Lithography aberration dynamic reference anti-noise optimization method and system

Abstract

The invention provides a photoetching aberration dynamic reference anti-noise optimization method and a photoetching aberration dynamic reference anti-noise optimization system, and the photoetching aberration dynamic reference anti-noise optimization method based on a gating time sequence convolutional network comprises the following core processes: synchronously collecting real-time wavefront signals and dynamic reference signals of aspherical lens surface measurement points; intercepting latest 50-millisecond window data and inputting the latest 50-millisecond window data into the gating time sequence convolutional network; time sequence features are extracted through causal expansion convolution, an anti-noise weight vector is dynamically generated by means of a Sigmoid gating unit, the weight vector is generated through online learning parameter regulation and control, and the learning rate is set to be 0.01; performing iterative optimization on the dynamic reference signal according to the anti-noise weight vector, and outputting the dynamic reference signal to an aberration compensator to correct the aberration of the aspherical lens; if five continuous windows of the compensation residual standard deviation are larger than 0.05 times of the wavelength of the photoetching light source, only the weight of the last layer of the gating time sequence convolution network is updated, and an anti-noise weight vector is regenerated.

Description

Technical Field

[0001] This invention relates to the field of photolithography technology, specifically to a method and system for optimizing the dynamic reference noise reduction of photolithographic aberrations, and more particularly to a method for optimizing the dynamic reference noise reduction of photolithographic aberrations based on gated temporal convolutional networks. Background Technology

[0002] In the manufacturing and production of aspherical lenses using lithography machines, aberration compensation accuracy directly determines the lens's imaging quality and production yield. Temporal correlation modeling of dynamic reference signals and real-time wavefront signals is the core technological foundation for achieving accurate aberration compensation. Currently, aberration compensation technology in the field of lithography lens manufacturing generally relies on the aforementioned correlation modeling scheme, but significant technical bottlenecks exist in practical production applications, specifically as follows: a. Insufficient noise suppression capability, unable to adapt to high-frequency interference in production scenarios: Existing aberration compensation schemes used for lithography lens manufacturing (such as patent document CN114325987A) use fixed bandwidth filters (upper limit 1.2kHz) to process dynamic reference signals, which are difficult to effectively filter out high-frequency vibration noise (>1.5kHz) present in the lithography machine production environment. In the actual test scenario of producing aspherical lenses with the MT1217 lithography machine, when the stage acceleration of the lithography machine is >0.5g, the vibration noise energy is concentrated in the 1.8-2.2kHz frequency band (accounting for more than 65% of the total noise power), directly causing distortion of the time-domain correlation between the dynamic reference signal and the wavefront signal. This issue caused the standard deviation of the aberration compensation residual to exceed the standard by 30% (measured value 0.08λ vs target 0.05λ, where λ is the wavelength of the lithography light source in the lithography machine, with a value of 13.5nm), which in turn caused local curvature distortion of the aspherical lens, resulting in a 5.2% decrease in lens production yield (data source: IMEC 2022 Semiconductor Lens Manufacturing Industry Report).

[0003] b. Rigid window mechanism and delayed compensation affect production accuracy: Traditional iterative optimization schemes for aberration compensation (such as patent document CN112860021B) rely on long time window statistics and use a 200ms sliding window to process signals. Although this can smooth noise to some extent, it introduces a significant delay of >1.5ms. In the lithography production scenario of 10nm process aspherical lenses, the aberration disturbance response time must meet the stringent requirement of <800μs (reference standard: SPIE Vol. 12149, 2022). The delayed compensation action caused by the long window cannot cope with sudden airflow disturbances (step changes <50ms) during production, resulting in untimely correction of lens aberrations and affecting the stability of production accuracy.

[0004] c. Imbalance in noise resistance model complexity, making it difficult to match the real-time requirements of production: Deep learning-based aberration compensation schemes used for lithography lens manufacturing (such as the patent document with publication number US20220155482A1) use LSTM networks to build noise resistance models, with a parameter count >50K, inference latency >1.2ms, and require full-window retraining (time >5ms), which cannot meet the real-time requirements of aberration compensation in lens production; while lightweight noise resistance schemes (such as JP2021152786A) only use FIR filters, which cannot dynamically adapt to the dynamic changes in the noise spectrum in the production environment, resulting in zero suppression capability for noise above 2kHz, which also cannot guarantee the compensation accuracy of lens production.

[0005] In summary, in the practical application of lithography machine manufacturing and production of aspherical lenses, existing aberration compensation technologies cannot simultaneously meet the triple constraints of high-frequency noise suppression (>2kHz), ultra-low latency (≤800μs), and lightweight deployment (<5K parameters). This is because they lack a noise-sensing optimization mechanism for short time windows of 50ms, making it difficult to adapt to the dynamic and complex interference environment and the high-precision, high-real-time compensation requirements during lithography lens production. Therefore, there is an urgent need for an efficient aberration compensation optimization technology adapted to the lithography lens manufacturing scenario to overcome the aforementioned technical bottlenecks and ensure lens production yield and accuracy.

[0006] This invention achieves dynamic noise-resistant weight generation with microsecond-level response for the first time through the causal dilated convolution and gating mechanism of GTCN, breaking through the above-mentioned technical bottlenecks. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for optimizing the noise reduction of dynamic references for photolithographic aberrations.

[0008] A method for noise reduction optimization of dynamic reference for photolithographic aberrations according to the present invention includes the following steps: Step S1: Synchronously acquire the real-time wavefront signal and dynamic reference signal of the measurement point on the surface of the aspherical lens, and extract the real-time wavefront signal and dynamic reference signal of the latest preset time length window to form input window data; Step S2: Input the input window data obtained in step S1 into the gated temporal convolutional network GTCN, extract the temporal features of the input window data through the causal dilation convolution in the gated temporal convolutional network GTCN, and dynamically generate the noise-resistant weight vector W using the sigmoid gate unit in the gated temporal convolutional network GTCN. Step S3: Based on the noise-resistant weight vector W generated in step S2, perform element-wise multiplication iterative optimization on the dynamic reference signal acquired in step S1 to obtain the optimized dynamic reference signal. Step S4: Output the optimized dynamic reference signal obtained in step S3 to the aberration compensator to drive the aberration compensator to generate a compensation wavefront to correct the aberration of the aspherical lens. Step S5: Calculate the standard deviation of the compensation residual after aberration correction in step S4; when the standard deviation of the compensation residual corresponding to N consecutive preset time length windows is greater than the preset threshold, trigger online incremental update, and return to step S2 to regenerate the noise-resistant weight vector W, then execute steps S3 to S4 in sequence to complete a new round of dynamic reference signal optimization and aberration correction; when the standard deviation of the compensation residual corresponding to N consecutive preset time length windows is less than or equal to the preset threshold, keep the weight parameters of the current gated temporal convolutional network GTCN unchanged, and continue to execute the optimization and correction process of steps S1 to S4.

