Multi-target adaptive control system and method for plasma light source

Abstract

The invention relates to a plasma light source multi-target adaptive control system and method, and the system comprises a multi-dimensional data collection and fusion layer which is used for integrating data and target instructions from a theoretical simulation database, a real-time sensor array and a user interaction interface; the digital twinborn and AI model layer is used for constructing and operating a digital twinborn model of the light source based on the integrated data; the intelligent decision and real-time control layer is used for receiving the output of the deep neural network model and generating a control instruction based on a preset constraint and optimization algorithm; and the execution mechanism layer is used for responding to the control instruction and adjusting operation parameters of a light source. The method has the beneficial effects that through full-automatic intelligent optimization and control, a user does not need to have professional plasma physical knowledge, and the system can automatically complete complex parameter search and stable operation only by appointing an advanced performance target, so that the use threshold is greatly reduced, and the productivity is liberated.

Description

Technical Field

[0001] This invention belongs to the field of optoelectronic technology, and in particular relates to a multi-objective adaptive control system and method for a plasma light source. Background Technology

[0002] Plasma discharge technology, as a highly efficient means of generating 2-50nm radiation, has become a core light source in cutting-edge semiconductor manufacturing processes such as mask inspection, metrology, and photoresist research and development. Its working principle involves inducing a plasma circuit in a pre-filled gas using a pulsed high current, which then undergoes radial constriction under the influence of a self-generated magnetic field, thereby obtaining high-temperature, high-density plasma that radiates 2-50nm light.

[0003] The output performance of this light source mainly includes total power, brightness, spectral purity, and temporal stability. It is highly affected by the nonlinear coupling of multiple operating parameters. Key control parameters include: the static pressure and dynamic flow rate of the working gas (e.g., xenon), the doping ratio of auxiliary gases (e.g., nitrogen, argon, helium for stabilization or spectral modulation), the DC charging voltage and discharge repetition frequency of the pulse power supply, and the pulse waveform characteristics.

[0004] Currently, the performance tuning and optimization of such high-end light sources rely entirely on experienced application engineers. Engineers must manually adjust the aforementioned parameter combinations through a tedious trial-and-error process, based on the specific needs of the end application (e.g., high total power for photoresist exposure and high brightness for high-resolution imaging), to find a "recipe" that meets specific performance requirements. This approach has significant drawbacks:

[0005] 1. Inefficient and difficult to find the optimal solution: The high-dimensional parameter space and complex physical coupling make the manual optimization process time-consuming and prone to getting trapped in local optima, and cannot guarantee that a global Pareto optimal solution can be obtained.

[0006] 2. Reliance on expert experience, difficult to pass on: Debugging knowledge exists in the minds of individual engineers, with low standardization, which is not conducive to large-scale deployment and maintenance.

[0007] 3. Lack of adaptability: Light source components (such as coupling magnetic cores and optical windows) age over time, causing performance drift; manufacturing tolerances exist between different light source units. A fixed "formula" cannot adapt to this time-varying nature and individual differences, leading to performance degradation.

[0008] 4. Passive response, unable to predict instability: The inherent magnetohydrodynamic instabilities of plasma (such as magnetic Rayleigh-Taylor instability) can lead to spot jitter, power fluctuations, and even discharge failure. Existing technologies only make passive adjustments after the problem occurs, and cannot achieve predictive suppression.

[0009] Although numerous studies have explored the physics of light sources through theoretical simulations (such as the Flychk atomic physics code and CESZAR radiation magnetohydrodynamic models) and experimental diagnostics (such as measuring magnetic field strength with a Bdot probe and capturing light spots with a high-speed CCD), these results are isolated and have failed to form a closed-loop intelligent system capable of processing multi-source information in real time, learning complex correlations, and making optimal control decisions. Therefore, developing an adaptive control system capable of intelligent sensing, autonomous decision-making, and precise execution is crucial to overcoming the current bottlenecks in the application of 2-50nm light sources, maximizing their performance, and democratizing their operation. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-objective adaptive control system and method for plasma light sources.

[0011] Firstly, a multi-objective adaptive control system for a plasma source is provided, including:

[0012] A multi-dimensional data acquisition and fusion layer is used to integrate data and target instructions from theoretical simulation databases, real-time sensor arrays, and user interaction interfaces. The target instructions are used to define the output performance target of the light source.

[0013] The digital twin and AI model layer, connected to the data acquisition and fusion layer, is used to construct and run a digital twin model of the light source based on the integrated data. The digital twin model includes a deep neural network model for learning and characterizing the physical behavior of the light source.

[0014] The intelligent decision-making and real-time control layer, connected to the digital twin and AI model layer, is used to receive the output of the deep neural network model and generate control commands based on preset constraints and optimization algorithms. The control commands are used to perform composite control of the light source that combines feedforward and feedback.

