Semiconductor photoetching system based on information fusion and intelligent regulation and control
By employing technologies such as multi-sensor fusion, physical field monitoring, deep learning, and reinforcement learning, the problems of positioning accuracy, pattern transfer, and coating uniformity in industrial robot lithography systems have been solved, achieving high-precision lithography and stable production.
Patent Information
- Application Number
- CN202510864631.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-26
- Publication Date
- 2025-10-21
AI Technical Summary
Existing industrial robot lithography systems have shortcomings in lithography positioning accuracy, pattern transfer accuracy, and photoresist coating uniformity, leading to unstable chip performance and increased manufacturing costs.
A semiconductor lithography system based on information fusion and intelligent control is adopted. Through multi-sensor fusion and real-time compensation, multi-physics field monitoring and adaptive control, multi-modal data fusion and deep learning, multi-robot collaboration of reinforcement learning, and digital twin and virtual simulation technologies, high-precision positioning, accurate pattern transfer and uniform coating are achieved.
It improves the lithography positioning accuracy and pattern transfer accuracy, enhances the uniformity of photoresist coating, optimizes the lithography process flow, and improves chip yield and production efficiency.
Smart Images
Figure CN120821157A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of information technology, and in particular relates to a semiconductor lithography system based on information fusion and intelligent control. Background Art
[0002] Semiconductor photolithography is a critical step in transferring the circuit pattern on the mask onto the surface of a semiconductor wafer coated with photoresist. It determines the chip's feature size and performance. With the continuous advancement of semiconductor technology, chip integration is increasing, and the precision and quality requirements of the photolithography process are becoming increasingly stringent. Industrial robots play a vital role in the operation and control of semiconductor photolithography equipment. However, existing industrial robot technology still has many problems in the photolithography process, which seriously restricts the further development of semiconductor photolithography technology.
[0003] Lithography positioning accuracy is difficult to guarantee: During the lithography process, the pattern on the mask must be precisely aligned to a specific location on the wafer. This positioning accuracy directly impacts chip performance and yield. Traditional industrial robotic lithography positioning systems rely primarily on simple visual positioning technology and preset alignment parameters, which cannot accurately adapt to factors such as microscopic fluctuations on the wafer surface, thermal expansion, and mechanical vibration. Furthermore, due to long-term operation of the lithography equipment and changes in environmental factors, the positioning system's accuracy gradually decreases, resulting in increased alignment deviation of the lithography pattern, affecting chip performance and reliability.
[0004] Poor photoresist pattern transfer accuracy: Photoresist pattern transfer accuracy refers to the ability to accurately replicate the pattern on the mask onto the photoresist layer. Existing industrial robotic photoresist systems are susceptible to factors such as uneven photoresist thickness, inconsistent exposure dose, and fluctuations in development process parameters during the pattern transfer process, leading to problems such as deformation, blurring, and distortion of the photoresist pattern. These issues not only affect chip performance but also increase chip manufacturing costs. Photoresist coating uniformity is difficult to control: Photoresist coating uniformity is critical to the quality of the photoresist pattern and the stability of the photoresist process. Currently, industrial robots primarily use fixed coating parameters and preset coating paths during the photoresist coating process, which cannot be adjusted in real time based on the size, shape, and surface characteristics of the wafer. Furthermore, due to the structural and process limitations of the coating equipment, the photoresist coating thickness on the wafer surface can easily become uneven, affecting the quality of the photoresist pattern and the stability of the photoresist process. Summary of the Invention
[0005] The present invention provides a semiconductor lithography system based on information fusion and intelligent control, comprising:
[0006] The lithography high-precision positioning algorithm module based on multi-sensor fusion and real-time compensation obtains multi-source information of lithography positioning through multiple sensors, fuses and processes it, and combines it with a real-time compensation algorithm to achieve high-precision positioning;
[0007] The lithography pattern transfer optimization algorithm module based on multi-physics field monitoring and adaptive control uses multi-physics field monitoring sensors to obtain parameters and achieves accurate pattern transfer through information fusion and adaptive control;
[0008] A photoresist coating uniformity control algorithm module based on multimodal data fusion and deep learning uses a variety of detection equipment to obtain multimodal data from wafers, and then uses a deep learning algorithm to achieve uniform coating after fusion;
[0009] A multi-robot collaborative lithography scheduling algorithm module based on reinforcement learning treats multiple robots as a multi-agent system, defines related spaces and functions, and uses reinforcement learning to achieve efficient collaboration;
[0010] Based on the semiconductor lithography process optimization algorithm module of digital twin and virtual simulation, a digital twin model is established and virtual simulation is used to optimize the lithography process.
