Intelligently-regulated industrial robot semiconductor cleaning system

Abstract

The invention belongs to the technical field of information, and particularly relates to an impurity targeted cleaning algorithm focused on multispectral detection and intelligent identification, which collects wafer impurity data through a multispectral detection sensor, identifies impurity categories by using CNN, and specifically adjusts cleaning parameters, so that the impurity residual rate is reduced from 30% to less than 5%, the cleaning effect is remarkably improved, and the cleaning efficiency is improved. And the high-cleanliness requirement is met. The cleaning efficiency improvement algorithm based on multi-objective optimization and reinforcement learning optimizes parameters by using the cleaning efficiency, the effect and the cost as indexes by means of the deep Q network, the cleaning efficiency is improved by more than 30%, the cost is reduced by more than 20%, and the production benefit is greatly improved while the cleaning quality is ensured.

Description

Technical Field

[0001] The present invention belongs to the field of information technology, and in particular relates to an intelligently controlled industrial robot semiconductor cleaning system. Background Art

[0002] The semiconductor cleaning process is an essential step in the semiconductor manufacturing process. Its purpose is to remove various impurities from the surface of semiconductor wafers, such as particulate contaminants, organic residues, and metal ions, to ensure the smooth execution of subsequent processes and the performance and reliability of semiconductor devices. With the continuous advancement of semiconductor technology, chip integration has become increasingly higher, and the requirements for wafer surface cleanliness have become increasingly stringent. Industrial robots, due to their automation, high precision, and programmability, have been widely used in the operation and control of semiconductor cleaning equipment. However, existing industrial robot semiconductor cleaning technology still has some urgent issues that need to be addressed, seriously hindering the further development of semiconductor cleaning processes.

[0003] Cleaning results struggle to meet high cleanliness requirements: Advanced semiconductor manufacturing processes place extremely high demands on wafer surface cleanliness, and any trace of impurities can impact chip performance and reliability. Traditional industrial robotic cleaning systems primarily operate based on pre-set cleaning procedures and fixed cleaning parameters, making them unable to accurately identify and remove impurities of varying types, sizes, and adhesion strengths. During the cleaning process, factors such as uneven concentration distribution of the cleaning fluid, inaccurate spray angle and pressure control of the cleaning equipment, and improper cleaning time settings can easily lead to some impurities remaining, failing to meet high cleanliness requirements and impacting the accuracy of subsequent processes and chip yield.

[0004] Inefficient cleaning: Cleaning efficiency is a significant factor affecting semiconductor manufacturing production costs and cycle times. Currently, industrial robots rely primarily on manual experience to set cleaning parameters and operating procedures during semiconductor cleaning, making it difficult to optimize the cleaning process. Furthermore, the limited speed and automation of cleaning equipment, as well as its inability to handle complex cleaning processes, result in a lengthy cleaning process and low production efficiency, making it unable to meet the demands of large-scale production. Furthermore, some cleaning processes require multiple repetitions to achieve the desired cleaning effect, further increasing cleaning time and costs.

[0005] Environmentally unfriendly and wasteful use of cleaning fluids: The semiconductor cleaning process typically requires large quantities of chemical cleaning fluids to remove impurities from the wafer surface. Existing industrial robotic cleaning systems lack effective monitoring and management of cleaning fluid usage, making it impossible to adjust the concentration and dosage of cleaning fluids in real time based on the type and degree of impurities on the wafer surface. This not only leads to significant waste of cleaning fluids and increases production costs, but also causes significant environmental pollution due to the large amounts of chemical cleaning fluids discharged into the environment, which violates environmental protection requirements. Summary of the Invention

[0006] The present invention provides an intelligently controlled industrial robot semiconductor cleaning system, comprising:

[0007] The impurity targeted cleaning algorithm module based on multispectral detection and intelligent identification obtains multispectral feature data of impurities on the wafer surface by installing a multispectral detection sensor on the cleaning equipment of the industrial robot. It uses the intelligent recognition algorithm and the intelligent adjustment mechanism of cleaning parameters to achieve targeted cleaning of different impurities. Assume that the multispectral feature data vector is

[0008] M=[m1,m2,…,mn], using convolutional neural network (CNN) for impurity identification, CNN outputs the probability distribution of impurity categories , where zCi is the impurity category of the CNN network

[0009] The output value of Ci, k is the total number of impurity categories, according to the recognition results, the cleaning liquid concentration adjustment amount ΔC and the impurity category Ci