[0015] The actuator layer, connected to the intelligent decision-making and real-time control layer, is used to respond to the control commands and adjust the operating parameters of the light source.

[0016] Preferably, the multi-dimensional data acquisition and fusion layer includes:

[0017] Theoretical simulation database is used to store data based on first-principles calculations and physical simulations;

[0018] A real-time sensor array is used to acquire electrical signals, plasma state signals, optical output signals, and gas and environmental parameter signals during the operation of the light source.

[0019] The user interaction and target definition module is used to receive performance targets defined by users in the form of semantic patterns or custom weighted combinations.

[0020] Preferably, the real-time sensor array includes:

[0021] An electrical diagnostic unit used to measure plasma loop current and its time derivative;

[0022] Plasma diagnostic unit for measuring plasma temperature and / or density;

[0023] An optical diagnostic unit used to measure output power in the 2-50nm band and / or acquire information on the spatial distribution of plasma spot;

[0024] An environmental sensing unit used to monitor gas flow rate, pressure, and temperature of critical components.

[0025] Preferably, the digital twin and AI model layer includes:

[0026] The feature engineering unit is used to preprocess the raw data from the data acquisition and fusion layer and extract feature parameters related to the plasma physical state.

[0027] A deep neural network model takes current control parameters, system state parameters and user target vector as input, and outputs at least one of the following: performance prediction value, stability risk assessment, control parameter adjustment suggestion and confidence level.

[0028] The model training and update engine is used to continuously train and update the deep neural network model using historical data and incremental data generated during online operation.

[0029] Preferably, the intelligent decision-making and real-time control layer includes:

[0030] An optimization decision-maker is used to solve the optimal control command sequence for multiple future control cycles based on the output of the deep neural network model, combined with system safety boundaries and user constraints, and employing a model predictive control algorithm.

[0031] A feedforward-feedback composite controller is used to generate and issue setpoint commands according to the optimal control command sequence, and to fine-tune the output based on real-time sensor feedback.

[0032] The anomaly detection and self-recovery unit is used to monitor the residuals between sensor data and model predictions. When an anomaly is detected, it triggers a safety protocol and initiates a recovery process.

[0033] Secondly, a multi-objective adaptive control method for a plasma source is provided, executed by any of the systems described in the first aspect, comprising:

[0034] S1. System initialization, loading pre-trained deep neural network models;

[0035] S2. Receive user-defined performance goals and constraints;

[0036] S3. Based on the current system state and the performance target, generate an initial set of control parameters through the deep neural network model and issue them for execution.

[0037] S4. During the operation of the light source, multi-dimensional sensor data is collected in real time and input into the deep neural network model to predict the performance and stability risks of the light source.

[0038] S5. Based on the prediction results, user objectives, and real-time feedback errors, dynamically calculate and execute the optimization adjustment of control parameters to achieve closed-loop control;

[0039] S6. Continuously collect operational data for online learning and updating of the deep neural network model.

[0040] Preferably, in step S5, the step of dynamically calculating and performing optimized adjustments to the control parameters includes:

[0041] Characteristic parameters characterizing compression dynamics are extracted from the real-time acquired plasma loop current waveform;

[0042] The feature parameters are compared with a stable benchmark predicted by the deep neural network model based on the current parameters;

[0043] Based on the comparison results, it is determined whether there is an instability risk in the plasma compression process, and corresponding gas pressure adjustment commands are generated to actively suppress instability.

[0044] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method as described in any of the second aspects.

[0045] Fourthly, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the method as described in any one of claims 7 to 9.

[0046] The beneficial effects of this invention are:

[0047] 1. This invention, through fully automatic intelligent optimization and control, eliminates the need for users to possess professional plasma physics knowledge. Users only need to specify advanced performance targets, and the system can automatically complete complex parameter searches and stable operation, greatly reducing the barrier to entry and liberating productivity.

[0048] 2. The system provided by this invention can deeply understand and quantify the inherent trade-offs between multiple performance indicators such as "power" and "brightness". Based on user-defined weights or combined objectives, it can autonomously find the global optimal balance point on the Pareto front, achieving true performance precision customization and output optimization.

[0049] 3. This invention possesses predictive stability control and strong adaptive capabilities. The system proactively identifies subtle signs of instability before they occur (such as enhancement of specific frequency components in the current waveform or minor distortions in the shape of the light spot) through an AI model, intervening in advance to mitigate risks (e.g., fine-tuning air pressure to alter compression dynamics) and ensuring stable output. Furthermore, it can automatically compensate for performance drift caused by equipment aging and individual differences through online learning. Simultaneously, the high-performance "recipes" and data generated during operation are structured and stored in a reusable knowledge base, making implicit experience explicit and persistent. Attached Figure Description

[0050] Figure 1 A schematic diagram of the overall architecture of the multi-objective adaptive control system for plasma light sources provided by the present invention;

[0051] Figure 2 A schematic diagram illustrating the structure and learning mechanism of the multi-source deep neural network model provided by this invention;

[0052] Figure 3 A flowchart of the adaptive control process based on multi-source data fusion provided by this invention;

[0053] Figure 4 The flowchart of the multi-objective adaptive control method for plasma light sources provided by the present invention is shown below.