[0009] Preferably, in step S1, the preset time length window is a 50ms window.

[0010] Preferably, in step S2, the elements of the noise reduction weight vector satisfy w t =σ(α·x t +β·Δx t ); Where σ(.) is the Sigmoid function, used to restrict the noise-resistant weights to the interval 0~1; α and β are the online learning parameters of the gated temporal convolutional network GTCN; t is the discrete time step index within the current preset time length window; x t The input signal takes the value at time step t; Δx t Let Δx be the first-order difference of the input signal. t =x t -x t-1 .

[0011] Preferably, the learning rate η of the online learning parameters of the gated temporal convolutional network GTCN is 0.01.

[0012] Preferably, in step S5, N is 5; the preset threshold is 0.05λ, where λ is the wavelength of the light source used for photolithography.

[0013] Preferably, the gated temporal convolutional network GTCN contains 3 layers of causal dilated convolutions, with dilation rates of 1, 2, and 4 for each layer.

[0014] Preferably, the total number of parameters of the gated temporal convolutional network GTCN is <5000, and the inference latency is ≤300μs.

[0015] Preferably, in step S2, vibration noise in the input window data is detected simultaneously during the process of dynamically generating the noise-resistant weight vector W. When the vibration noise is detected to be greater than 1.5 kHz, the bandpass filter is automatically switched to high Q mode, and the Q value in high Q mode is set to 8.

[0016] Preferably, in step S3, exponential smoothing is applied during the iterative optimization process, and the attenuation factor γ of the exponential smoothing is 0.95.

[0017] Preferably, when the standard deviation of the compensation residual corresponding to five consecutive 50ms windows is less than 0.05λ, the iterative optimization process in step S3 is paused, and the current optimized dynamic reference signal is continuously output to the aberration compensator.

[0018] Preferably, in the online incremental update triggered when the standard deviation of the compensation residual is greater than a preset threshold, the number of consecutive windows N is 5, the preset threshold is 0.05λ, and λ is the wavelength of the lithography source. η is the learning rate / step size for online incremental updates. η is 0.01 and is used to control the update magnitude when only updating the weights of the last layer of GTCN.

[0019] Preferably, updating only the weights of the last layer of the gated temporal convolutional network GTCN is achieved using the standard gradient descent method: Define the weight parameters of the last layer of GTCN as follows The residual loss function is L, and the incremental update formula for the weights of the last layer is... The loss function is: T represents the total number of time steps in a single preset time-length window, and r t Let be the compensation residual corresponding to time step t, and η be the learning rate for online incremental updates. This is the gradient result after gradient clipping.

[0020] Preferably, the gradient result after gradient clipping Through formula The calculation yielded, where For the loss function L, apply to the weight parameters of the last layer The partial derivatives, , This is the gradient clipping threshold, used to suppress gradient explosion. gradient The norm of .

[0021] Preferably, α and β are scalars or vectors, and are updated online according to the learning rate η, wherein the value of the learning rate η is consistent with the learning rate of the online incremental update of the weights of the last layer of GTCN.

[0022] Preferably, the online learning parameters α and β of the Sigmoid gated unit are initialized to α=0.8 and β=0.3, and the learning rate η=0.01.

[0023] Preferably, the three causal dilation convolutional layers of the gated temporal convolutional network GTCN all have a kernel length of 3, and the number of channels in each layer is set to 16, 32, and 1 respectively, so as to output a noise-reducing weight vector W=[w1, w2, ..., w] that corresponds one-to-one with the time steps within the current preset time length window. T ], where T is the total number of time steps in a single preset time length window.

[0024] Preferably, the total number of control parameters P of the gated temporal convolutional network GTCN is less than 5000, and the inference latency of the GTCN is less than 300μs. The inference latency is implemented on an embedded GPU or FPGA hardware platform.

[0025] Preferably, the dominant frequency f of the vibration noise n It can be obtained by detecting vibration sensors or by estimating the spectrum of the input window data; When f is detected n When the bandpass filter is switched to high-Q mode at frequencies above 1.5kHz, the center frequency f0 of the bandpass filter is set to the dominant frequency f. n Or its neighboring frequencies.

[0026] Preferably, the nearest frequency point to the center frequency f0 of the bandpass filter is approximately 2kHz.

[0027] The present invention also provides a dynamic reference noise reduction optimization system for photolithographic aberrations, comprising the following modules: Module M1: Synchronously acquires real-time wavefront signals and dynamic reference signals from measurement points on the surface of the aspherical lens, and extracts the real-time wavefront signals and dynamic reference signals within the latest preset time length window to form input window data; Module M2: Inputs the input window data obtained from Module M1 into the gated temporal convolutional network GTCN, extracts the temporal features of the input window data through the causal dilation convolution in the gated temporal convolutional network GTCN, and dynamically generates the noise-resistant weight vector W using the Sigmoid gate unit in the gated temporal convolutional network GTCN. Module M3: Based on the noise-resistant weight vector W generated by module M2, the dynamic reference signal collected by module M1 is optimized by element-wise multiplication iteration to obtain the optimized dynamic reference signal; Module M4: Outputs the optimized dynamic reference signal obtained from module M3 to the aberration compensator, driving the aberration compensator to generate a compensation wavefront to correct the aberrations of the aspherical lens; Module M5: Calculates the standard deviation of the compensation residuals after aberration correction in Module M4; when the standard deviation of the compensation residuals is greater than a preset threshold for N consecutive preset time length windows, an online incremental update is triggered, updating only the weights of the last layer of the gated temporal convolutional network GTCN, and returning to Module M2 to regenerate the noise-resistant weight vector W. Then, Modules M3 to M4 are executed sequentially to complete a new round of dynamic reference signal optimization and aberration correction; when the standard deviation of the compensation residuals corresponding to N consecutive preset time length windows is less than or equal to the preset threshold, the weight parameters of the current gated temporal convolutional network GTCN remain unchanged, and the optimization and correction process of Modules M1 to M4 continues to be executed.

