Machine Learning-Based Dosing Control Device and Method for Gas Target Laser Plasma Extreme Ultraviolet Source

Abstract

This invention discloses a dose control device and method for a gas target laser plasma extreme ultraviolet (EUV) source based on machine learning. The device includes a vacuum system, a laser system, a nozzle system, a gas supply system, a timing system, a monitoring system, and a control system. The vacuum system ensures the transmission of EUV light; the laser system provides the driving laser for plasma generation; the nozzle system injects target gas into the vacuum environment and can control the gas position; the gas supply system delivers gas to the nozzles and can control the gas flow time; the timing system controls the start-up time of the driving laser and the injection time of the target gas; the monitoring system records EUV energy and plasma images in real time; and the control system adjusts control parameters through data analysis to maintain the stability of EUV light output. This invention also proposes a dose control method using the above device, including feedforward control based on energy prediction-target setting-control inference and fast-response feedback control.

Description

Technical Field

[0001] This invention relates to the field of extreme ultraviolet lithography machines and light source equipment control, specifically to a dose control device and method for a gas target laser-generated plasma extreme ultraviolet light source based on machine learning, and an EUV light source system using gas targets such as Xe, O2, and CO2. Background Technology

[0002] As integrated circuit manufacturing processes continue to advance to the 7-nanometer node, traditional deep ultraviolet (DUV) lithography, limited by its optical resolution, can no longer meet the linewidth and pattern precision requirements of advanced manufacturing. According to the Rayleigh Criterion, shorter exposure wavelengths can achieve higher optical resolution. EUV lithography, with its short wavelength of 13.5nm, offers a significant improvement in lithographic resolution and has become one of the core technologies for logic device and memory manufacturing.

[0003] In the photolithography process, the EUV light source irradiates the photoresist on the wafer surface to complete the exposure reaction. The cumulative EUV energy received by the photoresist directly determines the degree of photoresist reaction and the final development profile. EUV lithography systems typically use laser-generated plasma (LPP) as the light source, generating plasma by bombarding a target material such as tin or xenon with a high-power laser to obtain the required EUV light. Due to the randomness of plasma generation, the EUV output energy exhibits significant inter-pulse fluctuations, which places extremely high demands on the stability of the photolithography process. If the cumulative EUV energy (dose) deviates from the target value per unit time, it will lead to underexposure or overexposure, thus affecting the pattern accuracy. For mass production, three times the standard deviation under a 50-pulse window needs to be less than 0.2% of the average (Suzuki K, Miyake A and Harned N, EUVL Symposium Proceedings, 2009).

[0004] The energy output of the LPP light source is affected by multiple factors, including but not limited to laser input, target condition, changes in the vacuum environment, and optical system contamination caused by long-term operation. Therefore, real-time, stable, and precise control of the light source output dose (Dose Control) has become one of the key technologies in EUV lithography systems. ASML, in its published patent document "EUV LPP Source with Improved Dose Control by Combining PulseModulation and Pulse Control Mode," proposed a control scheme based on cumulative dose closed-loop adjustment. By comparing the difference between the cumulative dose and the target dose, combined with pulse adjustment and pulse control mode adjustment, the stability of continuous multi-pulse operation is improved to some extent. However, this scheme mainly relies on passive compensation based on feedback loops, and its ability to coordinate multi-parameter adjustment is limited, leaving room for improvement. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a dose control device and method for generating plasma extreme ultraviolet light sources using Xe, O2, CO2, and other gas target lasers based on machine learning. By introducing energy prediction, target energy setting, and control quantity inference mechanisms, high-precision, stable, and real-time control of accumulated pulse energy is achieved.

[0006] The technical solution adopted in this invention is as follows: A machine learning-based dose control device for a gas target laser-generated plasma extreme ultraviolet light source, characterized in that it includes: Vacuum system, used to provide the vacuum environment required for extreme ultraviolet light transmission; Laser systems are used to generate driving lasers to bombard target gas and produce plasma; A nozzle system for injecting target gas into the vacuum system; A gas supply system, connected to the nozzle system via a gas pipeline, is used to supply the target gas to the nozzle system and control the gas supply. A timing system, electrically connected to the laser system and the gas supply system respectively, is used to control the start-up time of the driving laser and the injection time of the target gas; The monitoring system is used to monitor the energy and plasma state parameters of extreme ultraviolet light in real time; and The control system is connected to the vacuum system, laser system, nozzle system, gas supply system, timing system, and monitoring system, respectively. The control system is a machine learning-based feedforward control system. The feedforward control includes: using a machine learning prediction model to make multi-step predictions of future extreme ultraviolet pulse energy; using a control strategy to generate a multi-step target energy sequence based on the prediction results and target energy; and using a machine learning inference model to generate adjustment amounts of multiple control parameters to perform feedforward adjustment on the subsequent target firing process.

