Surface uniformity processing method and system for semiconductor components

Abstract

The application discloses a surface consistency processing method and system of semiconductor parts, and relates to the technical field of semiconductor manufacturing. The method comprises the following steps: uploading the thickness of the polishing pad, the groove depth, the pH value of the grinding fluid and the abrasive concentration, synchronously collecting the part surface features, and setting the polishing state atlas; generating a compensation vector based on the atlas combined with the accuracy requirement; dynamically adjusting the standard polishing path according to the compensation vector, determining the area unit composed of M target and N adjacent coordinates; deploying a digital twin simulation pre-rehearsal compensation scene, combining reinforcement learning to establish a parameter adaptive mechanism, and optimizing the regional resource allocation. The technical problem of local uneven polishing and poor surface consistency caused by traditional fixed path polishing in the process of precise polishing of semiconductor parts is solved, the technical effects of realizing adaptive adjustment of regional machining accuracy and optimization of surface consistency, improving the uniformity of part surface morphology and polishing efficiency are achieved through dynamic polishing compensation based on digital twin and reinforcement learning.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and more specifically to a method and system for surface uniformity treatment of semiconductor parts. Background Technology

[0002] As semiconductor devices become increasingly integrated and chip sizes shrink, higher demands are placed on the flatness and uniformity of wafer and semiconductor component surfaces. In semiconductor manufacturing, chemical mechanical polishing (CMP) is a key process for achieving global planarization, and its quality directly impacts the accuracy and yield of subsequent processes such as photolithography, deposition, and etching. Existing CMP methods largely rely on fixed parameter settings and empirical models, failing to adequately respond to dynamic changes in parameters such as polishing pad thickness, surface trench depth, slurry pH, and abrasive concentration. This leads to over- or under-polishing in localized areas, resulting in poor surface uniformity and significant morphological deviations. Furthermore, traditional polishing path planning is typically statically preset, failing to incorporate real-time monitoring data for adaptive adjustment of the polishing trajectory. When the polishing pad wears or the slurry performance deteriorates, the system lacks an effective feedback compensation mechanism, making it difficult to guarantee the timeliness and repeatability of polishing accuracy. Summary of the Invention

[0003] This application provides a method and system for surface uniformity processing of semiconductor parts, which solves the technical problem of uneven polishing and poor surface uniformity caused by traditional fixed-path polishing in the precision polishing process of semiconductor parts.

[0004] A first aspect of this application provides a surface uniformity treatment method for semiconductor parts, the method comprising:

[0005] Upload the polishing pad thickness and surface trench depth corresponding to the mechanical polishing parameters, along with the abrasive concentration and pH value of the polishing slurry corresponding to the mechanical polishing parameters. Simultaneously collect the surface features of the semiconductor parts and set a polishing processing state map. Based on the polishing processing state map and the polishing accuracy requirements of the semiconductor parts, generate a polishing compensation vector. According to the polishing compensation vector, dynamically adjust the standard polishing path corresponding to the surface treatment controller to determine M target polishing coordinates and N adjacent polishing coordinates under the constraints of the part polishing process, forming multiple regional polishing units, where M is less than or equal to N / 3. At the same time, deploy a digital twin simulation engine to pre-simulate the polishing compensation scenario, and establish an adaptive mechanism for polishing parameters by combining reinforcement learning to optimize and adjust the process parameter combinations of the chemical mechanical polishing process within the multiple regional polishing units.

[0006] A second aspect of this application provides a surface conformation processing system for semiconductor parts, the system comprising:

[0007] Data Acquisition Module: Uploads data on polishing pad thickness and surface trench depth corresponding to mechanical polishing parameters, as well as pH value and abrasive concentration data of the polishing slurry corresponding to the mechanical polishing parameters. Simultaneously, it collects surface features of the semiconductor parts and sets a polishing processing state map. Vector Generation Module: Based on the polishing processing state map and the polishing precision requirements of the semiconductor parts, it generates a polishing compensation vector. Path Adjustment Module: According to the polishing compensation vector, it dynamically adjusts the standard polishing path corresponding to the surface treatment controller, determining M target polishing coordinates and N adjacent polishing coordinates under the constraints of the part polishing process, forming multiple regional polishing units, where M is less than or equal to N / 3. Parameter Optimization Module: Simultaneously, it deploys a digital twin simulation engine to pre-simulate the polishing compensation scenario, and establishes an adaptive polishing parameter mechanism using reinforcement learning to optimize and adjust the process parameter combinations for the chemical mechanical polishing process within the multiple regional polishing units.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] First, mechanical polishing-related data such as polishing pad thickness, surface trench depth, polishing slurry pH value, and abrasive concentration are collected, and surface feature information of the parts is acquired simultaneously to construct a polishing state map. Then, based on this state map and accuracy requirements, a polishing compensation vector is generated to dynamically adjust the original standard polishing path, dividing the area into multiple polishing units composed of target coordinates and neighboring coordinates to achieve more precise local control. Simultaneously, a digital twin simulation engine is used to pre-simulate the compensation scenario, and a reinforcement learning algorithm is introduced to establish an adaptive adjustment mechanism for polishing parameters. This allows for optimal configuration of process parameters in different regions, improving the accuracy and surface consistency of the polishing process and adapting to the demanding process environment of semiconductor manufacturing. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic flowchart of a surface consistency processing method for semiconductor parts provided in an embodiment of this application.

