Plasma power supply pulse control method and device

Abstract

The invention belongs to the technical field of semiconductor manufacturing, and provides a plasma power supply pulse control method and device, and the method comprises the steps: carrying out the remapping of a benchmark reference waveform when a pulse parameter changes, and generating a new reference waveform; and the iterative learning device is updated in parallel independently in each time period, and the optimal adaptive learning gain and the update control quantity are determined. According to the method, a single pulse period is decomposed into a plurality of time periods, and through combination of iterative learning control and an annealing algorithm, gain and learning weight are controlled in a time-segmented differentiation manner, so that precise compensation of a nonlinear region is realized; independent parallel iterative learning is carried out in each time period, so that rising / falling edge overshoot-free, flat-top period rapid convergence and extremely low fluctuation are finally realized, and high-precision and high-consistency tracking of a full-period waveform is realized; the time duty ratio (DC%) of each stage is finely adjusted or the power gain is modulated into a multi-stage pulse, so that the free radical density and the ion bombardment intensity can be independently controlled, and accurate control on the plasma in the etching process is realized.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and specifically to a plasma power supply pulse control method and apparatus. Background Technology

[0002] In precision processes such as semiconductor dry etching and deposition, the generation and maintenance of plasma depend on a stable power supply from an RF power source. With the continuous miniaturization of integrated circuit process nodes (e.g., below 3nm), traditional continuous wave (CW) RF modes are no longer sufficient to meet the stringent requirements of modern etching processes for anisotropy, high aspect ratios, and no damage to the wafer surface. In CW mode, RF power is continuously applied to the plasma. This results in electrons and ions being excited throughout the entire discharge cycle, leading to excessively long bombardment of the wafer surface by high-energy ions, which can easily cause wafer damage or bottom notch effects. Furthermore, in structures with large aspect ratios, it is difficult to independently control ion energy and flux, limiting the process window.

[0003] To obtain an ideal square wave signal (steep rising edge, stable flat top, and clean falling edge) on a microsecond timescale, high aspect ratio etching processes must overcome complex load reflection and loop oscillation problems. This places extremely high demands on the dynamic response speed and closed-loop control accuracy of the power supply. However, existing plasma RF power supply pulse control technology is limited by a combination of technical challenges, such as nonlinear distortion of the power amplifier, difficulty in high-speed timing synchronization, and control loop delay. As a result, it is difficult to output pulse RF power with accurate waveforms, stable parameters, and high consistency, which cannot meet the stringent requirements for process accuracy and reliability in advanced semiconductor manufacturing.

[0004] Therefore, there is an urgent need for an innovative control scheme to systematically address the aforementioned shortcomings. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a plasma power supply pulse control method to solve the problem in current semiconductor processes where RF power supplies struggle to output pulsed RF power with precise waveforms, stable parameters, and high consistency.

[0006] In a first aspect, the present invention provides a plasma power supply pulse control method, comprising: The waveform of the controlled target within a single pulse cycle is divided into time periods, and the target reference trajectory for each time period is determined. The divided time periods Including rising edge Flat-top period Falling edge and rest period ; Pre-generate a set of reference waveforms And when the pulse parameters change, the reference waveform is changed. Remapping the time axis by time period generates a new reference waveform. Otherwise, maintain the reference waveform. Pulse parameters include target pulse width, amplitude, and period. Each time period The iterative learner is updated independently and in parallel to obtain the updated control variable; the updated control variable is specifically... , This represents the feedforward control variable, where i represents the time period and j represents the pulse period. This indicates that the forgetting factor is a preset constant. For adaptive learning gain, For the learning filter, q represents a complex variable in the continuous domain. To learn steps ahead of time, This represents the tracking error during that period. As the target reference trajectory, This is the actual output; Obtain the optimal adaptive learning gain according to the Metropolis criterion. Optimal adaptive learning gain The corresponding update control quantity; the update control quantity acts on the waveform or The actual output waveform is obtained.

[0007] Optionally, the step of dividing the waveform within the pulse period into time segments and determining the time index set and target reference trajectory for each time segment includes: Based on the characteristics of the pulse waveform, determine the start and end points of each time period; For each time period, determine the set of time indices. and target reference trajectory .