[0011] Furthermore, in the lithography high-precision positioning algorithm module based on multi-sensor fusion and real-time compensation, the multiple sensors include a laser interferometer displacement measurement accuracy of ±0.01μm, a capacitive sensor surface undulation detection accuracy of ±0.001μm, and a visual sensor image resolution of 0.1μm. The deployment position is determined by finite element simulation and actual testing.
[0012] Furthermore, multi-sensor data fusion adopts a fusion network based on the attention mechanism, and the fusion formula is: Where xi represents the sensor data, ωi represents the attention weight, and the number of neurons dynamically calculated using a three-layer fully connected network is 128, 64, and 32, respectively, optimized using a backpropagation algorithm. The real-time compensation algorithm calculates the positioning parameter adjustment ΔP based on the fused data S, satisfying: ΔP = f(S), where f is a function constructed based on the photolithography process.
[0013] Furthermore, in the lithography pattern transfer optimization algorithm module based on multi-physical field monitoring and adaptive control, the multi-physical field monitoring sensors include a temperature sensor with an accuracy of ±0.1°C, a pressure sensor with an accuracy of ±0.5kPa, an electric field sensor with an accuracy of ±0.1V / m, a magnetic field sensor with an accuracy of ±0.01mT, and a spectrometer exposure dose detection accuracy of ±1%.
[0014] Furthermore, multi-source information fusion utilizes a deep learning-based fusion algorithm to construct a multiphysics model. Based on this multiphysics model and real-time photoresist changes, the adaptive control algorithm calculates the photolithography process parameter adjustment ΔE, satisfying the following equation: ΔE = g(M, R), where M represents the multiphysics model data, R represents the real-time photoresist changes, and g is a function determined based on the adaptive control rule.
[0015] Furthermore, in the photoresist coating uniformity control algorithm module based on multimodal data fusion and deep learning, the detection equipment includes a laser rangefinder with a ranging accuracy of ±0.05μm, an optical microscope with a magnification of 50-1000 times, and an infrared thermal imager with a temperature resolution of ±0.5°C.
[0016] Furthermore, multimodal data fusion uses a deep learning algorithm to construct a comprehensive information model of the wafer surface. The deep learning model combines a convolutional neural network with a recurrent neural network, and is trained to establish a mapping relationship between wafer surface characteristics C and photoresist coating parameters P: P = h(C), where h is a function determined by training.
[0017] Furthermore, in the multi-robot collaborative lithography scheduling algorithm module based on reinforcement learning, the state space includes the robot position L(x, y), task progress Tp, working state Sw
[0018] and workshop environment information E; the action space includes robot motion Ma and task allocation decision Da; the reward function
[0019] R comprehensively considers factors such as lithography efficiency Ef, pattern transfer accuracy At, and photoresist coating uniformity Uc, satisfying R=α·Ef+β·At+γ·Uc, where α, β, and γ are corresponding weights with a value range of 0-1.
[0020] Furthermore, in the semiconductor lithography process optimization algorithm module based on digital twins and virtual simulation, the digital twin model is constructed using professional modeling software, and the model update frequency is consistent with the sensor data acquisition frequency.