[0010] Satisfy ΔC=f(Ci), where f is an adjustment function pre-set according to the impurity characteristics to achieve targeted cleaning;

[0011] The cleaning efficiency improvement algorithm module based on multi-objective optimization and reinforcement learning takes cleaning efficiency, cleaning effect and cleaning cost as indicators of multi-objective optimization, defines the state space, action space and reward function, and uses the reinforcement learning algorithm to achieve the improvement of cleaning efficiency while ensuring the cleaning effect and controlling the cleaning cost. In reinforcement learning, the proximal policy optimization algorithm (PPO) is adopted. The policy network πθ(a|s) outputs the probability of taking action a in state s, and the advantage function A π (s, a) = Q π (s, a)-V π (s), by maximizing the objective function Update the policy network parameters θ, where θold is the old policy network parameters and T is the time step. During the optimization process, balance the cleaning efficiency, effect and cost;

[0012] An algorithm module for environmentally friendly use of cleaning fluids based on multi-parameter monitoring and adaptive control. This module installs multiple parameter monitoring sensors in the cleaning equipment to obtain information on various parameters of the cleaning fluid. It utilizes multi-parameter monitoring technology and adaptive control algorithms to achieve environmentally friendly use of the cleaning fluid.

[0013] The cleaning process pre-planning and optimization algorithm module based on digital twin and virtual simulation establishes a digital twin model of the semiconductor cleaning system and uses virtual simulation technology to pre-plan and simulate the semiconductor cleaning process. At the same time, an equipment status assessment model is established to achieve pre-planning and optimization of the cleaning process.

[0014] The large-scale cleaning production optimization algorithm module based on multi-robot collaboration and intelligent scheduling regards multiple industrial robots in the semiconductor cleaning workshop as a multi-agent system, defines the state space, action space and reward function, and uses reinforcement learning algorithms to achieve efficient collaboration between multiple robots, thereby improving the overall efficiency and resource utilization of large-scale cleaning production.

[0015] Furthermore, in the impurity targeted cleaning algorithm module based on multispectral detection and intelligent identification, the multispectral detection sensor covers a wavelength range of 200-1100nm, with a detection accuracy of ±0.5nm. The sensor deployment position is determined by combining optical simulation and actual testing to ensure comprehensive and accurate detection of impurities on the wafer surface. The data acquisition frequency is 5Hz in the cleaning preparation stage and is increased to 20Hz in the cleaning execution stage, and is dynamically adjusted according to the cleaning process.

[0016] Furthermore, in the cleaning efficiency improvement algorithm module based on multi-objective optimization and reinforcement learning, the state space covers information such as cleaning liquid concentration, spray angle, pressure, cleaning time, temperature, etc. during the cleaning process. The action space includes adjustment actions such as increasing or decreasing the cleaning parameters by 10%-20% and adjusting ±15°. The reward function is constructed with the cleaning efficiency improvement ratio α, the cleaning effect improvement ratio β, and the cleaning cost reduction ratio γ as quantitative indicators, that is, the reward value

[0017] R=w1α+w2β+w3γ, and the hierarchical analysis method determined that w1=0.4, w2=0.3, and w3=0.3.

[0018] Furthermore, in the cleaning fluid environmentally friendly use algorithm module based on multi-parameter monitoring and adaptive control, the cleaning fluid concentration sensor has an accuracy of ±0.5%, the flow sensor has an accuracy of ±2mL / min, and the pH sensor has an accuracy of ±0.05pH. The sensors work together to provide high-frequency and precise cleaning fluid parameters for the adaptive control algorithm, and the data transmission of each sensor adopts a low-latency, high-bandwidth wireless communication protocol.

[0019] Furthermore, in the cleaning process pre-planning and optimization algorithm module based on digital twins and virtual simulation, professional modeling software such as ANSYS is used to build a digital twin model. The model update frequency is consistent with the sensor data acquisition frequency to ensure real-time synchronization between the physical entity and the virtual model. By simulating the flow field distribution, cleaning liquid injection trajectory, etc. of the cleaning process under different process parameters, potential problems are predicted in advance, and Monte Carlo simulation is used to perform uncertainty analysis on the prediction results. The number of simulations is set to 800 times.

[0020] Furthermore, in the large-scale cleaning production optimization algorithm module based on multi-robot collaboration and intelligent scheduling, the communication delay between intelligent agents is less than 10ms, the communication reliability reaches 99.9%, and real-time information interaction is achieved through a distributed communication network. When using the reinforcement learning algorithm to train the model, the experience replay buffer capacity is set to 8,000 data, and the target network update frequency is once every 80 iterations.