[0028] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention improves the anti-interference capability and yield of lithography machine lens production. By dynamically generating an anti-noise weight vector through GTCN and combining it with a high-Q bandpass filter that can be adapted and switched, it can accurately filter out high-frequency vibration noise of 1.8-2.2kHz in the lithography machine production environment, avoid distortion of the time-domain correlation between dynamic reference signal and wavefront signal, control the standard deviation of aberration compensation residual within the target range of ≤0.05λ, effectively reduce the problem of local curvature distortion of aspherical lens, and significantly improve the production yield of lithography machine lens manufacturing.

[0029] 2. This invention can ensure the precision and stability of lithography lens production. It adopts a 50ms short time window data input design, combined with GTCN's causal dilation convolution temporal feature extraction technology, and a lightweight model architecture (parameter count <5K) to achieve microsecond-level noise reduction optimization and compensation response of ≤800μs. It completely solves the compensation lag defect of traditional long time window schemes, and can respond to sudden airflow disturbances during the lithography lens production process in a timely manner, ensuring timely and accurate aberration correction and guaranteeing the stability of lens production precision.

[0030] 3. This invention adapts to the real-time dynamic requirements of lithography machine lens production. By using a targeted update mechanism that only updates the weights of the last layer of GTCN, combined with exponential smoothing in iterative optimization, it ensures that the noise-resistant model can dynamically adapt to the noise spectrum changes in the lithography machine production environment, while avoiding the time-consuming problem of full-window retraining. It balances lightweight deployment with real-time compensation performance, perfectly matching the dynamic and complex interference environment and the high-precision, high-real-time production requirements in the process of lithography machine lens manufacturing.

[0031] 4. This invention can improve the noise suppression performance of lithography machines, with a noise attenuation of -25dB in the 2kHz band (compared to -15dB in traditional solutions), and the standard deviation of the compensation residual is reduced to 0.038λ (measured).

[0032] 5. This invention breaks through the delay control, with a total loop delay of ≤780μs (including GTCN inference 300μs + signal optimization 300μs + compensation execution 180μs).

[0033] 6. This invention achieves engineering deployment feasibility, with fewer than 5000 parameters, and can be deployed on a Jetson AGX Xavier GPU card (power consumption <30W), meeting the space limitations of the lithography machine control cabinet. Attached Figure Description

[0034] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the aberration compensation system based on GTCN. Figure 2 Flowchart for GTCN noise reduction optimization operation; Figure 3 This is a timing interaction diagram for signal processing. Detailed Implementation

[0035] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0036] Example 1

[0037] This embodiment provides a dynamic benchmark noise reduction optimization method for lithography aberrations adapted to the manufacturing and production of aspherical lenses in lithography machines. It aims to solve problems such as insufficient aberration compensation accuracy and response lag caused by high-frequency vibrations and sudden airflow disturbances during the lens production process, ensuring lens production yield and accuracy stability. The method specifically includes the following steps: Step S1: During the photolithography process of the aspherical lens, the wavefront sensor and reference signal acquisition module on the lithography machine are used to simultaneously acquire the real-time wavefront signal and dynamic reference signal of the measurement point on the surface of the aspherical lens, and the real-time wavefront signal and dynamic reference signal of the latest preset time length window are captured to form input window data. In step S1, considering the real-time requirements and interference suppression requirements of lithography machine lens production, the preset time length window is set to a 50ms window. This window length can accurately capture the signal characteristics of sudden interference during production, while avoiding response delays introduced by excessively long windows.

[0038] Step S2: Input the input window data obtained in step S1 into a preset gated temporal convolutional network (GTCN). Extract the temporal features of the input window data through the causal dilation convolution in the GTCN (the temporal features can accurately reflect the dynamic change law of the signal in the lithography machine production process). At the same time, use the sigmoid gate unit in the GTCN to dynamically generate the noise-resistant weight vector W. In step S2, the elements of the noise-resistant weight vector satisfy w t =σ(α·x t +β·Δx t ); where σ(.) is the Sigmoid function, used to restrict the noise-resistant weights to the range of 0~1; α and β are the online learning parameters of the gated temporal convolutional network GTCN, which can be adaptively adjusted according to the real-time interference changes in the lithography machine production environment; t is the discrete time step index within the current preset time length window; x t The input signal takes the value at time step t; Δx t Let Δx be the first-order difference of the input signal. t =x t -x t-1 The online learning rate η=0.01 for the gated temporal convolutional network GTCN ensures the stability and fast convergence of the parameter learning. t represents the discrete time step index (or sampling point number) within the current 50ms window; x t The value of the input signal at time step t (which can be a dynamic reference signal, wavefront signal, or its characteristic quantity), Δx t Describing the first-order difference (Δx) t =x t -x t-1 σ(·) is the Sigmoid function, used to restrict the gating weights to the interval 0~1; α and β are the online learning parameters (which can be scalars or vectors) of the Sigmoid gating unit, used to balance x t With Δx t The influence on the weights can be determined and updated online according to the learning rate η; for example, it can be initialized as α=0.8, β=0.3, η=0.01, but is not limited to this.

[0039] To meet the real-time requirements of lithography lens production, the gated temporal convolutional network GTCN adopts a lightweight architecture design, specifically including 3 layers of causal dilated convolutions, with dilation rates of 1, 2, and 4 for each layer. The total number of parameters of the gated temporal convolutional network GTCN is <5000, and the inference latency is ≤300μs, which can meet the compensation requirements for microsecond-level response during production. GTCN can include 3 layers of causal dilated convolutions with dilation rates of 1, 2, and 4 respectively; the kernel length can be 3; the number of channels can be selected as 16, 32, or 1 to output a noise-reducing weight vector W=[w1, w2, ..., w...] that corresponds one-to-one with each time step. T ], where T is the number of sampling points in the current window. Through lightweight design, the total number of parameters P of GTCN can be controlled to be less than 5000, thereby enabling inference latency to be controlled to ≤300μs (e.g., implemented on embedded GPU / FPGA).

