阅读说明：本技术 基于深度gru的准分子激光器剂量控制方法及装置 (Depth GRU-based excimer laser dose control method and device ) 是由 冯泽斌 梁赛 刘广义 江锐 徐向宇 刘斌 于 2021-04-15 设计创作，主要内容包括：本发明公开了一种基于深度GRU的准分子激光器剂量控制方法及装置。该方法选取在准分子激光器每发出一个爆发模式的激光脉冲能量序列后的时间间隔中,利用深度门控循环网络判断当前爆发模式与上一个爆发模式的激光脉冲能量损失函数的滑动平均值差的绝对值是否小于阈值；若大于则更新深度门控循环网络的训练参数,得出剂量控制参数的更新值,应用到下一个激光爆发模式的剂量精度控制中；若小于则将当前爆发模式剂量精度控制使用的剂量控制参数,应用到下一个激光爆发模式的剂量精度控制中。本发明针对不同的目标能量、重复频率具有很强的适应性,能够很好的控制准分子激光器的剂量精度,有效控制激光脉冲能量的剂量稳定性。(The invention discloses a depth GRU-based excimer laser dose control method and device. In the method, in a time interval after an excimer laser emits a laser pulse energy sequence of one burst mode, a depth gating cycle network is used for judging whether the absolute value of the difference of the sliding average values of the laser pulse energy loss functions of the current burst mode and the last burst mode is smaller than a threshold value or not; if the current value is larger than the preset value, updating the training parameter of the deep gating circulation network to obtain an updated value of the dose control parameter, and applying the updated value to dose precision control of the next laser burst mode; and if the current burst mode dose precision is smaller than the preset burst mode dose precision, applying the dose control parameters used for controlling the dose precision of the current burst mode to the dose precision of the next laser burst mode. The invention has strong adaptability aiming at different target energies and repetition frequencies, can well control the dose precision of the excimer laser and effectively control the dose stability of laser pulse energy.)

1. A depth GRU-based excimer laser dose control method is characterized by comprising the following steps:

in the process that an excimer laser emits a laser pulse energy sequence of an explosion mode, the laser pulse energy sequence of the current explosion mode is preprocessed in a discharge high voltage-pulse energy combined mode and then is input into a pre-established deep gating circulation network one by one;

in the time interval after the laser pulse energy sequence of the current explosion mode is sent out, the deep gating circulation network judges whether the absolute value of the difference of the sliding average values of the laser pulse energy loss functions of the current explosion mode and the last explosion mode is smaller than a threshold value;

and if the current burst mode is larger than the threshold, updating the training parameters of the deep gating circulation network, performing forward calculation based on the preprocessed laser pulse energy sequence of the current burst mode to obtain an updated value of the dose control parameter, and applying the updated value to dose precision control of the next laser burst mode.

2. The method of depth GRU based excimer laser dose control of claim 1, wherein:

and if the current burst mode dose precision is smaller than the threshold value, stopping updating the training parameters of the gating cycle network, finishing the training of the gating cycle network, and applying the dose control parameters used for controlling the dose precision of the current burst mode to the dose precision of the next laser burst mode.

3. The method of depth GRU based excimer laser dose control of claim 1, wherein:

the deep gating circulation network comprises an input layer, a hidden layer and an output layer, wherein the input layer is connected with the hidden layer, and the hidden layer is connected with the output layer.

4. The method of depth GRU based excimer laser dose control of claim 1, wherein:

the input of the deep gated loop network is represented as:

wherein, x (t) represents the scale conversion of the discharge high voltage of the preprocessed single laser pulse and the energy of the pulse, HV (t) represents the discharge high voltage value corresponding to the energy value of each laser pulse, E (t) represents the energy value of each laser pulse, W (t) represents the energy value of each laser pulse_inTo input the scaling matrix, b represents the shift of the amplitude of each laser pulse, and t is 1,2, …, n, n represents the number of laser pulses in one burst mode.

5. The method of depth GRU based excimer laser dose control of claim 4, wherein:

the output of the output layer of the depth-gated circulation network is a dose control parameter expressed as:

wherein Y (n) represents the value of the deep gate control circulation network generated by the last pulse output network of the laser pulse energy sequence in a burst mode, K_EdProportional parameter, KI, representing a dose-controlling parameter_EdRepresenting dose control parametersIntegral parameter, W_pRepresenting an output scaling matrix;

wherein h (n) represents the state output of the output hidden layer at the moment corresponding to the last pulse of the laser pulse energy sequence in a burst mode, W_yA transformation matrix representing the output states.

6. The method of depth GRU based excimer laser dose control of claim 1, wherein:

the loss function deep gated loop network for the training is represented as:

therein, Dose_TIndicating a target value, Dose, to be controlled for each laser pulse_m,kAnd the laser pulse energy sequence represents the dosage value corresponding to the kth laser pulse in the laser pulse energy sequence of the mth burst mode, the dosage value corresponding to each laser pulse is the sum of the preset number of laser pulse energies, and n represents the number of laser pulses of one burst mode.

7. The method of depth GRU based excimer laser dose control of claim 6, wherein:

and when the deviation between the dose corresponding to the current laser pulse and the corresponding target dose is calculated, if the quantity of the laser pulse energy corresponding to the current laser pulse and the laser pulses in front of the current laser pulse does not reach the preset quantity, the target dose is used for supplementing.