[0021] Furthermore, in the impurity targeted cleaning algorithm module based on multispectral detection and intelligent identification, the CNN network structure includes 3 convolutional layers, 2 pooling layers and 2 fully connected layers, the convolution kernel sizes are 3×3, 5×5, and 7×7 respectively, the pooling layer adopts maximum pooling, the window size is 2×2, and the network is trained with a large amount of labeled data for 1500 times to improve the accuracy of impurity identification.

[0022] Furthermore, in the cleaning efficiency improvement algorithm module based on multi-objective optimization and reinforcement learning, during the actual cleaning process, the industrial robot selects actions according to the real-time cleaning status and adjusts the strategy based on the reward feedback after executing the action. After learning more than 500 cleaning tasks, the cleaning efficiency is improved by at least 30%, the cleaning effect is maintained at more than 95%, and the cleaning cost is reduced by at least 20%.

[0023] Furthermore, in the algorithm module for environmentally friendly use of cleaning fluid based on multi-parameter monitoring and adaptive control, the adaptive control algorithm adopts a fuzzy control strategy to construct a fuzzy rule base based on the cleaning fluid concentration, flow rate, pH value and impurity information on the wafer surface, such as "if the cleaning fluid concentration is high and the impurities are organic matter, increase the dilution ratio of the cleaning fluid", etc. The cleaning fluid parameter adjustment amount is obtained through fuzzy reasoning and defuzzification to achieve accurate and environmentally friendly use of the cleaning fluid.

[0024] Furthermore, in the cleaning process pre-planning and optimization algorithm module based on digital twins and virtual simulation, virtual simulation technology is used to predict and analyze in advance problems such as poor cleaning effect, low cleaning efficiency, and waste of cleaning fluid in the semiconductor cleaning process. Based on the prediction results, process parameters such as the cleaning fluid injection pressure are adjusted within the range of ±15%, the cleaning time is adjusted within the range of ±20%, and the process sequence in the production process are optimized to improve the reliability and stability of the semiconductor cleaning process. The optimization process adopts a multi-objective optimization algorithm to balance the mutual influence between various process parameters. The multi-objective optimization algorithm adopts the non-dominated sorting genetic algorithm (NSGA-II).

[0025] Beneficial effects:

[0026] Significantly improve cleaning effects: The impurity-targeted cleaning algorithm based on multispectral detection and intelligent identification can accurately identify different impurities on the wafer surface and intelligently adjust cleaning parameters according to the impurity type, achieving targeted cleaning of different impurities, improving cleaning effects, meeting high-cleanliness cleaning requirements, providing a good foundation for subsequent processes, and improving chip performance and yield.

[0027] Significantly improve cleaning efficiency: The cleaning efficiency improvement algorithm based on multi-objective optimization and reinforcement learning can achieve improved cleaning efficiency through reinforcement learning and multi-objective optimization, while ensuring cleaning effects and controlling cleaning costs. Under the premise of ensuring product quality, it can increase production speed, reduce production costs and improve production efficiency.

[0028] Achieve environmentally friendly use of cleaning fluid: The environmentally friendly cleaning fluid usage algorithm based on multi-parameter monitoring and adaptive control can monitor various parameters of the cleaning fluid in real time, and adaptively adjust the concentration, dosage and recycling strategy of the cleaning fluid according to the impurity information on the wafer surface, thereby reducing the waste and emission of cleaning fluid, reducing pollution to the environment, and complying with environmental protection requirements.

[0029] Advance planning and optimization of cleaning processes: The cleaning process pre-planning and optimization algorithm based on digital twins and virtual simulation can predict and analyze possible problems in the cleaning process in advance through virtual simulation technology, optimize and adjust the parameters and processes of the cleaning process, and use the equipment status assessment model to promptly identify potential problems and optimize them, thereby improving equipment reliability and stability and reducing production interruptions.