[0007] Furthermore, the control system is configured to: switch to a safe mode of feedback control when the deviation between the actual value and the target value of the extreme ultraviolet pulse energy exceeds a preset threshold; in the safe mode, it is used to calculate only the target energy of the next pulse and perform energy adjustment judgment and missed firing mechanism.

[0008] The vacuum system includes a vacuum chamber, a dry pump, a vacuum pump, a barometer, and a thermometer. The vacuum chamber creates a stable plasma generation space and optical path transmission environment while isolating air, thus constructing the vacuum environment required for EUV generation. The dry pump and vacuum pump are located outside the vacuum chamber and are connected to it via pipes. The combined use of the dry pump and vacuum pump can reduce the pressure inside the vacuum chamber to 10. -4 Pa to 10 -5 The system continuously evacuates residual gas from the chamber at the Pa level. The barometer and thermometer are connected to the vacuum chamber via flanges to monitor the chamber pressure (Pchamber) and temperature (Tchamber) in real time, respectively.

[0009] Furthermore, the laser system includes a laser, a focusing lens, and a collecting mirror. The laser, located outside the vacuum cavity, generates driving laser pulses to provide energy for plasma generation. The laser can be a laser of different wavelengths, such as Nd:YAG or CO2, wherein the driving voltage (V... laser The output energy is adjustable, typically ranging from 100mJ to 1000mJ, and the laser pulse width (W) is adjustable. laser The laser pulse energy is adjustable, typically from 1 to 20 ns, and the laser frequency is adjustable, typically from 0.1 to 10 Hz. The focusing lens is located inside the vacuum cavity and can focus the pulse energy generated by the laser onto the gas target to generate plasma. The collecting mirror is located inside the vacuum cavity and is arranged around the plasma generation point to collect EUV radiation and focus it to the central focal point.

[0010] Furthermore, the gas supply system includes a target gas cylinder, a buffer gas cylinder, a pressure reducing valve, and a solenoid valve. The target gas cylinder is located outside the vacuum chamber; the stored target gas can generate high-temperature, high-density plasma under laser bombardment, and can be gases such as Xe, O2, and CO2. The buffer gas cylinder is also located outside the vacuum chamber; the stored buffer gas improves EUV conversion efficiency by restricting the diffusion of the target gas, and is typically He gas. The target gas and buffer gas are delivered to the nozzle through flexible hoses passing through the vacuum chamber wall. Each flexible hose is equipped with its respective pressure reducing valve and solenoid valve. The pressure reducing valve can adjust the target gas pressure (P... Target ) and buffer gas pressure (P Buffer The pressure is typically 1-20 bar. The solenoid valve controls the gas passage time (T) of the target material. Target ) and buffer gas ventilation time (T Buffer ).

[0011] Furthermore, the nozzle system includes a triaxial displacement stage and a nozzle. The triaxial displacement stage is installed inside the vacuum chamber, and the nozzle is fixedly installed on the triaxial displacement stage, capable of spraying target gas and buffer gas. The triaxial displacement stage can adjust the nozzle position X (X... nozzle ), Nozzle position Y (Y nozzle ) and nozzle position Z (Z nozzle This allows the driving pulse laser to bombard the gas target at the appropriate position.

[0012] Furthermore, the timing system includes a signal generator and a BNC connector. The signal generator is located outside the vacuum chamber and is connected to the laser and solenoid valve via the BNC connector, allowing for independent control of the laser and solenoid valve's start-up time to ensure the driving pulse laser bombards the gas target at the appropriate time. The signal generator is used to control the start-up delay (T) between the target gas and the laser. Target-laser ) and the start-up delay (T) between the buffer gas and the laser. Buffer-laser ).

[0013] Furthermore, the monitoring system includes an EUV energy meter (E-mon) and an EUV camera (E-cam). The EUV energy meter is connected to the vacuum chamber via a flange and is located at the focal point of the collecting mirror. It receives extreme ultraviolet photons through a photodiode and converts them into electrical signals, enabling real-time monitoring of EUV energy (E). The EUV camera is also connected to the vacuum chamber via a flange. Its observation window is aligned with the plasma generation area, and it obtains plasma images using the pinhole imaging principle. The plasma length (L) can be obtained through image analysis. plasma ) and plasma diameter (D plasma ).

