Real-time monitoring and feedback of FIB equipment processing parameter optimization methods and systems

Abstract

This invention provides a method and system for optimizing processing parameters of FIB (Film Injection Blade) equipment with real-time monitoring and feedback. It generates a real-time monitoring data set by collecting real-time monitoring data of the processing environment and processing effect of the FIB equipment. Based on the processing environment and processing effect feature parameter sets in the real-time monitoring data set, parameter feature extraction is performed on the real-time monitoring data set to generate a parameter feature vector set. This parameter feature vector set is input into a pre-trained dynamic parameter optimization model to generate a set of optimization parameter adjustment instructions. Based on these instructions, the current processing parameters of the FIB equipment are adjusted to generate an adjusted processing parameter set. Real-time processing is then performed based on the adjusted parameter set, and the real-time acquisition process is repeated cyclically to update the real-time monitoring data set. This invention can meet the requirements of high-precision, complex structure processing for real-time performance, stability, and energy efficiency.

Description

Technical Field

[0001] This invention relates to the fields of data processing and intelligent control, and more specifically, to a method and system for optimizing the processing parameters of FIB equipment through real-time monitoring and feedback. Background Technology

[0002] Dynamic optimization of processing parameters in focused ion beam (FIB) equipment is a core technology for improving the precision of micro and nano-fabrication. It achieves nanoscale structure processing by real-time control of ion beam intensity, scanning path, and focusing parameters. Existing technologies typically employ environmental parameter calibration or offline compensation methods, such as adjusting the ion beam intensity based on a preset temperature compensation table or triggering fixed-step parameter correction through a material removal rate threshold. However, these methods can only perform linear compensation for simple environmental variables and cannot analyze the multi-dimensional dynamic coupling relationship between temperature distribution, electric field gradient, and magnetic field deviation in real time. This makes it difficult to quickly suppress ion beam focusing distortion caused by thermal expansion during processing, and there is a lack of a collaborative compensation mechanism for scanning path offset caused by electromagnetic environmental disturbances. The processing precision is significantly affected by instantaneous fluctuations in environmental parameters. In addition, traditional single-factor feedback control strategies suffer from parameter adjustment lag, large fluctuations in material removal rate, and a lack of multi-objective collaborative mechanisms in balancing processing efficiency and equipment energy consumption. These methods fail to meet the stringent requirements of real-time performance, stability, and energy efficiency in high-precision complex structure processing. Summary of the Invention

[0003] This invention provides a method and system for optimizing the processing parameters of FIB equipment through real-time monitoring and feedback.

[0004] In a first aspect, embodiments of the present invention provide a method for optimizing processing parameters of a FIB (Fiber Injection) equipment based on real-time monitoring and feedback. The method includes: real-time acquisition of processing environment monitoring data and processing effect monitoring data of the FIB equipment to generate a real-time monitoring data set; the processing environment monitoring data includes temperature distribution data, electric field intensity distribution data, and magnetic field intensity distribution data; the processing effect monitoring data includes ion beam focusing accuracy data and material removal rate data; based on the processing environment feature parameter set and the processing effect feature parameter set in the real-time monitoring data set, parameter feature extraction processing is performed on the real-time monitoring data set to generate a parameter feature vector set; each feature vector in the parameter feature vector set contains… The process involves combining characteristic values ​​of temperature fluctuations, electric field intensity gradients, and magnetic field intensity deviations. This set of parameter feature vectors is input into a pre-trained dynamic parameter optimization model to generate a set of optimized parameter adjustment instructions. The dynamic parameter optimization model is trained using the mapping relationship between environmental parameters and processing parameters in historical processing data. The set of optimized parameter adjustment instructions includes ion beam intensity adjustment values, scanning path offsets, and focusing compensation coefficients. Based on the set of optimized parameter adjustment instructions, the current processing parameters of the FIB equipment are adjusted to generate an adjusted processing parameter set. Real-time processing is performed based on the adjusted processing parameter set, and the real-time acquisition process is cyclically executed to update the real-time monitoring data set.

[0005] Secondly, embodiments of the present invention provide a computer system, comprising: a memory storing a computer program; and a processor for loading the computer program to implement the real-time monitoring and feedback method for optimizing FIB equipment processing parameters as described above.

[0006] This invention provides a real-time monitoring and feedback-based method for optimizing FIB (Fiber to the Source) equipment processing parameters. By fusing the spatiotemporal dynamic characteristics of multi-dimensional environmental parameters such as temperature distribution, electric field gradient, and magnetic field deviation in real time, and combining the processing effect feedback of material removal rate and ion beam focusing accuracy, a set of feature vectors characterizing complex processing environments and process states is constructed. A pre-trained deep learning model intelligently analyzes the nonlinear coupling relationship between environmental and processing parameters, generating collaborative optimization instructions for ion beam intensity, scanning path offset, and focusing compensation coefficients, achieving closed-loop dynamic control of processing parameters. This method overcomes the limitations of traditional parameter compensation. Through multi-source data feature enhancement technology, it effectively suppresses ion beam focusing distortion and positioning deviation caused by temperature drift and electromagnetic interference, significantly improving processing accuracy stability. Based on a real-time feedback incremental learning mechanism, it continuously optimizes parameter adjustment strategies, adapting to fluctuations in material removal rate under different operating conditions, ensuring efficient and controllable processing. Simultaneously, a multi-objective optimization algorithm is introduced to dynamically balance the synergistic relationship between processing accuracy, efficiency, and equipment energy consumption, achieving simultaneous optimization of multi-dimensional performance indicators in complex processing tasks. This provides an intelligent solution for high-precision micro / nano manufacturing that combines real-time response capabilities, environmental adaptability, and multi-objective collaboration. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a flowchart of a method for optimizing the processing parameters of a FIB (Fill in the Blank) equipment through real-time monitoring and feedback, provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] Please see Figure 1 , Figure 1The flowchart illustrates a method for optimizing FIB equipment processing parameters based on real-time monitoring and feedback, as provided in an embodiment of the present invention. This method can be executed by a computer system and may include the following steps:

[0012] Step S100: Real-time acquisition of processing environment monitoring data and processing effect monitoring data of FIB equipment to generate a real-time monitoring data set; the processing environment monitoring data includes temperature distribution data, electric field intensity distribution data and magnetic field intensity distribution data, and the processing effect monitoring data includes ion beam focusing accuracy data and material removal rate data.