[0021] Beneficial effects:
[0022] Improve lithography positioning accuracy: The high-precision lithography positioning algorithm based on multi-sensor fusion and real-time compensation can comprehensively utilize information from multiple sensors, adjust the positioning parameters of the industrial robot in real time, and compensate and correct errors in the lithography positioning process in real time, thereby achieving high-precision lithography positioning and improving chip performance and yield.
[0023] Optimizing the accuracy of lithography pattern transfer: The lithography pattern transfer optimization algorithm based on multi-physical field monitoring and adaptive control can comprehensively utilize information from multiple sensors, adjust the lithography process parameters of the industrial robot in real time, adapt to various changes in the lithography process, achieve accurate transfer of lithography patterns, and improve the quality of lithography patterns and the stability of the lithography process.
[0024] Improving photoresist coating uniformity: The photoresist coating uniformity control algorithm based on multimodal data fusion and deep learning can comprehensively utilize information from multiple detection devices, adjust the photoresist coating parameters in real time according to the surface characteristics of the wafer, achieve uniform photoresist coating, and improve the quality of the photolithography pattern and the stability of the photolithography process.
[0025] Optimizing multi-robot collaborative lithography scheduling: The multi-robot collaborative lithography scheduling algorithm based on reinforcement learning can achieve efficient collaborative work among multiple robots, dynamically adjust task allocation and production plans according to real-time production conditions and equipment status, and improve the overall efficiency and resource utilization of lithography.
[0026] Improve semiconductor lithography process: The semiconductor lithography process optimization algorithm based on digital twins and virtual simulation can predict and analyze possible problems in the lithography process in advance through virtual simulation technology, optimize and adjust the parameters and processes of the semiconductor lithography process, and formulate the optimal lithography process plan to achieve optimization and improvement of the semiconductor lithography process and improve the efficiency and quality of lithography. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 Module flow diagram. DETAILED DESCRIPTION
[0028] Example 1:
[0029] A semiconductor lithography system based on information fusion and intelligent control, comprising:
[0030] The lithography high-precision positioning algorithm module based on multi-sensor fusion and real-time compensation obtains multi-source information of lithography positioning through multiple sensors, fuses and processes it, and combines it with a real-time compensation algorithm to achieve high-precision positioning;
[0031] The lithography pattern transfer optimization algorithm module based on multi-physics field monitoring and adaptive control uses multi-physics field monitoring sensors to obtain parameters and achieves accurate pattern transfer through information fusion and adaptive control;
[0032] A photoresist coating uniformity control algorithm module based on multimodal data fusion and deep learning uses a variety of detection equipment to obtain multimodal data from wafers, and then uses a deep learning algorithm to achieve uniform coating after fusion;
[0033] A multi-robot collaborative lithography scheduling algorithm module based on reinforcement learning treats multiple robots as a multi-agent system, defines related spaces and functions, and uses reinforcement learning to achieve efficient collaboration;
[0034] A semiconductor lithography process optimization algorithm module based on digital twins and virtual simulation is used to establish a digital twin model and use virtual simulation to optimize the lithography process. In the lithography high-precision positioning algorithm module based on multi-sensor fusion and real-time compensation, the multiple sensors include a laser interferometer with a displacement measurement accuracy of ±0.01μm, a capacitive sensor with a surface undulation detection accuracy of ±0.001μm, and a visual sensor with an image resolution of 0.1μm. The deployment position is determined by finite element simulation and actual testing. Multi-sensor data fusion uses a fusion network based on the attention mechanism, and the fusion formula is:
[0035] Among them, xi is the data of each sensor, ωi is the attention weight, and the number of neurons dynamically calculated by the three-layer fully connected network is 128, 64, and 32 respectively, and optimized by the back propagation algorithm. The real-time compensation algorithm calculates the positioning parameter adjustment amount ΔP based on the fusion data S, satisfying: ΔP=f(S)f is a function constructed based on the lithography process. In the lithography pattern transfer optimization algorithm module based on multi-physical field monitoring and adaptive control, the multi-physical field monitoring sensors include temperature sensor accuracy ±0.1℃, pressure sensor accuracy ±0.5kPa, electric field sensor accuracy