[0054] Figure 5 This is a schematic diagram of the intelligent decision-making and real-time control layer provided by the present invention. Detailed Implementation

[0055] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0056] Example 1:

[0057] To address the technical problems of low efficiency, suboptimal performance, lack of self-adaptability, and difficulty in suppressing plasma instability caused by the reliance on manual adjustment of existing 2-50nm plasma light sources, this application provides a multi-objective adaptive control system for 2-50nm plasma light sources based on deep learning and digital twins. Its core lies in constructing a "digital twin" that can simulate, learn, and optimize the physical behavior of real light sources, and achieving precise control of the light source through a deep intelligent control loop.

[0058] Specifically, Embodiment 1 of this application provides a multi-objective adaptive control system for a plasma source, such as... Figure 1 As shown, it includes:

[0059] The multi-dimensional data acquisition and fusion layer is responsible for gathering all information related to the light source's state and performance, forming the data foundation for AI model training and decision-making. Specifically, this layer integrates data from theoretical simulation databases, real-time sensor arrays, and user interfaces with target instructions, which define the light source's output performance targets.

[0060] The multi-dimensional data acquisition and fusion layer includes: a theoretical simulation database, a real-time sensor array, and a user interaction and target definition module.

[0061] Theoretical simulation database, used to store data based on first-principles calculations and physical simulations; including but not limited to:

[0062] Atomic physics databases (such as NIST's FLYCHK and HULLAC) provide the ionization state distribution, spectral line emission coefficients, and radiation losses of different elements (Xe, Ne, etc.) at different electron temperatures and densities, as well as the ionization state / radiation coefficients of atomic physics models.

[0063] Results of plasma compression dynamics, instability growth rate, and energy conversion processes obtained from magnetohydrodynamic (MHD) simulation software (such as CESZAR, TRAC-II, and Helios).

[0064] A real-time sensor array is used to connect to the light source and beamline to acquire multi-dimensional physical signals in real time. Specifically, the real-time sensor array includes:

[0065] Electrical diagnostic unit: Includes Rogowski coil or Bdot probe for non-contact measurement of plasma loop current I(t) and its time derivative dI / dt; voltage probe for measuring pinch component voltage V(t), thereby deriving plasma loop inductance L and resistance R.

[0066] Plasma diagnostic unit: contains Langmuir probes to measure plasma temperature and plasma density.

[0067] Optical diagnostic unit: Includes multilayer reflectors and filters as well as photodiodes, used for precise measurement of transient and average power in the 2-50nm band (or other target bands); equipped with a pinhole camera with EUV or X-ray CCD, used to acquire the two-dimensional spatial intensity distribution, size (e.g., full width at half maximum, FWHM), shape ellipticity, and centroid position of the plasma spot; visible light CCD is used to monitor the discharge area and identify abnormal arcs and sputtering.

[0068] Gas and environment sensing unit: including actual gas flow rate fed back by high-precision mass flow controller (MFC), static and dynamic pressure of discharge chamber measured by capacitive thin-film vacuum gauge, temperature of key components (such as magnetic core and window) measured by thermocouple, and cooling water flow rate and temperature.

[0069] The user interaction and target definition module receives performance targets defined by the user using semantic patterns or custom weighted combinations. Specifically, it provides a graphical or programmable interface, allowing users to output targets using advanced semantic definitions. Targets can be preset to "high power mode," "high brightness mode," or defined by the user as a weighted combination of specific performance metrics (e.g., "0.7"). Power +0.3 Brightness), and can be subject to additional constraints (such as "spot size <300 μm", "power fluctuation <2%RMS") and can provide performance feedback scoring.

[0070] Furthermore, the data sources for the multi-dimensional data acquisition and fusion layer include: theoretical simulation data (obtained through the theoretical simulation database), real-time sensing data (obtained through the real-time sensor array), user interaction data (obtained through the user interaction and target definition module), and historical experimental data (including laboratory benchmark test data, equipment factory calibration data, and user historical operation data).

[0071] The digital twin and AI model layer, connected to the data acquisition and fusion layer, is used to construct and run a digital twin model of the light source based on the integrated data. The digital twin model includes a deep neural network model for learning and characterizing the physical behavior of the light source.