[0013] In this embodiment of the invention, the FIB device, or focused ion beam device, is used for micro-nano fabrication and analysis. Processing environment monitoring data refers to data related to the environment in which the FIB device operates during the processing. Temperature distribution data reflects the spatial distribution of temperature in and around the processing area. For example, in semiconductor chip processing, temperature differences at different locations can affect material performance and processing accuracy. Electric field intensity distribution data describes the intensity changes of the electric field within the processing space; the presence of the electric field affects the trajectory of the ion beam. Magnetic field intensity distribution data reflects the intensity distribution of the magnetic field in the processing environment; the magnetic field can also interfere with the path of the ion beam. Processing effect monitoring data is a key indicator for measuring the processing quality and efficiency of the FIB device. Ion beam focusing accuracy data indicates the degree to which the ion beam can be precisely focused on the target location; higher focusing accuracy results in better processing resolution and accuracy. Material removal rate data refers to the amount of material removed from the processed material per unit time, reflecting the processing efficiency.

[0014] To acquire this data in real time, various sensors can be used. For example, an infrared thermal imager can be used to acquire temperature distribution data, enabling rapid and non-contact measurement of the temperature in the processing area; electric field strength sensors and magnetic field strength sensors can measure the intensity distribution of electric and magnetic fields, respectively; ion beam focusing accuracy can be monitored using an ion beam imaging system, which can display the focusing state of the ion beam in real time; and the material removal rate can be determined by measuring the change in material mass before and after processing or by observing changes in the material surface using an optical microscope. By acquiring and integrating this data in real time, a real-time monitoring dataset can be generated, providing a foundation for subsequent parameter optimization.

[0015] Step S200: Based on the processing environment feature parameter set and processing effect feature parameter set in the real-time monitoring data set, perform parameter feature extraction processing on the real-time monitoring data set to generate a parameter feature vector set; each feature vector in the parameter feature vector set includes a combination of temperature fluctuation feature values, electric field intensity gradient feature values, and magnetic field intensity deviation feature values.

[0016] In this step, the processing environment characteristic parameter set is a set of parameters extracted from the processing environment monitoring data that reflect the characteristics of the processing environment, such as the temperature fluctuation range and the gradient change of the electric field; the processing effect characteristic parameter set is a set of parameters related to the processing effect extracted from the processing effect monitoring data, such as the stability of the ion beam focusing accuracy and the consistency of the material removal rate. Parameter feature extraction processing refers to extracting representative and distinguishable features from the real-time monitoring data set in order to better describe the state of the processing process.

[0017] Temperature fluctuation characteristic values ​​reflect the temperature change over a period of time. For example, temperature fluctuation characteristic values ​​can be obtained by calculating the standard deviation of the temperature; the larger the standard deviation, the more drastic the temperature fluctuation. Electric field intensity gradient characteristic values ​​represent the rate of change of electric field intensity in space. These characteristic values ​​can be obtained by differentiating the electric field intensity distribution data. Magnetic field intensity deviation characteristic values ​​refer to the degree of deviation of the magnetic field intensity from a preset reference magnetic field. The magnetic field intensity deviation characteristic value is obtained by comparing the actual measured magnetic field intensity with the reference magnetic field intensity and calculating their difference.

[0018] These feature values ​​are combined into feature vectors, forming a set of parameter feature vectors. For example, a feature vector can be represented as [temperature fluctuation feature value, electric field intensity gradient feature value, magnetic field intensity deviation feature value]. This set can more comprehensively and accurately describe the processing status of the FIB equipment, providing more effective input for subsequent parameter optimization.

[0019] As one implementation method, step S200, based on the processing environment feature parameter set and processing effect feature parameter set in the real-time monitoring data set, performs parameter feature extraction processing on the real-time monitoring data set to generate a parameter feature vector set, which can be specifically implemented as follows:

[0020] Step S210: Extract the time series fluctuation characteristics of temperature distribution data from the processing environment monitoring data to generate a temperature fluctuation characteristic value sequence.

[0021] The time-series fluctuation characteristics of temperature distribution data refer to the patterns and characteristics of temperature changes over time. In actual processing, temperature may fluctuate due to factors such as equipment operation and the influence of the external environment. To extract these fluctuation characteristics, time series analysis methods can be used. For example, the moving average method can be used to smooth the temperature distribution data, and then the temperature difference between adjacent time points can be calculated. By performing statistical analysis on these differences, such as calculating the mean and standard deviation, the temperature fluctuation characteristic values ​​can be obtained. Arranging these characteristic values ​​in chronological order generates a temperature fluctuation characteristic value sequence. This sequence can intuitively reflect the temperature fluctuation at different time points, providing an important reference for subsequent parameter optimization.

[0022] Step S220: Extract spatial gradient variation features from electric field intensity distribution data to generate an electric field intensity gradient feature value matrix.

[0023] Electric field intensity distribution data describes the intensity distribution of the electric field within the processing space. Spatial gradient variation characteristics refer to the rate of change of the electric field intensity in space, reflecting the non-uniformity of the electric field. Numerical differentiation methods can be used to extract these characteristics. For example, the processing space can be divided into several small grids, and the gradient of the electric field intensity can be calculated at each grid point. The gradient can be calculated by taking the partial derivatives of the electric field intensity in different directions. Arranging the calculated gradient values ​​according to the grid point positions generates an electric field intensity gradient eigenvalue matrix. This matrix clearly shows the spatial variation of the electric field intensity, which helps in analyzing the influence of the electric field on the ion beam motion.

[0024] Step S230: Extract the deviation features relative to the preset reference magnetic field from the magnetic field intensity distribution data to generate a set of magnetic field intensity deviation feature values.

[0025] The preset reference magnetic field is an ideal magnetic field state determined according to the design requirements and processing technology of the FIB equipment. Magnetic field strength distribution data is the distribution of the actual measured magnetic field strength within the processing space. The deviation characteristic relative to the preset reference magnetic field refers to the difference between the actual magnetic field strength and the reference magnetic field strength. To extract these characteristics, the actual measured magnetic field strength and the reference magnetic field strength are compared point-by-point, and their differences are calculated. These differences are then compiled into a set, generating the magnetic field strength deviation characteristic value set. This set reflects the degree of deviation between the actual and ideal state of the magnetic field and is of great significance for evaluating the impact of the magnetic field on ion beam focusing and processing accuracy.