[0036] ±0.1V / m, magnetic field sensor accuracy ±0.01mT, and spectrometer exposure dose detection accuracy ±1%. Multi-source information fusion utilizes a deep learning-based fusion algorithm to construct a multi-physics model. The adaptive control algorithm calculates the photolithography process parameter adjustment ΔE based on the multi-physics model and real-time photoresist change information, satisfying the following: ΔE = g(M,R), where M is the multi-physics model data, R is the real-time photoresist change information, and g is a function determined based on the adaptive control rule. The photoresist coating uniformity control algorithm module based on multimodal data fusion and deep learning includes detection equipment including a laser rangefinder with a ranging accuracy of ±0.05μm, an optical microscope with a magnification of 50-1000x, and an infrared thermal imager with a temperature resolution of ±0.5°C. Multimodal data fusion utilizes a deep learning algorithm to construct a comprehensive information model of the wafer surface. The deep learning model convolutional neural network is combined with the recurrent neural network to establish the mapping relationship between the wafer surface characteristics C and the photoresist coating parameters P through training: P = h (C) h is the function determined by training. In the multi-robot collaborative lithography scheduling algorithm module based on reinforcement learning, the state space covers the robot position L (x, y), task progress Tp, working status Sw
[0037] and workshop environment information E; the action space includes robot motion Ma and task allocation decision Da; the reward function
[0038] R integrates the factors of lithography efficiency Ef, pattern transfer accuracy At, and photoresist coating uniformity Uc, satisfying R = α·Ef + β·At + γ·Uc, where α, β, and γ are corresponding weights ranging from 0 to 1. In the semiconductor lithography process optimization algorithm module based on digital twins and virtual simulation, the digital twin model is constructed using professional modeling software, and the model update frequency is consistent with the sensor data acquisition frequency.
[0039] Implementation of high-precision positioning algorithm for lithography based on multi-sensor fusion and real-time compensation
[0040] Sensor Deployment and Data Acquisition: Laser interferometers, capacitive sensors, and visual sensors are strategically deployed within the industrial robot's lithography positioning system to ensure accurate acquisition of various parameters during the lithography positioning process and relevant information about wafers and masks. Sensors collect data in real time at a set frequency and transmit it to the data processing unit via the network. Multi-Sensor Data Fusion and State Model Construction: The data processing unit utilizes deep learning-based multi-sensor data fusion algorithms, such as an attention-based fusion network, to fuse data collected from different sensors and construct a real-time state model of the lithography positioning process. Deep learning algorithms, such as convolutional neural networks (CNNs), are used to analyze and process the fused data and extract key feature information.
[0041] Real-time compensation algorithm implementation and positioning execution: Based on the constructed real-time state model, the real-time compensation algorithm calculates the adjustment amount for the industrial robot's positioning parameters. The robot control system adjusts the robot's positioning parameters in real time based on the adjustment amount to complete the lithography positioning task. During the positioning process, sensor data is continuously monitored in real time and positioning parameters are dynamically adjusted to ensure lithography positioning accuracy.
[0042] Implementation of lithography pattern transfer optimization algorithm based on multi-physics field monitoring and adaptive control
[0043] Sensor Deployment and Data Collection: Temperature sensors, pressure sensors, electric field sensors, magnetic field sensors, and spectrometers are strategically deployed throughout the lithography equipment to accurately capture multi-physics parameters during the lithography process, as well as information such as the photoresist exposure dose and development level. The sensors collect data in real time at a set frequency and transmit it to the data processing unit via the network.
[0044] Multi-source information fusion and multi-physics model construction: The data processing unit uses a deep learning-based multi-source information fusion algorithm to fuse data collected by different sensors and construct a multi-physics model of the lithography pattern transfer process. Deep learning algorithms, such as convolutional neural networks (CNNs), are used to analyze and process the fused data and extract key feature information.
