Its input is Output multi-scale features and partition them by angle Independent decoding involves upsampling multi-scale features to the same resolution as the partitioned map tiles during the decoding process, followed by concatenation, and then using three sets of... Convolutional layers and a set The convolutional classification layer outputs three types of pixel-level log-probability maps, which correspond to particulate contamination, drug residue, and residual film, respectively.

[0047] To meet the requirements of the few-shot training mechanism, the system establishes a partitioned benchmark feature library based on normal washing images. ,in Diagonal partition numbering To index and store normal feature prototype vectors and feature dispersion in five categories of zones, the normal feature prototype vectors are obtained by pixel averaging of multi-scale features of corresponding partitions and zones in the normal dataset and stored separately at each scale. The feature dispersion is obtained by the channel variance of features within the same set and used for subsequent metric normalization.

[0048] Subsequently, a small number of abnormal images labeled with particulate contamination, drug residue, and residual film were used to analyze the regional benchmark feature library. Metric learning correction was performed, with the amount of few-shot outlier data set at 60 circular partition state maps, each with pixel-level annotations, and the maps were frozen during correction. Only one set of affine correction parameters shared by band area is updated, so that the corrected normal prototype and abnormal pixel features can be separated in the metric space, and the corrected feature distribution is used as... The input prior is used to train the few-sample outlier segmentation head;

[0049] The training of the few-shot outlier segmentation head uses the following total loss function. ,in This indicates that during the training phase, each angular partition... Total loss, This represents the pixel-level cross-entropy loss for three categories: particulate contamination, drug residue, and residual film, as well as the background category. This represents the Dice loss for three types of masks to suppress class imbalance. The term represents a triplet metric for learning loss, constructed using the corrected normal prototype as the anchor point and the features of anomalously labeled pixels as positive samples and the features of normal pixels as negative samples. The weights represent the cross-entropy loss. Indicates the Dice loss weights. This represents the triplet loss weights; during training, partitioned image patches with a batch size of 16 are batched and processed. Using the Adam optimizer, the learning rate is set to... Furthermore, the training rounds are set to 50 epochs, and the circular partition state graph is used during inference. enter and Get each angular partition The particulate contamination mask, the drug residue mask, and the residual film mask are used together as the contamination mask output for the partition and then used for state fusion in the subsequent step S3.

[0050] In this specific embodiment, S3 includes:

[0051] The particulate contamination mask, drug residue mask, and residual liquid film mask are aligned with the process signal at the partition level and the residual liquid thickness, backsplash probability, and clamping and blocking compensation coefficient of each partition are output through the partition state fusion network.

[0052] The alignment objects include the angular partition numbers. and Particulate pollution mask Drug residue mask Residual liquid film mask and the flow signal recorded by the synchronous sampling module. Pressure signal Negative pressure signal Spindle current signal Harmony and acoustic signals ;

[0053] Time alignment is based on the partition timestamp written to the circular partition state diagram in step S1. The system aligns each angular partition... Take the sampling time corresponding to the center of the encoder angle window of the partition as the partition alignment time. and will exist The value is resampled to a partition scalar using linear interpolation. At the same time, the three types of masks in the same partition are arranged according to the partition number. The direct index is combined with the aforementioned partition scalars to form a partition fusion input;

[0054] Acoustic signals The original microphone waveform was obtained by bandpass filtering and root mean square calculation. The bandpass frequency was set to 8kHz to 40kHz, and the root mean square window length was set to 5ms, so that the short-time energy change caused by splash impact could be stably characterized on the partition scale.

[0055] The partitioned state fusion network is denoted as And the parameter set is denoted as Its structure consists of mask branches and signal branches. The mask branches will... Superimposed as a three-channel tensor and uniformly resampled as After the pixels are input, two convolutional encoder layers are used. The first convolutional kernel is... And the number of output channels is 16, and the second convolutional kernel is... Furthermore, the number of output channels is 32, and ReLU activation is used between the two convolutional layers, and global average pooling is used after the second layer to obtain a 32-dimensional mask feature vector;

[0056] Signal branch will Normalization to The interval is then input into a two-layer fully connected network to obtain a 16-dimensional signal feature vector. The output dimensions of the two fully connected layers are 32 and 16 respectively, and both are activated by ReLU.

[0057] The mask feature vector and the signal feature vector are concatenated to form a 48-dimensional fused feature, which is then input into the three-head regression and classification output layer. The first output head is the residual liquid thickness output head, which outputs the estimated value of the residual liquid thickness in that region. The supervisory labels are derived from the calibration results of the edge liquid film thickness measurement station for the same angular partition and are written into the training data in micrometers.