[0072] The digital twin and AI model layer is the intelligent core of the system, responsible for constructing and running the "digital twin" of the light source, which specifically includes:

[0073] The feature engineering unit preprocesses the raw sensing data and extracts features with clear physical meaning. For example: extracting peak current, current rise time, and the onset of the "inductance sag" (the point at which the current begins to drop sharply) from the current waveform; extracting FWHM, centroid coordinates, and image moments (for shape characterization) from the spot image; and calculating the instantaneous input power. ; Calculate plasma loop parameters The plasma density and temperature were calculated using the Langmuir probe, and the pinch current was then deduced. .

[0074] Deep neural network models are multi-task learning models trained on massive amounts of data. Their inputs include: the current set of adjustable control parameters, current system state parameters (estimated aging coefficients, temperature, etc.), and the user's target vector. Their outputs include: (a) performance predictions (total power, brightness, spot size); (b) stability risk assessments (outputting one or more instability indices); (c) recommended control parameter adjustments (ΔV, ΔF, ΔP, etc.) to optimize the current state; and (d) confidence in the current prediction. This model learns a complex nonlinear mapping from the input space to the output space. The architecture of a deep neural network model includes a theoretical knowledge encoder, an experimental data encoder, a real-time sensor encoder, a multimodal feature fusion layer, and a multi-task prediction and decision layer.

[0075] The model training and update engine is responsible for the continuous evolution of the DNN model. Specifically, it uses historical experimental data, theoretical simulation data, and incremental data generated during online system operation to periodically or trigger-based retraining and fine-tuning of the model, enabling the "digital twin" to track the physical state drift of the real light source and maintain prediction accuracy.

[0076] The data fusion processing methods performed by the digital twin and AI model layers include: multi-source data alignment and time synchronization / spatial registration; feature extraction and fusion of physical and statistical features; data quality assessment and anomaly detection / completion.

[0077] The digital twin and AI model layer also features a continuous learning mechanism, including incremental learning modules, federated learning support, knowledge distillation mechanisms, and model version management.

[0078] A schematic diagram of the learning mechanism of a multi-source deep neural network model is shown below. Figure 2 As shown, for example, in Figure 2In this training, the data source adopts a four-dimensional fusion approach, which includes theoretical physics data, simulation data, experimental history data, and user operation data. Theoretical calculation data includes the principle physics database of ionization cross sections such as Xe, Ni, and Ne, radiation transport models, and instability growth models of the magnetohydrodynamic equations. Simulation data is based on particle simulation of microscopic particle dynamics, fluid simulation of macroscopic plasma behavior, and radiation transfer simulation of spectra and spatial distribution. Experimental history data includes laboratory calibration data, benchmark performance curves, diagnostic equipment data, multi-sensor correlation, and fault case library with abnormal pattern recognition. User operation data includes application mode data detection / calculation, performance feedback data, user satisfaction scores, and environmental adaptation data for different environments.

[0079] Figure 2 The continuous learning process includes an offline pre-training phase, an online fine-tuning phase, a federated learning phase, and a knowledge transfer phase. The offline pre-training phase includes: training basic physical understanding using TD1+TD2+TD3; self-supervised learning and masked data reconstruction; and contrastive learning to compare positive and negative samples. The online fine-tuning phase includes: real-time data stream learning and continuous input of TD4, adaptive learning rate to balance new and old data; catastrophic forgetting prevention and elastic weight consolidation. The knowledge transfer phase includes cross-application scenario transfer (e.g., from metrology to testing), cross-device type transfer (e.g., from model A to model B), and cross-condition transfer (e.g., from laboratory to production line).

[0080] The architecture of the multi-source deep neural network model includes a multi-task prediction module, physics-guided learning, a feature fusion mechanism, and parallel encoding channels. The multi-task prediction module comprises a performance prediction head regression network, a stability prediction head classification network, a control recommendation head policy network, and an uncertainty estimation head variance network. The physics-guided learning mechanism incorporates physical constraints such as energy / momentum conservation, symmetry constraints such as spatiotemporal invariance, causal constraints such as temporal causality, and uncertainty quantification via Bayesian neural networks. The feature fusion mechanism includes cross-modal attention, feature map concatenation, knowledge distillation fusion, and adaptive weighting. The parallel encoding channels include a theoretical data encoder (CNN + Attention), a simulation data encoder (LSTM + GraphNN), an experimental data encoder (Transformer), and a user data encoder (MLP + Embedding).

[0081] Furthermore, model outputs and applications include high-precision prediction, robust decision-making, adaptive control, and knowledge accumulation. Updated models are used to control applications in real time, and the acquired performance data (such as new experimental / user data) is then fed back into the aforementioned continuous learning process.

[0082] The intelligent decision-making and real-time control layer, connected to the digital twin and AI model layer, is used to receive the output of the deep neural network model and generate control commands based on preset constraints and optimization algorithms. The control commands are used to perform composite control of the light source that combines feedforward and feedback.