[0045] Adaptive control algorithm implementation and lithography execution: Based on the constructed multi-physics model and real-time information about photoresist changes, the adaptive control algorithm calculates the required adjustments for the industrial robot's lithography process parameters. The robot control system adjusts the robot's lithography process parameters in real time based on these adjustments to complete the lithography task. During the lithography process, sensor data is continuously monitored in real time, and lithography process parameters are dynamically adjusted to ensure accurate transfer of the lithography pattern.
[0046] Implementation of a photoresist coating uniformity control algorithm based on multimodal data fusion and deep learning. Laser rangefinders, optical microscopes, and infrared thermal imagers are strategically deployed on the photoresist coating equipment to ensure comprehensive multimodal data capture on the wafer surface. The equipment collects data in real time at a set frequency and transmits it to the data processing center via the network.
[0047] Multimodal Data Fusion and Feature Extraction: The data processing center uses a deep learning-based multimodal data fusion algorithm to fuse data collected by different testing devices and construct a comprehensive information model of the wafer surface. Deep learning algorithms, such as convolutional neural networks (CNNs), are used to analyze and process the fused data and extract key features of the wafer surface.
[0048] Deep learning model training and coating execution: Collect a large amount of different types of wafer surface data and photoresist coating data to build a training data set. Utilize deep learning algorithms, such as a combination of convolutional neural networks (CNN) and recurrent neural networks (RNN), to learn and train the training data set, and establish a mapping relationship model between wafer surface characteristics and photoresist coating parameters. In the actual coating process, the extracted feature information is input into the trained model, and the photoresist coating parameters of the industrial robot are adjusted in real time according to the output results of the model to complete the photoresist coating task. During the coating process, the detection equipment data is continuously monitored in real time, and the coating parameters are dynamically adjusted to ensure the uniformity of the photoresist coating. Implementation of multi-robot collaborative lithography scheduling algorithm based on reinforcement learning
[0049] Agent Definition and Communication Mechanism Establishment: Multiple industrial robots in a semiconductor lithography 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.
[0050] Reinforcement Learning Model Training and Application: Clarify the state space, action space, and reward function for reinforcement learning. Collect historical operational data from semiconductor lithography workshops 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. In actual production, the intelligent 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 lithography scheduling strategy and enables efficient collaboration among multiple robots.
[0051] Implementation of semiconductor lithography process optimization algorithm based on digital twin and virtual simulation
[0052] Digital twin model construction: Using modeling software and sensor data, a digital twin model of the semiconductor lithography system is established, mapping entities such as industrial robots, lithography equipment, and wafers in the physical world into the virtual space, enabling real-time interaction and synchronization between physical entities and virtual models.
[0053] Virtual simulation and process optimization: Utilizing virtual simulation technology, semiconductor photolithography processes are simulated and optimized, enabling the early prediction and analysis of potential issues in the process within a virtual environment. Based on the simulation results, process parameters and procedures are optimized and adjusted to develop an optimal photolithography solution. This optimized solution is then applied to actual production, achieving optimization and improvement of the semiconductor photolithography process.
[0054] Example 2
[0055] Photolithography positioning accuracy example
[0056] In a 10nm process semiconductor chip production workshop, two photolithography production lines were selected for comparative testing, with each production line photolithography 100 wafers per day.
[0057] The traditional lithography positioning technology group used a single vision positioning technology, with a visual sensor image resolution of 0.5μm. The positioning system relied on preset alignment parameters. After one week of operation, statistics showed that the average lithography positioning deviation reached ±0.2μm, resulting in a chip yield of only 70%. Positioning deviations required manual intervention and adjustment five times a day, seriously affecting production efficiency.