[0026] Step S240: Normalize the temperature fluctuation feature value sequence, the electric field intensity gradient feature value matrix, and the magnetic field intensity deviation feature value set to generate a standardized environmental feature dataset.

[0027] Normalization transforms data from different ranges and scales into a unified range for comparison and analysis. For temperature fluctuation characteristic sequences, electric field intensity gradient characteristic matrices, and magnetic field intensity deviation characteristic sets, direct comparison and fusion can be difficult due to differences in their dimensions and value ranges. Therefore, normalization is necessary. Common normalization methods include min-max normalization and Z-score normalization. Taking min-max normalization as an example, for each element in the dataset, the transformation is performed using the formula (x-min) / (max-min), where x is the original data, min is the minimum value in the dataset, and max is the maximum value. After normalization, the value ranges of these datasets are mapped to the [0,1] interval, generating a standardized environmental feature dataset. This dataset eliminates scale differences between different data, providing a more accurate data foundation for subsequent weighted fusion.

[0028] Step S250: Based on the correlation between ion beam focusing accuracy data and material removal rate data in the processing effect monitoring data, generate processing effect weight coefficients, and weight and fuse the standardized environmental feature dataset with the processing effect weight coefficients to generate a set of parameter feature vectors.

[0029] There is a correlation between ion beam focusing accuracy data and material removal rate data. During FIB (Film Injection Bypass) processing, an increase in ion beam focusing accuracy may lead to a change in the material removal rate, and vice versa. To quantify this correlation, statistical analysis methods, such as correlation analysis, can be used. By calculating the correlation coefficient between ion beam focusing accuracy data and material removal rate data, the degree of their correlation can be determined. Based on this correlation, a processing effect weighting coefficient is generated. This coefficient represents the importance of the processing effect in parameter optimization.

[0030] The standardized environmental feature dataset is weighted and fused with the processing effect weight coefficient. This involves multiplying each element in the standardized environmental feature dataset by its corresponding processing effect weight coefficient, and then summing the results. This method effectively combines processing environment features and processing effect features, generating a set of parameter feature vectors. This set can more comprehensively reflect the processing status of the FIB equipment, providing more accurate input for subsequent parameter optimization.

[0031] Step S300: Input the set of parameter feature vectors into the pre-trained dynamic parameter optimization model to generate a set of optimization parameter adjustment instructions; the dynamic parameter optimization model is generated by training the mapping relationship between environmental parameters and processing parameters in historical processing data, and the set of optimization parameter adjustment instructions includes ion beam intensity adjustment value, scanning path offset and focusing compensation coefficient.

[0032] In this embodiment of the invention, the pre-trained dynamic parameter optimization model is a trained machine learning model that can predict the optimal processing parameter adjustment instructions based on the input parameter feature vector set. This model is trained using the mapping relationship between environmental parameters and processing parameters in historical processing data. The historical processing data contains a large number of processing environmental parameters (such as temperature, electric field, magnetic field, etc.) and corresponding processing parameters (such as ion beam intensity, scanning path, focus compensation, etc.). By analyzing and learning from this data, the model can establish the mapping relationship between environmental parameters and processing parameters.

[0033] When the set of parameter feature vectors is input into the dynamic parameter optimization model, the model predicts parameter adjustment instructions that can optimize the processing effect based on the learned mapping relationship. The set of optimized parameter adjustment instructions includes ion beam intensity adjustment value, scan path offset, and focus compensation coefficient. The ion beam intensity adjustment value is used to adjust the intensity of the ion beam to control the material removal rate and processing accuracy; the scan path offset is used to adjust the scan path of the ion beam to avoid deviations during processing; and the focus compensation coefficient is used to compensate for the focusing error of the ion beam and improve the focusing accuracy.

[0034] As one implementation method, step S300, which involves inputting the set of parameter feature vectors into a pre-trained dynamic parameter optimization model to generate a set of optimization parameter adjustment instructions, can be specifically implemented as follows:

[0035] Step S310: Obtain the multilayer time series analysis network in the dynamic parameter optimization model; the multilayer time series analysis network includes an LSTM layer for extracting the temporal dependencies of environmental parameters and a fully connected layer for generating parameter adjustment instructions.

[0036] Multilayer time series analysis networks are a core component of dynamic parameter optimization models, capable of handling data with time-series characteristics. During the processing of FIB equipment, environmental parameters (such as temperature, electric field, and magnetic field) change over time, and these changes exhibit temporal dependencies. LSTM (Long Short-Term Memory) layers are a special type of recurrent neural network that can effectively handle long-range dependencies in time-series data. Through LSTM layers, the characteristics and patterns of environmental parameter changes over time can be extracted.

[0037] Fully connected layers are used to convert the features extracted by the LSTM layer into specific parameter adjustment instructions. Each neuron in a fully connected layer is connected to all neurons in the previous layer. By performing linear transformations and nonlinear activations on the feature vector output by the LSTM layer, parameter adjustment instructions are generated. For example, a fully connected layer can map the feature vector output by the LSTM layer to parameter spaces such as ion beam intensity adjustment values, scan path offsets, and focus compensation coefficients.

[0038] Step S320: Divide the parameter feature vector set into multiple subsequences according to the time window, and input each subsequence into the LSTM layer to generate the corresponding hidden state vector set.

[0039] To better process the time-series information in the parameter feature vector set, it needs to be divided into multiple subsequences according to time windows. A time window is a fixed-length period of time, within which data is considered a subsequence. For example, dividing the parameter feature vector set into windows of 10 time points yields multiple subsequences.

[0040] Each subsequence is input into an LSTM layer. The LSTM layer calculates the hidden state for the current time step based on the input of the current subsequence and the hidden state of the previous time step. The hidden state is a vector within the LSTM layer that contains all the information up to the current time step. By processing each subsequence, the LSTM layer generates a corresponding set of hidden state vectors. This set reflects the changing characteristics and patterns of environmental parameters within different time windows.

[0041] Step S330: Perform a product operation between each vector in the hidden state vector set and the weight matrix of the fully connected layer to generate a preliminary set of parameter adjustment values.