[0058] The second output head is the clamping and occlusion compensation coefficient output head and outputs... Defined as the ratio of the effective washable area of ​​the clamping contact zone to the nominal washable area, and the value range is [value missing]. The supervision label is calculated by the clamping occlusion projection area determined by the CAD geometry of the clamping mechanism and the camera calibration mapping, and written into the training data along with the partition number.

[0059] The third output head is a splash probability output head and outputs... The backsplash probability is determined based on the statistical frequency of backsplash droplets returning to the wafer surface. This statistical frequency is achieved by counting backsplash droplet hit events in the aligned partitioned image sequence. The system performs this count at the partition alignment time. Corresponding continuity Within a frame partition map, inter-frame difference and connected component filtering are performed to detect bright droplets that are "newly appearing and lasting for no more than 3 frames, with a connected component area between 4 and 80 pixels, and located in a non-clamping occluded channel region" hitting the connected component and counting them as a hit event, thereby obtaining the hit count. And calculate ,in Indicates angular partitioning The estimated splash probability, This indicates the number of hit events where splashed droplets return to the wafer surface within the statistics window. This indicates the number of partition tile frames included in the statistics within the statistics window. Indicates the angular partition number;

[0060] network The training uses a fixed ( Input and corresponding ( ) tag pairs, where This is the calibration value for liquid film thickness. The occlusion compensation coefficient calibration value is obtained from geometric projection. To determine the statistical frequency calibration value obtained according to the above hit counting rules, the training is completed and the value is output in real time for each angular partition during online operation. and This information is used for risk prediction, screening, and closed-loop control in subsequent steps S4 and S5.

[0061] In this specific embodiment, S4 includes:

[0062] Estimate the residual liquid thickness in each angular zone Splash probability estimate and clamping occlusion compensation coefficient The candidate execution parameters are input into the digital twin model to predict the risk at the next moment and to screen candidate execution parameters under preset process constraints;

[0063] The set of candidate execution parameters is denoted as Each candidate execution parameter vector is denoted as ,and From nozzle flow Spray angle wafer rotation speed Edge auxiliary nozzle start / stop and the negative pressure drainage control values ​​for each angular zone Composition, angular partition numbering is and The candidate set is generated within each control round using a defined discrete combination rule, and the nozzle flow rate set is fixed. The set of spray angles is fixed as follows The wafer rotation speed set is fixed. The edge auxiliary nozzle start / stop assembly is fixed as follows: The negative pressure drainage control values ​​for each angular zone are normalized scalars with a fixed set of values. ,in This indicates that the negative pressure drainage in this zone is closed. This indicates that the negative pressure drainage of the zone reaches the set percentage corresponding to the maximum available negative pressure capacity of the equipment, and the normalization corresponds one-to-one with the opening degree of the equipment's negative pressure regulating valve and is fixed by the equipment calibration table;

[0064] To ensure that the size of the candidate set is computable within the control rounds, the system performs a certain number of tests on each group. Only two defined negative pressure drainage distribution patterns are configured. The first is a uniform pattern, which satisfies the following for all angular zones. The second type is the risk-focused model, which applies to models that meet certain conditions. Located at the top of all partition residual liquid thickness estimates or Located at the top of all partition splash probability estimates Partition settings Configure the remaining partitions And when the partition belongs to the clamping contact area channel, the partition's The value was forcibly set to 0.8 to compensate for poor drainage caused by clamping obstruction, resulting in... The candidate size is Group;

[0065] Digital twin model is denoted as And the parameter set is denoted as Its input consists of the current partition state and candidate execution parameters. The current partition state is concatenated into a tensor along the angular dimension and contains... Three channels, candidate execution parameters are broadcast in the angular dimension to the input channels consistent with the partition and include and ;

[0066] The model structure consists of a angular one-dimensional convolutional encoder and a time recursion unit. The angular one-dimensional convolutional encoder uses two layers of one-dimensional convolution to characterize the spraying and splashing coupling between adjacent angular partitions. The first convolutional layer has a kernel length of 5 and 32 output channels, and the second convolutional layer has a kernel length of 5 and 64 output channels. The time recursion unit uses a single gated loop unit with a hidden dimension of 64. The control wheel time step is fixed at 20ms, and three types of prediction results are output during recursion: the predicted residual liquid thickness of each angular partition at the next time step. Heat value at the splash point Watermark risk value ,in Indicates in candidate execution parameters Action of the lower corner partition The predicted value of the residual liquid thickness at the next moment, in micrometers. Indicates in candidate execution parameters Action of the lower corner partition The normalized heat value of the expected landing point density of the splashed droplets and the range of values ​​is: Indicates in candidate execution parameters Action of the lower corner partition The normalized risk value for watermark formation risk and its range is [missing value]. ,and The training data consists of 12,000 control wheel samples collected by the same model of equipment under mass production conditions. The sample labels are obtained by offline edge residual liquid thickness measurement, cavity back splash landing point statistics and water mark detection after drying, and are written into the training set in alignment with the angular partition number.