[0058] The Multi-Sensor Fusion and Real-Time Compensation Technology Group deployed laser interferometers (displacement measurement accuracy of ±0.01μm), capacitive sensors (surface undulation detection accuracy of ±0.001μm), and visual sensors (image resolution of 0.1μm). Multi-sensor data fusion employed a fusion network based on an attention mechanism, and the real-time compensation algorithm calculated the positioning parameter adjustment ΔP = f(S) based on the fused data S. After a week of operation, the mean lithography positioning deviation was controlled at ±0.03μm, the chip yield increased to 90%, and only one manual intervention adjustment was required within a week, significantly improving lithography positioning accuracy and production stability. A month-long test was conducted on the 300mm wafer lithography process in the lithography workshop of a semiconductor manufacturer.
[0059] Traditional lithography pattern transfer technology: Lithography equipment lacks multi-physics field monitoring and relies solely on fixed lithography process parameters. Testing revealed an average lithography pattern transfer deviation of ±0.3μm. Pattern deformation and blurring frequently occurred, resulting in poor chip performance consistency and a defective rate of 15%.
[0060] The Multi-Physical Field Monitoring and Adaptive Control Technology Team installed temperature sensors (accuracy ±0.1°C), pressure sensors (accuracy ±0.5kPa), electric field sensors (accuracy ±0.1V / m), magnetic field sensors (accuracy ±0.01mT), and a spectrometer (exposure dose detection accuracy ±1%). Multi-source information fusion was used to construct a multi-physics model. The adaptive control algorithm calculated the lithography process parameter adjustment ΔE = g(M,R) based on the model and real-time photoresist changes. During testing, the mean deviation of the lithography pattern transfer was reduced to ±0.08μm, and the defective rate dropped to 5%, significantly improving the accuracy of lithography pattern transfer and chip quality.
[0061] In a semiconductor photoresist coating workshop, two groups of coating equipment of the same model were selected to conduct photoresist coating tests on 100 wafers.
[0062] Traditional photoresist coating technology group: Using fixed coating parameters and preset coating paths, it was measured that the photoresist coating thickness uniformity deviation reached ±15%. Due to the uneven coating, the quality of the photolithography pattern was poor, and the rework rate was 20%.
[0063] Multimodal data fusion and deep learning technology team: equipped with a laser rangefinder (distance measurement accuracy ±0.05μm), an optical microscope (magnification 50-1000 times), and an infrared thermal imager (temperature resolution ±0.5℃). Multimodal data fusion builds a comprehensive information model of the wafer surface, and a deep learning model establishes the relationship between wafer surface characteristics C and photoresist coating parameters.
[0064] The mapping relationship of P is P = h(C). Test results show that the photoresist coating thickness uniformity deviation is controlled within ±3%, and the rework rate is reduced to 5%, effectively improving the photoresist coating uniformity and photolithography pattern quality.
[0065] In a semiconductor lithography workshop with 10 industrial robots, a week of lithography production tasks is simulated.
[0066] Traditional multi-robot collaborative technology groups lack effective task allocation and coordination mechanisms, leading to frequent conflicts between robots. Statistics show that the average daily photolithography task completion time is 10 hours, and equipment utilization is only 60%.
[0067] The Reinforcement Learning Multi-Robot Collaborative Lithography Scheduling Technology Team treats the robots as a multi-agent system. The state space encompasses robot position L(x, y), task progress Tp, work status Sw, and workshop environment information E. The action space includes robot motion Ma and task allocation decisions Da. The reward function R = α·Ef+β·At+γ·Uc is used. After one week of operation, the daily lithography task completion time was reduced to 6 hours, and equipment utilization increased to 85%. This system achieves efficient multi-robot collaboration and significantly improves lithography efficiency.