[0040] In step S2, considering the interference of high-frequency vibration noise on the signal in the lithography machine production environment, vibration noise in the input window data is detected simultaneously during the dynamic generation of the anti-noise weight vector W. When the vibration noise is detected to be >1.5kHz (i.e., the frequency band of high-frequency vibration noise of the worktable commonly used in lithography machine production), the bandpass filter is automatically switched to high Q value mode, and the Q value in this high Q value mode is set to 8 to achieve accurate filtering of high-frequency vibration noise.

[0041] Vibration noise can be obtained by using vibration sensors or by spectral estimation of the input window data, and its dominant frequency f can be obtained. n When f is detected n At frequencies above 1.5kHz, the bandpass filter switches to high-Q mode (Q=8) to improve suppression of the dominant narrowband; the center frequency f0 can be set to f n Or its adjacent frequency (e.g., about 2kHz) to improve narrowband suppression without weakening the effective components of the dynamic reference signal as much as possible.

[0042] Step S3: Based on the noise-resistant weight vector W generated in step S2, perform element-wise multiplication iterative optimization on the dynamic reference signal acquired in step S1 to obtain the optimized dynamic reference signal. In step S3, to improve the stability of the optimized reference signal, exponential smoothing is applied during the iterative optimization process, with an attenuation factor γ=0.95. Meanwhile, to avoid ineffective optimization consuming production resources, when the standard deviation of the compensation residual corresponding to five consecutive 50ms windows is <0.05λ, the iterative optimization process of step S3 is paused, and the current optimized dynamic reference signal is continuously output to the aberration compensator to ensure the continuity and stability of lens production.

[0043] Step S4: Output the optimized dynamic reference signal obtained in step S3 to the aberration compensator of the lithography machine, drive the aberration compensator to generate a compensation wavefront corresponding to the aberration, so as to accurately correct the aberration of the aspherical lens and ensure the accuracy of lithography processing.

[0044] Step S5: The standard deviation of the compensation residual after aberration correction in step S4 is calculated using the residual detection module of the lithography machine. When the standard deviation of the compensation residual is greater than the preset threshold for N consecutive preset time length windows, it indicates that the current noise-resistant weight vector can no longer adapt to the changes in interference in the production environment. At this time, an online incremental update is triggered, and only the weights of the last layer of the gated temporal convolutional network GTCN are updated (to avoid response delay caused by full layer update). Then, the process returns to step S2 to regenerate the noise-resistant weight vector W, and steps S3 to S4 are executed sequentially to complete a new round of dynamic reference signal optimization and aberration correction. When the standard deviation of the compensation residual corresponding to N consecutive preset time length windows is less than or equal to the preset threshold, it indicates that the current noise-resistant optimization effect meets the production requirements. The weight parameters of the current gated temporal convolutional network GTCN remain unchanged, and the optimization and correction process of steps S1 to S4 continues to be executed. Online incremental updates are triggered when the standard deviation of the compensation residual is greater than a preset threshold for N consecutive preset time length windows; where N is the number of consecutive windows (preferably N=5); the threshold can be set to 0.05λ, where λ is the wavelength of the lithography source (e.g., EUV corresponds to λ=13.5nm); η is the learning rate / step size of the online incremental updates (preferably η=0.01), used to control the update magnitude when only the weights of the last layer of GTCN are updated.

[0045] Let the weight parameters of the last layer be... The residual loss function is L (for example) r t Let the compensation residual be the time step t. Then the incremental update can be written as: ,in gradient The result after gradient clipping; gradient clipping can be used... , The pruning threshold is used to suppress gradient explosion. This incremental update only applies to the weights of the last layer to reduce the computational cost of online updates, and through optimization, the latency of a single update can be controlled to the microsecond level (e.g., ≤200μs).

[0046] In step S5, N is set to 5 in accordance with the precision requirements of lithography lens production; the preset threshold is set to 0.05λ, where λ is the wavelength of the light source used for lithography (e.g., 13.5nm for EUV lithography). This threshold ensures that the aberration correction accuracy meets the production quality standards of aspherical lenses.

[0047] The present invention also provides a dynamic reference noise reduction optimization system for lithographic aberrations. The dynamic reference noise reduction optimization system for lithographic aberrations can be implemented by executing the process steps of the dynamic reference noise reduction optimization method for lithographic aberrations. That is, those skilled in the art can understand the dynamic reference noise reduction optimization method for lithographic aberrations as a preferred embodiment of the dynamic reference noise reduction optimization system for lithographic aberrations.

[0048] Example 2

[0049] This embodiment provides a dynamic reference noise reduction optimization system for lithography aberrations, adapted for lithography machine manufacturing and production of aspherical lenses. It aims to solve problems such as insufficient aberration compensation accuracy and response lag caused by high-frequency vibrations and sudden airflow disturbances during lithography lens production, ensuring lens production yield and accuracy stability. Specifically, it includes the following modules: Module M1: During the photolithography process of aspherical lenses, the wavefront sensor and reference signal acquisition module on the photolithography machine synchronously acquire the real-time wavefront signal and dynamic reference signal of the measurement point on the surface of the aspherical lens, and capture the real-time wavefront signal and dynamic reference signal of the latest preset time length window to form input window data; In module M1, considering the real-time requirements and interference suppression requirements of lithography machine lens production, the preset time length window is set to a 50ms window. This window length can accurately capture the signal characteristics of sudden interference during production, while avoiding the response delay introduced by an excessively long window.