[0012] Figure 2 This is a schematic diagram of the surface uniformity processing system for semiconductor components provided in an embodiment of this application.

[0013] Figure labeling: Data acquisition module 1, vector generation module 2, path adjustment module 3, parameter optimization module 4. Detailed Implementation

[0014] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0015] Example 1, as Figure 1 As shown, this application provides a surface uniformity treatment method for semiconductor parts, the method including:

[0016] Upload the polishing pad thickness and surface groove depth corresponding to the mechanical polishing parameters, as well as the pH value and abrasive concentration data of the polishing slurry corresponding to the mechanical polishing parameters. Simultaneously collect the surface characteristics of the semiconductor parts and set the polishing processing status map.

[0017] In this embodiment, process data corresponding to the mechanical polishing parameters is first acquired in real time from the polishing equipment, including the initial thickness of the polishing pad and the surface trench depth parameters, to reflect the wear state and surface morphology characteristics of the polishing pad. Simultaneously, chemical characteristic data of the polishing slurry matching the mechanical polishing parameters are collected, including the pH value, abrasive concentration, flow rate, and temperature of the slurry, to reflect the intensity of the chemical effect during the polishing process. During data acquisition, high-precision surface detection devices, such as white light interferometers or laser confocal microscopes, are used to simultaneously acquire feature information of the semiconductor component surface, including surface roughness distribution, flatness error, and micro-trench morphology. The acquired multi-source heterogeneous data are synchronized with time stamps and spatial coordinates to form the surface features of the semiconductor component. Then, based on the obtained mechanical polishing parameters, polishing slurry parameters, and semiconductor component surface features, a polishing processing state map is established. This polishing processing state map uses a time series as the main line, mapping the physical and chemical states during the polishing process in multiple dimensions to reflect the overall operating state of the current polishing process, providing basic data support for the subsequent generation of polishing compensation vectors and path optimization.

[0018] Based on the polishing process state map, a polishing compensation vector is generated in combination with the polishing accuracy requirements of semiconductor parts.

[0019] In one embodiment, after obtaining the polishing process state map, based on multi-dimensional data recorded in the polishing process state map, such as polishing pad thickness, surface trench depth, polishing slurry pH value, abrasive concentration, and surface feature distribution, and combined with the target polishing accuracy requirements of the semiconductor component, the deviation value between the current polishing state and the target accuracy is calculated. Subsequently, this deviation value is input into a pre-established multi-parameter coupling model. This multi-parameter coupling model is constructed from a deep neural network and has been iteratively trained using historical deviation data and historical polishing compensation data, employing steps such as forward propagation, loss calculation, backpropagation, and parameter optimization. Afterward, the multi-parameter coupling model analyzes and calculates the main factors affecting the surface removal rate based on the received deviation data, generating a set of compensation parameters including polishing pressure, polishing time, polishing slurry ratio, polishing pad rotation speed, and feed rate. By splicing these compensation parameters according to a preset vector template, a polishing compensation vector is formed. This polishing compensation vector is used to dynamically correct the polishing process in different areas, so that each process parameter is adjusted in real time according to the local surface condition, thereby realizing adaptive feedback control of the polishing process, ensuring that semiconductor parts achieve the required surface consistency and processing accuracy at both the overall and local scales, and improving the product yield and polishing uniformity.

[0020] Based on the polishing compensation vector, the standard polishing path corresponding to the surface treatment controller is dynamically adjusted to determine M target polishing coordinates and N adjacent polishing coordinates under the constraints of the part polishing process, forming multiple regional polishing units, where M is less than or equal to N / 3.

[0021] In one embodiment, based on the generated polishing compensation vector, the system dynamically adjusts the preset standard polishing path in the surface treatment controller to achieve refined regional control of the semiconductor component surface. Specifically, the polishing compensation vector, including parameters such as polishing pressure, polishing time, slurry ratio, rotation speed, and feed rate, is first combined with the polishing process state map. An attention mechanism is used to analyze the polishing characteristics of each region under the current process conditions. Based on the analysis results, M key location points are aggregated on the standard polishing path as target polishing coordinates. Simultaneously, N neighboring auxiliary coordinate points are automatically selected around each target polishing coordinate to construct local polishing units containing spatial correlation features. Each regional polishing unit consists of a target polishing coordinate and several neighboring coordinates around it. The value of M does not exceed one-third of N to ensure that each regional unit has sufficient spatial sampling density, thereby achieving comprehensive analysis of local surface characteristics. Through the above dynamic path adjustment mechanism, the foundation for subsequent multi-region adaptive polishing control on a global scale can be laid to improve the surface consistency of complex curved surfaces or non-homogeneous material parts and extend the service life of polishing pads.

[0022] Furthermore, the method further includes dynamically adjusting the standard polishing path corresponding to the surface treatment controller.

[0023] Based on the mechanical polishing parameters and polishing pad wear time series data, as well as the mechanical polishing parameters and polishing slurry parameter change data, cross-dimensional polishing features are extracted through an attention mechanism; based on a federated learning framework, the M target polishing coordinates located on the standard polishing path are aggregated, and the surface accuracy compensation parameters are dynamically optimized in combination with the cross-dimensional polishing features.