[0030] Optimize large-scale cleaning production scheduling: The large-scale cleaning production optimization algorithm based on multi-robot collaboration and intelligent scheduling can achieve efficient collaborative work among multiple robots, dynamically adjust task allocation and production plans according to real-time production conditions and equipment status, improve the overall efficiency and resource utilization of large-scale cleaning production, and meet the needs of large-scale production. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Module flow diagram. DETAILED DESCRIPTION

[0032] Example 1

[0033] The impurity targeted cleaning algorithm module based on multispectral detection and intelligent identification obtains multispectral feature data of impurities on the wafer surface by installing a multispectral detection sensor on the cleaning equipment of the industrial robot. It uses the intelligent recognition algorithm and the intelligent adjustment mechanism of cleaning parameters to achieve targeted cleaning of different impurities. Assume that the multispectral feature data vector is

[0034] M=[m1,m2,…,mn], using convolutional neural network (CNN) for impurity identification, CNN outputs the probability distribution of impurity categories , where zCi is the impurity category of the CNN network

[0035] The output value of Ci, k is the total number of impurity categories. Based on the identification results, the cleaning solution concentration adjustment amount ΔC and the impurity category Ci satisfy ΔC = f(Ci), where f is an adjustment function pre-set according to the impurity characteristics to achieve targeted cleaning;

[0036] The cleaning efficiency improvement algorithm module based on multi-objective optimization and reinforcement learning takes cleaning efficiency, cleaning effect and cleaning cost as indicators of multi-objective optimization, defines the state space, action space and reward function, and uses the reinforcement learning algorithm to achieve the improvement of cleaning efficiency while ensuring the cleaning effect and controlling the cleaning cost. In reinforcement learning, the proximal policy optimization algorithm (PPO) is adopted. The policy network πθ(a|s) outputs the probability of taking action a in state s, and the advantage function A π (s,a)=Q π (s,a)-V π (s), by maximizing the objective function Update the policy network parameters θ, where θold is the old policy network parameters and T is the time step. During the optimization process, balance the cleaning efficiency, effect and cost;

[0037] An algorithm module for environmentally friendly use of cleaning fluids based on multi-parameter monitoring and adaptive control. This module installs multiple parameter monitoring sensors in the cleaning equipment to obtain information on various parameters of the cleaning fluid. It utilizes multi-parameter monitoring technology and adaptive control algorithms to achieve environmentally friendly use of the cleaning fluid.

[0038] The cleaning process pre-planning and optimization algorithm module based on digital twin and virtual simulation establishes a digital twin model of the semiconductor cleaning system and uses virtual simulation technology to pre-plan and simulate the semiconductor cleaning process. At the same time, an equipment status assessment model is established to achieve pre-planning and optimization of the cleaning process.

[0039] The large-scale cleaning production optimization algorithm module based on multi-robot collaboration and intelligent scheduling regards multiple industrial robots in the semiconductor cleaning workshop as a multi-agent system, defines the state space, action space and reward function, and uses reinforcement learning algorithms to achieve efficient collaboration between multiple robots, thereby improving the overall efficiency and resource utilization of large-scale cleaning production.

[0040] Implementation of impurity targeted cleaning algorithm based on multispectral detection and intelligent identification

[0041] Multispectral detection sensor deployment and data collection: Multispectral detection sensors are rationally deployed on the cleaning equipment of industrial robots to ensure that the multispectral characteristic data of impurities on the wafer surface can be accurately obtained. The sensor collects data in real time according to the set frequency.

[0042] The data is transmitted to the data processing unit through the network.

[0043] Multispectral feature data processing and impurity identification: The data processing unit uses deep learning-based intelligent recognition algorithms, such as convolutional neural networks (CNN), to analyze and process multispectral feature data, accurately identifying different types of impurities, their distribution, adhesion strength, and other information.

[0044] Intelligent Cleaning Parameter Adjustment and Execution: Based on impurity identification results, the industrial robot intelligently adjusts cleaning parameters, such as the type, concentration, spray angle, and pressure of the cleaning fluid. The robot control system controls the cleaning equipment according to these adjusted parameters. During the cleaning process, it continuously monitors data from the multispectral detection sensor in real time and dynamically adjusts cleaning parameters to ensure complete impurity removal.

[0045] Implementation of cleaning efficiency improvement algorithm based on multi-objective optimization and reinforcement learning

[0046] State space, action space, and reward function definitions: Define the state space, action space, and reward function for reinforcement learning. The state space includes various parameters and state information during the cleaning process, such as cleaning fluid concentration, spray angle, pressure, cleaning time, and temperature. The action space includes various adjustments to cleaning parameters, such as increasing or decreasing cleaning fluid concentration, adjusting spray angle and pressure, and changing cleaning time. The reward function comprehensively considers factors such as cleaning efficiency, cleaning results, and cleaning cost, and rewards or penalizes the robot's actions.