[0008] Optionally, the reference waveform is changed when the pulse parameters change. Remapping the time axis by time period generates a new reference waveform. ,include: After applying anti-aliasing filtering to the waveform, the reference waveform is extracted. Using cubic spline interpolation method Resampling to the current control cycle Corresponding discrete time point index Above, generate the current reference waveform; Correct the phase deviation of the starting point and key feature points of each time period to align them on the time axis.

[0009] Optionally, learn the filter Using the inverse model filtering method, the transfer function is expressed as follows: , To learn the filter gain coefficients, , The coefficients of the denominator polynomial, , , These are the coefficients of the numerator polynomial.

[0010] Optionally, the step of obtaining the optimal adaptive learning gain according to the Metropolis criterion... ,include: Initialize parameters and set the initial temperature. Termination temperature and Current temperature ; State Apply perturbation Generate a new state Then, the cost function It has become ; , It is the model prediction value Compared with the true value The losses between N It is the number of samples; like The next state value is ,like but ; Check if the iteration count has been reached; if not, then check... To obtain optimized adaptive learning gain .

[0011] Optionally, a reference waveform Pre-generation is achieved through low-speed, high-precision measurement or model simulation.

[0012] In a second aspect, the present invention provides a plasma power supply pulse control device for executing the control method of any possible implementation of the first aspect.

[0013] By adopting the above technical solution, this application has the following beneficial effects: (1) This invention decomposes a single pulse cycle into multiple time periods, including rising edge, flat top, falling edge, and rest period. By combining iterative learning control and annealing algorithm, it adopts differentiated control gain and learning weight for different time periods to achieve accurate compensation in nonlinear regions. Each time period is independently and in parallel iteratively learned, and finally achieves no overshoot on rising / falling edge, fast convergence on flat top, and extremely low fluctuation, thus achieving high-precision and high-consistency tracking of the full-cycle waveform. (2) When the pulse parameters change, a variable sampling reconstruction algorithm is adopted. Through anti-aliasing decimation and high-order interpolation, the time axis is automatically aligned to maintain microsecond-level accuracy and ensure the robustness of the algorithm under variable parameters. (3) By finely adjusting the time duty cycle (DC%) or power gain modulation of each stage into multi-level pulses, the "free radical density, ion bombardment intensity, and pumping degree" can be independently controlled to achieve precise control of the plasma during the etching process. Attached Figure Description

[0014] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0015] Figure 1 This invention provides a flowchart of one embodiment of a plasma power supply pulse control method. Figure 2 This is a second flowchart of a plasma power supply pulse control method provided by an embodiment of the present invention; Figure 3 A schematic diagram of the pulse waveform features provided in an embodiment of the present invention is shown; Figure 4 A flowchart of S4 provided in an embodiment of the present invention is shown; Figure 5 The third flowchart of a plasma power supply pulse control method provided by an embodiment of the present invention is shown; Figure 6 A schematic diagram of the actual output waveform before optimization provided in an embodiment of the present invention is shown; Figure 7 A schematic diagram of the optimized actual output waveform provided by an embodiment of the present invention is shown; Figure 8 A schematic diagram of the optimized multi-stage pulse waveform provided in an embodiment of the present invention is shown; Figure 9 A schematic diagram of a plasma power supply pulse control device provided in an embodiment of the present invention is shown. Detailed Implementation

[0016] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention. It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0017] To obtain radio frequency power with accurate waveforms, stable parameters, and high consistency to meet the technical requirements of semiconductor processes, in one embodiment, such as... Figure 1-2 As shown, a plasma power supply pulse control method is provided, including: S1. Divide the controlled target waveform within a single pulse cycle into time periods and determine the target reference trajectory for each time period. The divided time periods Including rising edge Flat-top period Falling edge and rest period .

[0018] Based on the waveform of the controlled target, one pulse period T is divided into four consecutive and potentially overlapping time periods: , , and .in, The rising edge period is the phase from the start of triggering until the output reaches the target amplitude. This is the plateau period, the main stage for maintaining the stability of the target amplitude; The falling edge period is the phase from the end of the flat top to near zero. This is a rest period.

[0019] For each time period Define its time index set and target reference trajectory ,in k For discrete time points, a set of time indices It is a set of discrete time points k, representing a time period. The time range within the pulse period.