[0024] Preferably, to achieve dynamic optimization of polishing accuracy, the model is based on mechanical polishing parameters and polishing pad wear data over time, combined with data on the variation of polishing slurry parameters. This data typically includes multi-dimensional data such as polishing pressure, rotation speed, polishing pad thickness variation sequence, polishing slurry pH value, and abrasive concentration. Subsequently, these multi-dimensional data are aligned to the same scale using a maximum-minimum normalization method. The aligned multi-dimensional data is then input into an attention mechanism model. This model uses a feature encoding layer to vectorize the sequence corresponding to each parameter, and then uses an attention weight calculation module, such as a multi-head self-attention mechanism, to calculate the correlation of parameters in each dimension, identifying the key parameters that have the greatest impact on surface quality changes in the current time window and assigning them higher weights. After completing the attention weight allocation, the attention mechanism model fuses the dynamic dependencies between different parameter dimensions, generating a cross-dimensional polishing feature vector with time sensitivity and cross-dimensional coupling. This cross-dimensional polishing feature vector not only includes the variation trend of each individual parameter but also reflects the synergistic effect and nonlinear influence law between parameters. After completing cross-dimensional feature extraction, multiple points on the part surface are uniformly or non-uniformly sampled according to a preset standard polishing path. Each sampling point collects mechanical polishing parameters, polishing slurry parameters, and surface feature information, forming an original polishing path point dataset with temporal features and spatial coordinates. Then, based on the process line or equipment layout, the standard polishing path data is distributed to multiple federated learning nodes, such as different production batches or different equipment control units. Each node independently performs preliminary feature aggregation and coordinate filtering operations on the path points locally. That is, it calculates the similarity of the cross-dimensional polishing features extracted by the attention mechanism and performs local aggregation of path points based on similarity to form a set of candidate target coordinates. Then, through the central node, candidate coordinates with similar features are federated and merged using federated averaging or parameter-weighted fusion to obtain the M target polishing coordinates with the strongest feature representation. Then, the M target polishing coordinates and cross-dimensional polishing features are input into the polishing compensation optimization model. This polishing compensation optimization model can be constructed using a long short-term memory network to dynamically calculate the surface accuracy compensation parameters of each region, such as real-time correction of polishing pressure distribution, adjustment of polishing time window, and optimization of polishing slurry ratio, thereby continuously improving the flatness, consistency and processing controllability of semiconductor parts surface.

[0025] Furthermore, the methods include:

[0026] A dynamic polishing pressure compensation strategy is set up to generate pressure adjustment commands based on the real-time thickness of the polishing pad and the depth of the surface trenches. Combined with LSTM to predict the trend of surface accuracy changes during the polishing process, the M target polishing coordinates are deployed. Based on the M target polishing coordinates, the pH value of the polishing slurry and the calibration parameters of the abrasive concentration are processed in parallel to ensure that the dynamic compensation response time meets the time limit of semiconductor polishing accuracy.

[0027] Optionally, to ensure precise control of the polishing process based on real-time operating conditions, a dynamic polishing pressure compensation strategy is first established. This strategy continuously monitors the instantaneous thickness and surface groove depth changes of the polishing pad through sensors on the polishing equipment, encapsulates the obtained real-time pad thickness and surface groove depth, and generates corresponding pressure adjustment commands. Subsequently, a time-series prediction model based on a Long Short-Term Memory (LSTM) network is constructed using historical compensation data to dynamically predict the surface accuracy change trend during polishing. This time-series prediction model identifies potential surface morphology deviation trends in advance through real-time data in the pressure adjustment commands, proactively adjusts the polishing strategy before predicting the risk of accuracy degradation, and automatically deploys M target polishing coordinates as key compensation areas, pre-adjusting the pressure, rotation speed, and polishing time parameters of these areas. During the polishing execution phase, chemical parameters related to the polishing slurry, including slurry pH value and abrasive concentration calibration, are processed in parallel based on the M target polishing coordinates. By monitoring the chemical state of the polishing slurry in real time and using LSTM to automatically correct the slurry ratio and supply rate based on feedback from polishing temperature rise, flow rate, and material removal rate, the chemical reaction intensity is matched with the mechanical action, and pressure adjustment and chemical parameter calibration are carried out simultaneously, thereby meeting the timeliness requirements of semiconductor polishing precision.

[0028] Furthermore, the methods include:

[0029] The polishing precision weight is determined by connecting to the semiconductor component process database, and a zoned polishing reference curve is formulated by combining the performance data of the polishing equipment. Based on the zoned polishing reference curve, a blockchain storage unit is integrated to record the polishing compensation parameter update log, and the zoned polishing unit is mapped and associated with the blockchain storage unit.