[0014] Furthermore, the control system includes a computer and data cables. The computer is located outside the vacuum chamber and is connected via the data cables to a pressure reducing valve, a laser, an EUV camera, an EUV energy meter, a barometer, a thermometer, a signal generator, and a three-axis displacement stage. The computer can record the aforementioned data, including V... laser W laser , P Target , P Buffer , T Target ,T Buffer X nozzle , Y nozzle Z nozzle , T Target-laser , and T Buffer-laser Control parameter data and including E, L plasma D plasma , P chamber , and T chamber The monitoring parameter data is used to perform the energy prediction-target energy setting-control quantity inference.

[0015] Furthermore, the multi-step pulse energy prediction is performed by a machine learning prediction model. This model predicts the EUV pulse energy corresponding to multiple future laser pulses based on recent historical control and monitoring parameter data. Deployed in a computer, the machine learning prediction model learns the implicit mapping relationship between control parameters, monitoring parameters, and EUV pulse energy output through pre-training on a large amount of historical operating data. It then fits and optimizes this implicit mapping function to achieve the prediction of future multi-step EUV pulse energy. The model can employ a regression tree model, a neural network model, a statistical regression model, or a combination of multiple models, selected based on the prediction accuracy.

[0016] Furthermore, the multi-step target energy setting employs an incremental proportional-integral-derivative (PID) strategy. This control strategy compares the average pulse energy within a time window with the user-defined nominal energy to obtain the window error. Then, by using a proportional term reflecting the current dose error change trend, an integral term eliminating long-term cumulative dose deviation, and a derivative term suppressing energy fluctuations, the future EUV target energy is set. During the generation of the multi-step target energy, the energy of pulses not yet emitted is replaced by the target energy value to calculate the average energy.

[0017] Furthermore, the control inference is performed by a machine learning inference model. This model generates adjustments for each control parameter based on predicted multi-step EUV pulse energy, target energy, control parameter states, and monitoring parameter data. Deployed in a computer, the machine learning model learns the implicit mapping between control parameter adjustments and EUV pulse energy changes to infer a control parameter adjustment scheme that makes the EUV pulse energy approach the target energy. During the inference process, the machine learning inference model aims to minimize the deviation between the predicted energy and the target energy, achieving convergence of the EUV pulse energy to the target energy with minimal adjustments while satisfying control parameter constraints. The model can employ a regression tree model, a neural network model, a statistical regression model, or a combination of multiple models, selected based on the control error.

[0018] On the other hand, the present invention also provides a dose control method for a gas target laser-generated plasma extreme ultraviolet light source applied to the above-mentioned device, comprising the following steps: S1. Setting and Startup: The user sets the target dose and initial control parameters V. laser W laser , P Target ,P Buffer , T Target , T Buffer X nozzle , Y nozzle Z nozzle , T Target-laser , and T Buffer-laser .

[0019] S2, Targeting: The Xe-LPP system generates EUV radiation by firing at a target according to control parameters.

[0020] S3. Measurement: Record monitoring parameter data E and L after target firing. plasma D plasma , P chamber , and T chamber When the monitored data exceeds its upper and lower thresholds, proceed directly to step S6.

[0021] S4. Mode Selection: The system defaults to normal mode upon initialization. In normal mode, if the deviation between the measured EUV pulse energy and the target energy (e...) is... i If ei remains above a certain threshold, the system switches to safe mode. If ei remains below the threshold while in safe mode, the system switches to normal mode.

[0022] In the normal mode, machine learning-based feedforward control is performed, specifically including: - Prediction Step: Using a machine learning prediction model deployed in the control system, the extreme ultraviolet pulse energy corresponding to multiple future laser pulses is predicted based on historical control parameter data and monitoring parameter data.

[0023] - Target energy setting steps: Using an incremental proportional-integral-derivative (PID) strategy, a target energy sequence for multiple future pulses is output based on historical extreme ultraviolet pulse energies. This strategy obtains the window error by comparing the average pulse energy within a time window with the nominal energy set by the user, and uses proportional, integral, and derivative terms to set the future target energy. For pulses that have not yet been emitted, their target energy values ​​are used to replace the actual measured values ​​to calculate the average energy.

[0024] - Reasoning steps: Using a machine learning reasoning model deployed in the control system, based on the predicted multi-step extreme ultraviolet pulse energy, the target energy sequence, the current control parameter status, and monitoring parameter data, with the goal of reducing the deviation between the predicted energy and the target energy, the adjustment amount of each control parameter that meets the control parameter constraints is generated.

[0025] In the aforementioned safety mode, feedback-based protective controls are implemented, specifically including: - Target calculation steps: Using the incremental PID strategy, the target energy of the next pulse is output based solely on the historical extreme ultraviolet pulse energy.