[0020] Assuming a pulse period of T, after discretization there are N sampling points, that is... k =0,1,...,N-1. The pulse period is divided into four time segments, each corresponding to a time index set. and target reference trajectory .

[0021] S110. Based on the characteristics of the pulse waveform, determine the start and end points of each time period and represent them using the sampling point index; like Figure 3As shown, the settings are as follows: Rising edge period: from k_rise_start to k_rise_end, a total of N_rise points; Flat-top period: from k_flat_start to k_flat_end, a total of N_flat points; Falling edge period: from k_fall_start to k_fall_end, a total of N_fall points; Rest period: from k_rest_start to k_rest_end, a total of N_rest points. That is, the rising edge period is from the start of the rising edge at the current time to the end of the rising edge at the current time; the flat-top period is from the end of the rising edge at the current time to the start of the falling edge at the current time; the falling edge period is from the start of the falling edge at the current time to the end of the falling edge at the current time; the rest period is from the end of the falling edge at the current time to the start of the rising edge at the next time. The four time periods are consecutive, i.e., k_rise_start=0, k_rise_end+1=k_flat_start, k_flat_end+1=k_fall_start, k_fall_end+1=k_rest_start, k_rest_end=N-1.

[0022] S120. For each time period, define its time index set. That is, the index of all sampling points included in that time period.

[0023] Define the time index set for each time period as follows: Ω1 = {k | k_rise_start ≤ k ≤ k_rise_end}; Ω2 = {k | k_flat_start ≤ k ≤ k_flat_end}; Ω3 = {k | k_fall_start ≤ k ≤ k_fall_end}; Ω4 = {k | k_rest_start ≤ k ≤ k_rest_end}.

[0024] S130. Define the target reference trajectory for each time period. That is, the expected output signal during that time period.

[0025] For each time period, define according to the desired waveform shape. The rising edge is set as an exponential rise, the flattening period is a constant, the falling edge is an exponential fall, and the resting period is the opposite constant of the flattening period. Therefore, the target reference trajectory... It can be defined as: rising edge ( ): (1) Flat-top period ( ): (2) Falling edge ( ): (3) Rest period ( ): (4) Wherein, time constant t r and t f The set reference rise and fall times.

[0026] S2. Pre-generate a set of reference waveforms When the pulse parameters change, the reference waveform is changed. Remapping the time axis by time period generates a new reference waveform. Otherwise, maintain the reference waveform. The pulse parameters include the target pulse width PW, amplitude A, and pulse period T.

[0027] S210. For different target pulse widths PW and amplitudes A, a set of reference waveforms are generated in advance through low-speed, high-precision measurement or model simulation. This is used as the initial goal and convergence criterion for iterative learning.

[0028] S220. When the pulse parameters (pulse width PW, amplitude A, period T) change, execute the variable sampling reconstruction algorithm to generate a reference waveform and ensure time axis alignment by remapping the time axis. This step includes: S221. After applying anti-aliasing filtering to the waveform, the reference waveform is extracted.

[0029] The reference waveform is decimated to reduce the amount of data collected, and anti-aliasing filtering is performed before decimation to prevent frequency aliasing. The design of an anti-aliasing filter is as follows: A low-pass filter is typically used, with a cutoff frequency of... An FIR filter was chosen because of its linear phase characteristic. After filtering, each D One point is extracted from each point to obtain a new reference waveform. The extraction factor is... (Round to the nearest integer). Wherein, As the reference sampling frequency, The new sampling frequency; its mathematical representation is: reference waveform At the baseline sampling rate Defined, length is Let the reference waveform be... The impulse response of the anti-aliasing filter is The filtered signal is: (5) The waveform at the new sampling rate was obtained: (6) S222. Using cubic spline interpolation method to... Resampling to the current control cycle Corresponding discrete time point index Above, generate the current reference waveform.

[0030] S222.1 Establish mapping relationship Let the period of the reference pulse be The new pulse period is Reference waveform Defined in interval Above. Map it to a new interval. Typically, the shape of a pulse waveform is determined by several key parameters: rise time. descent time Flat top width Rest time When these parameters change, the waveform is non-linearly scaled to maintain the relative proportions of each time segment.