8. The method of depth GRU based excimer laser dose control of claim 6, wherein:

the sliding average value of the laser pulse energy loss function of each burst mode is obtained according to the following formula;

wherein l_slit(m) represents a moving average of laser pulse energy loss functions of the mth burst mode, q represents the number of laser pulse energy loss functions of the burst mode adjacent to the laser pulse energy loss function of the mth burst mode, and l (i) represents a laser pulse energy loss function of the ith burst mode adjacent to the laser pulse energy loss function of the mth burst mode.

9. The method of depth GRU based excimer laser dose control of claim 1, wherein:

calculating the gradient of a loss function l for each parameter of each layer of the deep gated cyclic network according to a time back propagation algorithm, and then updating the corresponding parameter in the network according to the following formula;

Θ(t)＝Θ(t-1)+V

wherein V represents the updating speed, alpha represents a parameter of momentum, alpha is more than 0 and less than 1, theta represents a certain parameter of the deep gating circulation network, and lambda represents the learning rate.

10. A depth GRU based excimer laser dose control apparatus comprising a processor and a memory, the processor reading a computer program or instructions in the memory for performing the following operations:

in the process that an excimer laser emits a laser pulse energy sequence of an explosion mode, the laser pulse energy sequence of the current explosion mode is preprocessed in a discharge high voltage-pulse energy combined mode and then is input into a pre-established deep gating circulation network one by one;

in the time interval after the laser pulse energy sequence of the current explosion mode is sent out, the deep gating circulation network judges whether the absolute value of the difference of the sliding average values of the laser pulse energy loss functions of the current explosion mode and the last explosion mode is smaller than a threshold value;

and if the current burst mode is larger than the threshold, updating the training parameters of the deep gating circulation network, performing forward calculation based on the preprocessed laser pulse energy sequence of the current burst mode to obtain an updated value of the dose control parameter, and applying the updated value to dose precision control of the next laser burst mode.

Technical Field

The invention relates to an excimer laser dose control method based on a deep GRU (gated circulating network), and also relates to a corresponding excimer laser dose control device, belonging to the technical field of laser.

Background

In order to maintain proper operation of the excimer laser, it is necessary to maintain the energy stability and dose stability of the laser. The energy stability and dose stability of the pulses cannot be fully controlled based on existing excimer laser manufacturing processes and materials. Therefore, in this case, a control algorithm is required to be incorporated for the purpose of controlling the pulses.

In the prior art, a decision algorithm is usually adopted to control the dose stability of a laser, and the control effect of the decision algorithm is different along with the difference of energy set values, and the analysis is that the parameter values (PI parameters, namely dose control parameters) of the PI control algorithm used in the decision control are all constant, but the variation rate of energy along with voltage is different in different discharge high-voltage areas in the process of high-voltage discharge light emission of the excimer laser. In order to achieve a good control effect, the PI parameters need to be adjusted and optimized in real time.

Disclosure of Invention

The invention aims to provide a depth GRU-based excimer laser dose control method.

Another technical problem to be solved by the present invention is to provide a depth GRU-based excimer laser dose control apparatus.

In order to achieve the purpose, the invention adopts the following technical scheme:

according to a first aspect of embodiments of the present invention, there is provided a depth GRU-based excimer laser dose control method, including the steps of:

in the process that an excimer laser emits a laser pulse energy sequence of an explosion mode, the laser pulse energy sequence of the current explosion mode is preprocessed in a discharge high voltage-pulse energy combined mode and then is input into a pre-established deep gating circulation network one by one;

in the time interval after the laser pulse energy sequence of the current explosion mode is sent out, the deep gating circulation network judges whether the absolute value of the difference of the sliding average values of the laser pulse energy loss functions of the current explosion mode and the last explosion mode is smaller than a threshold value;

and if the current burst mode is larger than the threshold, updating the training parameters of the deep gating circulation network, performing forward calculation based on the preprocessed laser pulse energy sequence of the current burst mode to obtain an updated value of the dose control parameter, and applying the updated value to dose precision control of the next laser burst mode.

Preferably, if the threshold value is smaller than the threshold value, the updating of the training parameters of the gating circulation network is stopped, the training of the gating circulation network is ended, and the dosage control parameters used for controlling the dosage precision of the current burst mode are applied to the dosage precision of the next laser burst mode.

Preferably, the deep-gated loop network comprises an input layer, a hidden layer and an output layer, wherein the input layer is connected with the hidden layer, and the hidden layer is connected with the output layer.

Wherein preferably, the input of the deep gated loop network is represented as:

wherein, x (t) represents the scale conversion of the discharge high voltage of the preprocessed single laser pulse and the energy of the pulse, HV (t) represents the discharge high voltage value corresponding to the energy value of each laser pulse, E (t) represents the energy value of each laser pulse, W (t) represents the energy value of each laser pulse_inTo input the scaling matrix, b represents the shift of the amplitude of each laser pulse, and t is 1,2, …, n, n represents the number of laser pulses in one burst mode.

Preferably, the output of the output layer of the deep gated circulating network is a dose control parameter expressed as:

wherein Y (n) represents the value of the deep gate control circulation network generated by the last pulse output network of the laser pulse energy sequence in a burst mode, K_EdProportional parameter, KI, representing a dose-controlling parameter_EdIntegral parameter, W, representing a dose control parameter_pRepresenting an output scaling matrix;

wherein h (n) represents the state output of the output hidden layer at the moment corresponding to the last pulse of the laser pulse energy sequence in a burst mode, W_yA transformation matrix representing the output states.

Preferably, the trained loss function deep gated loop network is represented as:

therein, Dose_TIndicating a target value, Dose, to be controlled for each laser pulse_m,kAnd the laser pulse energy sequence represents the dosage value corresponding to the kth laser pulse in the laser pulse energy sequence of the mth burst mode, the dosage value corresponding to each laser pulse is the sum of the preset number of laser pulse energies, and n represents the number of laser pulses of one burst mode.