[0026] - Energy adjustment judgment step: If the target energy is higher than the modulation upper limit, the amplitude is limited; if it is lower than the modulation lower limit, the missed shot mechanism is triggered, the next shot is skipped, and the process returns to step S3.

[0027] - Alarm procedure: When the EUV pulse energy remains at the modulation limit for multiple consecutive times or is triggered multiple times in a short period of time, a system abnormality alarm is issued; the user can manually enter step S6 according to the system operation status.

[0028] -Safe adjustment procedure: Only adjust fast control parameters with clear mathematical relationships, and dynamically adjust their adjustment weights according to the parameter status.

[0029] S5. Adjustment: Adjust the corresponding control parameters according to the adjustment amount generated by the normal mode or the safety mode, and return to the target firing step to form a closed-loop control. S6. Shutdown: The control system stops starting the laser and solenoid valves, and waits for user inspection.

[0030] Compared with the prior art, the present invention has the following beneficial effects: 1) In the dose control of a gas target LPP-EUV light source, this invention organically integrates a machine learning prediction model, incremental PID target programming, and a machine learning inference model to form a complete three-stage feedforward control architecture of "prediction-planning-inference". This architecture can predict energy fluctuation trends multiple pulses in advance and actively coordinate the adjustment of multi-domain control parameters such as laser, gas, and timing, achieving proactive and predictive control. This is significantly better than the single-step feedback control of "post-compensation" in traditional schemes, and effectively overcomes the random disturbances and multivariate coupling effects in the plasma generation process.

[0031] 2) A dual-mode collaborative working mechanism is adopted. In normal mode, the precision advantage of feedforward control is fully utilized to achieve highly stable dose output. When abnormal deviations are detected, the system automatically switches to a feedback safety mode based on explicit mathematical relationships. Through mechanisms such as amplitude limiting, missed doses, and dynamic weight adjustment, the system's safety and controllability under extreme operating conditions are ensured. This architecture design guarantees high performance during normal operation while providing reliable fault protection.

[0032] 3) This invention can compress the energy fluctuation of EUV light source within a continuous multi-pulse window to an extremely narrow range, meeting or even exceeding the stringent requirements for dose stability in large-scale semiconductor manufacturing (such as three times the standard deviation within a 50-pulse window being less than 0.2% of the average value), providing key light source technology support for advanced process lithography. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the dose control device for a gas target laser-generated plasma extreme ultraviolet light source based on machine learning, according to the present invention. Figure 2 This is a schematic flowchart of the dose control method for gas target laser-generated plasma extreme ultraviolet light source based on machine learning according to the present invention. Figure 3 This is a comparative schematic diagram of the dosage control effect of embodiments of the present invention; Figure 4 This is a schematic diagram of the predicted pulse energy and the target pulse energy according to an embodiment of the present invention. Detailed Implementation

[0034] To clearly illustrate the embodiments of the present invention, the present invention will be described in detail below with reference to the schematic diagrams of the embodiments. Obviously, the accompanying drawings described below are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the present invention without creative effort are within the scope of protection of the present invention.

[0035] This embodiment provides a machine learning-based laser-driven xenon (Xe) target extreme ultraviolet (EUV) light source dose control device and method, aiming to provide stable EUV energy output. It should be noted that although this embodiment uses Xe gas as the target gas, the technical solution of this invention is also applicable to other gas targets such as O2 and CO2, as well as scenarios involving the mixed use of multiple target gases.

[0036] Please see Figure 1 , Figure 1 This is a schematic diagram of a laser-driven xenon target dose control device based on machine learning, according to an embodiment of the present invention. As shown in the figure, the device mainly consists of seven parts: a vacuum system, a laser system, a nozzle system, a gas supply system, a timing system, a monitoring system, and a control system. The collaborative working relationship between the various systems is as follows: Vacuum System: Since EUV light is easily absorbed, it is essential to ensure that the EUV light source is always in a vacuum environment. Vacuum chamber 1 is made of stainless steel with a polished inner wall, providing excellent airtightness and a low outgassing rate. This isolates the chamber from the external atmosphere, providing a stable spatial environment for plasma generation and optical transmission. Multiple flange interfaces are provided on the chamber for connecting various sensors, actuators, and optical windows. Dry pump 2 and molecular pump 3 combine to form a vacuum pumping unit, capable of gradually reducing the pressure within vacuum chamber 1 from atmospheric pressure to 10... -4 Pa to 10 -5 A high vacuum state on the order of Pa is maintained, and residual gases released from the inner wall of the cavity and generated by plasma reactions are continuously evacuated to ensure the transmittance of EUV light in the transmission path. A barometer 4 and a thermometer 5 are installed on the side wall of the vacuum cavity 1, which can monitor the air pressure and temperature inside the vacuum cavity 1 in real time and transmit the data to the computer 22 via data cable 23.