[0067] After obtaining each candidate execution parameter After obtaining the prediction results, the system performs screening under preset process constraints, including the upper limit of wafer rotation speed. Surface shear threshold Negative pressure capacity threshold Clamping stability threshold and liquid volume threshold , of which surface shear Based on the fixed shear lookup table model The query retrieves and uses the peak shear rate of the edge region as the judgment value, along with the negative pressure capacity occupancy value. Negative pressure drainage control values ​​from each corner to the zone The maximum capacity percentage occupied by the entire zone is obtained by converting the pump capacity calibration curve and used as the judgment value, along with stability indicators. Based on the clamping stability lookup table model The query returned the result and was used to retrieve the result. Normalized representation of liquid volume per unit time From nozzle flow Calculated in milliliters per second;

[0068] The selection rule retains candidate execution parameters that satisfy all constraints as the source of candidate parameters for subsequent control steps and outputs them along with their predicted risk results. The retained set is denoted as... And from the formula:

[0069] ;

[0070] in This represents the set of candidate execution parameters retained after filtering. This represents the set of candidate execution parameters before filtering. Indicates the first A candidate execution parameter vector, Indicates the first One candidate wafer rotation speed, Indicates the upper limit of wafer rotation speed. Indicates the first One candidate surface shear value. Indicates the surface shear threshold. Indicates the first The negative pressure capacity utilization value of each candidate This indicates the threshold for negative pressure capability. Indicates the first One candidate clamping stability index, This indicates the clamping stability threshold. Indicates the first The volume of liquid used per unit time for each candidate This indicates the threshold for liquid volume.

[0071] In this specific embodiment, S5 includes:

[0072] Based on the filtered candidate execution parameter set The corresponding prediction results are subject to constrained reinforcement learning closed-loop control, and local re-examination and conservative action control are performed when the identification uncertainty is triggered. At the same time, the time-series residual liquid memory vector is incorporated into the state vector to improve control stability.

[0073] The system will assign each candidate execution parameter vector Predicted residual liquid thickness for each angular partition at the next moment Heat value at the splash point Watermark risk value First, candidate-level aggregation is performed to construct the state vector. The aggregation rule is to aggregate the same candidate... The maximum value of all partitions is taken as one of the three risk characteristics and denoted as . and ,in Indicates candidate Predicted maximum residual liquid thickness in the all-angle zone under action. Indicates candidate The maximum splash point heat value of the full-angle zone under the action, Indicates candidate The maximum risk value of watermarks in the full-angle zone under the action;

[0074] The system will include all candidates Concatenate the candidate risk matrices according to their candidate indices and then construct state vectors. ,in Indicates the current control wheel number;

[0075] To meet the requirements of time series modeling, the system addresses each angular partition. and Maintaining the time-series residual fluid memory vector The timing length is fixed. The rotation should fall within the range of 3 to 10 revolutions. By this partition in the most recent The internal transfer is obtained from the already executed action. and After performing a first-in-first-out update, the sliding mean is calculated, thus enabling... It also carries the residual liquid evolution trend, splash accumulation trend and water mark accumulation trend of this zone;

[0076] The system will store the time-series residual memory vectors of all partitions. The clamping occlusion compensation coefficient obtained in step S3 Concatenate them together into the extended state vector and denote them as The extended state vector serves as the sole input to the constrained reinforcement learning controller;

[0077] The constrained reinforcement learning controller uses a discrete action set solution and is consistent with claim 6. As a set of actions, each action corresponds to a set of nozzle flow rates to be executed. Spray angle wafer rotation speed Edge auxiliary nozzle start / stop and negative pressure drainage control values ​​for each angular zone ;