[0047] Reinforcement Learning Model Training and Application: We collect extensive cleaning experiment data and use reinforcement learning algorithms, such as the Deep Q-Network (DQN) or Proximal Policy Optimization (PPO), to train and optimize the reinforcement learning model. During the actual cleaning process, the industrial robot selects actions based on the real-time cleaning status and adjusts its strategy based on reward feedback after executing the action. This gradually optimizes the cleaning parameter adjustment strategy, improving cleaning efficiency while ensuring cleaning results and controlling cleaning costs.

[0048] Implementation of an environmentally friendly cleaning fluid usage algorithm based on multi-parameter monitoring and adaptive control

[0049] Parameter Monitoring Sensor Deployment and Data Collection: We rationally deploy cleaning fluid concentration sensors, flow sensors, pH sensors, and other sensors throughout the cleaning equipment to ensure accurate acquisition of various cleaning fluid parameter information. The sensors collect data in real time at a set frequency and transmit it to the data processing center via the network.

[0050] Multi-parameter data fusion and real-time status model construction: The data processing center uses deep learning-based multi-parameter monitoring technology to fuse data collected by different sensors and build a real-time status model of cleaning fluid usage. Deep learning algorithms, such as convolutional neural networks (CNNs), are used to analyze and process the fused data and extract key feature information.

[0051] Adaptive control algorithm implementation and cleaning fluid usage adjustment: Based on a real-time state model and impurity information on the wafer surface, the adaptive control algorithm calculates the cleaning fluid concentration, dosage, and adjustments to the recycling strategy. The robot control system adjusts the relevant cleaning fluid parameters in real time based on these adjustments. During the cleaning process, the robot continuously monitors the data from parameter monitoring sensors in real time and dynamically adjusts the cleaning fluid usage to achieve environmentally friendly use of the cleaning fluid.

[0052] Cleaning process pre-planning and optimization algorithm implementation based on digital twin and virtual simulation

[0053] Digital twin model construction: Using modeling software and sensor data, a digital twin model of the semiconductor cleaning system is established, mapping entities such as industrial robots, cleaning equipment, and wafers in the physical world into the virtual space, enabling real-time interaction and synchronization between physical entities and virtual models.

[0054] Virtual simulation and process pre-planning: Utilizing virtual simulation technology, we pre-plan and simulate semiconductor cleaning processes, predicting and analyzing potential cleaning process issues in a virtual environment. Based on the simulation results, we optimize and adjust semiconductor cleaning process parameters and procedures to develop an optimal cleaning process plan.

[0055] Equipment status monitoring and optimization: Utilizing digital twin models, the operating status of cleaning equipment is monitored and analyzed in real time. Historical operating and fault data is collected to establish an equipment status assessment model. During actual operation, this real-time monitoring data is fed into the equipment status assessment model to promptly identify potential issues and implement optimizations, thereby improving equipment reliability and stability.

[0056] Implementation of large-scale cleaning production optimization algorithm based on multi-robot collaboration and intelligent scheduling

[0057] Agent Definition and Communication Mechanism Establishment: Multiple industrial robots in a semiconductor cleaning workshop are defined as multiple agents, each with independent decision-making capabilities and goals. A communication network is established between these agents to enable real-time information exchange.

[0058] Reinforcement learning model training and application: clarify the state space, action space and reward function of reinforcement learning. Collect historical operation data of semiconductor cleaning workshops and use reinforcement learning algorithms such as deep Q network (DQN) or proximal policy optimization algorithm to train and apply reinforcement learning models.

[0059] The reinforcement learning model is trained and optimized using the PPO method. In actual production, the agent selects actions based on real-time state information and adjusts its strategy based on reward feedback after executing the action. This gradually optimizes the collaborative cleaning production scheduling strategy, enabling efficient collaboration among multiple robots and improving the overall efficiency and resource utilization of large-scale cleaning production.

[0060] Beneficial effects

[0061] Significantly improve cleaning effects: The impurity-targeted cleaning algorithm based on multispectral detection and intelligent identification can accurately identify different impurities on the wafer surface and intelligently adjust cleaning parameters according to the impurity type, achieving targeted cleaning of different impurities, improving cleaning effects, meeting high-cleanliness cleaning requirements, providing a good foundation for subsequent processes, and improving chip performance and yield.