[0031] Assume the time length of each segment of the reference waveform is: rising edge Flat-top period Falling edge Rest period ,satisfy The duration of each segment of the new waveform: rising edge Flat top Falling edge Rest period ,satisfy Establish from the reference time To the new time mapping function Perform piecewise linear mapping: At the rising edge ( ): ,when ; During the flat-top period ( ): ,when ; At the falling edge ( ): ,when ; During the rest period ( ): ,when ; S222.2 Resampling Once the mapping relationship is determined, sampling needs to be performed on the new time axis. New time axis discrete points: Corresponding time . The number of sampling points for the new cycle. Let be the number of sampling points in the reference period; assuming the reference waveform has At each point, the new reference waveform needs... One point. Need to be added for each new index. Find the corresponding reference time Then, the magnitude at that point is obtained by interpolation; Depend on Through the inverse function of the mapping Obtain the corresponding reference time, and then refer to the reference waveform. exist The value at that point is used as the new reference waveform. The value at that point. Since the mapping is piecewise linear, the inverse mapping is also easy to obtain. For example, in the rising segment, when Sometimes, The other paragraphs are similar; S222.3 Higher-order interpolation In obtaining Later, due to The time interval may not be an integer multiple of the reference discrete time, requiring interpolation of the reference waveform. Cubic spline interpolation, linear interpolation, etc., can be used. Cubic spline interpolation is typically used to ensure waveform smoothness.

[0032] Assume the reference waveform at discrete time points The value on Corresponding time We need to ask for time now. The value at that point can be obtained using the cubic spline interpolation function. To calculate: .

[0033] The specific steps are as follows: 1. Construct a cubic spline interpolation function based on the reference waveform. .

[0034] (7) in, , , , These are the interpolation coefficients; 2. For each new index ,calculate .

[0035] 3. Obtained through inverse mapping .

[0036] 4. Calculation , For the new amplitude, The reference amplitude; S223. Correct the phase deviation of the starting point and key feature points of each time period to align them on the time axis.

[0037] After timeline mapping, it is also necessary to ensure that the starting points and key feature points of each time period are aligned. This is because even after time mapping, numerical errors or inaccurate mapping may cause slight deviations in the positions of feature points (such as the midpoint of the rising edge, the starting point of the falling edge, etc.). Such deviations will lead to inaccurate time period division in subsequent hierarchical control, thus requiring alignment.

[0038] 1) Calculate the theoretical time position of each feature point based on the new pulse parameters.

[0039] The theoretical position under the new parameters is: , , ;in , , These are the midpoint times for the rising edge, the flat top, and the falling edge, respectively. 2) Detect the time position of the actual feature points from the newly generated reference waveform.

[0040] For example, the midpoint of the rising edge can be detected by finding the point where the amplitude is 50%; (8) 3) Calculate the time offset for each feature point. (9) in The actual time at the midpoint of each time period; 4) Time-domain spline interpolation is used to correct the time shift of the entire waveform.

[0041] (10) in It is the cubic spline interpolation function of the original waveform; 5) After correction, the time index sets for each time period are re-divided. .

[0042] S3. Time Periods Independent and parallel update iterative learners based on benchmark reference waveforms or Obtain the updated waveform; Each time period Each has a corresponding iterative learner (ILC) that stores the optimal feedforward control input for that time period. The optimal feedforward control input for this time period is the learning result output from the previous cycle. Each time period updates its own iterative learner independently and in parallel, using an algorithm with a forgetting factor and adaptive learning gain for the update law. (11) It is the forgetting factor, usually close to 1, to ensure learning stability. To accommodate the learning gain, the plateau region (e.g., the flat-top segment) can be increased for faster convergence; the transition region (e.g., rising / falling edges) can be decreased to avoid overshoot. This is a learning filter used to correct the nonlinear phase of the system and ensure learning convergence. To advance the learning steps and compensate for system latency. This represents the tracking error during that period.

[0043] Learning Filter The design is crucial, as it needs to compensate for the system's phase lag and amplitude attenuation. To avoid high-frequency noise amplification, the introduced learning filter employs an inverse model filtering method, and its transfer function can be expressed as: (12) S4. Adjust the adaptive learning gain in the update law for the rising edge, flat-top period, and falling edge according to the Metropolis criterion. ; This represents the adaptive learning gain for the i-th time period in the j-th pulse cycle.