[0083] The actuator layer, connected to the intelligent decision-making and real-time control layer, is used to respond to the control commands and adjust the operating parameters of the light source.

[0084] Example 2:

[0085] Based on Example 1, Example 2 of this application provides a more specific multi-objective adaptive control system for an inductively coupled plasma source, including:

[0086] A multi-dimensional data acquisition and fusion layer is used to integrate data and target instructions from theoretical simulation databases, real-time sensor arrays, and user interfaces. The target instructions are used to define the output performance target of the light source.

[0087] The digital twin and AI model layer, connected to the data acquisition and fusion layer, is used to construct and run a digital twin model of the light source based on the integrated data. The digital twin model includes a deep neural network model for learning and characterizing the physical behavior of the light source.

[0088] The intelligent decision-making and real-time control layer, connected to the digital twin and AI model layer, is used to receive the output of the deep neural network model and generate control commands based on preset constraints and optimization algorithms. The control commands are used to perform composite control of the light source that combines feedforward and feedback.

[0089] The intelligent decision-making and real-time control layer transforms the "suggestions" from the AI ​​model into safe, feasible, and optimal control actions. This layer includes: an optimization decision-maker, a feedforward-feedback composite controller, and an anomaly detection and self-recovery unit.

[0090] An optimization decision maker is used to solve for the optimal control command sequence over multiple future control cycles based on the output of the deep neural network model, combined with system safety boundaries and user constraints, using a model predictive control algorithm.

[0091] Specifically, the optimization decision-maker receives the "recommended control input" and "predictive performance" output by the AI ​​model. Combining predefined system safety hard boundaries (such as maximum allowable voltage and maximum safe air pressure) and user-defined soft constraints, it uses model predictive control (MPC) or other fast optimization algorithms to solve for a series of optimal control commands over several future control cycles, ensuring that the control commands are as close as possible to the user-defined target while satisfying all constraints.

[0092] The formula for optimizing the decision-maker can be expressed as:

[0093]

[0094] The constraint function is:

[0095]

[0096]

[0097]

[0098]

[0099] The feedback correction mechanism is as follows:

[0100]

[0101] When constraints conflict, slack variables are introduced.

[0102]

[0103]

[0104] The feedforward-feedback composite controller is used to generate and issue setpoint commands according to the optimal control command sequence, and to fine-tune the output based on real-time sensor feedback.

[0105] Specifically, feedforward control involves directly and rapidly calibrating actuators such as pulse power supplies and gas valves based on the setpoint provided by the optimization decision-maker, thereby achieving proactive response to changes in the target.

[0106] Feedback control involves continuously fine-tuning the feedforward output by utilizing the deviation between real-time sensor data (especially the photodiode power signal and the spot centroid position) and the setpoint to suppress random disturbances and slow drift not covered by the model. For example, a closed-loop controller can be designed to dynamically fine-tune the trigger timing based on the spot centroid offset.

[0107] The anomaly detection and self-recovery unit continuously monitors the residuals between all sensor readings and AI model predictions. When the residual exceeds a dynamic threshold (e.g., a visible light CCD detects strong abnormal luminescence, or a sudden drop in power signal), the unit immediately triggers a safety protection protocol (such as pausing discharge) and initiates a root cause analysis process. The AI ​​model is then used to infer the most likely cause of the failure (e.g., breakdown caused by a tiny debris) and generate a sequence of attempted recovery instructions (e.g., executing several low-energy cleaning pulses) or directly prompting for necessary maintenance operations.

[0108] The actuator layer, connected to the intelligent decision-making and real-time control layer, is used to respond to the control commands and adjust the operating parameters of the light source.

[0109] Specifically, the actuator layer includes a high-stability pulse power modulator, a valve-controlled gas delivery system, a vacuum pump control unit, and an auxiliary magnetic field coil drive, which are responsible for executing digital instructions from the control layer with high fidelity.