[0062] Significantly improve cleaning efficiency: The cleaning efficiency improvement algorithm based on multi-objective optimization and reinforcement learning can achieve improved cleaning efficiency through reinforcement learning and multi-objective optimization, while ensuring cleaning effects and controlling cleaning costs. Under the premise of ensuring product quality, it can increase production speed, reduce production costs and improve production efficiency.

[0063] Achieve environmentally friendly use of cleaning fluid: The environmentally friendly cleaning fluid usage algorithm based on multi-parameter monitoring and adaptive control can monitor various parameters of the cleaning fluid in real time, and adaptively adjust the concentration, dosage and recycling strategy of the cleaning fluid according to the impurity information on the wafer surface, thereby reducing the waste and emission of cleaning fluid, reducing pollution to the environment, and complying with environmental protection requirements.

[0064] Advance planning and optimization of cleaning processes: The cleaning process pre-planning and optimization algorithm based on digital twins and virtual simulation can predict and analyze possible problems in the cleaning process in advance through virtual simulation technology, optimize and adjust the parameters and processes of the cleaning process, and use the equipment status assessment model to promptly identify potential problems and optimize them, thereby improving equipment reliability and stability and reducing production interruptions.

[0065] Optimize large-scale cleaning production scheduling: The large-scale cleaning production optimization algorithm based on multi-robot collaboration and intelligent scheduling can achieve efficient collaborative work among multiple robots, dynamically adjust task allocation and production plans according to real-time production conditions and equipment status, improve the overall efficiency and resource utilization of large-scale cleaning production, and meet the needs of large-scale production.

[0066] Example 2

[0067] Implementation of an impurity targeted cleaning algorithm based on multispectral detection and intelligent identification

[0068] A semiconductor manufacturing facility has introduced a targeted impurity cleaning algorithm based on multispectral detection and intelligent identification in the wafer cleaning process. Multispectral detection sensors are precisely deployed on the industrial robot's cleaning equipment, according to locations determined by optical simulations and actual testing. These sensors cover a wavelength range of 200-1100nm and achieve a detection accuracy of ±0.5nm.

[0069] During the cleaning preparation stage, the sensor collects data at a frequency of 5Hz. Taking the cleaning of a batch of 100 wafers as an example, the traditional cleaning process is initially used. After cleaning, it is found that about 30% of the wafer surfaces have impurities remaining, mainly mixed impurities of metal ions and organic matter. After the introduction of this algorithm, the sensor acquisition frequency is increased to 20Hz during the cleaning execution stage. The multispectral detection sensor obtains the multispectral feature data vector M of impurities on the wafer surface, which is processed by a CNN network with 3 convolutional layers, 2 pooling layers and 2 fully connected layers. The convolution kernel sizes are 3×3, 5×5, and 7×7 respectively, and the pooling layer window is 2×2. The model trained 1500 times outputs the impurity category probability distribution P(Ci|M). For example, it is identified that there are a large number of metal ion impurities on the surface of a certain wafer. According to the pre-set adjustment function f, the cleaning solution concentration adjustment amount ΔC

[0070] By increasing the acid cleaning solution concentration by 15%, adjusting the spray angle by 10°, and increasing the pressure by 20%, the company demonstrated that the residual impurity rate on wafers in this batch of wafers was reduced to less than 5%. This significantly improved cleaning performance met the high-cleanliness cleaning requirements and provided a high-quality foundation for subsequent chip manufacturing processes.

[0071] Example of cleaning efficiency improvement algorithm based on multi-objective optimization and reinforcement learning

[0072] In another semiconductor cleaning workshop, a cleaning efficiency improvement algorithm based on multi-objective optimization and reinforcement learning was applied. The initial cleaning parameters for the cleaning equipment in this workshop were: 5% cleaning solution concentration, 30° spray angle, 0.5 MPa pressure, and 10 minutes of cleaning time.

[0073] The state space of reinforcement learning includes information such as cleaning fluid concentration, spray angle, pressure, cleaning time, temperature, etc. The action space is actions such as increasing or decreasing cleaning parameters by 10%-20% and adjusting ±15°. The reward function

[0074] R=0.4α+0.3β+0.3γ

[0075] The industrial robot learns based on the Deep Q Network (DQN). In the first 100 cleaning tasks, the cleaning efficiency was low, with an average of 10 wafers cleaned per hour and a cleaning effect of about 85%, resulting in high costs. As the number of cleaning tasks increases, the robot selects actions based on the real-time cleaning status and adjusts its strategy based on reward feedback. After learning more than 500 cleaning tasks, the cleaning efficiency has increased to more than 15 wafers per hour, an increase of at least 30%, and the cleaning effect has remained above 95%. At the same time, by optimizing cleaning parameters, such as reducing the concentration of the cleaning solution to 4%, the cleaning time has been shortened to 8 minutes, and the cleaning cost has been reduced by at least 20%. While ensuring the cleaning quality, the cleaning efficiency has been greatly improved, the production cost has been reduced, and the overall production efficiency has been improved.