[0078] The controller internally uses a policy evaluation network based on a DeepSets structure to encode the candidate risk matrix. The encoder encodes each candidate triplet. The algorithm maps the candidate embedding vectors to 16-dimensional candidate embedding vectors using two fully connected layers, and then performs max pooling on all candidates to obtain the candidate set representation vector. Simultaneously, it merges the candidate set representation vector with the representation vectors of all partitions. The cascaded input has two output branches; the first branch outputs the reward estimate for each candidate. The objectives are to reduce residual liquid, reduce backsplash, and reduce watermarks. Depend on By weight The weighted summation and negativeing ​​yields a result that maximizes reward while minimizing risk. The second branch outputs the constraint cost estimate for each candidate. And with the goal of meeting process constraints, among which The deviations consist of rotational speed deviation, surface shear deviation, negative pressure deviation, clamping stability deviation, and liquid volume deviation. Each deviation is determined as described in claim 6 by the difference between the corresponding candidate value and the corresponding threshold value, and then normalized and weighted. We get the result by weighted summation;

[0079] The controller maintains the cumulative constraint cost and sets the preset constraint budget. In each control round, the remaining budget is calculated based on the accumulated costs. Then, select action vectors from the action set, and the selection rules are given by the following formula:

[0080] ;

[0081] in This represents the action vector output by the current control wheel. Represents the first action in the action set. A candidate execution parameter vector, This represents the set of candidate execution parameters retained after S4 filtering. Indicates candidate The estimated value of the constraint cost, This represents the remaining constraint budget available for the current control wheel. Indicates candidate The estimated return This means that the constrained Lagrange multipliers are adjusted in steps of 0.01 during online operation based on whether the cumulative constraint cost approaches the budget, thereby achieving closed-loop constraint satisfaction in constrained reinforcement learning.

[0082] Offline training of the controller in the digital twin model The simulation environment is designed to complete and cover the full flushing cycle of the same equipment. During training, each control wheel uses a time step of 20ms and the cumulative constraint cost of the control wheels does not exceed a certain limit. The objective is a hard constraint, which allows the above selection rules to be used directly to output action vectors on the discrete candidate set during the online inference stage;

[0083] To meet the uncertainty safety mechanism, the few-sample anomaly segmentation model simultaneously outputs the identification uncertainty for each angular partition in step S2. ,in The partitioning probability entropy and the confidence difference of the partitioning boundary are calculated together and normalized to... The system sets an uncertainty threshold. When a partition satisfies Immediately trigger a local re-examination of the corresponding partition. The local re-examination involves shrinking the sampling windows of the two edge cameras to the encoder angle window of that partition and re-acquiring local multi-view images with a higher exposure time over three consecutive revolutions. At the same time, freeze the pre-trained visual feature extraction network of step S2 and perform forward inference only once for that partition to recalculate. and Then, the updated mask and uncertainty are written back to the annular partition state diagram, and steps S3 and S4 are re-executed within the current control wheel to update. And the corresponding predicted risk results;

[0084] While triggering a partial re-check, the system switches to a preset conservative parameter set to narrow the range of complete candidate execution parameters for the current control wheel. The conservative parameter set fixes the upper limit of the nozzle flow rate at [value missing]. The upper limit of the wafer rotation speed is fixed at 1900 rpm, and the lower limit of the negative pressure drainage control value corresponding to the trigger zone is fixed at 0.5, so that the candidate set generated by S4 only contains candidates that meet the above shrinkage conditions in the control wheel;

[0085] Simultaneously, the constrained reinforcement learning controller generates the complete action vector of the current control wheel within the shrinkage candidate set and limits the action vector components of the triggering partition to conservative action components selected from the conservative action set. The conservative action set limits the negative pressure drainage control value of the partition to... Furthermore, the start / stop of the edge auxiliary nozzles is limited to "on" to improve recovery capacity and reduce the risk of residual liquid retention in the clamping and blocking area;

[0086] When action vector After generation, the controller sets the nozzle flow rate within it. The data is sent to the mass flow controller to stabilize the flow rate in a closed loop, and the injection angle is adjusted. The control is sent to the nozzle servo mechanism to achieve angle positioning and control the wafer rotation speed. The command is sent to the spindle drive to achieve a speed closed loop, enabling the edge auxiliary nozzles to start and stop. The control is sent to the solenoid valve to complete the start / stop control, and the negative pressure drainage control values ​​for each angular zone are also set. The system sends data to the segmented annular drainage negative pressure module valve array, which corresponds one-to-one with the angular partitions, to generate negative pressure drainage settings for each partition and execute corresponding flushing and recycling. This achieves closed-loop anti-fouling flushing control for the edge area, the outer ring on the back, and the clamping contact area while meeting the preset process constraint budget.