Preferably, when the deviation between the dose corresponding to the current laser pulse and the corresponding target dose is calculated, if the number of the laser pulse energies corresponding to the current laser pulse and the preceding laser pulses does not reach the preset number, the target dose is used for supplementing.

Preferably, the sliding average value of the energy loss function of the laser pulse of each burst mode is obtained according to the following formula;

wherein l_slit(m) represents a moving average of laser pulse energy loss functions of the mth burst mode, q represents the number of laser pulse energy loss functions of the burst mode adjacent to the laser pulse energy loss function of the mth burst mode, and l (i) represents a laser pulse energy loss function of the ith burst mode adjacent to the laser pulse energy loss function of the mth burst mode.

Preferably, the gradient of the loss function l is calculated for each parameter of each layer of the deep gated loop network according to a time back propagation algorithm, and then the corresponding parameter in the network is updated according to the following formula;

Θ(t)＝Θ(t-1)+V

wherein V represents the updating speed, alpha represents a parameter of momentum, alpha is more than 0 and less than 1, theta represents a certain parameter of the deep gating circulation network, and lambda represents the learning rate.

According to a second aspect of embodiments of the present invention, there is provided a depth GRU based excimer laser dose control apparatus comprising a processor and a memory, the processor reading a computer program or instructions in the memory for performing the following operations:

in the process that an excimer laser emits a laser pulse energy sequence of an explosion mode, the laser pulse energy sequence of the current explosion mode is preprocessed in a discharge high voltage-pulse energy combined mode and then is input into a pre-established deep gating circulation network one by one;

in the time interval after the laser pulse energy sequence of the current explosion mode is sent out, the deep gating circulation network judges whether the absolute value of the difference of the sliding average values of the laser pulse energy loss functions of the current explosion mode and the last explosion mode is smaller than a threshold value;

and if the current burst mode is larger than the threshold, updating the training parameters of the deep gating circulation network, performing forward calculation based on the preprocessed laser pulse energy sequence of the current burst mode to obtain an updated value of the dose control parameter, and applying the updated value to dose precision control of the next laser burst mode.

The method and the device for controlling the dose of the excimer laser based on the depth GRU provided by the invention select the PI parameter for controlling the laser pulse dose of the excimer laser to be adjusted in real time by utilizing the depth gating circulation network in the time interval after the excimer laser emits a laser pulse energy sequence of an explosion mode. The invention has strong adaptability aiming at different target energies and repetition frequencies, can well control the dose precision of the excimer laser, effectively control the dose stability of laser pulse energy, realize that the dose control precision of the excimer laser can be less than 0.5 percent under any working condition, and meet the photoetching requirement.

Drawings

FIG. 1 is a flow chart of a method for depth GRU-based excimer laser dose control according to an embodiment of the present invention;

fig. 2 is an architecture diagram of a depth-gated circulation network in a depth GRU-based excimer laser dose control method according to an embodiment of the present invention;

FIG. 3 is a graph of a loss function variation of a depth-gated circulation network in a method for deep GRU-based excimer laser dose control according to an embodiment of the present invention;

fig. 4 is a graph illustrating a variation of a value of a scaling parameter with a depth-gated cyclic network learning in the depth GRU-based excimer laser dose control method according to the embodiment of the present invention;

fig. 5 is a graph illustrating a variation of a value of an integration parameter with a depth-gated cyclic network learning in the depth GRU-based excimer laser dose control method according to the embodiment of the present invention;

FIG. 6 is a graph showing the variation of the maximum deviation and the minimum deviation of the dose precision in each Burst mode in the depth GRU-based excimer laser dose control method according to the embodiment of the present invention;

FIG. 7 is a graph of the energy stability change in each Burst mode in a depth GRU-based excimer laser dose control method provided by an embodiment of the present invention;

fig. 8 is a graph comparing dose control effects of an excimer laser dose control method based on a depth GRU according to an embodiment of the present invention with dose control methods based on decision control and PID dual closed-loop dose control methods based on PID for setting targets at different energies;

fig. 9 is a graph comparing dose control effects for different repetition frequencies of the excimer laser dose control method based on depth GRU provided in the embodiment of the present invention with the existing dose control method based on decision control and the dose control method based on PID dual closed loop;

fig. 10 is a structural diagram of an excimer laser dose control apparatus based on a depth GRU according to an embodiment of the present invention.

Detailed Description

The technical contents of the present invention will be further described in detail with reference to the accompanying drawings and specific embodiments.

When the working state of the excimer laser changes, such as the change of set target energy, the change of self working gas state or the change of working repetition frequency, because the change of energy along with the change rate of voltage mainly affects the laser pulse dose precision, in order to achieve good control effect when controlling the dose stability of the laser, as shown in fig. 1, the embodiment of the invention provides a depth GRU-based excimer laser dose control method for adjusting the PI parameter of the laser pulse dose control of the excimer laser in real time. The method comprises the following steps:

and step S1, in the process that the excimer laser emits a laser pulse energy sequence of one explosion mode, preprocessing the laser pulse energy sequence of the current explosion mode in a discharge high voltage-pulse energy combination mode, and inputting the preprocessed laser pulse energy sequence into a pre-established deep gating circulation network one by one.