[0110] In addition, through Figure 3 This paper describes the adaptive control workflow of multi-source data fusion in this application. First, system initialization is performed, including data source connection (theoretical database connection, historical data loading, sensor calibration, and user configuration reading); loading a pre-trained multi-source DNN model; and establishing a digital twin baseline. Next, multi-objective definitions are performed, including application scenario selection, performance weight setting, constraint definition, and mask detection metrology. The total power weight is set to 0.7, brightness weight to 0.3, stability weight to 0.5, power ≥ 15W, fluctuation < 3%, and spot size < 300μm. Then, initial parameter optimization is performed, including multi-source AI inference, physical constraint verification, and safety boundary detection. Theoretical models, historical data, and user preferences are fused, and energy conservation / magnetic diffusion / radiation constraints are verified. Finally, the safe ranges for voltage / air pressure / temperature are checked. Afterward, the adaptive control main loop is executed using the above initial parameters, including real-time data acquisition and fusion, digital twin state update, AI multi-task prediction, multi-objective decision optimization, adaptive control execution, and performance evaluation learning. Real-time data acquisition and fusion include multi-sensor synchronization, theoretical data correlation, historical data comparison, and user feedback integration. Digital twin state updates include plasma state estimation, equipment aging assessment, and environmental change supplementation. AI multi-task prediction includes short-term performance prediction (next pulse), stability risk assessment (calculating instability index), long-term trend prediction (performance degradation), and control strategy recommendation (parameter adjustment). Multi-objective decision optimization includes model predictive control, Pareto optimality search, constraint satisfaction adjustment, and risk balance prediction. Adaptive control execution includes model-based feedforward, error-based feedback, learning-based adaptation, and disturbance rejection robustness. Performance evaluation learning includes data collection and annotation, online incremental learning, model performance evaluation, and knowledge base updates.

[0111] Furthermore, the aforementioned adaptive control workflow also includes anomaly handling and maintenance upgrades. Anomaly handling includes: anomaly detection and classification, meme analysis, recovery strategy execution, and experience learning. Anomaly detection includes: sensor anomalies / performance mutations; meme analysis is based on an AI fault diagnosis model; recovery strategy execution includes safety protocols and recovery procedures; and experience learning includes adding to the fault database. Maintenance and upgrades include: regular maintenance, model retraining, system performance verification, and knowledge transfer application.

[0112] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.

[0113] Example 3:

[0114] Based on Example 2, Example 3 of this application provides a multi-objective adaptive control method for a plasma source, including:

[0115] S1. System initialization: Load the pre-trained deep neural network model and establish data links with all hardware.

[0116] S2. Receive user-defined performance goals and constraints.

[0117] S3. Based on the current system state sensor readings, estimated parameters, and the performance target, generate an initial set of control parameters through the deep neural network model and issue them for execution.

[0118] S4. During the operation of the light source, multi-dimensional sensor data is collected in real time and input into the deep neural network model to predict the performance and stability risks of the light source.

[0119] Specifically, the light source starts operating. A multi-dimensional data acquisition layer synchronously and at high speed collects all sensor data. The feature engineering unit processes the raw data in real time, extracts key features, and inputs them into the AI ​​model. Based on these real-time features, the AI ​​model periodically predicts the light source's output performance and stability trends within a short future timeframe (such as the next pulse or a few milliseconds).

[0120] S5. Based on the prediction results, user objectives, and real-time feedback errors, dynamically calculate and execute the optimization adjustment of control parameters to achieve closed-loop control.

[0121] In S5, the step of dynamically calculating and performing optimized adjustments to the control parameters includes:

[0122] Characteristic parameters characterizing compression dynamics are extracted from the real-time acquired plasma loop current waveform;

[0123] The feature parameters are compared with a stable benchmark predicted by the deep neural network model based on the current parameters;

[0124] Based on the comparison results, it is determined whether there is an instability risk in the plasma compression process, and corresponding gas pressure adjustment commands are generated to actively suppress instability.

[0125] Specifically, the optimization decision-maker compares the model's predicted values ​​with the user's target values, while taking into account real-time feedback errors, calculates new and better control parameter adjustment instructions, and issues them for execution through the feedforward-feedback controller to achieve dynamic closed-loop optimization of plasma size and light source output.

[0126] For example, the plasma current waveform is analyzed to extract the onset time feature of the "inductance drop". If the onset time is earlier than the stable reference time predicted by the AI ​​model based on the current parameters, it is determined that there is a risk of excessive compression, and an adjustment command to increase the working gas pressure is generated; otherwise, an adjustment command to decrease the pressure is generated. The magnitude of the adjustment command is determined by the AI ​​model based on the current time deviation, historical unstable event correlation data, and the jitter amplitude of the spot image.

[0127] S6. Continuously collect operational data for online learning and updating of the deep neural network model.

[0128] Specifically, if an operational anomaly is detected or the user actively changes the target, the system will immediately switch to the anomaly handling process or return to step S2 / S3 to achieve target reconfiguration and rapid re-optimization.

[0129] The following two more specific examples illustrate the method provided in this application.

[0130] In one example, a plasma light source for detecting mask defects in the 2-50nm range is used to illustrate how the system of the present invention can achieve seamless switching from "high-throughput screening mode" to "high-resolution detailed inspection mode".

[0131] The plasma source initially operates in "high-throughput screening mode," which is automatically optimized by the system's AI to target maximum total power. Current operating parameters are: voltage 380V, frequency 1900 Hz, xenon pressure 85 mTorr, and output power stable at 18.5W.