[0050] Module M2: Inputs the input window data obtained from module M1 into a preset gated temporal convolutional network (GTCN). The temporal features of the input window data are extracted through the causal dilation convolution in the GTCN (the temporal features can accurately reflect the dynamic change law of the signal in the lithography machine production process). At the same time, the noise-resistant weight vector W is dynamically generated using the sigmoid gate unit in the GTCN. In module M2, the elements of the noise-resistant weight vector satisfy w t =σ(α·x t +β·Δx t ); where σ(.) is the Sigmoid function, used to restrict the noise-resistant weights to the range of 0~1; α and β are the online learning parameters of the gated temporal convolutional network GTCN, which can be adaptively adjusted according to the real-time interference changes in the lithography machine production environment; t is the discrete time step index within the current preset time length window; x t The input signal takes the value at time step t; Δx t Let Δx be the first-order difference of the input signal. t =x t -x t-1The online learning rate η=0.01 for the gated temporal convolutional network GTCN ensures the stability and fast convergence of the parameter learning. To meet the real-time requirements of lithography lens production, the gated temporal convolutional network GTCN adopts a lightweight architecture design, specifically including 3 layers of causal dilated convolutions, with dilation rates of 1, 2, and 4 for each layer. The total number of parameters of the gated temporal convolutional network GTCN is <5000, and the inference latency is ≤300μs, which can meet the compensation requirements for microsecond-level response during production. In module M2, considering the interference of high-frequency vibration noise on the signal in the lithography machine production environment, vibration noise in the input window data is detected simultaneously during the dynamic generation of the anti-noise weight vector W. When the vibration noise is detected to be >1.5kHz (i.e., the frequency band of high-frequency vibration noise of the worktable commonly used in lithography machine production), the bandpass filter is automatically switched to high Q value mode, and the Q value in this high Q value mode is set to 8 to achieve accurate filtering of high-frequency vibration noise.

[0051] Module M3: Based on the noise-resistant weight vector W generated by module M2, the dynamic reference signal collected by module M1 is optimized by element-wise multiplication iteration to obtain the optimized dynamic reference signal; In module M3, to improve the stability of the optimized reference signal, exponential smoothing is applied during the iterative optimization process, with an attenuation factor γ=0.95. Simultaneously, to avoid ineffective optimization consuming production resources, when the standard deviation of the compensation residual corresponding to five consecutive 50ms windows is <0.05λ, the iterative optimization process of module M3 is paused, maintaining the continuous output of the currently optimized dynamic reference signal to the aberration compensator, thus ensuring the continuity and stability of lens production.

[0052] Module M4: Outputs the optimized dynamic reference signal obtained from module M3 to the aberration compensator of the lithography machine, drives the aberration compensator to generate a compensation wavefront corresponding to the aberration, so as to accurately correct the aberration of the aspherical lens and ensure the accuracy of lithography processing.

[0053] Module M5: Calculates the standard deviation of the compensation residual after aberration correction in Module M4 using the residual detection module of the lithography machine. When the standard deviation of the compensation residual is greater than a preset threshold for N consecutive preset time length windows, it indicates that the current noise-resistant weight vector can no longer adapt to changes in production environment interference. At this time, an online incremental update is triggered, updating only the weights of the last layer of the gated temporal convolutional network GTCN (avoiding response delay caused by full layer updates), and returning to Module M2 to regenerate the noise-resistant weight vector W. Then, Modules M3 to M4 are executed sequentially to complete a new round of dynamic reference signal optimization and aberration correction. When the standard deviation of the compensation residual corresponding to N consecutive preset time length windows is less than or equal to the preset threshold, it indicates that the current noise-resistant optimization effect meets production requirements. The weight parameters of the current gated temporal convolutional network GTCN remain unchanged, and the optimization and correction process of Modules M1 to M4 continues. In module M5, N is set to 5 in accordance with the precision requirements of lithography lens production; the preset threshold is set to 0.05λ, where λ is the wavelength of the light source used for lithography (e.g., 13.5nm for EUV lithography). This threshold ensures that the aberration correction accuracy meets the production quality standards of aspherical lenses.

[0054] In online incremental updates triggered when the standard deviation of the compensation residual exceeds a preset threshold, the number of continuous windows N is 5, the preset threshold is 0.05λ, where λ is the wavelength of the lithography source; η is the learning rate / step size of the online incremental update, and η is 0.01, used to control the update magnitude when only the weights of the last layer of GTCN are updated.

[0055] Only the weights of the last layer of the gated temporal convolutional network GTCN are updated using the standard gradient descent method: the weight parameters of the last layer of GTCN are defined as follows. The residual loss function is L, and the incremental update formula for the weights of the last layer is... The loss function is: T represents the total number of time steps in a single preset time-length window, and r t Let be the compensation residual corresponding to time step t, and η be the learning rate for online incremental updates. This is the gradient result after gradient clipping.

[0056] The gradient result after gradient clipping Through formula The calculation yielded, where For the loss function L, apply to the weight parameters of the last layer The partial derivatives, , This is the gradient clipping threshold, used to suppress gradient explosion. gradient The norm of .

[0057] α and β are scalars or vectors, and are updated online according to the learning rate η. The value of the learning rate η is consistent with the learning rate of the online incremental update of the weights of the last layer of GTCN. The online learning parameters α and β of the Sigmoid gate unit are initialized to α=0.8 and β=0.3, and the learning rate η=0.01.

[0058] The gated temporal convolutional network GTCN has three layers of causal dilated convolutions, each with a kernel length of 3. The number of channels in each layer is set to 16, 32, and 1 respectively, to output a noise-reducing weight vector W=[w1, w2, ..., w] that corresponds one-to-one with the time steps within the current preset time window. T ], where T is the total number of time steps in a single preset time length window. The total number of control parameters P of the gated temporal convolutional network GTCN is less than 5000, and the inference latency of the GTCN is ≤300μs, which is implemented on an embedded GPU or FPGA hardware platform.

[0059] The dominant frequency f of the vibration noise n It is obtained through vibration sensor detection or through spectral estimation of input window data; when f is detected n When the bandpass filter is switched to high-Q mode at frequencies above 1.5kHz, the center frequency f0 of the bandpass filter is set to the dominant frequency f. n Or its nearest frequency. The nearest frequency to the center frequency f0 of the bandpass filter is approximately 2kHz.

[0060] Example 3

[0061] Those skilled in the art can understand this embodiment as a more specific description of Embodiment 1 and Embodiment 2.