In the light emitting process of the excimer laser, a Burst mode laser pulse energy sequence is obtained through an energy detector, and a factory generation vector D (t) is generated through a vector in a discharge high voltage-pulse energy combination mode, namely the energy value of each laser pulse and the corresponding discharge high voltage value generate a factory generation vector D (t) through the vector, and the method is specifically represented as follows:

hv (t) represents a discharge high voltage value corresponding to an energy value of each laser pulse, e (t) represents an energy value of each laser pulse, t is a time corresponding to each laser pulse in 1,2, … n laser pulses, and n represents a number of laser pulses in a Burst mode.

In the laser pulse energy sequence of the current Burst mode, the energy value of each laser pulse and the vector after the pre-processing of the corresponding discharge high voltage value are input into a pre-established deep gating circulation network one by one.

The deep gating circulation network comprises an input layer, a hidden layer and an output layer, wherein the input layer is connected with the hidden layer, and the hidden layer is connected with the output layer. In the embodiment of the present invention, in order to improve the generalization capability of the deep gated loop network, as shown in fig. 2, the input layer of the deep gated loop network is configured to receive the preprocessed laser pulse energy sequence of the current burst mode and perform the scaling. The hidden layer comprises a middle hidden layer and an output hidden layer, wherein the middle hidden layer is formed by overlapping a plurality of layers of depth gating circulating network layers output in sequence. The output hidden layer is the last hidden layer and is formed by a depth gating circulation network layer finally output by the sequence. Wherein the excitation of the current burst pattern input to the last hidden layerAnd the last pulse energy in the light pulse energy sequence and the corresponding discharge high voltage pass through an output layer of the whole depth gating circulation network and pass through the whole depth gating circulation network to obtain PI parameters, and the PI parameters are applied to dose precision control of the next laser explosion mode. The PI parameters comprise proportional parameters and integral parameters. W in FIG. 2 denotes the state transition matrix between the various levels of the deep gated loop network, h_x(t) denotes the t-th hidden neuron in the x-th layer, i.e., a GRU unit cell.

Specifically, the relevant quantities for the adaptive adjustment of the PI parameter are known from the analysis to be the discharge high voltage of each laser pulse and the energy of that pulse. Therefore, when the deep-gated loop network is established, its input is related to the discharge high voltage and the energy of the single pulse, i.e., the input of the deep-gated loop network can be expressed as equation (2).

Where, t is 1,2, …, n, x (t) represents the discharge high voltage of the preprocessed single laser pulse and the energy of the pulse, which are input at each moment, and the scale conversion is carried out, hv (t) represents the discharge high voltage value corresponding to the energy value of each laser pulse, e (t) represents the energy value of each laser pulse, W (t) represents the energy value of each laser pulse_inTo input the scaling matrix, b represents the shift in amplitude of each laser pulse.

For the output layer of the deep gated loop network, the output form of the network takes the form of the last output, i.e., the value produced by the network of the last pulse output of the laser pulse energy sequence in a Burst mode. Because the purpose of establishing the deep gating circulation network is to enable the PI parameter of the dose accuracy control to be automatically adjusted in real time to achieve self-adaptation, the output of the output layer of the deep gating circulation network is the PI parameter of the dose accuracy control, and the output layer can be expressed as an expression (3).

Wherein P represents PI parameter of dose precision control, Y (n) represents value generated by last pulse output network of laser pulse energy sequence of a Burst mode of the depth gating circulation network, and K_EdRepresenting a proportional parameter, KI_EdDenotes an integral parameter, W_pRepresenting the output scaling matrix, the main purpose being to obtain a suitable scaling of Y (n), so W_pEach element selects a real number greater than 0.

As can be seen from the formula (3), the PI parameters output by the output layer of the deep gated loop network are all positive numbers, so that when the deep gated loop network performs forward calculation, it is ensured that each element in y (n) is a positive number, and thus when the hidden layer is output to output y (n) for transformation, the activation function of the neuron selects an activation function that always keeps a positive number. In the present invention, the softplus function is chosen as the activation function of the output neuron, i.e. the softplus function

Wherein h (n) represents the state output at the last moment of outputting the hidden layer, i.e. the state output at the moment corresponding to the last pulse of the laser pulse energy sequence in a Burst mode, W_yA transformation matrix representing the output states.

The learning target of the deep gating circulation network is a PI parameter for dose accuracy control, but the optimal PI parameter is unknown, so that when a loss function for network training is constructed, the optimal value of the PI parameter cannot be directly used as the target, and an indirect loss function needs to be constructed. Because the maximum deviation of the dose is concerned in the dose accuracy control of the excimer laser, a loss function needs to be constructed according to the Integral (ISE) of the controlled deviation square of the finally selected laser dose, wherein the integral time is a laser pulse energy sequence of a Burst mode, and the functional form is shown as a formula (5).

Therein, Dose_TIndicating a target value, Dose, to be controlled for each laser pulse_m,kThe method comprises the steps of representing a dosage value corresponding to a kth laser pulse in a laser pulse energy sequence of an mth Burst mode, wherein the dosage value corresponding to each laser pulse is the sum of laser pulse energies of a preset number, namely the sum of the laser pulse and a plurality of laser pulse energies in front of the laser pulse; n represents the number of laser pulses of one Burst mode. The formula represents the deviation of the doses corresponding to all laser pulses in the laser pulse energy sequence of a Burst mode and the corresponding target doses, and the smaller the value of the deviation is, the smaller the sum of the dose deviation and the target dose value is, namely, the more stable the sum is, the closer the PI parameter used in the control process is to the optimal value. It should be emphasized that, when calculating the deviation between the current laser pulse corresponding dose and the corresponding target dose, if the current laser pulse and the number of laser pulse energies corresponding to the preceding laser pulses do not reach the preset number, the target dose is used for supplementation. For example, taking the sum of the laser pulse energies corresponding to a single laser pulse as an example, if the total number of laser pulse energies corresponding to the current laser pulse and the preceding laser pulses is 25, and the total number of laser pulse energies corresponding to the current laser pulse and the preceding laser pulses is less than 30, then 5 target doses are compensated to obtain the dose corresponding to the current laser pulse.