[0042] The weight matrix of the fully connected layer is learned during model training and represents the mapping between hidden state vectors and parameter adjustment values. Multiplying each vector in the hidden state vector set with the weight matrix of the fully connected layer yields a preliminary set of parameter adjustment values. For example, assuming the hidden state vectors have dimension n and the weight matrix of the fully connected layer has dimension n×m, where m is the dimension of the parameter adjustment values ​​(such as ion beam intensity adjustment, scan path offset, and focus compensation coefficients), multiplying the hidden state vectors with the weight matrix using matrix multiplication results in an m-dimensional vector, which represents the preliminary parameter adjustment values. Performing this operation on each hidden state vector generates the preliminary set of parameter adjustment values.

[0043] Step S340: Perform constraint filtering on the preliminary parameter adjustment value set to generate a candidate parameter adjustment instruction set that conforms to the physical limitations of the FIB equipment; the constraints include the maximum threshold of ion beam intensity and the threshold of the rate of change of focusing compensation coefficient.

[0044] The initial set of parameter adjustment values ​​is obtained through model prediction, but these values ​​may exceed the physical limitations of the FIB equipment. Therefore, constraint filtering is required. Constraints are limitations determined based on the design requirements and safety standards of the FIB equipment, such as the maximum threshold for ion beam intensity and the threshold for the rate of change of the focusing compensation coefficient. The maximum threshold for ion beam intensity refers to the maximum value that the ion beam intensity cannot exceed; exceeding this value may damage the equipment and processing materials. The threshold for the rate of change of the focusing compensation coefficient refers to the maximum rate at which the focusing compensation coefficient changes; excessively rapid changes may lead to instability in ion beam focusing.

[0045] Each value in the initial parameter adjustment set is checked, and if a value exceeds the constraint range, it is adjusted or removed. In this way, a candidate parameter adjustment instruction set that conforms to the physical limitations of the FIB equipment is generated. This set of instructions satisfies both the model's prediction requirements and the equipment's physical limitations, ensuring the safety and stability of the processing.

[0046] Step S350: Based on the predicted processing effect score of each instruction in the candidate parameter adjustment instruction set, select the instruction with the highest score as the set of optimization parameter adjustment instructions.

[0047] Each instruction in the candidate parameter adjustment instruction set corresponds to a predicted processing effect score, which is obtained by evaluating the instruction's impact on the processing effect. The predicted processing effect score can consider multiple factors, such as processing accuracy, material removal rate, and equipment energy consumption. The predicted processing effect score for each instruction is obtained by comprehensively evaluating these factors.

[0048] The instructions with the highest scores are selected as the set of instructions for optimizing parameters. This ensures that the selected instructions can maximize the processing effect. For example, if an instruction has the highest predicted processing effect score, it means that the instruction has good overall performance in improving processing accuracy, increasing material removal rate, and reducing equipment energy consumption. Selecting this instruction as the set of instructions for optimizing parameters can improve the processing efficiency and quality of the FIB equipment.

[0049] As one implementation method, step S350, based on the predicted processing effect scores corresponding to each instruction in the candidate parameter adjustment instruction set, selects the instruction with the highest score as the set of optimization parameter adjustment instructions, which can be specifically implemented as follows:

[0050] Step S351: Extract the ion beam intensity adjustment value, scan path offset and focusing compensation coefficient corresponding to each instruction from the candidate parameter adjustment instruction set to generate a candidate instruction parameter triplet set.

[0051] Each instruction in the candidate parameter adjustment instruction set contains three parameters: ion beam intensity adjustment value, scan path offset, and focus compensation coefficient. These three parameters for each instruction are extracted and combined into a triplet. For example, if an instruction corresponds to an ion beam intensity adjustment value of 'a', a scan path offset of 'b', and a focus compensation coefficient of 'c', then the generated triplet is (a, b, c). Combining the triplets corresponding to all instructions generates the candidate instruction parameter triplet set. This set clearly displays the specific parameter information of each candidate instruction, providing a foundation for subsequent processing effect prediction.

[0052] Step S352: Input the set of candidate instruction parameter triples into the pre-trained processing effect prediction model, and output the processing accuracy prediction value, material removal rate prediction value and equipment energy consumption prediction value corresponding to each triple, and generate the target item prediction value set.

[0053] The pre-trained processing effect prediction model is a trained machine learning model that predicts the processing effect corresponding to each triplet of input candidate instruction parameters. Processing effects include aspects such as processing accuracy, material removal rate, and equipment energy consumption. When the set of candidate instruction parameter triplets is input into the processing effect prediction model, the model outputs predicted values ​​for processing accuracy, material removal rate, and equipment energy consumption for each triplet, based on learned knowledge and patterns. Combining these predicted values ​​generates a set of predicted values ​​for the target item. This set provides specific data support for subsequent scoring.

[0054] Step S353: Perform error inverse normalization on the processing accuracy prediction values ​​in the target item prediction value set to generate a processing accuracy score sequence; perform rate classification mapping on the material removal rate prediction values ​​to generate a removal rate score sequence; and perform energy efficiency ratio conversion on the equipment energy consumption prediction values ​​to generate an energy consumption score sequence.

[0055] Error inverse normalization is a method to convert predicted machining accuracy values ​​into scores. Predicted machining accuracy values ​​are typically error values; the smaller the error, the higher the machining accuracy. Through error inverse normalization, the error value is converted into a score, with higher scores indicating higher machining accuracy. For example, a linear mapping method can be used to map the error value to a score range of [0, 100].

[0056] Rate grading mapping is a method for converting predicted material removal rates into scores. Material removal rates can be graded according to different ranges, with each grade corresponding to a score. For example, material removal rates can be divided into three levels: low, medium, and high, each corresponding to a different score. Rate grading mapping converts the predicted material removal rates into a sequence of removal rate scores.

[0057] Energy efficiency ratio (EER) conversion is a method of converting predicted energy consumption values ​​of equipment into a score. The EER is the ratio of a device's output power to its input power; a higher EER indicates lower energy consumption. By converting predicted energy consumption values ​​into EERs and then assigning scores based on these EERs, an energy consumption score sequence is generated.

[0058] Step S354: Linearly superimpose the machining accuracy score sequence, removal rate score sequence, and energy consumption score sequence according to the preset score weight ratio to generate a comprehensive score value sequence for each candidate instruction.

[0059] The preset scoring weights are determined based on the needs and priorities of the processing task, representing the importance of processing accuracy, material removal rate, and equipment energy consumption in the overall score. For example, for tasks with high processing accuracy requirements, the scoring weight for processing accuracy can be set higher; for tasks with high processing efficiency requirements, the scoring weight for material removal rate can be set higher.