[0037] The laser system provides the energy source for driving the plasma generation. In this embodiment, laser 6 is an Nd:YAG solid-state laser with an output wavelength of 1064 nm, a frequency of 0.1 Hz-10 Hz, a pulse width of 1-20 ns, and an energy of 100-1000 mJ. The driving laser 7 generated by laser 6 is focused by focusing lens 8 into a very small focal spot area, with its focal position precisely aligned with a specific area of ​​the gas target 9 (Xe gas in this embodiment) ejected by nozzle 13. When the high-power-density laser pulse interacts with the gas target 9, the gas atoms of the target are rapidly ionized, forming a high-temperature, high-density plasma. During the de-excitation process, this plasma radiates EUV light with a wavelength of 13.5 nm. The EUV radiation is collected by collecting mirror 10 and focused onto EUV energy meter 11.

[0038] The nozzle system is responsible for controlling the injection position of the target gas within the vacuum chamber. In this embodiment, the nozzle 13 is fixedly mounted on a three-axis displacement stage 12. The three-axis displacement stage 12 is controlled by a computer 22 and can move the spatial position of the nozzle 13 with sub-micron precision in the three orthogonal directions of X, Y, and Z, respectively, ensuring that the focusing point of the driving laser 7 always accurately acts on the optimal density region of the gas target 9, thereby obtaining the highest and most stable EUV conversion efficiency.

[0039] The gas supply system is responsible for providing the nozzle with target gas and buffer gas with precise and controllable flow, pressure and timing.

[0040] The target gas cylinder 14 stores high-purity Xe gas, serving as the core working medium for generating EUV plasma. The buffer gas cylinder 15 stores high-purity He gas. He gas acts as a buffer gas, its function being to limit the diffusion range of Xe gas within the vacuum chamber by colliding with Xe gas molecules, thereby increasing the local density of the gas jet in the laser focal region and effectively improving EUV conversion efficiency. Both gas streams are delivered to the corresponding inlets of the nozzle 13 via stainless steel hoses 16. A pressure reducing valve 17 and a solenoid valve 18 are installed in series on each gas delivery line. The pressure reducing valve 17 is controlled by analog signals or digital commands from the computer 22, allowing for precise adjustment of the output gas pressure. The solenoid valve 18 is controlled by timing signals, with opening and closing response times on the order of milliseconds, enabling precise control of the gas flow time. Through the coordination of pressure and on / off time, the amount of target gas participating in the reaction during each laser pulse can be stably controlled. The timing system is the core synchronization unit for generating stable pulses in the entire device. Because the duration of the driving laser pulse is extremely short (on the order of nanoseconds), and there is a millisecond-level transport delay from the opening of the solenoid valve to the arrival at the laser focal point, sub-microsecond or even higher precision timing coordination is required between the two. One channel of the signal generator 19 is connected to the trigger input of the laser 6 via a BNC cable 20 to control the emission time of the laser pulse. The other two channels of the signal generator 19 are connected to the control terminals of the two solenoid valves 18 via BNC cables 20 respectively, controlling the opening time of the two gas streams. The start-up delay between the target gas and the laser, and between the buffer gas and the laser, are defined as the time difference between the opening time of the solenoid valve and the laser pulse emission time. These delay parameters can be set by the computer 22, and the signal generator 19 executes precisely according to the set values.

[0041] The monitoring system provides real-time plasma status and EUV energy feedback information to the control system. The EUV energy meter 11 detects the EUV energy, and the EUV camera 21 captures plasma images.

[0042] The control system is responsible for data acquisition, model calculation, and command issuance. Computer 22 is connected via data cable 23 to pressure reducing valve 17, laser 6, EUV camera 21, EUV energy meter 11, barometer 4, thermometer 5, signal generator 19, and three-axis displacement stage 12. Computer 22 operates on the repetition frequency of laser pulses, performing the following cyclical tasks within each pulse cycle: reading and recording all control parameter data and monitoring parameter data; running machine learning prediction models, incremental PID target energy algorithms, and machine learning inference models; generating the latest control parameter adjustments; and issuing adjustment commands to each actuator. All historical data is stored in a local database on computer 22, serving as the data foundation for offline training and online updates of the machine learning model.