Considering that when the number of the hidden layers of the deep gating circulating network is too large, not only is gradient explosion or disappearance easily caused in the training process, but also more software and hardware resources are consumed in application, so that the network training and running are slow; therefore, by adopting a simulation experiment method in the embodiment of the invention, the deep gated circulation network can meet the requirements of excimer laser dose precision control on the dynamic performance of an adjustment period in the future, can meet the final dose control precision, and can also meet the requirement of saving computing resources, the hidden layer of the deep gated circulation network is set to be 3 layers, and the hidden state dimension is set to be 4 dimensions.

Specifically, a simulation experiment method is adopted to determine that a hidden layer of a deep gating circulation network is set to be 3 layers, and the hidden state dimension is set to be 4 dimensions, and the specific process is as follows:

in the test of the number of layers of the deep gated circulation network, 4 dimensions are selected for the dimensions of hidden layers of the deep gated circulation network. The initial values of the adjustment parameters of the deep gated cyclic network are set as random values, the network with the structure of one layer, two layers and three layers is tested for 10 times, and the results are shown in table 1, table 2 and table 3 respectively.

By comparing the results of the convergence periods in table 1, table 2 and table 3, it can be seen that the convergence periods of more than 20 bursts in table 1 are 4, the convergence periods of more than 20 bursts in table 2 are 2, all the convergence periods in table 3 are less than 20 bursts, and the convergence periods are significantly reduced as the number of layers increases. Therefore, the convergence period is gradually reduced and tends to be stable along with the increase of the hidden layer of the deep gating circulation network. As can be seen by comparing the loss function values in the three tables, the value of the loss function gradually decreases as the hidden layer of the depth-gated loop network increases. The Dose precision variance in the three tables is compared to obtain that the Dose precision variance is slightly reduced but not obvious with the increase of the number of hidden layers of the depth-gated cyclic network. By comparing the worst Dose precision of the three tables, it can be found that the value of the worst Dose precision is slightly reduced as the number of hidden layers of the depth-gated cyclic network is increased. Finally, the worst Dose precision in the three tables is less than 5%, the dynamic effect of Dose precision control is greatly improved along with the increase of the number of layers of the deep gating circulating network, and the steady-state control effect is slightly increased. Therefore, in the depth GRU-based excimer laser dose control method, a network with three hidden layers can achieve a good dynamic effect, and meanwhile, the steady-state precision can meet the requirement of integrated circuit photoetching

Table 1 test results of deep gated cyclic network with one hidden layer

Table 2 deep gated loop network test results with two hidden layers

Table 3 test results of deep gated cyclic network with three hidden layers

The structure of the deep gated loop network is not only related to the number of layers of the hidden layer, but also possibly to the dimensions of the hidden states in the hidden layer. In order to verify whether the hidden layer state dimension designed by the invention is proper, the hidden layer state dimension of the deep gated cyclic network is set to be 10 and 20, 10 simulation experiments are respectively carried out, and the obtained results are shown in tables 4 and 5. Comparing table 3, table 4 and table 5, it can be found that there is no obvious difference whether the convergence period of the dynamic performance is characterized, the loss function and the Dose precision variance of the stable performance are represented, or the worst Dose precision of the final effect of Dose control is measured. Therefore, in order to save computing resources in application, the hidden layer state dimension of the deep gating circulation network is selected to be 4 dimensions.

Table 4 deep gated loop network test results with hidden layer state dimension of 10

TABLE 5 deep gated round robin network test results with hidden layer state dimension of 20

Step S2, in the time interval after the laser pulse energy sequence in the current burst mode is sent out, the deep gating cycle network determines whether the absolute value of the difference between the moving average values of the laser pulse energy loss functions in the current burst mode and the previous burst mode is smaller than a threshold.

In the embodiment of the invention, the PI parameter controlled by the laser pulse dose of the alignment molecule laser is adjusted in real time by using the deep gating circulation network, and meanwhile, the network needs to be trained in real time, so that the network can give a better PI parameter to ensure that the PI parameter is always kept in an optimal state. In order to keep the real-time property of the deep gating circulation network training and the adaptability of dose precision control, the selection is carried out in the time interval after the excimer laser emits a laser pulse energy sequence of a Burst mode.

Specifically, the training of the deep gated loop network is an on-line training that accompanies the entire laser pulse dose accuracy control lifecycle. Since the laser pulse energy data generated during the laser discharge process always accompanies with noise, the precision of laser pulse dose control cannot approach the set value of laser pulse dose infinitely, i.e. the result of the laser pulse energy loss function of a burst mode cannot approach zero wirelessly, and when the loss function decays to a certain value, it will not decay. If the network continues to be trained by using the loss function which does not attenuate any more at this time, the deep gating circulation network is overfitting, so that a pathological PI parameter is generated, and the laser pulse dose precision is influenced. Therefore, it is necessary to set the condition that the parameters of the deep gated loop network are not updated so as to avoid the occurrence of the overfitting.