[0030] Optionally, the process first accesses a semiconductor component process database to retrieve historical process data related to the current component model, material, structural characteristics, and target surface precision, including information such as target surface roughness, flatness requirements, and key geometric region weights. Based on these process parameters, a normalized weighted average is used to calculate the polishing precision weight of each region during the polishing process; that is, the sensitivity index of different regions to the final product performance. Subsequently, real-time performance data of the polishing equipment is retrieved, including polishing head load capacity, polishing pad dynamic response characteristics, spindle speed range, and upper limit of polishing slurry flow rate. A nonlinear fitting method is used to partition the component surface spatially. For each region, the least squares method is used to fit the corresponding polishing precision weight and equipment performance data, solving for the regression coefficients to minimize the sum of squared errors between the actual data and the fitted curve, thereby generating a corresponding partitioned polishing reference curve as a standard baseline for subsequent dynamic compensation and process control. After the reference curve is established, a blockchain-based notarization unit is introduced to ensure the immutable recording of polishing compensation parameters. Whenever the compensation parameters of a regional polishing unit change, a log message is automatically generated containing a timestamp, regional identifier, parameter content, and operation source. This message is then hash-encrypted and written to the blockchain storage system. This process ensures the traceability and tamper-proof nature of the compensation parameters throughout the entire processing cycle. Simultaneously, a one-to-one mapping relationship is established between the regional polishing units and the blockchain storage units. Each regional unit corresponds to a unique blockchain record node, facilitating rapid retrieval of the compensation history and actual execution status of a specific region during subsequent product quality traceability, process optimization, or anomaly analysis. Through this process, not only is regional precision control of the polishing process achieved, but a reliable parameter traceability system is also established using blockchain technology, providing technical support for process transparency, quality control, and continuous optimization in semiconductor manufacturing.

[0031] Simultaneously, a digital twin simulation engine is deployed to pre-simulate polishing compensation scenarios, and a polishing parameter adaptive mechanism is established by combining reinforcement learning to optimize and adjust the process parameter combinations of the chemical mechanical polishing process in the multiple region polishing units.

[0032] In one embodiment, to achieve intelligent prediction and adaptive control of the polishing process, a digital twin simulation engine is first deployed. This engine synchronizes bidirectionally with the actual polishing equipment via a data interface, receiving real-time operational data from sensors, such as polishing pad thickness, surface stress, material removal rate, slurry pH, and abrasive concentration. A high-fidelity polishing compensation scenario, consistent with the physical system, is established in a virtual environment to simulate the surface removal behavior of each polishing unit under different pressures, times, speeds, and chemical ratios. Simultaneously, a reinforcement learning control framework is introduced, treating each polishing unit as an independent agent. Through a defined function, optimization targets are set for indicators such as surface consistency, polishing efficiency, and polishing pad life. The reinforcement learning algorithm continuously receives feedback from the digital twin simulation results, automatically adjusting control parameters such as polishing pressure, rotation speed, slurry flow rate, and ratio, iteratively training to determine the optimal strategy under different operating conditions. After multiple rounds of simulation training, multiple sets of stable adaptive process parameters are generated, each corresponding to the chemical mechanical polishing process of a polishing unit. When process deviations or equipment status changes are detected in the production environment, the parameters can be automatically corrected based on these adaptive process parameters, dynamically optimizing the process configuration within each polishing unit to ensure that the surface consistency of semiconductor parts remains in the optimal state over the long term.

[0033] Furthermore, the method involves deploying a digital twin simulation engine to pre-simulate polishing compensation scenarios and establishing an adaptive mechanism for polishing parameters using reinforcement learning.

[0034] In the polishing compensation scenario, a limited optimization space is set by the thickness threshold under the polishing pad replacement cycle, the upper limit of the number of times the polishing slurry can be recycled, and the surface accuracy loss limit; based on the limited optimization space, a multi-objective optimization solution is performed with the objectives of surface consistency, polishing efficiency, and polishing pad life as objectives.

[0035] Optionally, in the deployed polishing compensation scenario, key process boundary conditions are set based on the equipment operation records and material characteristics of each polishing pad replacement cycle. These include a polishing pad thickness threshold, an upper limit for the number of polishing slurry cycles, and a surface accuracy loss limit. The polishing pad thickness threshold is a set lower limit; when the polishing pad thickness is below this value, pressure compensation is no longer increased or polishing time extended to avoid process failure or equipment damage. The upper limit for the number of polishing slurry cycles is the set maximum number of cycles; exceeding this maximum value automatically triggers replacement or slurry replenishment to avoid affecting the uniformity of surface removal. The surface accuracy loss limit is the set maximum allowable surface error tolerance range; exceeding this range automatically abandons the current strategy and initiates a new optimization iteration. By setting these constraints, a parameter optimization space can be defined. Subsequently, within this limited optimization space, the optimization problem is formalized into a multi-objective function, with surface consistency, polishing efficiency, and polishing pad life as optimization objectives. That is, minimizing error, shortening processing time, and minimizing equipment consumption are the goals. The digital twin simulation engine simulates the process response behavior under different parameter combinations within the optimization space and automatically explores the optimal combination of multi-dimensional parameters such as pressure, time, rotational speed, and polishing slurry ratio through reinforcement learning algorithms. During the optimization process, a batch of initial parameter combinations is generated based on the polishing pad thickness threshold, the upper limit of polishing slurry usage, and surface precision limitations. These combinations include process elements such as pressure, time, rotational speed, polishing slurry flow rate, and ratio. Then, the digital twin simulation engine performs simulation calculations on each set of parameters to obtain corresponding performance indicators such as surface consistency, processing efficiency, and polishing pad life. By weighted summing of these indicators, the merits of all initial parameter combinations can be compared, with better-performing combinations being classified as frontier solutions and poorer-performing ones being gradually eliminated. Simultaneously, diversity is maintained among the frontier solutions to prevent getting trapped in local optima. Then, through evolutionary steps such as selection, crossover, and mutation, new parameter combinations are continuously generated, and the simulation and selection process is repeated. After multiple iterations, a set of optimal solutions is obtained, representing the best balance between different objectives. Finally, based on actual process priorities, such as prioritizing surface consistency or extending polishing pad life, the most suitable strategy is automatically selected from the optimal solution set and sent to the polishing controller for execution in real time. This achieves adaptive optimization control of parameters, providing a highly stable and reliable process control foundation for semiconductor manufacturing.