[0043] Please see Figure 2 , Figure 2 This is a flowchart illustrating the dose control method for a gas target laser-generated plasma extreme ultraviolet light source based on machine learning, as shown in the figure. The method operates on a control system and includes two working states: normal operation mode and fault protection safety mode. The two can be automatically switched according to the deviation of EUV pulse energy.

[0044] Setup and Startup 24: Before the system is officially put into operation, the operator inputs the following parameters through the human-computer interaction interface of computer 22: target dose (i.e., the cumulative EUV energy that the wafer is expected to receive per unit time) and initial control parameter set. This initial control parameter set includes: laser drive voltage V. laser Laser pulse width W laser Target gas pressure P Target Buffer gas pressure P Buffer Target gas ventilation time T Target Buffer gas ventilation time T Buffer Nozzle X-axis position nozzle Nozzle Y-axis position nozzle Nozzle Z-axis position Z nozzle Target material-laser delay T Target-laser Buffer gas-laser delay T Buffer-laser These initial values ​​can be set based on historical experience data or the parameter states remembered at the end of the last run. After setting, the system enters the standby state, and the operator issues the start command.

[0045] Target firing 25: The control system generates a target firing command sequence based on a set of currently valid control parameters. The signal generator 19 triggers the corresponding solenoid valves 18 and laser 6 sequentially according to the set timing logic. Target gas and buffer gas are first injected into the vacuum chamber 1, and then... Target-laser and T Buffer-laserAfter a delay, the driving laser 7 pulses out, focusing and bombarding the gas target 9 to generate plasma, which then radiates EUV light.

[0046] Measurement 26: After the target impact is completed, the EUV energy meter 11 collects and transmits the EUV energy E data of this pulse. Simultaneously, the EUV camera 21 captures plasma images, which are then processed in real-time by the computer 22 to extract the plasma length L. plasma and diameter D plasma Barometer 4 and thermometer 5 respectively transmit the internal air pressure P. chamber and cavity temperature T chamber .

[0047] All the above monitoring parameters are recorded and timestamped, and stored in the historical database for use in subsequent models. If any monitoring parameter exceeds the preset safe operating threshold range (such as a sudden increase in the chamber pressure exceeding the alarm value), the system will determine that a serious abnormality has occurred and immediately proceed to step 35 to execute the shutdown procedure to protect the equipment and product safety.

[0048] Mode selection 27: The system initially defaults to normal mode. Safety mode serves as a protective backup strategy, automatically taking over when control accuracy decreases or abnormal fluctuations occur in normal mode. The logic for mode switching is as follows: When in normal mode, if e i If the value remains above a certain threshold, the system will switch to safe mode. When in safe mode, if e i If the value remains below the threshold, switch to normal mode.

[0049] Prediction 28: The prediction model outputs multi-step predictions of EUV pulse energy based on recent historical control parameter data and monitoring parameter data.

[0050] Target Energy Setting 29: The PID strategy outputs multi-step target EUV pulse energy based on historical EUV pulse energy.

[0051] Inference 30: The inference model generates the adjustment amount of each control parameter based on the predicted multi-step EUV pulse energy, target energy, control parameter status, and monitoring parameter data.

[0052] Target Calculation 31: In safe mode, the PID strategy outputs the target EUV pulse energy based on the historical EUV pulse energy.

[0053] Energy adjustment judgment 32: When the target energy is greater than the modulation upper limit, the target energy is limited to the modulation upper limit; when the target energy is less than the modulation lower limit, a missed shot is triggered, and adjustment 34 and target shooting 25 are skipped.

[0054] Alarm 33: A system abnormality alarm is issued when the EUV pulse energy remains at the modulation limit for multiple consecutive times or when missed pulses are triggered multiple times in a short period of time. The user can manually enter shutdown 35 according to the system operating status.

[0055] Adjustment 34: In normal mode, adjust the control parameters according to the adjustment amount generated by the inference model; in safe mode, adjust the fast control parameters according to the adjustment weights and mathematical relationships. Repeat target practice 25 after adjustment.

[0056] Stop 35: The control system stops starting the laser and solenoid valve, awaiting user inspection. After the user has completed the inspection, they can manually enter setting and start 24.