[0132] Target switching: The operator switches the target from "high power mode" to "high brightness mode" in the software interface, aiming to perform high-resolution defect imaging.

[0133] Intelligent decision-making involves an immediate response from the intelligent decision-maker. The AI ​​model rapidly extrapolates within the digital space based on the current system state (coupled core temperature, number of pulses already executed) and the new "high brightness" target. The model retrieves historical data and, combined with physical laws, outputs a recommended formula within milliseconds: voltage 395V, frequency 2500 Hz, xenon pressure 115 mTorr. After the optimized decision-maker verifies that these parameters are within the equipment's safe range, it issues the command.

[0134] The execution and monitoring steps are as follows: The actuator layer completes parameter switching within 100 milliseconds. The light source begins operating under the new parameters. The current waveform monitored in real time by the Bdot probe shows that the negative peak of dI / dt occurs slightly earlier than the ideal time point predicted by the model. At the same time, the CCD shows that the spot size (FWHM) is 280 μm.

[0135] Dynamic optimization involves the AI ​​model receiving two real-time features: "premature compression" and "spot size." The instability risk prediction module within the model is activated, assessing that maintaining the current state increases the probability of spot jitter in subsequent pulses. The optimization decision-maker then calculates a fine-tuning scheme: while keeping the voltage frequency constant, the xenon pressure is slightly reduced to 110 mTorr. This adjustment aims to slightly slow down the compression rate of the plasma column to achieve more stable clamping.

[0136] Stable output: After air pressure adjustment, the current waveform characteristics returned to normal, and the spot size stabilized at around 300μm. The brightness output in the 2-50nm band monitored by the photodiode reached and stabilized at the highest level in the system's history, fully meeting the requirements of high-resolution imaging. The entire mode switching and optimization process was completed automatically within seconds, requiring no manual intervention.

[0137] In another example, the system is described as combining real-time diagnostic signals with historical user feedback to create and solidify a "custom recipe" for a specific long-term stability requirement through online learning.

[0138] This example specifically includes:

[0139] S1. Complex requirements: A user is developing a photoresist process and has put forward a demanding requirement: while ensuring that the power of 2-50nm is not less than 15W, the user wants to achieve the highest possible brightness to improve the roughness of the line edges, and requires that the power fluctuation (RMS) be less than 3% during continuous 8 hours of operation.

[0140] In S1, after the user inputs this custom target, the system selects several similar "high brightness" recipes from the historical database to begin exploration. Running under an initial recipe (V=392V, F=2450Hz, P_Xe=108 mTorr), the data acquisition layer synchronously records the following:

[0141] Bdot signal: The calculated start time of the "inductance drop" was 2.5 ns earlier than the model-predicted stable baseline (Δt = -2.5 ns).

[0142] CCD image: The standard deviation of the centroid of the continuous pulse spot is 6 μm, indicating obvious jitter.

[0143] User database association: The system retrieved three past feedback records for this user account regarding "excessive short-term power fluctuations".

[0144] S2. AI Comprehensive Assessment and Risk Prediction: The AI ​​model simultaneously inputs three features—negative Δt, spot jitter, and historical user feedback—into its stability prediction subnetwork. The model outputs the conclusion that, under the current parameters, the plasma compression dynamics are on the verge of instability, and the probability of long-term operation leading to performance degradation (brightness decrease, excessive fluctuations) is as high as 70%.

[0145] S3. Multi-objective Collaborative Optimization Decision: The optimization decision-maker faces a multi-objective optimization problem: maximizing brightness under the hard constraint of power ≥15W and the soft objective of minimizing fluctuations. The decision-maker performs rapid calculations and proposes an innovative solution: increasing the xenon gas pressure to 112 mTorr to stabilize compression dynamics, while slightly reducing the frequency to 2400 Hz to compensate for potential power loss due to the increased gas pressure, and introducing a very small voltage feedback gain for dynamic fine-tuning of ±5V based on real-time power readings.

[0146] S4. Closed-loop execution and verification: The system executes the new scheme. After adjustment, Δt is improved to -0.8 ns, and spot jitter is reduced to 2 μm. Although the peak brightness of a single pulse is slightly reduced due to the slight decrease in frequency, the overall performance of long-term average brightness and power stability is significantly better than before the adjustment due to the greatly improved stability. Power is stable at 15.3W ± 2.5%.

[0147] S5. Knowledge Accumulation and Model Evolution: This successful "stabilization" adjustment was automatically marked as a high-quality data sample by the system. Its complete "initial state - adjustment action - final result" data chain was sent to the model training engine. The engine initiated a round of online incremental learning, specifically reinforcing the rule in the DNN model that "when early signs of compression (negative Δt) and high user stability requirements occur simultaneously, the strategy of 'increasing air pressure and fine-tuning frequency for power compensation' should be prioritized." From then on, a custom "high-brightness long-term stability" formula optimized for this user's long-term stability needs was officially incorporated into the system's knowledge base and can be used for all similar tasks of this user in the future, or as a high-quality starting point for other users seeking similar performance. This marks the system's transition from the "automation" stage of executing preset rules to the "intelligent" stage capable of creating new strategies.