[0036] Furthermore, the methods also include:

[0037] Based on the generative adversarial network to simulate the interference mode of the polishing environment, the acquisition control parameters of the semiconductor part surface features are adaptively optimized; a digital twin simulation engine is used to correlate historical fault data with the acquisition control parameters to conduct a surface treatment risk assessment.

[0038] Optionally, to improve the adaptability and risk warning capability of the polishing process to external interference, common environmental interference data in actual production are first collected, such as fluctuations in polishing slurry temperature, humidity changes, airborne impurity content, equipment vibration, and unstable slurry flow. Then, based on this interference data, a GAN (Generative Adversarial Network) is used to construct an interference simulation model to simulate common polishing environmental interference patterns in actual production. The generator of the interference simulation model generates environmental disturbance signals that conform to the real distribution, while the discriminator evaluates the similarity between the generated signals and actual historical data. After multiple rounds of adversarial training, the model can simulate various complex disturbance scenarios with high accuracy. After obtaining the interference simulation model, it is integrated with a digital twin simulation engine for the polishing process. When disturbances such as temperature and humidity fluctuations occur in the simulated environment, the digital twin simulation engine adaptively adjusts the surface feature acquisition control parameters according to the dynamic changes in the interference effect, such as increasing the sampling frequency, increasing the density of monitoring points in specific areas, or changing the optical measurement mode, thereby ensuring that high-precision surface feature data can still be obtained under interference conditions. Finally, a risk assessment model is established by combining the digital twin simulation engine with a historical fault database. This model uses the relationship between past fault records and corresponding collected parameters to perform real-time risk scoring for potential surface anomalies, processing failures, or quality deviations under the current disturbance situation. When the risk exceeds the set risk threshold, a risk warning instruction is automatically issued, providing reliable process assurance for high-precision semiconductor surface processing.

[0039] Furthermore, performing weight fusion based on the lightweight routing network on the identification matrix group includes:

[0040] The N neighboring polishing coordinates are distributed around the M target polishing coordinates; taking the M target polishing coordinates as the core, and combining the N neighboring polishing coordinates, a regional consistency assessment is performed to correct the confidence interval of the partitioned polishing reference curve; through the corrected confidence interval, the surface characteristics of the semiconductor part are compared with a preset risk threshold, and if it exceeds the threshold, a risk warning instruction is issued, and the risk assessment result is fed back to the surface treatment controller.

[0041] Optionally, to achieve high-precision regional control and risk warning in the polishing process, M target polishing coordinates are first set as the core points for regional polishing compensation. Then, based on a preset spatial neighborhood range, N neighboring polishing coordinates are automatically selected around the target polishing coordinates to form a spatially correlated regional polishing unit. The N neighboring coordinates are evenly distributed or distributed according to gradient density around the M target points to comprehensively reflect the actual processing state and surface characteristics of the target area. Subsequently, with the M target coordinates as the center, combined with the real-time measurement data of the N neighboring coordinates, a consistency assessment of the material removal rate, pressure distribution, and surface roughness changes within the area is performed through weighted statistics to calculate the regional consistency index. Based on these regional consistency indices, the confidence interval of the original partitioned polishing reference curve is dynamically corrected. If the regional consistency index between the neighboring coordinates and the target coordinates is detected to exceed the index range, the confidence interval is tightened according to the difference in deviation to improve the sensitivity of process control in that area; conversely, the confidence interval is widened to improve process stability and tolerance. Next, the corrected confidence interval and surface feature data of the semiconductor components are input into the risk assessment model for analysis, classifying the current risk type and scoring the risk in real time. When the real-time risk score exceeds a preset risk threshold, a risk alert instruction is immediately generated, including the risk type, area coordinates, and suggested compensation measures. Simultaneously, the risk assessment results are fed back to the surface treatment controller in the form of a data packet. The controller adjusts parameters such as pressure, path, and polishing slurry ratio accordingly, implementing real-time compensation and process correction to ensure that the surface quality of the components meets high-precision manufacturing requirements.

[0042] Furthermore, an adaptive polishing parameter mechanism is established to optimize and adjust the combination of process parameters for the chemical mechanical polishing process within the multiple region polishing units. The method also includes:

[0043] Based on the energy efficiency ratio of the surface treatment controller, and combined with the wear progress of the polishing pad and the polishing process of the parts, the execution frequency of surface feature detection in the polishing units of the multiple regions is dynamically adjusted; at the same time, the surface consistency of the parts after being processed by the polishing units of the multiple regions is monitored, and the polishing parameter adaptive mechanism is updated.