The main factor of the on-line training of the deep gating circulation network is the change of a loss function, and when the PI parameter of the laser pulse dose control precision calculated by the deep gating circulation network is optimal, the loss function is stable and should be kept basically unchanged. The value obtained by the loss function in each Burst mode fluctuates due to the presence of noise in the pulse energy emitted by the excimer laser. In order to eliminate the fluctuation in the loss function, the loss function used for judging when the PI parameter of the laser pulse dose control of the alignment molecule laser reaches the optimum is adopted, and the sliding average value of a plurality of recent loss functions is adopted. The formula for calculating the running average of the loss function is as follows.

In the above formula, /)_slit(m) represents a moving average of laser pulse energy loss functions of the mth Burst mode, q represents the number of laser pulse energy loss functions of the Burst mode adjacent to the laser pulse energy loss function of the mth Burst mode, and l (i) represents a laser pulse energy loss function of the ith Burst mode adjacent to the laser pulse energy loss function of the mth Burst mode.

The conditions that the established deep gating circulation network does not update the parameters are set as follows: the absolute value of the difference of the moving average values of the energy loss functions of the laser pulses of the current burst mode and the previous burst mode is smaller than the threshold epsilon, and the expression is shown in formula (7).

|l_slit(m)-l_slit(m-1)|＜ε (7)

Wherein the threshold is set according to the practical situation of the deep gating circulation network training. Therefore, in the time interval after each burst mode of laser pulse energy sequence is emitted by the excimer laser, the deep gating cycle network firstly judges whether the absolute value of the difference of the moving average values of the laser pulse energy loss functions of the current burst mode and the last burst mode is smaller than the threshold value so as to determine whether to update the training parameters.

And step S3, if the value is larger than the threshold value, updating the pre-established training parameters of the deep gating circulation network, and performing forward calculation based on the preprocessed laser pulse energy sequence of the current burst mode to obtain the updated value of the dose control parameter, and applying the updated value to the dose precision control of the next laser burst mode.

The loss function of the deep gated cyclic network is shown in formula (5), wherein Dose_m,kAnd the discharge high voltage value of the laser pulse, which is unknown due to the lack of an accurate discharge energy dynamics model. This relationship is expressed as an implicit function, as shown in equation (8):

Dose_m,t＝f(HV_Dm,t) (8)

wherein, HV_Dm,tAnd when the dose value corresponding to the t-th laser pulse in the laser pulse energy sequence of the m-th Burst mode is represented, the high voltage value calculated by a dose control algorithm in the discharge voltage composition corresponding to the last laser pulse is represented.

Since the function in the formula (8) is an implicit function, and a specific function form cannot be given, the derivative of the function is solved by a numerical method, as shown in the formula (9):

for HV_Dm,tCan be expressed as the output of the dose controller as shown in equation (10):

wherein, Delta Dose_m,t-1Shows the deviation of the dose value corresponding to the t-1 th laser pulse in the laser pulse energy sequence of the m-th Burst mode and the corresponding target dose,and the sum of the deviation between the accumulated dose value corresponding to the t laser pulses and the corresponding target dose at the time corresponding to the t laser pulse in the laser pulse energy sequence of the mth Burst mode is shown.

Writing equation (9) as a matrix multiplication form can be expressed as equation (11):

HV_Dm,t＝DError_m,tP_m (11)

wherein, P_mPI parameter, vector DER representing dose accuracy control used in the laser pulse energy sequence of the mth Burst mode_m,tCan be represented by formula (12):

in the deep gated circulation network, the output hidden layer is the state h (t) output at the moment corresponding to the last pulse of a laser pulse energy sequence in a Burst mode, so that the derivative of the loss function for the output is only related to the output at the moment corresponding to the last pulse, and net is set_y(t)＝W_yH (t), then:

then it can be found that:

when t is equal to n, there are:

when t < n, there are:

in the process of error back propagation, each hidden layer respectively utilizes a corresponding algorithm to calculate the partial derivative of the loss function aiming at each parameter needing to be updated in the network.

In the dose control of the excimer laser, the performance of each laser energy pulse and dose needs to be kept good, and if the laser pulse energy is unstable or the dose accuracy does not reach the standard, the photoetching processing of an integrated circuit is seriously influenced. Therefore, the real-time property of parameter adjustment and the stability and control precision of laser dose after parameter update are both considered when the PI parameter adjustment of dose precision control is performed by using the depth gating circulation network.

In the invention, a parameter updating algorithm is adopted as a random gradient descent method (SGD with momentum, SGDM) carrying momentum, and in the SGDM algorithm, parameter updating relates to calculation of speed and momentum. The update speed is defined as V, the parameter of the momentum is set to α (0 < α < 1), a certain parameter in the network is set to Θ, and the learning rate is set to λ, and the speed update formula can be expressed as formula (17).

Then, the update of each set of parameters in the deep gated loop network can be expressed in the form of equation (17).

Θ(t)＝Θ(t-1)+V (18)

Therefore, when the depth-gated cyclic network judges that the absolute value of the difference of the moving average values of the energy loss functions of the laser pulses of the current burst mode and the last burst mode is larger than the threshold value, the gradient of the loss function l, namely the partial derivative of l with respect to the parameter, is calculated according to a Back Propagation Through Time (BPTT) for each parameter of each layer of the depth-gated cyclic network, and then the corresponding parameter in the network is updated according to the formulas (17) and (18), so that the purpose of network training is achieved.

In the time-back-propagation algorithm,the first order partial derivative is a partial derivative, and if the partial derivative is ill-conditioned, it directly causes the loss function to ill-condition the partial derivative of each parameter in the network, so the value obtained by this derivative is limited. The limiting method is to specify the maximum and minimum values as shown in formula (19), where v is_minTo setLower limit of value, v_maxTo setThe upper limit of the value.