[0060] The processing accuracy score sequence, material removal rate score sequence, and energy consumption score sequence are linearly superimposed according to preset score weight ratios. This involves multiplying corresponding elements in each sequence by their respective weights and then summing the results. This generates a comprehensive score sequence for each candidate instruction. This sequence comprehensively considers multiple factors such as processing accuracy, material removal rate, and equipment energy consumption, providing a more accurate basis for selecting the optimal instruction.

[0061] Step S355: Based on the maximum threshold of ion beam intensity and the threshold of the rate of change of focusing compensation coefficient in the constraints, perform dynamic stability verification on the candidate instruction parameter triplet corresponding to the comprehensive score value sequence, remove instructions that cause the equipment state to exceed the historical safe range, and generate a filtered score instruction set.

[0062] Dynamic stability verification ensures that selected instructions will not cause the equipment state to exceed the historical safe range. The maximum threshold for ion beam intensity and the threshold for the rate of change of the focusing compensation coefficient in the constraints are determined based on the equipment's safety requirements. The candidate instruction parameter triplets corresponding to the comprehensive score value sequence are checked. If the ion beam intensity adjustment value corresponding to a triplet exceeds the maximum threshold for ion beam intensity, or the rate of change of the focusing compensation coefficient exceeds the threshold for the rate of change of the focusing compensation coefficient, then the instruction is considered to cause the equipment state to exceed the historical safe range and is discarded. In this way, a filtered scoring instruction set is generated. Instructions in this set have both high comprehensive scores and can guarantee the safe and stable operation of the equipment.

[0063] Step S356: Extract the top N instructions with the highest comprehensive score from the filtered scoring instruction set, generate a high-scoring candidate instruction subset, and compare the compatibility of the high-scoring candidate instruction subset with the current processing parameter set to calculate the smooth transition index of the parameter adjustment range.

[0064] The top N instructions with the highest overall scores are selected from the filtered scoring instruction set to form a high-scoring candidate instruction subset. The value of N can be determined according to actual needs, for example, selecting the top 5 instructions with the highest overall scores. The high-scoring candidate instruction subset is then compared with the current processing parameter set to check the degree of difference between each instruction and the current processing parameters. The smoothness transition index of parameter adjustment amplitude is an indicator used to measure the smoothness of parameter adjustment. For example, the differences between each instruction in the high-scoring candidate instruction subset and the current processing parameter set in terms of ion beam intensity, scan path offset, and focus compensation coefficient can be calculated. These differences are then normalized to eliminate the influence of different parameter dimensions. Next, weights are assigned to each parameter, such as ion beam intensity (0.4), scan path offset (0.3), and focus compensation coefficient (0.3). Finally, the normalized differences are multiplied by their corresponding weights and summed; the result is the smoothness transition index of parameter adjustment amplitude. The smaller the value, the smoother the adjustment. If the parameter adjustment amplitude is too large, it may lead to instability in the processing; if the parameter adjustment amplitude is too small, it may not achieve the goal of optimizing the processing effect. By calculating the smooth transition index of parameter adjustment range, instructions that can effectively optimize processing results and ensure processing stability can be selected.

[0065] Step S357: Select the instruction with the lowest smooth transition index and the highest comprehensive score as the final optimization parameter adjustment instruction, and integrate the ion beam intensity adjustment value, scan path offset and focus compensation coefficient of the instruction into an optimization parameter adjustment instruction set.

[0066] From the subset of high-scoring candidate instructions, the instruction with the lowest smooth transition index and the highest overall score is selected as the final optimized parameter adjustment instruction. This ensures that the selected instruction maximizes the processing effect while guaranteeing a smooth transition in parameter adjustments, avoiding excessive impact on the processing. Integrating the ion beam intensity adjustment value, scan path offset, and focus compensation coefficient of this instruction generates the optimized parameter adjustment instruction set. This set of instructions, obtained through multiple rounds of screening and evaluation, provides the optimal solution for adjusting the processing parameters of the FIB equipment.

[0067] As one implementation method, the training process of a dynamic parameter optimization model may include the following steps:

[0068] Step S301: Obtain historical processing environment dataset and historical processing parameter adjustment record set; wherein, the historical processing environment dataset includes multiple sets of temperature distribution sequences, electric field intensity distribution sequences and magnetic field intensity distribution sequences arranged in chronological order, and the historical processing parameter adjustment record set includes ion beam intensity adjustment sequences, scan path offset sequences and focus compensation coefficient sequences corresponding to each set of historical processing environment data.

[0069] Historical processing environment datasets refer to collections of data related to the processing environment gathered during past processing processes. They include multiple sets of temperature distribution sequences, electric field intensity distribution sequences, and magnetic field intensity distribution sequences arranged chronologically. These sequences record the changes in the processing environment at different points in time. For example, the temperature distribution sequence reflects the temperature changes in the processing area over time, the electric field intensity distribution sequence shows the trend of electric field intensity changes within the processing space, and the magnetic field intensity distribution sequence reflects the fluctuations in magnetic field intensity over time.

[0070] The historical processing parameter adjustment record set refers to the collection of records showing adjustments to processing parameters during past processing operations. It includes ion beam intensity adjustment sequences, scan path offset sequences, and focus compensation coefficient sequences corresponding to each set of historical processing environment data. These sequences record the adjustments made to parameters such as ion beam intensity, scan path, and focus compensation coefficients under different processing environments to achieve optimal processing results. Acquiring this historical data provides a rich sample for training dynamic parameter optimization models.

[0071] Step S302: Perform time-domain filtering on the temperature distribution sequence in the historical processing environment dataset to remove high-frequency noise components and generate a smooth temperature distribution sequence set. Perform spatial alignment correction on the electric field intensity distribution sequence to generate a corrected electric field intensity distribution set. Calculate the difference between the magnetic field intensity distribution sequence and the equipment's reference magnetic field configuration to generate a magnetic field intensity deviation sequence set.

[0072] Time-domain filtering is a method used to remove high-frequency noise components from a signal. In temperature distribution sequences, there may be high-frequency noise caused by measurement errors, external interference, or other factors. Time-domain filtering, such as using a low-pass filter, can remove this high-frequency noise, making the temperature distribution sequence smoother. The resulting smooth temperature distribution sequence set can more accurately reflect the true trend of temperature change.