[0087] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0088] This invention organically combines multi-view imaging recognition, contamination and residual liquid status assessment, digital twin risk prediction, and constrained reinforcement learning control to form a continuous technology chain from perception and prediction to execution in the rinsing process of wafer edges and clamping areas, thereby more effectively suppressing residual liquid retention and secondary contamination from splashing back.

[0089] This invention improves the control accuracy of high-risk zones and clamping / obstruction areas through candidate execution parameter screening, local re-inspection, and conservative action control mechanisms, making the technical solution more suitable for high-cleanliness rinsing scenarios at the edges of semiconductor wafers.

Claims

1. A method for preventing contamination at the edge of a wafer, characterized in that, include: S1. Acquire multi-view edge images, encoder angles, and flow rate, pressure, negative pressure, spindle current, and acoustic signals synchronized with the high-speed rotating wafer. Expand the wafer edge region and clamping contact area into a ring-shaped partition state map based on the encoder angle. S2. Input the ring-shaped partition state map into the trained few-sample anomaly segmentation model to obtain the contamination mask and residual liquid film mask for each partition. S3. Fuse the contamination mask, residual liquid film mask, and flow rate, pressure, negative pressure, spindle current, and acoustic signals to obtain the residual liquid thickness, splash probability, and clamping obstruction compensation coefficient for each partition. S4. Input the residual liquid thickness, backsplash probability, clamping and shielding compensation coefficient and candidate execution parameters into the digital twin model. Under the preset process constraints, predict the residual liquid state and risk results of each partition at the next moment for each candidate execution parameter, and select the candidate execution parameters that meet the preset process constraints. S5. Input the risk results and preset process constraints corresponding to the selected candidate execution parameters into the constrained reinforcement learning controller. Select and output the action vectors from the selected candidate execution parameters, including nozzle flow rate, spray angle, wafer rotation speed, edge auxiliary nozzle start / stop and negative pressure drainage control values ​​of each zone. Control the adjustable angle nozzle module and the segmented annular drainage negative pressure module to perform the corresponding flushing and recycling.

2. The wafer edge anti-fouling rinsing method according to claim 1, characterized in that, S1 includes: The edge top chamfer, edge vertex, bottom chamfer, back outer ring, and each clamping contact area are divided into multiple angular partitions according to polar coordinates. The edge images from at least two viewpoints are registered and expanded according to the encoder angle window corresponding to each angular partition to generate the annular partition state diagram.

3. The wafer edge anti-fouling rinsing method according to claim 1, characterized in that, S2 includes: extracting multi-scale features of the annular partition state map using a pre-trained visual feature extraction network, inputting the multi-scale features into a few-sample anomaly segmentation head for partition anomaly segmentation, and outputting the contamination mask and the residual liquid mask based on the segmentation results, wherein the contamination mask includes a particulate contamination mask and a drug residue mask.

4. The wafer edge anti-fouling rinsing method according to claim 3, characterized in that, The contamination mask includes a particulate contamination mask and a liquid residue mask. S3 includes: aligning the particulate contamination mask, the liquid residue mask, the residual liquid film mask with the flow rate, pressure, negative pressure, spindle current, and acoustic signal according to the partition number; and outputting the residual liquid thickness, the backsplash probability, and the clamping and shielding compensation coefficient through a partition state fusion network, wherein the backsplash probability is determined based on the statistical frequency of backsplash droplets returning to the wafer surface.

5. The wafer edge anti-fouling rinsing method according to claim 1, characterized in that, S4 includes: The residual liquid thickness, the backsplash probability, the clamping and shielding compensation coefficient, and the candidate execution parameters consisting of nozzle flow rate, spray angle, wafer rotation speed, edge auxiliary nozzle start / stop, and negative pressure drainage control values ​​for each zone are input into the digital twin model. The predicted residual liquid thickness, backsplash landing point heat value, and water mark risk value for each zone at the next moment are calculated under the action of each candidate execution parameter. Candidate execution parameters that do not meet the wafer rotation speed upper limit, surface shear threshold, negative pressure capability threshold, clamping stability threshold, and liquid consumption threshold are eliminated. The retained candidate execution parameters are used as the source of candidate parameters for subsequent control steps.