If the denominator HV in equation (9) is calculated_Dm,t-HV_Dm,t-1Is a value close to zero or very small, which may cause operation errors or the partial derivative of the final network parameter is close to infinity, thereby causing a ill-conditioned state for adjusting the network parameter. In addition, the memory may overflow during actual operation, so that it must be limited. As can be seen from analyzing equation (9), if there is no significant difference in the values of the discharge high voltage calculated by the dose controller for the two laser pulses before and after, it is meaningless to calculate equation (9), which cannot represent the derivative of the dose with the discharge high voltage, and if equation (9) is calculated in this case, a large noise is introduced. Equation (9) can be written as equation (20), where Δ HV_DlimIs a suitably positive number, representing a limit value. The formula is shown as HV_Dm,t-HV_Dm,t-1Is less than a certain limit, the partial derivative of the dose with respect to the discharge high voltage is set to zero.

And performing forward calculation by adopting a deep gating circulation network after updating the training parameters based on the preprocessed laser pulse energy sequence of the current burst mode to obtain an updated value of the PI parameter, and applying the updated value to the dose precision control of the next laser burst mode, namely enabling the dose controller to calculate the discharge high voltage corresponding to each laser pulse of the next laser burst mode according to a formula (10).

And step S4, if the threshold value is smaller than the threshold value, stopping updating the pre-established training parameters of the gating circulation network, finishing the training of the gating circulation network, and applying the dosage control parameters used for the dosage precision control of the current burst mode to the dosage precision control of the next laser burst mode.

The excimer laser dose control method based on the depth GRU of the invention is used for carrying out control verification on the molecular laser energy simulation model. An excimer laser energy simulation model is set to work at the repetition frequency of 4KHz, and 250 laser pulse energies exist in a laser pulse energy sequence of a Burst mode. The number of pulses for calculating the dose value of each laser pulse was set to 30 laser pulse energies.

Setting target energy in an excimer laser dose controller to be 10mJ, controlling an excimer laser energy simulation model, and observing loss function change, maximum deviation and minimum deviation change of dose precision of each Burst mode laser energy, proportional parameter value change of dose precision control and integral parameter value change of dose precision control.

The laser pulse energy was controlled for 1000 Burst modes, and the loss function for each Burst mode varied as shown in fig. 3. As can be seen from the figure, the loss function of the deep gating circulation network is rapidly attenuated along with the operation of a simulation model of the network, the loss function is attenuated to the limit after reaching the 9 th Burst mode, no obvious attenuation change exists, and the parameter convergence of the network is proved.

The proportional parameter P and the integral parameter I in the dose control vary with the decay of the loss function as shown in fig. 4 and 5. It can be seen from fig. 4 that the value of the scaling parameter P converges to 71.09 as the loss function decays, and no longer changes. It can be seen from fig. 5 that the value of the integral parameter I converges to 7.22 as the loss function decays.

The maximum and minimum variation of the laser energy dose precision degradation for each Burst mode, accompanied by the operation of the depth GRU based excimer laser dose control method of the present invention, is shown in fig. 6. As can be seen from the figure, when the loss function does not converge, the dose accuracy deviation of each Burst mode is large, and as the loss function converges, the dose accuracy deviation of the Burst mode gradually decreases to be within plus or minus 0.5%. From the stopping convergence of the deep gating circulation network to the ending of the energy control operation, the maximum deviation value of the dosage precision is-0.42%, the deviation of all dosage precisions is within plus or minus 0.5%, and the requirement that the dosage precision is less than 0.5% is met. Therefore, the depth GRU-based excimer laser dose control method can be proved to be capable of well controlling the dose precision of the excimer laser.

With the operation of the depth GRU based excimer laser dose control method of the present invention, the variation of the laser pulse energy stability per Burst mode is shown in fig. 7, and the energy stability per Burst mode is measured by 3 δ. It can be seen from the figure that with the convergence of the depth-gated circular network, the energy stability of the excimer laser gradually becomes stable, and the 3 δ stability thereof is below 5.5%, which is far less than 15% of the lithography requirement. Therefore, the depth GRU-based excimer laser dose control method disclosed by the invention can effectively control the energy stability of laser pulse energy.

In order to verify the adaptability of the depth GRU-based excimer laser dose control method, the control effect of the algorithm is verified under different set energy values and different repetition frequencies respectively.

In order to verify the adaptability of the depth GRU-based excimer laser dose control method of the invention to different set energy values, an excimer laser energy simulation model is set to work under the working condition of repetition frequency of 4KHz, and control target energy values are respectively set to be 8mJ, 8.5mJ, 9mJ, 9.5mJ and 10mJ, and the control effect is shown in Table 6 by utilizing the method of the invention to control. From the table, it can be seen that no matter the dynamic performance or the steady-state performance, the control result of the method has no obvious difference for different set values, and the method is proved to have strong adaptability for the target energy.

TABLE 6 control effect of deep GRU energy control method for different target energy values

Aiming at a simulation model with the repetition frequency of 4KHz and aiming at different set target energy values, the depth GRU-based excimer laser dose control method is compared with the conventional decision control-based dose control method and PID double-closed-loop-based dose control method, and the comparison result is shown in FIG. 8. In the figure, the abscissa is the different energy setting targets and the ordinate is the dose control accuracy, expressed as the maximum deviation H of all dose control accuracies. The comparison in the figure can show that, in the three methods, the method has the strongest adaptability to different target energy set values, and when the target energy set value is changed, the dosage control precision is basically kept unchanged, and the requirements that the maximum deviation of the dosage control precision is less than 0.5 percent are met. Compared with the other two control methods, the obtained dosage control precision is highest.