Claims

1. An intelligently controlled industrial robot semiconductor cleaning system, characterized in that: include: The impurity targeted cleaning algorithm module based on multispectral detection and intelligent identification obtains multispectral feature data of impurities on the wafer surface by installing a multispectral detection sensor on the cleaning equipment of the industrial robot. It uses the intelligent recognition algorithm and the intelligent adjustment mechanism of cleaning parameters to achieve targeted cleaning of different impurities. Assume that the multispectral feature data vector is M=[m1,m2,…,mn], using convolutional neural network (CNN) for impurity identification, CNN outputs the probability distribution of impurity categories , where zCi is the impurity category of the CNN network The output value of Ci, k is the total number of impurity categories. Based on the identification results, the cleaning solution concentration adjustment amount ΔC and the impurity category Ci satisfy ΔC = f(Ci), where f is an adjustment function pre-set according to the impurity characteristics to achieve targeted cleaning; The cleaning efficiency improvement algorithm module based on multi-objective optimization and reinforcement learning takes cleaning efficiency, cleaning effect and cleaning cost as indicators of multi-objective optimization, defines the state space, action space and reward function, and uses the reinforcement learning algorithm to achieve the improvement of cleaning efficiency while ensuring the cleaning effect and controlling the cleaning cost. In reinforcement learning, the proximal policy optimization algorithm (PPO) is used. The policy network πθ(a|s) outputs the probability of taking action a in state s, and the advantage function A π (s,a)=Q π (s,a)-V π (s), by maximizing the objective function Update the policy network parameters θ, where θold is the old policy network parameters and T is the time step. During the optimization process, balance the cleaning efficiency, effect and cost; An algorithm module for environmentally friendly use of cleaning fluids based on multi-parameter monitoring and adaptive control. This module installs multiple parameter monitoring sensors in the cleaning equipment to obtain information on various parameters of the cleaning fluid. It utilizes multi-parameter monitoring technology and adaptive control algorithms to achieve environmentally friendly use of the cleaning fluid. The cleaning process pre-planning and optimization algorithm module based on digital twin and virtual simulation establishes a digital twin model of the semiconductor cleaning system and uses virtual simulation technology to pre-plan and simulate the semiconductor cleaning process. At the same time, an equipment status assessment model is established to achieve pre-planning and optimization of the cleaning process. The large-scale cleaning production optimization algorithm module based on multi-robot collaboration and intelligent scheduling regards multiple industrial robots in the semiconductor cleaning workshop as a multi-agent system, defines the state space, action space and reward function, and uses reinforcement learning algorithms to achieve efficient collaboration between multiple robots, thereby improving the overall efficiency and resource utilization of large-scale cleaning production.

2. The key technology of industrial robot semiconductor cleaning based on multi-dimensional perception and intelligent control according to claim 1 is characterized in that: In the impurity targeted cleaning algorithm module based on multispectral detection and intelligent identification, the multispectral detection sensor covers a wavelength range of 200-1100nm, with a detection accuracy of ±0.5nm. The sensor deployment position is determined by combining optical simulation and actual testing to ensure comprehensive and accurate detection of impurities on the wafer surface. The data acquisition frequency is 5Hz in the cleaning preparation stage and is increased to 20Hz in the cleaning execution stage, and is dynamically adjusted according to the cleaning process.

3. The key technology of industrial robot semiconductor cleaning based on multi-dimensional perception and intelligent control according to claim 1 is characterized in that: In the cleaning efficiency improvement algorithm module based on multi-objective optimization and reinforcement learning, the state space covers information such as cleaning liquid concentration, spray angle, pressure, cleaning time, temperature, etc. during the cleaning process. The action space includes adjustment actions such as increasing or decreasing the cleaning parameters by 10%-20% and adjusting ±15°. The reward function is constructed with the cleaning efficiency improvement ratio α, the cleaning effect improvement ratio β, and the cleaning cost reduction ratio γ as quantitative indicators, that is, the reward value R=w1α+w2β+w3γ, and the hierarchical analysis method determined that w1=0.4, w2=0.3, and w3=0.