[0044] like Figure 4 As shown, S4 employs a differentiated strategy to adjust the adaptive learning gain. : at the rising / falling edge ( S 1, S 3) The time period should be set to a smaller value to suppress pulse overshoot or undershoot. In the flat-top region ( S 2) Time period: Set a larger time period β It can quickly converge fluctuations to within ±1% high precision.

[0045] The Simulated Annealing (SA) algorithm was first proposed by Metropolis et al. Utilizing the Metropolis criterion within the annealing algorithm, adaptive learning gains can be effectively found at various stages, including the rising / falling edges and the flattening phase. The goal is to find the global optimum to improve recognition accuracy and avoid local optima. The steps are as follows: S410. Initialize parameters, set initial temperature. Termination temperature and current temperature T Therefore, the initial and final temperatures during the rising / falling edge and the flattening period can be set to [0.1, 0.3] and [0.5, 0.9] respectively, which is... The training range; d This is the temperature decay parameter, and its value is... ,in The sampling period; the randomly generated initial state. (in ω Including adaptive learning gain The number of iterations for each temperature is L; S420. Status Apply perturbation Generate a new state ,but Disturbance It is a very small perturbation that follows a uniform distribution, and the corresponding cost function is given by... Become Calculation formula: (13) Cost function optimization problems are typically represented as minimizing function values. The cost function is usually a loss function, which can be expressed as: (14) in, It is the model prediction value Compared with the true value The loss between the two samples, where N is the number of samples. Loss function f ( x ) is a complex non-convex function that may contain multiple local minima.

[0046] S430. If Δ F If ≤0, then the next state value is If Δ F >0, then: (15) Then with the generated random number (0< <1) Compare, if < p The next state value is Conversely, the next state value is still... ω This is the Metropolis criterion; S440. Determine if the iteration count has been reached. If not, repeat steps S420 and S430. If the iteration count has been reached, determine the tracking error for the current time period. Is it less than or equal to the threshold of the objective function? That is, the condition is Until T = T final And obtain optimized adaptive learning gain. .

[0047] If convergence fails, then j = j + 1, return to step S4, and enter the j-th pulse cycle for the next round of learning. Conversely, if convergence occurs, the system enters a steady-state operating mode.

[0048] In summary, the method provided in this embodiment is as follows: Figure 5 As shown.

[0049] Figure 6 Before optimization, the pulse output waveform exhibits a distinct sawtooth oscillation characteristic in its transition section, accompanied by significant voltage overshoot and overshoot, leading to decreased energy transfer efficiency and potentially causing load loss; in contrast, Figure 7 The optimized pulse output waveform features smooth transitions on both the rising and falling edges, completely eliminating overshoot and demonstrating excellent dynamic pulse tracking performance. Under high-power operation, the flat-top power fluctuation range is precisely controlled, ensuring high consistency of output energy and fully meeting the requirements of high-frequency applications, significantly improving waveform consistency and repeatability. Figure 8 The output waveform is a single pulse. By adjusting the DC voltage output of the power supply module or the amplification gain of the power amplifier output module, and configuring different duty cycles, multi-level pulse waveforms can be output. It can be seen that the rising and falling edges are smoothly modulated, without overshoot or undershoot.

[0050] In one embodiment, a plasma power pulse control device is provided for performing the steps of the plasma power pulse control method described above.

[0051] In practical applications, such as Figure 9 As shown, the plasma radio frequency power supply includes a power module, a power circuit, a power amplifier module, a matching circuit, a reaction chamber, and a plasma power pulse control device. The power module is responsible for converting AC power to DC power. It typically includes circuits such as rectifiers, filters, and voltage regulators to provide a stable DC voltage output. The power circuit is the core component for generating pulse signals. It uses a reference source to generate a stable frequency. Power circuits typically have the ability to perform amplitude modulation, frequency modulation, or phase modulation on signals as needed. Among them, the DDS direct digital frequency synthesizer is a type of radio frequency generator. DDS can achieve dynamic frequency modulation and amplitude modulation according to the processor's commands.