6. The wafer edge anti-fouling rinsing method according to claim 5, characterized in that, S5 includes: The predicted value of residual liquid thickness at the next moment corresponding to the retained candidate execution parameters, the heat value of the splash landing point, and the watermark risk value are constructed into a state vector. The retained candidate execution parameters are constructed into an action set. The rotation speed deviation, surface shear deviation, negative pressure deviation, clamping stability deviation, and liquid consumption deviation are determined according to the wafer rotation speed, surface shear amount, negative pressure capacity occupancy value determined by the negative pressure drainage control value of each partition, clamping stability index, and the difference between the liquid consumption per unit time and the corresponding wafer rotation speed upper limit, surface shear threshold, negative pressure capacity threshold, clamping stability threshold, and liquid consumption threshold. The preset constraint budget is the upper limit set for the cumulative constraint cost corresponding to each of the above deviations. The constrained reinforcement learning controller selects the action vector from the action set that ensures the cumulative constraint cost does not exceed the preset constraint budget, and controls the adjustable angle nozzle module and the segmented annular drainage negative pressure module according to the action vector, wherein the negative pressure drainage control value of each partition in the action vector is used to determine the negative pressure drainage setting value corresponding to each partition.

7. The wafer edge anti-fouling rinsing method according to claim 3, characterized in that, The training process of the few-shot anomaly segmentation model includes: establishing a partitioned benchmark feature library based on normal rinsing images; using few-shot anomaly images with particle contamination, drug residue, and residual film annotations to perform metric learning correction on the partitioned benchmark feature library; and using the corrected feature distribution to train the few-shot anomaly segmentation head.

8. The wafer edge anti-fouling rinsing method according to claim 6, characterized in that, The contamination mask includes a particulate contamination mask and a drug residue mask. The few-sample anomaly segmentation model outputs the identification uncertainty of each partition simultaneously with the particulate contamination mask, the drug residue mask, and the residual film mask. The identification uncertainty is determined based on the segmentation probability entropy and the segmentation boundary confidence difference. When the identification uncertainty exceeds a preset uncertainty threshold, a local re-examination of the corresponding partition is triggered. This local re-examination involves re-acquiring local multi-view images of the corresponding partition and recalculating the particulate contamination mask, the drug residue mask, the residual film mask, and the identification uncertainty. The updated calculation results are then written back to the annular segmentation model. The system generates a state diagram of the zone and, based on the updated particulate contamination mask, the residual drug mask, and the residual liquid film mask, redetermines the residual liquid thickness, backsplash probability, and clamping and shielding compensation coefficient of the corresponding zone within the current control wheel. It also switches the upper limit of nozzle flow rate, the upper limit of wafer rotation speed, and the lower limit of the corresponding zone's negative pressure drainage control value to a preset conservative parameter set. This preset conservative parameter set is used to narrow the complete candidate execution parameter range of the current control wheel. The constrained reinforcement learning controller generates the complete action vector of the current control wheel within the narrowed candidate execution parameter range and limits the action vector components of the corresponding zone to conservative action components selected from the conservative action set.

9. A wafer edge anti-fouling rinsing method according to claim 6, characterized in that, The state vector also includes a temporal residual liquid memory vector established for each angular partition. The temporal residual liquid memory vector is updated based on the predicted residual liquid thickness, splash landing heat value, and watermark risk value for 3 to 10 consecutive rotations in the same angular partition. The temporal residual liquid memory vector and the clamping occlusion compensation coefficient together constitute an extended state vector and are input to the constrained reinforcement learning controller to determine the action vector.

10. A wafer edge anti-fouling rinsing system, used to perform the wafer edge anti-fouling rinsing method according to any one of claims 1 to 6, characterized in that, include: The synchronous sampling module is used to acquire multi-view edge images, encoder angles, and flow, pressure, negative pressure, spindle current, and acoustic signals synchronized with the high-speed rotating wafer, and to generate a ring partition state diagram. The anomaly segmentation module is used to output the contamination mask and residual liquid mask for each partition; The status assessment module is used to output the residual liquid thickness, backsplash probability, and clamping and shielding compensation coefficient of each zone based on the contamination mask, the residual liquid film mask, and the flow rate, pressure, negative pressure, spindle current, and acoustic signals; the twin prediction and screening module is used to predict the residual liquid status and risk results corresponding to each candidate execution parameter based on the digital twin model, and screen the candidate execution parameters that meet the preset process constraints. The constrained control module is used to input the selected candidate execution parameters and their corresponding prediction results into the constrained reinforcement learning controller, and control the adjustable angle nozzle module and the segmented annular drainage negative pressure module to perform corresponding flushing and recycling according to the solution results of the constrained reinforcement learning controller.

Images