[0073] Spatial alignment correction is used to process electric field intensity distribution sequences to ensure spatial consistency of electric field intensity data at different time points. In actual measurements, due to factors such as the position and orientation of the measuring equipment, electric field intensity distribution data may exhibit spatial deviations. Spatial alignment correction can correct these deviations, generating a corrected electric field intensity distribution set.

[0074] The difference calculation between the magnetic field strength distribution sequence and the equipment's reference magnetic field configuration is to obtain the deviation of the magnetic field strength relative to the reference magnetic field. The equipment's reference magnetic field configuration is an ideal magnetic field state determined according to the equipment's design requirements and manufacturing process. By comparing the actual measured magnetic field strength distribution sequence with the reference magnetic field configuration point by point and calculating the difference, a set of magnetic field strength deviation sequences is generated. This set can reflect the degree of deviation between the actual and ideal magnetic field state, providing more valuable information for model training.

[0075] Step S303: Align and merge the smooth temperature distribution sequence set, the corrected electric field intensity distribution set, and the magnetic field intensity deviation sequence set according to the timestamp to generate a historical environmental feature matrix set, and convert the ion beam intensity adjustment sequence, scanning path offset sequence, and focusing compensation coefficient sequence in the historical processing parameter adjustment record set into a parameter adjustment vector set.

[0076] Aligning and merging the smoothed temperature distribution sequence, the corrected electric field intensity distribution sequence, and the magnetic field intensity deviation sequence by timestamp is to unify different types of environmental data along the time dimension. Timestamps represent the acquisition time of each data point; aligning by timestamps ensures that data points in different sequences correspond in time. Merging the aligned data generates a historical environmental feature matrix. This matrix comprehensively reflects the characteristics of the processing environment at different points in time.

[0077] Converting the ion beam intensity adjustment sequences, scan path offset sequences, and focus compensation coefficient sequences from the historical processing parameter adjustment record set into a parameter adjustment vector set is to transform these sequences into a vector form suitable for model training. For example, the elements in each sequence can be arranged sequentially to form a vector. Combining the vectors corresponding to all sequences generates the parameter adjustment vector set. The vectors in this set represent the adjustments made to the processing parameters at different time points.

[0078] Step S304: Use the historical environmental feature matrix set as the input sample set and the parameter adjustment vector set as the output label set to form a supervised training dataset.

[0079] The supervised training dataset consists of an input sample set and an output label set. In this embodiment of the invention, the historical environment feature matrix set is used as the input sample set, which contains various feature information of the processing environment; the parameter adjustment vector set is used as the output label set, which represents the processing parameter adjustments that should be made in the corresponding environment. By combining these two sets, the supervised training dataset is formed. This dataset can be used to train a dynamic parameter optimization model, enabling the model to learn the mapping relationship between environmental parameters and processing parameters.

[0080] Step S305: Perform multiple rounds of iterative training on the initial neural network model using the supervised training dataset. In each round of iteration, calculate the mean square error between the model's predicted parameter adjustment vector and the actual parameter adjustment vector, and update the model weights based on backpropagation of the error until the prediction error stabilizes within the preset convergence interval, thereby generating a trained dynamic parameter optimization model.

[0081] The initial neural network model is an untrained model with a pre-defined structure (such as LSTM layers and fully connected layers) and parameters. Multiple iterations of training on the initial neural network model using a supervised training dataset allow the model to gradually learn the mapping relationship between environmental parameters and processing parameters. In each iteration, the input sample set (the historical environmental feature matrix set) is fed into the model, and the model outputs a predicted parameter adjustment vector. This predicted parameter adjustment vector is compared with the true parameter adjustment vector (the vector in the output label set), and their mean squared error (MSE) is calculated. MSE is a metric that measures the difference between the predicted and true values; a smaller error indicates a more accurate prediction.

[0082] Using the backpropagation algorithm, the mean squared error is propagated from the output layer to the input layer, and the gradient value of each neuron is calculated. The gradient value represents the direction and magnitude of the model parameters' change under the current error. The model weights are updated based on the gradient values ​​to make the model's predictions closer to the true values. This process is repeated for multiple rounds of iterative training until the prediction error stabilizes within a preset convergence interval. The preset convergence interval is a pre-defined error range; when the prediction error enters this interval and remains stable, it indicates that the model has converged and training is complete. At this point, a dynamically optimized model with optimized parameters is generated.

[0083] As one implementation method, step S305, which involves performing multiple rounds of iterative training on the initial neural network model using a supervised training dataset, can be specifically implemented as follows:

[0084] Step S3051: Extract subsets of historical environmental feature matrices and corresponding parameter adjustment vector subsets from the supervised training dataset in batches, and input the subsets of historical environmental feature matrices into the feature encoding layer of the initial neural network model to generate environmental feature encoding vector sets.

[0085] To improve training efficiency and model generalization ability, supervised training datasets are typically divided into batches. In each iteration, a subset of the historical environment feature matrix and the corresponding subset of parameter adjustment vectors for each batch are extracted from the supervised training dataset. The subset of historical environment feature matrix is ​​a part of the historical environment feature matrix set, containing the processing environment feature information of that batch; the subset of parameter adjustment vectors is a part of the parameter adjustment vector set, representing the actual processing parameter adjustments for that batch.

[0086] A subset of the historical environment feature matrix is ​​input into the feature encoding layer of the initial neural network model. The role of the feature encoding layer is to encode the input historical environment feature matrix and extract key features. Through the processing of the feature encoding layer, an environment feature encoded vector set is generated. The vectors in this set can more effectively represent the features of the processing environment, providing better input for subsequent model training.

[0087] Step S3052: Input the environmental feature encoding vector set into the LSTM layer of the initial neural network model, extract the dependency pattern of environmental parameters changing over time, and generate a time-series feature vector set.

[0088] The LSTM layer is a crucial component of the initial neural network model, capable of handling data with time-series characteristics. When the environmental feature encoding vector set is input into the LSTM layer, it calculates the hidden state for the current time step based on the current input and the hidden state of the previous time step. By processing the environmental feature encoding vector set, the LSTM layer can extract the dependency patterns of environmental parameters over time. For example, it can learn the relationships and changing patterns of environmental parameters such as temperature, electric field, and magnetic field at different points in time.