[0052] The power amplification module amplifies the modulated pulse signal, increasing its power to the required level. Its power amplification capability depends on the application requirements; single-stage or multi-stage amplifiers can be used to achieve different power levels. By adjusting the DC voltage output of the power supply module or the amplification gain of the power amplification module (i.e., adjusting the duty cycle of the pulse signal amplitude) and setting different duty cycle adjustment times, a single-pulse output can be changed to a multi-stage pulse output. The matching network is used to match the input impedance of the resonant converter with the actual load impedance. When the matching is successful, the impedance presented by the matching network is equal to the complex conjugate of the load impedance. At this time, the entire system behaves as a purely resistive load, and energy transfer reaches its optimal state. The reaction chamber is used to ionize gas molecules to form a high-density plasma. The plasma power supply pulse control device is used to monitor and control various parameters of the radio frequency power supply system. It uses the changes in pulse width PW, amplitude A, and pulse period T to dynamically adjust the pulse waveform by outputting control quantities. Each pulse period is divided into multiple characteristic segments, which are modeled and controlled separately. Waveform self-correction is achieved through iterative learning.

[0053] The plasma power supply pulse control device provided in this application embodiment adopts the same inventive concept as the above-described plasma power supply pulse control method and can achieve the same beneficial effects, so it will not be described again here.

[0054] The above embodiments are only used to provide a detailed description of the technical solutions of this application. However, the descriptions of the above embodiments are only for helping to understand the methods of the embodiments of the present invention and should not be construed as limiting the embodiments of the present invention. Variations or substitutions that can be easily conceived by those skilled in the art should be covered within the protection scope of the embodiments of the present invention.

Claims

1. A plasma power supply pulse control method, characterized in that, include: The waveform of the controlled target within a single pulse cycle is divided into time periods, and the target reference trajectory for each time period is determined. The divided time periods Including rising edge Flat-top period Falling edge and rest period ; Pre-generate a set of reference waveforms And when the pulse parameters change, the reference waveform is changed. Remapping the time axis by time period generates a new reference waveform. Otherwise, maintain the reference waveform. ; Pulse parameters include target pulse width, amplitude, and period; Each time period The iterative learner is updated independently and in parallel to obtain independent update control variables for each time period; the update control variables are specifically... , This represents the feedforward control variable, where i represents the time period and j represents the pulse period. This indicates that the forgetting factor is a preset constant. For adaptive learning gain, For the learning filter, q represents a complex variable in the continuous domain. To learn steps ahead of time, For time period Tracking error, As the target reference trajectory, This is the actual output; Determine the optimal adaptive learning gain based on the Metropolis criterion. To obtain the optimal adaptive learning gain The corresponding update control quantity; the update control quantity acts on the waveform or The actual output waveform is obtained.

2. The method according to claim 1, characterized in that, The process of dividing the waveform within the pulse period into time segments and determining the time index set and target reference trajectory for each time segment includes: Based on the characteristics of the pulse waveform, determine the start and end points of each time period; For each time period, determine the set of time indices. and target reference trajectory .

3. The method according to claim 2, characterized in that, The reference waveform is changed when the pulse parameters change. Remapping the time axis by time period generates a new reference waveform. ,include: After applying anti-aliasing filtering to the waveform, the reference waveform is extracted. Using cubic spline interpolation method Resampling to the current control cycle Corresponding discrete time point index Above, generate the current reference waveform; Correct the phase deviation of the starting point and key feature points of each time period to align them on the time axis.

4. The method according to claim 3, characterized in that, Learning Filter Using the inverse model filtering method, the transfer function is expressed as follows: , To learn the filter gain coefficients, , The coefficients of the denominator polynomial, , , These are the coefficients of the numerator polynomial.

5. The method according to claim 4, characterized in that, The optimal adaptive learning gain is obtained according to the Metropolis criterion. ,include: Initialize parameters and set the initial temperature. Termination temperature and Current temperature ; State Apply perturbation Generate a new state Then, the cost function It has become ; , It is the model prediction value Compared with the true value The losses between N It is the number of samples; like The next state value is ,like but ; Check if the iteration count has been reached; if not, then check... To obtain optimized adaptive learning gain .

6. The method according to claim 1, characterized in that, Reference waveform Pre-generation is achieved through low-speed, high-precision measurement or model simulation.

7. A plasma power supply pulse control device, characterized in that, The steps for performing the method according to any one of claims 1 to 6.

Images