[0057] Figure 3 This section demonstrates a comparison of the dose control performance of the apparatus and method in actual operation. The horizontal axis represents the number of pulses, and the vertical axis represents the normalized EUV pulse energy. During the first 1500 pulses, the dose control function of this invention was not enabled, and all control parameters remained at their initial settings. It can be clearly observed that the EUV pulse energy exhibits significant random fluctuations, with a significant difference between the peak and trough values, which is unacceptable for nanoscale lithography. Starting from the 1501st pulse, the dose control method of this invention was enabled, and the system initially operated in normal mode. It can be seen that the pulse energy rapidly converged to near the set energy value after a very short transition period (approximately several tens of pulses). In the subsequent 1500 pulses, the energy remained highly stable, and the fluctuation range was compressed to an extremely narrow interval. This indicates that the feedforward prediction and multi-parameter collaborative inference mechanism of this invention can effectively suppress the inherent randomness in the plasma generation process, achieving excellent dose stability.

[0058] Figure 4 The micro-decision-making process at the 1500th pulse (i.e., the moment control was just initiated) was captured. The first solid line in the figure represents the energy of the 5 historical pulses emitted before control was initiated, showing significant oscillations. The middle dashed line represents the predicted energy of the next 5 pulses based on recent data at that moment, indicating that without intervention, the subsequent energy will likely continue to fluctuate significantly. The bottom dotted line represents the target energy sequence of the next 5 pulses calculated by the incremental PID strategy based on the current cumulative dose error and its trend. It is clear that the target energy is not simply expected to reach the set value immediately, but rather a smooth and rapid convergence path from the current state to the set energy is planned, with each adjustment considering the system's inertia and stability. This "advance planning + predictive control" strategy is the essence of this invention compared to traditional single-step feedback methods.

[0059] The above embodiments illustrate the principles and implementation methods of specific examples and are only used to explain the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the core ideas of the present invention, there will be changes in the specific real-time devices and methods. Therefore, the content of this document should not be construed as limiting the present invention.

Claims

1. A dose control device for a gas target laser-generated plasma extreme ultraviolet light source based on machine learning, characterized in that, include: Vacuum system, used to provide the vacuum environment required for extreme ultraviolet light transmission; Laser systems are used to generate driving lasers to bombard target gas and produce plasma; A nozzle system for injecting target gas into the vacuum system; A gas supply system, connected to the nozzle system via a gas pipeline, is used to supply the target gas to the nozzle system and control the gas supply. A timing system, electrically connected to the laser system and the gas supply system respectively, is used to control the start-up time of the driving laser and the injection time of the target gas; The monitoring system is used to monitor the energy of extreme ultraviolet light and plasma state parameters in real time. as well as The control system is connected to the vacuum system, laser system, nozzle system, gas supply system, timing system, and monitoring system, respectively. The control system is a machine learning-based feedforward control system. The feedforward control includes: using a machine learning prediction model to make multi-step predictions of future extreme ultraviolet pulse energy; using a control strategy to generate a multi-step target energy sequence based on the prediction results and target energy; and using a machine learning inference model to generate adjustment amounts of multiple control parameters to perform feedforward adjustment on the subsequent target firing process.

2. The dose control device for gas target laser-generated plasma extreme ultraviolet light source according to claim 1, characterized in that, The control system is configured to switch to a safe mode of feedback control when the deviation between the actual value and the target value of the extreme ultraviolet pulse energy exceeds a preset threshold. In the safe mode, the target energy of the next pulse is calculated only, and an energy adjustment judgment and missed firing mechanism are executed.

3. The dose control device for gas target laser-generated plasma extreme ultraviolet light source according to claim 1, characterized in that, The vacuum system includes a vacuum chamber, a dry pump and a vacuum pump connected to the vacuum chamber via vacuum tubing, and a barometer and a thermometer mounted on the vacuum chamber. The pressure inside the vacuum chamber is maintained at 10. -4 Pa to 10 - 5 Pa level.

4. The dose control device for gas target laser-generated plasma extreme ultraviolet light source according to claim 1, characterized in that, The laser system includes a laser, a focusing lens disposed on the output optical path of the laser, and a collecting mirror disposed on the extreme ultraviolet radiation path; the laser is an Nd:YAG laser or a CO2 laser with adjustable driving voltage, laser pulse width and laser frequency.

5. The dose control device for gas target laser-generated plasma extreme ultraviolet light source according to claim 1, characterized in that, The gas supply system includes a target gas cylinder for storing target gas, a buffer gas cylinder for storing buffer gas, and a pressure reducing valve and a solenoid valve connected between the gas cylinder and the nozzle system; the target gas is Xe, O2 or CO2, and the buffer gas is He; the pressure reducing valve is used to adjust the target gas pressure and the buffer gas pressure, and the solenoid valve is used to control the target gas ventilation time and the buffer gas ventilation time.