[0062] like Figures 1 to 3 As shown, this embodiment proposes a dynamic benchmark noise reduction optimization method based on GTCN, the core technical solution of which is: 1) 50ms short window driving mechanism: Only the latest 50ms wavefront signal is captured (1000 sampling points, 20kHz sampling rate), avoiding long window delay; 2) GTCN Lightweight Noise Reduction Engine: Employs a 3-layer causal dilated convolution (dilation rates of 1, 2, and 4 respectively), covering a receptive field of 40ms temporal features; embeds a Sigmoid gate unit to dynamically suppress noise in the >1.2kHz frequency band; outputs a noise reduction weight vector W=[w1,w2,…,w 1000 ], where w t =σ(α·x t +β·Δx t (α / β online learning update, η=0.01); 3) Closed-loop validation and incremental retraining: When the residual standard deviation is >0.05λ, only the weights of the last layer of GTCN are updated (time ≤200μs) to avoid training the entire network.

[0063] Specifically, this embodiment provides a method for optimizing the dynamic benchmark for lithographic aberration based on gated temporal convolutional networks, including the following steps: Simultaneously acquire real-time wavefront signals and dynamic reference signals from measurement points on the surface of the aspherical lens, and extract the latest 50ms window data; The window data is input into a gated temporal convolutional network (GTCN), where temporal features are extracted through causal dilated convolution, and a noise-resistant weight vector W is dynamically generated using a Sigmoid gate unit, where w t =σ(α·x t + β·Δx t ), where α and β are online learning parameters, and the learning rate η = 0.01; The dynamic reference signal is optimized by element-wise multiplication iterative optimization based on the noise-resistant weight vector W to obtain the optimized dynamic reference signal. The optimized dynamic reference signal is output to the aberration compensator, which drives it to generate a compensation wavefront to correct the aberrations of the aspherical lens. When the standard deviation of the compensated residual is greater than 0.05λ for 5 consecutive windows, only the weights of the last layer of GTCN are updated and the noise-resistant weight vector W is regenerated.

[0064] Furthermore, the GTCN contains 3 layers of causal dilated convolutions with dilation rates of 1, 2, and 4 respectively, a total number of parameters < 5,000, and an inference latency ≤ 300 μs.

[0065] Furthermore, in the step of dynamically generating the noise reduction weight vector W, when vibration noise > 1.5 kHz is detected, the bandpass filter is automatically switched to high Q value mode, Q=8.

[0066] Furthermore, the iterative optimization step applies exponential smoothing with a decay factor γ=0.95, and optimization is paused when the residual standard deviation of 5 consecutive windows is <0.05λ.

[0067] More specifically, this implementation uses the MT1217 lithography machine as an application scenario to describe in detail the implementation details of the technical solution. All components strictly adhere to physical constraints: signal transmission line ≤15cm, loop delay ≤800μs, and power consumption ≤30W.

[0068] I. Hardware Deployment and Connection.

[0069] 1) Real-time wavefront signal acquisition unit: Employs a 4×4 Shack-Hartmann sensor array (Hamamatsu S9993-01), physically deployed inside the aspherical lens support ring, 0.2mm from the lens surface. Measurement point grid spacing is 0.5mm (covering the central area of ​​the lens), sampling rate is 20kHz (synchronized via IEEE 1588v2 protocol, timing error <5μs). Output interface: Differential LVDS signal line, cable length strictly controlled to 14.8cm (measured attenuation <2.8dB@2kHz), end connected to a 50Ω impedance matching circuit (eliminating high-frequency reflection noise <-20dB).

[0070] 2) Dynamic reference signal source: Integrated into the light source control module (Cymer XLR 700e), it outputs the original reference signal (bandwidth 1.5kHz) through a single-mode fiber. The fiber is 10cm long and is connected to the GTCN module via a photoelectric converter to ensure phase jitter <0.1°.

[0071] 3) GTCN Module: Deployed on an NVIDIA Jetson AGX Xavier GPU card (32 TOPS computing power), installed in a dedicated slot within the lithography machine control cabinet. Input Interface: Dual-channel PCIe Gen3 direct-connect acquisition unit (delay <100ns), receiving wavefront and reference signals. Key Circuitry: Front-end bandpass filter (adjustable from 0.5-2.5kHz), automatically switches to high Q-value mode (Q=8) when the vibration sensor detects noise >1.5kHz.

[0072] 4) Reference Update Unit and Optimized Signal Generation: Based on a Xilinx Artix-7 FPGA (XC7A50T), mounted on the compensator driver board, 4.7cm from the aberration compensator. Pipeline design: Input buffer (50ns) → weighted multiplication (150ns) → exponential smoothing (100ns) → output latch (50ns), total delay ≤350ns. Optimized signal generation: A 16-bit DAC (TIDAC8830) outputs ±10V analog voltage, which is controlled by a driver amplifier circuit (bandwidth >5kHz) to control the PZT.

[0073] 5) Aberration Compensator: Piezoelectric ceramic actuator (PI P-753.3CD), bonded to the support structure on the back of the lens, physically aligned with the measurement point grid. Response characteristics: displacement range ±5μm, rise time <150μs (meets 800μs loop constraint).

[0074] II. Detailed Explanation of Operation Procedures

[0075] Step 1: Synchronous signal acquisition (t=0~5μs).

[0076] The acquisition unit captures the dynamic reference signal (source signal) and the real-time wavefront signal (measurement signal) at a rate of 20kHz, and extracts the latest 50ms window (1000 sampling points).

[0077] Triggering mechanism: The light source control module outputs a synchronization pulse, which is distributed to all sensors via the IEEE 1588v2 protocol to eliminate timing jitter in multi-point acquisition (measured error 3.2μs).

[0078] Data preprocessing: The raw signal is filtered by a bandpass filter to remove DC offset and noise >3kHz, normalized to [-1,1], and written to the GTCN module DMA buffer (to avoid CPU interruption).

[0079] Step 2: GTCN noise reduction weight generation (t=5~305μs).