[0148] It should be noted that the method provided in this embodiment is the corresponding method of the system provided in embodiment 2. Therefore, the parts that are the same as or similar to those in embodiment 2 in this embodiment can be referred to each other, and will not be described again in this application.

Claims

1. A multi-objective adaptive control system for a plasma light source, characterized in that, include: A multi-dimensional data acquisition and fusion layer is used to integrate data and target instructions from theoretical simulation databases, real-time sensor arrays, and user interaction interfaces. The target instructions are used to define the output performance target of the light source. The digital twin and AI model layer, connected to the data acquisition and fusion layer, is used to construct and run a digital twin model of the light source based on the integrated data. The digital twin model includes a deep neural network model for learning and characterizing the physical behavior of the light source. The intelligent decision-making and real-time control layer, connected to the digital twin and AI model layer, is used to receive the output of the deep neural network model and generate control commands based on preset constraints and optimization algorithms. The control commands are used to perform composite control of the light source that combines feedforward and feedback. The actuator layer, connected to the intelligent decision-making and real-time control layer, is used to respond to the control commands and adjust the operating parameters of the light source.

2. The multi-objective adaptive control system for a plasma source according to claim 1, characterized in that, The multi-dimensional data acquisition and fusion layer includes: Theoretical simulation database is used to store data based on first-principles calculations and physical simulations; A real-time sensor array is used to acquire electrical signals, plasma state signals, optical output signals, and gas and environmental parameter signals during the operation of the light source. The user interaction and target definition module is used to receive performance targets defined by users in the form of semantic patterns or custom weighted combinations.

3. The multi-objective adaptive control system for a plasma source according to claim 2, characterized in that, The real-time sensor array includes: An electrical diagnostic unit used to measure plasma loop current and its time derivative; Plasma diagnostic unit for measuring plasma temperature and / or density; An optical diagnostic unit used to measure output power in the 2-50nm band and / or acquire information on the spatial distribution of plasma spot; An environmental sensing unit used to monitor gas flow rate, pressure, and temperature of critical components.

4. The multi-objective adaptive control system for a plasma source according to claim 3, characterized in that, The digital twin and AI model layer includes: The feature engineering unit is used to preprocess the raw data from the data acquisition and fusion layer and extract feature parameters related to the plasma physical state. A deep neural network model takes current control parameters, system state parameters and user target vector as input, and outputs at least one of the following: performance prediction value, stability risk assessment, control parameter adjustment suggestion and confidence level. The model training and update engine is used to continuously train and update the deep neural network model using historical data and incremental data generated during online operation.

5. The plasma source multi-objective adaptive control system according to claim 4, characterized in that, The intelligent decision-making and real-time control layer includes: An optimization decision-maker is used to solve the optimal control command sequence for multiple future control cycles based on the output of the deep neural network model, combined with system safety boundaries and user constraints, and employing a model predictive control algorithm. A feedforward-feedback composite controller is used to generate and issue setpoint commands according to the optimal control command sequence, and to fine-tune the output based on real-time sensor feedback. The anomaly detection and self-recovery unit is used to monitor the residuals between sensor data and model predictions. When an anomaly is detected, it triggers a safety protocol and initiates a recovery process.

6. A multi-objective adaptive control method for a plasma source, characterized in that, Performed by the system according to any one of claims 1 to 5, comprising: S1. System initialization, loading pre-trained deep neural network models; S2. Receive user-defined performance goals and constraints; S3. Based on the current system state and the performance target, generate an initial set of control parameters through the deep neural network model and issue them for execution. S4. During the operation of the light source, multi-dimensional sensor data is collected in real time and input into the deep neural network model to predict the performance and stability risks of the light source. S5. Based on the prediction results, user objectives, and real-time feedback errors, dynamically calculate and execute the optimization adjustment of control parameters to achieve closed-loop control; S6. Continuously collect operational data for online learning and updating of the deep neural network model.

7. The multi-objective adaptive control method for a plasma source according to claim 6, characterized in that, In S5, the step of dynamically calculating and performing optimized adjustments to the control parameters includes: Characteristic parameters characterizing compression dynamics are extracted from the real-time acquired plasma loop current waveform; The feature parameters are compared with a stable benchmark predicted by the deep neural network model based on the current parameters; Based on the comparison results, it is determined whether there is an instability risk in the plasma compression process, and corresponding gas pressure adjustment commands are generated to actively suppress instability.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 6 to 7.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 7 to 9.

Images