Claims
1. A semiconductor lithography system based on information fusion and intelligent control, characterized in that: include: The lithography high-precision positioning algorithm module based on multi-sensor fusion and real-time compensation obtains multi-source information of lithography positioning through multiple sensors, fuses and processes it, and combines it with a real-time compensation algorithm to achieve high-precision positioning; The lithography pattern transfer optimization algorithm module based on multi-physics field monitoring and adaptive control uses multi-physics field monitoring sensors to obtain parameters and achieves accurate pattern transfer through information fusion and adaptive control; A photoresist coating uniformity control algorithm module based on multimodal data fusion and deep learning uses a variety of detection equipment to obtain multimodal data from wafers, and then uses a deep learning algorithm to achieve uniform coating after fusion; A multi-robot collaborative lithography scheduling algorithm module based on reinforcement learning treats multiple robots as a multi-agent system, defines related spaces and functions, and uses reinforcement learning to achieve efficient collaboration; Based on the semiconductor lithography process optimization algorithm module of digital twin and virtual simulation, a digital twin model is established and virtual simulation is used to optimize the lithography process.
2. The key technology according to claim 1 is characterized in that: In the lithography high-precision positioning algorithm module based on multi-sensor fusion and real-time compensation, the multiple sensors include a laser interferometer with a displacement measurement accuracy of ±0.01μm, a capacitive sensor with a surface undulation detection accuracy of ±0.001μm, and a visual sensor with an image resolution of 0.1μm. The deployment position is determined by finite element simulation and actual testing.
3. The algorithm module according to claim 2, characterized in that: Multi-sensor data fusion adopts a fusion network based on the attention mechanism, and the fusion formula is: Where xi is the sensor data, ωi is the attention weight, and the number of neurons is dynamically calculated by a three-layer fully connected network, which is 128, 64, and 32 respectively, and optimized by the back propagation algorithm. The real-time compensation algorithm calculates the positioning parameter adjustment ΔP based on the fused data S, satisfying: ΔP=f(S)f is a function constructed based on the photolithography process.
4. The key technology according to claim 1 is characterized in that: In the lithography pattern transfer optimization algorithm module based on multi-physical field monitoring and adaptive control, the multi-physical field monitoring sensor includes a temperature sensor accuracy ±0.1℃, pressure sensor accuracy ±0.5kPa, electric field sensor accuracy ±0.1V / m, magnetic field sensor accuracy ±0.01mT, spectrometer exposure dose detection accuracy ±1%.
5. The algorithm module according to claim 4, characterized in that: Multi-source information fusion utilizes a deep learning-based fusion algorithm to construct a multiphysics model. Based on this multiphysics model and real-time photoresist changes, the adaptive control algorithm calculates the photolithography process parameter adjustment ΔE, satisfying the following equation: ΔE = g(M, R), where M represents the multiphysics model data, R represents the real-time photoresist changes, and g is a function determined based on the adaptive control rule.
6. The key technology according to claim 1 is characterized in that: In the photoresist coating uniformity control algorithm module based on multimodal data fusion and deep learning, the detection equipment includes a laser rangefinder with a ranging accuracy of ±0.05μm, an optical microscope with a magnification of 50-1000 times, and an infrared thermal imager with a temperature resolution of ±0.5°C.
7. The algorithm module according to claim 6, characterized in that: Multimodal data fusion uses a deep learning algorithm to construct a comprehensive information model of the wafer surface. The deep learning model combines a convolutional neural network with a recurrent neural network, and is trained to establish a mapping relationship between wafer surface characteristics C and photoresist coating parameters P: P = h(C), where h is the training function.
8. The key technology according to claim 1 is characterized in that: In the reinforcement learning-based multi-robot collaborative lithography scheduling algorithm module, the state space covers the robot position L(x, y), task progress Tp, working status Sw and workshop environment information E; the action space includes the robot motion Ma and task allocation decision Da; the reward function R comprehensively considers the factors of lithography efficiency Ef, pattern transfer accuracy At, and photoresist coating uniformity Uc, satisfying R=α·Ef+β·At+γ·Uc, where α, β, and γ are corresponding weights with a value range of 0-1.
9. The key technology according to claim 1 is characterized in that: In the semiconductor lithography process optimization algorithm module based on digital twins and virtual simulation, the digital twin model is constructed using professional modeling software, and the model update frequency is consistent with the sensor data acquisition frequency.