[0080] Input processing: 1000-point signal input to the GTCN module, processed by 3 layers of causal dilated convolution: Layer 1: dilation rate = 1, kernel size = 3, output channels = 16 (extracting local features); Layer 2: dilation rate = 2, kernel size = 3, output channels = 32 (expanding the receptive field to 40ms); Layer 3: dilation rate = 4, kernel size = 3, output channels = 1 (generating noise-resistant weights).

[0081] Gating mechanism: Sigmoid unit calculates w t =σ(0.8·x t + 0.3·Δx t ), where Δx t It is a first-order difference. When |Δx t The threshold (dynamically updated by online learning) automatically suppresses noise in the >1.2kHz frequency band.

[0082] Lightweight design: fewer than 5000 parameters, inference latency ≤300μs (e.g., can be controlled within 300μs through hardware).

[0083] Step 3: Iterative optimization of the reference signal (t=305~605μs).

[0084] Benchmark update unit execution: Optimized signal = Original dynamic benchmark signal ⊙ W (element-by-element multiplication). Exponential smoothing: S t = γ·S t-1 + (1-γ)·Optimize signal, attenuation factor γ=0.95, suppress high-frequency mutations. Convergence judgment: If the standard deviation of the residuals for 5 consecutive windows is <0.05λ, pause updates to save computing power (actual trigger rate 68%).

[0085] Step 4: Compensation execution and verification (t=605~780μs).

[0086] The optimized signal is converted into an analog voltage via a DAC, driving the PZT to generate a compensated wavefront (displacement accuracy ±0.5nm). The optical feedback loop completes residual verification within 780μs: RMS value = √(Σ(real-time wavefront - optimized signal)² / N).

[0087] Step 5: Incremental retraining triggered (only when residual exceeds the limit).

[0088] If RMS > 0.05λ, the acquisition unit triggers retraining: only the weights of the GTCN layer 3 are updated, with a learning rate η = 0.01; the new weights are calculated based on the current window of data: ΔW = η· L, where L is the residual loss function; retraining time ≤200μs (measured at 198μs), avoiding training delay across the entire network.

[0089] (III) Verification of key parameters

[0090] This invention belongs to the field of semiconductor lithography technology, specifically relating to real-time aberration compensation technology in the lithography exposure process of aspherical lenses. More specifically, this invention focuses on solving the problem of dynamic reference signal mismatch under high-frequency vibration noise interference greater than 1kHz. It achieves noise-aware adaptive reference update by embedding a lightweight gated temporal convolutional network (GTCN), which is applicable to EUV lithography scenarios at 10nm and below nodes, especially for transient aberration compensation caused by mechanical vibration and airflow disturbance of lithography machines.

[0091] This invention aims to solve the technical problem of mismatch between dynamic reference signal and real-time wavefront signal caused by high-frequency noise greater than 1kHz during photolithography exposure. Specifically, it achieves the following: increasing the noise suppression bandwidth from 1.2kHz to 2kHz, making the standard deviation of the compensation residual ≤0.04λ; optimizing and compressing the reference signal delay to ≤800μs to meet the real-time requirements of 10nm photolithography; and achieving noise spectrum adaptation under the condition of <5K parameters to avoid the overhead of full network retraining.

[0092] This invention provides a method and system for optimizing the noise reduction of a dynamic reference for lithographic aberrations. The core process of the method based on a gated temporal convolutional network is as follows: Real-time wavefront signals and dynamic reference signals from measurement points on the surface of an aspherical lens are simultaneously acquired; the latest 50-millisecond window of data is extracted and input into the gated temporal convolutional network; temporal features are extracted through causal dilation convolution; a noise reduction weight vector is dynamically generated using a Sigmoid gate unit, wherein the weight vector is generated by online learning parameters with a learning rate set to 0.01; the dynamic reference signal is iteratively optimized based on this noise reduction weight vector and then output to an aberration compensator to correct the aberrations of the aspherical lens; if the standard deviation of the compensation residual is greater than 0.05 times the wavelength of the lithographic source for five consecutive windows, only the weights of the last layer of the gated temporal convolutional network are updated and a new noise reduction weight vector is generated.

[0093] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0094] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for optimizing the noise resistance of a dynamic reference for photolithographic aberrations, characterized in that, Includes the following steps: Step S1: Synchronously acquire the real-time wavefront signal and dynamic reference signal of the measurement point on the surface of the aspherical lens, and extract the real-time wavefront signal and dynamic reference signal of the latest preset time length window to form input window data; Step S2: Input the input window data obtained in step S1 into the gated temporal convolutional network GTCN, extract the temporal features of the input window data through the causal dilation convolution in the gated temporal convolutional network GTCN, and dynamically generate the noise-resistant weight vector W using the sigmoid gate unit in the gated temporal convolutional network GTCN. Step S3: Based on the noise-resistant weight vector W generated in step S2, perform element-wise multiplication iterative optimization on the dynamic reference signal acquired in step S1 to obtain the optimized dynamic reference signal. Step S4: Output the optimized dynamic reference signal obtained in step S3 to the aberration compensator to drive the aberration compensator to generate a compensation wavefront to correct the aberration of the aspherical lens. Step S5: Calculate the standard deviation of the compensation residual after aberration correction in Step S4; when the standard deviation of the compensation residual is greater than the preset threshold for N consecutive preset time length windows, trigger online incremental update, update only the weights of the last layer of the gated temporal convolutional network GTCN, and return to Step S2 to regenerate the noise-resistant weight vector W, and then execute Steps S3 to S4 in sequence to complete a new round of dynamic reference signal optimization and aberration correction; when the standard deviation of the compensation residual corresponding to N consecutive preset time length windows is less than or equal to the preset threshold, keep the weight parameters of the current gated temporal convolutional network GTCN unchanged, and continue to execute the optimization and correction process of Steps S1 to S4.

2. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 1, characterized in that, In step S1, the preset time length window is a 50ms window.

3. The method for optimizing the dynamic reference noise reduction of photolithographic aberrations according to claim 1, characterized in that, In step S2, the elements of the noise-resistant weight vector satisfy w t =σ(α·x t +β·Δx t ); Where σ(.) is the Sigmoid function, used to restrict the noise-resistant weights to the interval 0~1; α and β are the online learning parameters of the gated temporal convolutional network GTCN; t is the discrete time step index within the current preset time length window; x t The input signal takes the value at time step t; Δx t Let Δx be the first-order difference of the input signal. t =x t -x t-1 .