[0044] Preferably, to improve the detection efficiency and energy efficiency matching during the polishing process, energy consumption and efficiency data of the surface treatment controller under different polishing conditions are first collected, and the energy efficiency ratio is calculated, i.e., the amount of material removed or the improvement in accuracy corresponding to unit energy consumption. Simultaneously, the sensors corresponding to the polishing pad record parameters such as polishing pad thickness change, groove depth, and wear rate in real time. Combined with the polishing process stage of the part, such as rough polishing, fine polishing, and final polishing, a set of energy efficiency curves matching the working conditions is generated. Subsequently, based on these energy efficiency curves and the polishing process, the execution frequency of surface feature detection is dynamically scheduled. For example, in the rough polishing stage, the material removal rate is high but the accuracy requirement is relatively relaxed, so the detection frequency can be reduced to save energy consumption and detection resources; in the fine polishing or final polishing stage, the surface error control requirements are higher, so the detection frequency and sampling density are increased, and high-resolution monitoring is performed on key areas to ensure that surface consistency meets process standards. Meanwhile, the surface consistency of multiple polishing units is continuously monitored. When the regional consistency index still does not meet the preset risk threshold, the reinforcement learning control framework involved in the polishing parameter adaptive mechanism will be iteratively optimized, and the process parameters will be re-optimized to enhance the precision stability and overall process efficiency of the polishing process.

[0045] Furthermore, the method for monitoring the surface consistency of the parts after processing by the multiple area polishing units and updating the adaptive polishing parameter mechanism includes:

[0046] By integrating polishing pressure compensation and polishing slurry ratio correction, the generation of surface treatment process parameter templates for semiconductor parts, including silicon wafers and silicon carbide, is supported. The polishing pad-polishing slurry combination ratio is optimized in the multiple area polishing units.

[0047] Optionally, to adapt to the precision polishing requirements of semiconductor components made of different materials, such as silicon wafers and silicon carbide, initial process parameter ranges are first set based on the physical properties of different material components recorded in the process database, such as hardness, brittleness, thermal conductivity, and surface chemical reactivity. These parameters include polishing pressure, slurry pH, abrasive concentration, flow rate, and polishing pad type. Subsequently, a polishing response model for the corresponding material is established in a digital twin simulation engine to simulate the material removal rate, surface roughness evolution, and thermal stress distribution under different process combinations, providing a reference for subsequent parameter optimization. During the actual polishing process, each polishing unit area is monitored in real time, collecting dynamic information including polishing pressure feedback signals, surface feature detection data, and slurry state parameters. When a deviation in material removal rate or surface error is detected in a region exceeding the set range, the aforementioned polishing pressure compensation mechanism is triggered, appropriately adjusting the polishing pressure distribution for that region to balance the material removal rate. Simultaneously, based on changes in the slurry state, a deep neural network is used for ratio correction to automatically adjust the abrasive concentration, pH, or flow rate to maintain the stability and uniformity of the chemical reaction. Building upon this foundation, the control logic for pressure compensation and polishing slurry ratio correction is integrated. A reinforcement learning algorithm is used to match the characteristics of the polishing pad with the polishing slurry formulation. For example, for softer materials like silicon wafers, a combination of low pressure and medium concentration polishing slurry is preferred to ensure flatness and reduce damage. Conversely, for high-hardness materials like silicon carbide, a combination of higher pressure and higher concentration abrasive is favored to improve removal rate and maintain uniformity. After multiple rounds of iterative calculations and feedback optimization, optimized polishing pad-slurry combination ratios for different materials and operating conditions are finalized and solidified into corresponding process parameter templates. These templates can be quickly invoked in different production batches and equipment, reducing manual debugging time and ensuring processing accuracy and consistency. Through this process, process adaptation for various semiconductor materials is achieved, improving the flexibility, adaptability, and process stability of the polishing process.

[0048] In summary, the embodiments of this application have at least the following technical effects:

[0049] First, the polishing pad thickness and surface trench depth corresponding to the mechanical polishing parameters, along with the pH value of the polishing slurry and the abrasive concentration data corresponding to the mechanical polishing parameters, are uploaded. Simultaneously, surface features of the semiconductor component are collected, and a polishing processing state map is set. Next, based on the polishing processing state map and the polishing precision requirements of the semiconductor component, a polishing compensation vector is generated. Then, according to the polishing compensation vector, the standard polishing path corresponding to the surface treatment controller is dynamically adjusted to determine M target polishing coordinates and N adjacent polishing coordinates under the constraints of the component polishing process, forming multiple regional polishing units, where M is less than or equal to N / 3. Finally, a digital twin simulation engine is deployed to pre-simulate the polishing compensation scenario, and a polishing parameter adaptive mechanism is established using reinforcement learning to optimize and adjust the process parameter combinations for the chemical mechanical polishing process within the multiple regional polishing units. This solves the technical problem of uneven local polishing and poor surface consistency caused by traditional fixed-path polishing in the precision polishing process of semiconductor components. It achieves the technical effect of adaptive adjustment of regional processing precision and optimization of surface consistency through dynamic polishing compensation based on digital twins and reinforcement learning, thereby improving the uniformity of component surface morphology and polishing efficiency.