[0089] Based on the extracted dependency patterns, the LSTM layer generates a set of temporal feature vectors. These vectors contain the temporal features of environmental parameters, better reflecting the dynamic changes in the processing environment.

[0090] Step S3053: Input the temporal feature vector set into the fully connected layer of the initial neural network model to generate the predicted parameter adjustment vector set, and calculate the error distribution between the predicted parameter adjustment vector set and the subset of the true parameter adjustment vector.

[0091] The fully connected layer is the last layer in the initial neural network model. Its role is to transform the temporal feature vector set into a prediction parameter adjustment vector set. Through the linear transformation and non-linear activation function of the fully connected layer, the temporal feature vectors are mapped to the parameter adjustment vector space, generating the prediction parameter adjustment vector set.

[0092] The predicted parameter adjustment vector set is compared with the subset of the true parameter adjustment vectors, and the error distribution between them is calculated. The error distribution can be calculated in various ways, such as mean squared error or cross-entropy loss. The error distribution reflects the degree of difference between the model's predictions and the true values, providing a basis for subsequent model optimization.

[0093] Step S3054: Calculate the gradient values ​​of the weights of each layer of the initial neural network model according to the error distribution, and use an adaptive learning rate algorithm to scale the gradient values ​​to generate the adjusted weight update amount.

[0094] Based on the error distribution, the backpropagation algorithm can be used to calculate the gradient values ​​of the weights of each layer in the initial neural network model. The gradient value represents the direction and magnitude of the change in model parameters under the current error. By calculating the gradient values, it is possible to determine how to adjust the model weights to reduce the error.

[0095] An adaptive learning rate algorithm is used to scale the gradient values. This algorithm automatically adjusts the learning rate based on the model's training progress, preventing the model from failing to converge due to an excessively large learning rate or from training too slowly due to a too small learning rate. Scaling the gradient values ​​generates adjusted weight updates. This update represents the magnitude of the adjustment to the model weights in the current iteration.

[0096] Step S3055: Iteratively correct the weight parameters of the initial neural network model according to the weight update amount, and verify the prediction accuracy of the model on the independent validation dataset after each iteration. When the improvement of the validation accuracy in five consecutive iterations is less than the preset threshold, terminate the training and generate the final optimized dynamic parameter optimization model.

[0097] The weight parameters of the initial neural network model are iteratively corrected based on the adjusted weight update. In each iteration, the updated weight is added to the current weight parameters to obtain new weight parameters. Through continuous iterative correction, the model's prediction results gradually approach the true values.

[0098] After each iteration, the model's prediction accuracy is validated using an independent validation dataset. The independent validation dataset is a different set of data than the training dataset; it is used to evaluate the model's performance on unseen data. The model's prediction accuracy on the independent validation dataset is calculated, including metrics such as precision and mean squared error.

[0099] When the improvement in validation accuracy over five consecutive iterations is less than a preset threshold, it indicates that the model's performance has stabilized, and further training may not yield significant improvements. At this point, training is terminated, and the final optimized dynamic parameter optimization model is generated. This model, having undergone multiple iterations of training and validation, exhibits good generalization ability and prediction accuracy, effectively predicting the optimal processing parameter adjustment commands based on processing environment parameters.

[0100] Step S400: Adjust the current processing parameters of the FIB equipment according to the set of optimized parameter adjustment instructions, and generate the adjusted processing parameter set.

[0101] After receiving the set of optimized parameter adjustment instructions, the current processing parameters of the FIB equipment are adjusted according to these instructions. The set of optimized parameter adjustment instructions includes information such as ion beam intensity adjustment values, scan path offsets, and focus compensation coefficients. Based on this information, the ion beam intensity, scan path, and focus compensation parameters of the FIB equipment are adjusted accordingly. For example, if the ion beam intensity adjustment value in the optimized parameter adjustment instruction set is positive, it indicates that the ion beam intensity needs to be increased; if it is negative, it indicates that the ion beam intensity needs to be decreased. The current ion beam intensity is added to the ion beam intensity adjustment value to obtain the adjusted ion beam intensity. Similarly, the scan path is adjusted according to the scan path offset, and the focus compensation is adjusted according to the focus compensation coefficient. The adjusted ion beam intensity, scan path, and focus compensation parameters are combined to generate an adjusted processing parameter set. The parameters in this set are optimized, which can improve the processing efficiency and quality of the FIB equipment.

[0102] Step S500: Perform real-time processing based on the adjusted processing parameter set, and repeatedly execute the real-time acquisition process to update the real-time monitoring data set.

[0103] Real-time machining is performed based on the adjusted machining parameter set, meaning the workpiece is machined using the adjusted machining parameters. During machining, a real-time data acquisition process is continuously executed, collecting machining environment monitoring data and machining effect monitoring data from the FIB equipment, and updating the real-time monitoring data set.

[0104] The real-time data acquisition process is the same as that in step S100, using various sensors to acquire real-time data such as temperature distribution, electric field intensity distribution, magnetic field intensity distribution, ion beam focusing accuracy, and material removal rate. The newly acquired data is then merged and updated with the previous real-time monitoring data set to obtain the latest real-time monitoring data set.

[0105] By cyclically executing the process of real-time data acquisition and parameter optimization, the processing parameters can be continuously adjusted according to changes in the processing environment and processing results, thereby achieving real-time optimization of the FIB equipment processing process and improving processing stability and quality.

[0106] As one implementation method, step S500, which involves performing real-time processing based on the adjusted processing parameter set and cyclically executing the real-time acquisition process to update the real-time monitoring data set, can be specifically implemented as follows:

[0107] Step S510: During the execution of the adjusted processing parameter set, real-time thermal imaging data and ion beam trajectory data of the processing area are collected simultaneously.

[0108] When performing real-time processing using the adjusted processing parameter set, it is necessary to simultaneously acquire real-time thermal imaging data and ion beam trajectory data of the processing area. Real-time thermal imaging data can be acquired using an infrared thermal imager, which can reflect the temperature distribution of the processing area in real time. Temperature changes affect material properties and processing accuracy; therefore, real-time thermal imaging data is crucial for monitoring the processing.

[0109] Ion beam trajectory data can be acquired through ion beam imaging systems or other trajectory monitoring equipment, displaying the movement trajectory of the ion beam within the processing area. The trajectory of the ion beam directly affects the processing position and accuracy; by monitoring the ion beam trajectory data, deviations can be detected and adjusted in a timely manner.