6. The dose control device for gas target laser-generated plasma extreme ultraviolet light source according to claim 1, characterized in that, The nozzle system includes a nozzle and a three-axis displacement stage for adjusting the three-dimensional spatial position of the nozzle. The nozzle is connected to the gas pipeline outlet of the gas supply system.

7. The dose control device for gas target laser-generated plasma extreme ultraviolet light source according to claim 1, characterized in that, The timing system includes a signal generator, which is connected to the laser and the solenoid valve via BNC connectors to control the laser start-up time, the target gas, and the buffer gas opening and closing times, respectively.

8. The dose control device for gas target laser-generated plasma extreme ultraviolet light source according to claim 1, characterized in that, The monitoring system includes an extreme ultraviolet energy meter for real-time monitoring of extreme ultraviolet light energy, and an extreme ultraviolet camera for acquiring plasma images and analyzing them to obtain the plasma length and diameter.

9. The dose control device for a gas target laser-generated plasma extreme ultraviolet light source according to any one of claims 1-8, characterized in that, The control parameters recorded by the control system include: laser driving voltage, laser pulse width, target gas pressure, buffer gas pressure, target gas ventilation time, buffer gas ventilation time, nozzle position coordinates, target-laser start-up delay, and buffer gas-laser start-up delay; the monitoring parameters include: extreme ultraviolet light energy, plasma length, plasma diameter, and gas pressure and temperature inside the vacuum chamber.

10. A dose control method for a gas target laser-generated plasma extreme ultraviolet light source, characterized in that, The apparatus applied to any one of claims 1 to 9 comprises the following steps: Setup steps: Set the target dose and initial control parameters; Target firing steps: Based on the current control parameters, the control timing system sequentially triggers the gas supply system to spray target gas and the laser system to generate a driving laser, which bombards the target gas to produce extreme ultraviolet radiation. Measurement steps: Record the monitoring parameter data after target firing using the monitoring system; Mode selection steps: Based on the deviation between the actual value and the target value of the extreme ultraviolet pulse energy, select to enter normal mode or safe mode; In the normal mode, machine learning-based feedforward control is performed, and the feedforward control includes: Prediction steps: Using a machine learning prediction model deployed in the control system, based on historical control parameter data and monitoring parameter data, the extreme ultraviolet pulse energy corresponding to multiple future laser pulses is predicted; Target energy setting steps: Utilizing an incremental proportional-integral-differential strategy, a target energy sequence for multiple future pulses is output based on historical extreme ultraviolet pulse energies. This strategy obtains the window error by comparing the average pulse energy within the time window with the user-defined nominal energy, and then uses proportional, integral, and differential terms to set the future target energy; and Reasoning steps: Using a machine learning reasoning model deployed in the control system, based on the predicted multi-step extreme ultraviolet pulse energy, the target energy sequence, the current control parameter status, and monitoring parameter data, with the goal of reducing the deviation between the predicted energy and the target energy, the adjustment amount of each control parameter that meets the control parameter constraints is generated. Adjustment steps: Adjust the corresponding control parameters according to the adjustment amount, and return to the target shooting step.

11. The dose control method for a gas target laser-generated plasma extreme ultraviolet light source according to claim 10, characterized in that, The security modes include: Target calculation steps: Using the incremental proportional-integral-differential strategy, the target energy of the next pulse is output based solely on the historical extreme ultraviolet pulse energy; Energy adjustment judgment step: If the target energy of the next pulse is higher than the modulation upper limit, then it is limited to the modulation upper limit; if it is lower than the modulation lower limit, then the missed shot mechanism is triggered to skip the next shot. Alarm procedure: When the extreme ultraviolet pulse energy is repeatedly at the modulation upper limit or the missed pulse mechanism is frequently triggered, a system abnormality alarm is issued; and Safety adjustment procedure: Only fast control parameters that have a clear mathematical relationship with the extreme ultraviolet energy output and can respond quickly are adjusted. The adjustment weight of each fast control parameter is dynamically adjusted according to its current state and distance from its modulation boundary, and the adjustment value is calculated according to the adjustment weight and the mathematical relationship.

12. The dose control method for a gas target laser-generated plasma extreme ultraviolet light source according to claim 10, characterized in that, In the mode selection step: When in normal mode, if the deviation between the actual value of the extreme ultraviolet pulse energy and the target value continues to be greater than a threshold, then switch to safe mode. When in safe mode, if the deviation remains below the threshold, switch back to normal mode.

13. The dose control method for a gas target laser-generated plasma extreme ultraviolet light source according to claim 10, characterized in that, The measurement step is followed by a shutdown step, in which the control system stops starting the laser and the solenoid valve when any of the monitored parameters exceeds its preset safety threshold.

Images