In order to verify the adaptability of the depth GRU-based excimer laser dose control method to different repetition frequencies of the excimer laser, the target value of the energy is set to 10mJ, and the adaptability of the method is verified under the working conditions that the work repetition frequency of an excimer laser energy simulation model is 500Hz and 1KHz, 2KHz, 3KHz and 4KHz respectively. The dynamic characteristics of the control effect and the steady-state control accuracy are shown in table 7. As can be seen from the table, the convergence period of the dynamic characteristics has no obvious difference under the working conditions of different repetition frequencies; the worst dose precision, the dose precision variance and the loss function value which represent the stable state characteristics have no obvious difference under different repetition frequency working conditions, and the steady state dose precision under all the working repetition frequencies meets the requirement of less than 0.5 percent, thereby proving that the method has strong adaptability aiming at different repetition frequencies of the molecular laser.

TABLE 7 deep GRU energy control method for different repetition frequency control effects

Aiming at the working conditions that the energy target is set to be 10mJ and the work repetition frequency of an excimer laser energy simulation model is 500Hz, 1KHz, 2KHz, 3KHz and 4KHz, the excimer laser dose control method based on the depth GRU is compared with the existing dose control method based on decision control and the dose control method based on the PID double closed loop, and the comparison result of the dose control precision is shown in figure 9. In the figure, the abscissa is different working repetition frequencies of the laser energy simulation model, and the ordinate is the dose control precision. The dose accuracy is expressed as the maximum deviation H of all dose control accuracies. The comparison in the figure can show that, in the three methods, the method has the strongest adaptability to different work repetition frequencies, and when the work repetition frequency is changed, the dosage control precision is basically kept unchanged, and the requirements that the maximum deviation of the dosage control precision is less than 0.5 percent are met. Compared with other two control methods, the method has the advantage that the dosage control precision is highest when the control reaches a steady state.

In addition, as shown in fig. 10, an embodiment of the present invention further provides a depth GRU-based excimer laser dose control apparatus, which includes a processor 32 and a memory 31, and may further include a communication component, a sensor component, a power supply component, a multimedia component, and an input/output interface according to actual needs. The memory, communication components, sensor components, power components, multimedia components, and input/output interfaces are all connected to the processor 32. As mentioned above, the memory 31 may be a Static Random Access Memory (SRAM), an Electrically Erasable Programmable Read Only Memory (EEPROM), an Erasable Programmable Read Only Memory (EPROM), a Programmable Read Only Memory (PROM), a Read Only Memory (ROM), a magnetic memory, a flash memory, etc.; the processor 32 may be a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Field Programmable Gate Array (FPGA), Application Specific Integrated Circuit (ASIC), Digital Signal Processing (DSP) chip, or the like. Other communication components, sensor components, power components, multimedia components, etc. may be implemented using common components found in existing smartphones and are not specifically described herein.

In addition, the present invention provides a depth GRU-based excimer laser dose control apparatus, which includes a processor 32 and a memory 31, wherein the processor 32 reads a computer program or instructions in the memory 31, and is configured to perform the following operations:

in the process that an excimer laser emits a laser pulse energy sequence of one burst mode, the laser pulse energy sequence of the current burst mode is preprocessed in a discharge high voltage-pulse energy combined mode and then is input into a pre-established deep gating circulation network one by one.

In the time interval after the laser pulse energy sequence of the current burst mode is sent out, the deep gating circulation network judges whether the absolute value of the difference of the sliding average values of the laser pulse energy loss functions of the current burst mode and the last burst mode is smaller than a threshold value.

And if the current burst mode is larger than the threshold, updating the pre-established training parameters of the deep gating circulation network, performing forward calculation based on the preprocessed laser pulse energy sequence of the current burst mode to obtain an updated value of the dose control parameter, and applying the updated value to dose precision control of the next laser burst mode.

And if the current burst mode dose precision is smaller than the threshold value, stopping updating the pre-established training parameters of the gating circulation network, finishing the training of the gating circulation network, and applying the dose control parameters used for controlling the dose precision of the current burst mode to the dose precision of the next laser burst mode.

In addition, an embodiment of the present invention further provides a computer-readable storage medium, where instructions are stored on the computer-readable storage medium, and when the instructions are executed on a computer, the computer is enabled to execute the method for providing a dose control of an excimer laser based on a depth GRU as described in fig. 1, and details of a specific implementation manner of the method are not repeated here.

In addition, an embodiment of the present invention further provides a computer program product including instructions, which when run on a computer, causes the computer to execute the method for providing depth GRU-based excimer laser dose control as described in fig. 1, and detailed implementation thereof is not repeated herein.

The method and the device for controlling the dose of the excimer laser based on the depth GRU provided by the invention select the PI parameter for controlling the laser pulse dose of the excimer laser to be adjusted in real time by utilizing the depth gating circulation network in the time interval after the excimer laser emits a laser pulse energy sequence of a Burst mode. The method has strong adaptability aiming at different target energies and repetition frequencies, can well control the dose precision of the excimer laser, effectively control the dose stability of laser pulse energy, realize that the dose control precision of the excimer laser is less than 0.5 percent under any working condition, and meet the photoetching requirement.

The method and apparatus for controlling the dose of an excimer laser based on a deep GRU provided by the present invention are described in detail above. It will be apparent to those skilled in the art that various modifications can be made without departing from the spirit of the invention.