3.

4. The key technology for industrial robot semiconductor cleaning based on multi-dimensional perception and intelligent control according to claim 1 is characterized in that: In the cleaning fluid environmentally friendly use algorithm module based on multi-parameter monitoring and adaptive control, the cleaning fluid concentration sensor has an accuracy of ±0.5%, the flow sensor has an accuracy of ±2mL / min, and the pH sensor has an accuracy of ±0.05pH. The sensors work together to provide high-frequency and precise cleaning fluid parameters for the adaptive control algorithm, and the data transmission of each sensor adopts a low-latency, high-bandwidth wireless communication protocol.

5. The key technology for industrial robot semiconductor cleaning based on multi-dimensional perception and intelligent control according to claim 1 is characterized in that: In the cleaning process pre-planning and optimization algorithm module based on digital twins and virtual simulation, professional modeling software such as ANSYS is used to build a digital twin model. The model update frequency is consistent with the sensor data acquisition frequency to ensure real-time synchronization between the physical entity and the virtual model. By simulating the flow field distribution, cleaning liquid injection trajectory, etc. of the cleaning process under different process parameters, potential problems are predicted in advance, and Monte Carlo simulation is used to perform uncertainty analysis on the prediction results. The number of simulations is set to 800.

6. The key technology for industrial robot semiconductor cleaning based on multi-dimensional perception and intelligent control according to claim 1 is characterized in that: In the large-scale cleaning production optimization algorithm module based on multi-robot collaboration and intelligent scheduling, the communication delay between intelligent agents is less than 10ms, the communication reliability reaches 99.9%, and real-time information interaction is achieved through a distributed communication network. When using the reinforcement learning algorithm to train the model, the experience replay buffer capacity is set to 8,000 data points, and the target network update frequency is once every 80 iterations.

7. The key technology for industrial robot semiconductor cleaning based on multi-dimensional perception and intelligent control according to claim 1 is characterized in that: In the impurity targeted cleaning algorithm module based on multispectral detection and intelligent identification, the CNN network structure includes 3 convolutional layers, 2 pooling layers and 2 fully connected layers. The convolution kernel sizes are 3×3, 5×5, and 7×7 respectively. The pooling layer adopts maximum pooling with a window size of 2×2. The network is trained with a large amount of labeled data for 1500 times to improve the accuracy of impurity identification.

8. The key technology for industrial robot semiconductor cleaning based on multi-dimensional perception and intelligent control according to claim 1 is characterized in that: In the cleaning efficiency improvement algorithm module based on multi-objective optimization and reinforcement learning, during the actual cleaning process, the industrial robot selects actions based on the real-time cleaning status and adjusts the strategy based on the reward feedback after executing the action. After learning more than 500 cleaning tasks, the cleaning efficiency is improved by at least 30%, the cleaning effect is maintained at more than 95%, and the cleaning cost is reduced by at least 20%.

9. The key technology for industrial robot semiconductor cleaning based on multi-dimensional perception and intelligent control according to claim 1 is characterized in that: In the algorithm module for environmentally friendly use of cleaning fluid based on multi-parameter monitoring and adaptive control, the adaptive control algorithm adopts a fuzzy control strategy to construct a fuzzy rule base based on the concentration, flow rate, pH value and impurity information of the cleaning fluid and the wafer surface. For example, "if the cleaning fluid concentration is high and the impurities are organic matter, increase the dilution ratio of the cleaning fluid", etc. The cleaning fluid parameter adjustment amount is obtained through fuzzy reasoning and defuzzification to achieve accurate and environmentally friendly use of the cleaning fluid.

10. The key technology for industrial robot semiconductor cleaning based on multi-dimensional perception and intelligent control according to claim 1 is characterized in that: In the cleaning process pre-planning and optimization algorithm module based on digital twins and virtual simulation, virtual simulation technology is used to predict and analyze in advance problems such as poor cleaning effect, low cleaning efficiency, and waste of cleaning fluid in the semiconductor cleaning process. Based on the prediction results, process parameters such as the cleaning fluid injection pressure are adjusted within the range of ±15%, the cleaning time is adjusted within the range of ±20%, and the process sequence in the production process are optimized to improve the reliability and stability of the semiconductor cleaning process. The optimization process adopts a multi-objective optimization algorithm to balance the mutual influence between various process parameters. The multi-objective optimization algorithm adopts the non-dominated sorting genetic algorithm (NSGA-II).

Images