4. The method for optimizing the dynamic reference noise reduction of photolithographic aberrations according to claim 3, characterized in that, The online learning rate η of the gated temporal convolutional network GTCN is 0.

01.

5. The method for optimizing the dynamic reference noise reduction of photolithographic aberrations according to claim 1, characterized in that, In step S5, N is 5; the preset threshold is 0.05λ, where λ is the wavelength of the light source used for photolithography.

6. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 1, characterized in that, The gated temporal convolutional network GTCN contains three layers of causal dilated convolutions, with dilation rates of 1, 2, and 4 for each layer, respectively.

7. The method for optimizing the dynamic reference noise reduction of photolithographic aberrations according to claim 6, characterized in that, The total number of parameters of the gated temporal convolutional network GTCN is <5000, and the inference latency is ≤300μs.

8. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 1, characterized in that, In step S2, during the process of dynamically generating the anti-noise weight vector W, vibration noise in the input window data is detected simultaneously. When the vibration noise is detected to be greater than 1.5 kHz, the bandpass filter is automatically switched to high Q mode, and the Q value in high Q mode is set to 8.

9. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 1, characterized in that, In step S3, exponential smoothing is applied during the iterative optimization process, and the attenuation factor γ of the exponential smoothing is 0.

95.

10. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 1, characterized in that, When the standard deviation of the compensation residuals corresponding to five consecutive 50ms windows is less than 0.05λ, pause the iterative optimization process in step S3 and keep the currently optimized dynamic reference signal continuously output to the aberration compensator.

11. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 1, characterized in that, In online incremental updates triggered when the standard deviation of the compensation residual is greater than a preset threshold, the number of consecutive windows N is 5, the preset threshold is 0.05λ, and λ is the wavelength of the lithography source. η is the learning rate / step size for online incremental updates. η is 0.01 and is used to control the update magnitude when only updating the weights of the last layer of GTCN.

12. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 11, characterized in that, Only the weights of the last layer of the gated temporal convolutional network GTCN are updated using the standard gradient descent method: Define the weight parameters of the last layer of GTCN as follows The residual loss function is L, and the incremental update formula for the weights of the last layer is... The loss function is: T represents the total number of time steps in a single preset time-length window, and r t Let be the compensation residual corresponding to time step t, and η be the learning rate for online incremental updates. This is the gradient result after gradient clipping.

13. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 12, characterized in that, The gradient result after gradient clipping Through formula The calculation yielded, where For the loss function L, apply to the weight parameters of the last layer The partial derivatives, , This is the gradient clipping threshold, used to suppress gradient explosion. gradient The norm of .

14. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 3, characterized in that, α and β are scalars or vectors, and are updated online according to the learning rate η. The value of the learning rate η is consistent with the learning rate of the online incremental update of the weights of the last layer of GTCN.

15. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 14, characterized in that, The online learning parameters α and β of the Sigmoid gated unit are initialized to α=0.8 and β=0.3, and the learning rate η=0.

01.

16. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 6, characterized in that, The gated temporal convolutional network GTCN has three layers of causal dilated convolutions, each with a kernel length of 3. The number of channels in each layer is set to 16, 32, and 1 respectively, to output a noise-reducing weight vector W=[w1, w2, ..., w] that corresponds one-to-one with the time steps within the current preset time window. T ], where T is the total number of time steps in a single preset time length window.

17. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 16, characterized in that, The total number of control parameters P of the gated temporal convolutional network GTCN is less than 5000, and the inference latency of the GTCN is less than 300μs. The inference latency is implemented on an embedded GPU or FPGA hardware platform.

18. The method for optimizing the dynamic reference noise reduction of photolithographic aberrations according to claim 8, characterized in that, The dominant frequency f of the vibration noise n It can be obtained by detecting vibration sensors or by estimating the spectrum of the input window data; When f is detected n When the bandpass filter is switched to high-Q mode at frequencies above 1.5kHz, the center frequency f0 of the bandpass filter is set to the dominant frequency f. n Or its neighboring frequencies.

19. The method for optimizing the dynamic reference noise resistance of photolithographic aberrations according to claim 18, characterized in that, The nearest frequency point of the center frequency f0 of the bandpass filter is approximately 2kHz.

20. A dynamic reference noise reduction optimization system for photolithographic aberrations, characterized in that, Includes the following modules: Module M1: Synchronously acquires real-time wavefront signals and dynamic reference signals from measurement points on the surface of the aspherical lens, and extracts the real-time wavefront signals and dynamic reference signals within the latest preset time length window to form input window data; Module M2: Inputs the input window data obtained from Module M1 into the gated temporal convolutional network GTCN, extracts the temporal features of the input window data through the causal dilation convolution in the gated temporal convolutional network GTCN, and dynamically generates the noise-resistant weight vector W using the Sigmoid gate unit in the gated temporal convolutional network GTCN. Module M3: Based on the noise-resistant weight vector W generated by module M2, the dynamic reference signal collected by module M1 is optimized by element-wise multiplication iteration to obtain the optimized dynamic reference signal; Module M4: Outputs the optimized dynamic reference signal obtained from module M3 to the aberration compensator, driving the aberration compensator to generate a compensation wavefront to correct the aberrations of the aspherical lens; Module M5: Calculates the standard deviation of the compensation residuals after aberration correction in Module M4; when the standard deviation of the compensation residuals is greater than a preset threshold for N consecutive preset time length windows, an online incremental update is triggered, updating only the weights of the last layer of the gated temporal convolutional network GTCN, and returning to Module M2 to regenerate the noise-resistant weight vector W. Then, Modules M3 to M4 are executed sequentially to complete a new round of dynamic reference signal optimization and aberration correction; when the standard deviation of the compensation residuals corresponding to N consecutive preset time length windows is less than or equal to the preset threshold, the weight parameters of the current gated temporal convolutional network GTCN remain unchanged, and the optimization and correction process of Modules M1 to M4 continues to be executed.

Images