[0050] Example 2, based on the same inventive concept as the surface consistency processing method for semiconductor parts in the foregoing examples, such as... Figure 2 As shown, this application provides a surface uniformity processing system for semiconductor components, the system comprising:

[0051] Data Acquisition Module 1: Uploads data on polishing pad thickness and surface trench depth corresponding to mechanical polishing parameters, as well as pH value and abrasive concentration data of the polishing slurry corresponding to the mechanical polishing parameters, and synchronously collects surface features of semiconductor parts to set a polishing processing state map; Vector Generation Module 2: Generates polishing compensation vectors based on the polishing processing state map and the polishing accuracy requirements of semiconductor parts; Path Adjustment Module 3: Dynamically adjusts the standard polishing path corresponding to the surface treatment controller according to the polishing compensation vectors, determines M target polishing coordinates and N adjacent polishing coordinates under the constraints of the part polishing process, and forms multiple regional polishing units, where M is less than or equal to N / 3; Parameter Optimization Module 4: Simultaneously, deploys a digital twin simulation engine to pre-simulate the polishing compensation scenario, establishes an adaptive mechanism for polishing parameters by combining reinforcement learning, and optimizes and adjusts the process parameter combinations of the chemical mechanical polishing process within the multiple regional polishing units.

[0052] Furthermore, the path adjustment module 3 is used to perform the following method:

[0053] Based on the mechanical polishing parameters and polishing pad wear time series data, as well as the mechanical polishing parameters and polishing slurry parameter change data, cross-dimensional polishing features are extracted through an attention mechanism; based on a federated learning framework, the M target polishing coordinates located on the standard polishing path are aggregated, and the surface accuracy compensation parameters are dynamically optimized in combination with the cross-dimensional polishing features.

[0054] Furthermore, the path adjustment module 3 is used to perform the following method:

[0055] A dynamic polishing pressure compensation strategy is set up to generate pressure adjustment commands based on the real-time thickness of the polishing pad and the depth of the surface trenches. Combined with LSTM to predict the trend of surface accuracy changes during the polishing process, the M target polishing coordinates are deployed. Based on the M target polishing coordinates, the pH value of the polishing slurry and the calibration parameters of the abrasive concentration are processed in parallel to ensure that the dynamic compensation response time meets the time limit of semiconductor polishing accuracy.

[0056] Furthermore, the path adjustment module 3 is used to perform the following method:

[0057] The polishing precision weight is determined by connecting to the semiconductor component process database, and a zoned polishing reference curve is formulated by combining the performance data of the polishing equipment. Based on the zoned polishing reference curve, a blockchain storage unit is integrated to record the polishing compensation parameter update log, and the zoned polishing unit is mapped and associated with the blockchain storage unit.

[0058] Furthermore, the parameter optimization module 4 is used to perform the following method:

[0059] In the polishing compensation scenario, a limited optimization space is set by the thickness threshold under the polishing pad replacement cycle, the upper limit of the number of times the polishing slurry can be recycled, and the surface accuracy loss limit; based on the limited optimization space, a multi-objective optimization solution is performed with the objectives of surface consistency, polishing efficiency, and polishing pad life as objectives.

[0060] Furthermore, the parameter optimization module 4 is used to perform the following method:

[0061] Based on the generative adversarial network to simulate the interference mode of the polishing environment, the acquisition control parameters of the semiconductor part surface features are adaptively optimized; a digital twin simulation engine is used to correlate historical fault data with the acquisition control parameters to conduct a surface treatment risk assessment.

[0062] Furthermore, the parameter optimization module 4 is used to perform the following method:

[0063] The N neighboring polishing coordinates are distributed around the M target polishing coordinates; taking the M target polishing coordinates as the core, and combining the N neighboring polishing coordinates, a regional consistency assessment is performed to correct the confidence interval of the partitioned polishing reference curve; through the corrected confidence interval, the surface characteristics of the semiconductor part are compared with a preset risk threshold, and if it exceeds the threshold, a risk warning instruction is issued, and the risk assessment result is fed back to the surface treatment controller.

[0064] Furthermore, the parameter optimization module 4 is used to perform the following method:

[0065] Based on the energy efficiency ratio of the surface treatment controller, and combined with the wear progress of the polishing pad and the polishing process of the parts, the execution frequency of surface feature detection in the polishing units of the multiple regions is dynamically adjusted; at the same time, the surface consistency of the parts after being processed by the polishing units of the multiple regions is monitored, and the polishing parameter adaptive mechanism is updated.

[0066] Furthermore, the parameter optimization module 4 is used to perform the following method:

[0067] By integrating polishing pressure compensation and polishing slurry ratio correction, the generation of surface treatment process parameter templates for semiconductor parts, including silicon wafers and silicon carbide, is supported. The polishing pad-polishing slurry combination ratio is optimized in the multiple area polishing units.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for surface uniformity treatment of semiconductor components, characterized in that, The method includes: Upload the polishing pad thickness and surface trench depth corresponding to the mechanical polishing parameters, as well as the pH value and abrasive concentration data of the polishing slurry corresponding to the mechanical polishing parameters, and simultaneously collect the surface characteristics of semiconductor parts to set the polishing processing status map. Based on the polishing process state map, a polishing compensation vector is generated in combination with the polishing accuracy requirements of semiconductor components. Based on the polishing compensation vector, the standard polishing path corresponding to the surface treatment controller is dynamically adjusted to determine M target polishing coordinates and N adjacent polishing coordinates under the constraints of the part polishing process, forming multiple area polishing units, where M is less than or equal to N / 3. Simultaneously, a digital twin simulation engine is deployed to pre-simulate polishing compensation scenarios, and a polishing parameter adaptive mechanism is established by combining reinforcement learning to optimize and adjust the combination of process parameters for chemical mechanical polishing processes in the multiple region polishing units; The method further includes dynamically adjusting the standard polishing path corresponding to the surface treatment controller. Based on the mechanical polishing parameters and polishing pad wear time series data, as well as the mechanical polishing parameters and polishing slurry parameter change data, cross-dimensional polishing features are extracted through an attention mechanism. Based on the federated learning framework, the M target polishing coordinates located on the standard polishing path are aggregated, and the surface accuracy compensation parameters are dynamically optimized by combining the cross-dimensional polishing features. The method includes: A dynamic polishing pressure compensation strategy is set up to generate pressure adjustment commands based on the real-time thickness of the polishing pad and the depth of the surface grooves. Combined with LSTM to predict the trend of surface accuracy changes during the polishing process, the M target polishing coordinates are deployed. Based on the M target polishing coordinates, the pH value of the polishing slurry and the calibration parameters of the abrasive concentration are processed in parallel to ensure that the dynamic compensation response time meets the time limit of semiconductor polishing accuracy. The method includes: Connect to the semiconductor component process database to determine the polishing precision weights, and combine the performance data of the polishing equipment to formulate a zoned polishing reference curve; Based on the partitioned polishing reference curve, an integrated blockchain evidence storage unit is used to record the polishing compensation parameter update log, and the regional polishing unit is mapped and associated with the blockchain evidence storage unit.

2. The surface uniformity treatment method for semiconductor parts as described in claim 1, characterized in that, Deploying a digital twin simulation engine to pre-simulate polishing compensation scenarios, and combining reinforcement learning to establish an adaptive mechanism for polishing parameters, the method includes: In the polishing compensation scenario, the optimization space is limited by setting the thickness threshold under the polishing pad replacement cycle, the upper limit of the number of times the polishing slurry can be recycled, and the surface accuracy loss limit. Based on the defined optimization space, a multi-objective optimization solution is performed with the objectives of surface consistency, polishing efficiency, and polishing pad life as the goals.

3. The surface uniformity treatment method for semiconductor parts as described in claim 2, characterized in that, The method further includes: Based on the simulation of polishing environment interference modes using generative adversarial networks, the acquisition and control parameters of the semiconductor component surface features are adaptively optimized. A digital twin simulation engine is used to correlate historical fault data with the acquired control parameters to conduct a surface treatment risk assessment.

4. The surface uniformity treatment method for semiconductor parts as described in claim 3, characterized in that, The N neighboring polishing coordinates are distributed around the M target polishing coordinates; Using the M target polishing coordinates as the core, and combining the N neighboring polishing coordinates, a regional consistency assessment is performed to correct the confidence interval of the partitioned polishing reference curve; By comparing the corrected confidence interval with the surface features of the semiconductor component and a preset risk threshold, a risk warning instruction is issued if the threshold is exceeded, and the risk assessment result is fed back to the surface treatment controller.

5. The surface uniformity treatment method for semiconductor parts as described in claim 1, characterized in that, The method further includes establishing an adaptive polishing parameter mechanism to optimize and adjust the combination of process parameters for chemical mechanical polishing processes within the multiple region polishing units. Based on the energy efficiency ratio of the surface treatment controller, and combined with the wear progress of the polishing pad and the polishing process of the parts, the execution frequency of the surface feature detection polishing unit in the multiple areas is dynamically adjusted. Simultaneously, the consistency of the part surface after processing by the multiple area polishing units is monitored, and the polishing parameter adaptive mechanism is updated accordingly.

6. The surface uniformity treatment method for semiconductor parts as described in claim 5, characterized in that, The method for monitoring the surface consistency of parts after processing by the multiple polishing units and updating the adaptive polishing parameter mechanism includes: By integrating polishing pressure compensation and polishing slurry ratio correction, the generation of surface treatment process parameter templates for semiconductor parts, including silicon wafers and silicon carbide, is supported. The polishing pad-polishing slurry combination ratio is optimized in the multiple area polishing units.

7. A surface uniformity treatment system for semiconductor components, characterized in that, The system is used to implement the surface uniformity treatment method for semiconductor parts according to any one of claims 1-6, the system comprising: Data acquisition module: Uploads polishing pad thickness and surface trench depth corresponding to mechanical polishing parameters, as well as grinding slurry pH value and abrasive concentration data corresponding to mechanical polishing parameters, synchronously collects surface features of semiconductor parts, and sets polishing processing status map; Vector generation module: Based on the polishing process state map, and combined with the polishing accuracy requirements of semiconductor parts, a polishing compensation vector is generated; Path adjustment module: Based on the polishing compensation vector, dynamically adjust the standard polishing path corresponding to the surface treatment controller, determine M target polishing coordinates and N adjacent polishing coordinates under the limit of the part polishing process, and form multiple area polishing units, where M is less than or equal to N / 3; Parameter optimization module: Simultaneously, a digital twin simulation engine is deployed to pre-simulate the polishing compensation scenario, and a polishing parameter adaptive mechanism is established by combining reinforcement learning to optimize and adjust the process parameter combination of the chemical mechanical polishing process in the multiple region polishing units.

Images