[0110] Step S520: Calculate the cumulative thermal effect based on real-time thermal imaging data, and trigger a processing interruption command when the cumulative thermal effect exceeds a preset safety threshold.

[0111] The cumulative thermal effect is calculated based on real-time thermal imaging data. The cumulative thermal effect refers to the degree of heat accumulation due to energy input during processing. It is related to factors such as temperature changes in the processing area and the thermal properties of the material. To calculate the cumulative thermal effect, the real-time thermal imaging data can be divided into multiple sub-regions, and the temperature rise rate and spatial temperature gradient of each sub-region can be calculated. A predicted local thermal deformation value is generated based on the product of the temperature rise rate and the material's coefficient of thermal expansion for each sub-region. The predicted local thermal deformation values ​​of all sub-regions are then weighted and summed to generate the overall cumulative thermal effect.

[0112] The preset safety threshold is a critical value determined based on the properties of the material, the processing technology, and the safety requirements of the equipment. When the cumulative thermal effect exceeds the preset safety threshold, it indicates that excessive heat is generated during processing, which may damage the material and equipment. At this point, a processing interruption command is triggered to stop the processing and ensure the safety and quality of the process.

[0113] Specifically, in step S520, calculating the cumulative thermal effect based on real-time thermal imaging data can be implemented as follows:

[0114] Step S521: Divide the real-time thermal imaging data into multiple sub-regions and calculate the temperature rise rate and spatial temperature gradient of each sub-region.

[0115] Dividing real-time thermal imaging data into multiple sub-regions allows for more detailed analysis of temperature variations within the processing area. These sub-regions can be divided into regular grid areas, such as square or rectangular areas, based on the shape and size of the processing area.

[0116] For each sub-region, calculate its rate of temperature rise and spatial temperature gradient. The rate of temperature rise refers to the rate of change of temperature within the sub-region over time, which can be obtained by dividing the temperature difference between adjacent time points by the time interval. The spatial temperature gradient refers to the rate of change of temperature within the sub-region in space, which can be obtained by dividing the temperature difference between different locations within the sub-region by the distance.

[0117] Step S522: Generate local thermal deformation prediction values ​​based on the product of the temperature rise rate of each sub-region and the material's thermal expansion coefficient.

[0118] The coefficient of thermal expansion is a physical property of a material, representing the degree of expansion of the material when the temperature changes. By multiplying the temperature rise rate of each sub-region by the coefficient of thermal expansion, the predicted local thermal deformation value for each sub-region is obtained. This predicted local thermal deformation value reflects the degree of deformation that may occur within the sub-region due to temperature changes.

[0119] Step S523: Weighted summation of the predicted local thermal deformation values ​​of all sub-regions to generate the overall cumulative thermal effect.

[0120] The weighted summation of the predicted local thermal deformation values ​​for all sub-regions requires assigning a weight to each sub-region. This weight can be determined based on factors such as the sub-region's area and importance. For example, a higher weight can be assigned to a sub-region closer to the processing center.

[0121] The predicted local thermal deformation value of each sub-region is multiplied by its corresponding weight, and then the results are summed to obtain the overall cumulative thermal effect. The overall cumulative thermal effect can comprehensively reflect the degree of thermal effect accumulation within the processing area.

[0122] Step S524: Compare the total accumulated thermal effect with the safety threshold corresponding to different material types. If the threshold is exceeded, generate a processing interruption command and corresponding cooling parameter adjustment suggestions.

[0123] Different material types have different thermal properties and safety requirements, and therefore correspond to different safety thresholds. The calculated cumulative thermal effect is compared with the safety thresholds corresponding to different material types.

[0124] If the cumulative overall thermal effect exceeds the safety threshold, it indicates that excessive heat is generated during processing, which may damage materials and equipment. In this case, a processing interruption command is generated to stop the process. Simultaneously, based on the magnitude of the cumulative thermal effect and the material properties, corresponding cooling parameter adjustment suggestions are generated. These suggestions include information such as cooling method and cooling rate, guiding subsequent cooling operations to reduce the temperature in the processing area and ensure processing safety and quality.

[0125] Step S530: During the processing interruption, generate path calibration parameters based on the deviation between the ion beam trajectory data and the target processing path, and update the adjusted processing parameter set.

[0126] During processing interruptions, the scanning path of the ion beam is calibrated based on the deviation between the ion beam trajectory data and the target processing path. The ion beam trajectory data records the actual movement trajectory of the ion beam, while the target processing path is a pre-set ideal processing path. The deviation between the two is calculated, such as the positional deviation and angular deviation of the ion beam.

[0127] Path calibration parameters are generated based on the deviation, including the offset and rotation angle of the scan path. These path calibration parameters are then updated in the adjusted machining parameter set, ensuring that the adjusted machining parameter set includes the calibrated scan path information.

[0128] Step S540: After restarting the machining process, input the updated and adjusted machining parameter set into the dynamic parameter optimization model for secondary optimization, and generate a corrected set of optimized parameter adjustment instructions.

[0129] After the machining process is interrupted and path calibration is completed, machining is restarted. The updated and adjusted machining parameter set is then input into the dynamic parameter optimization model for secondary optimization. The dynamic parameter optimization model will re-predict the optimal machining parameter adjustment instructions based on the updated machining parameters and the current machining environment.

[0130] Through secondary optimization, a revised set of optimization parameter adjustment instructions is generated. This set of instructions takes into account the situation after path calibration, further improving machining accuracy and efficiency.

[0131] Step S550: Adjust the instruction set according to the corrected optimization parameters and continue processing until all processing tasks are completed.

[0132] Based on the revised set of optimized parameter adjustment instructions, the processing parameters of the FIB equipment are readjusted. Processing continues using the adjusted parameters until all processing tasks are completed. During processing, the process of real-time data acquisition and parameter optimization is continuously repeated, adjusting the processing parameters based on changes in the processing environment and results to ensure processing quality and stability.

[0133] As one implementation method, the dynamic parameter optimization model further includes an anomaly detection module; the method provided in this embodiment of the invention may further include:

[0134] Step S600: After generating the set of optimization parameter adjustment instructions, the abnormal patterns in the parameter feature vector set are analyzed by the anomaly detection module.

[0135] The anomaly detection module is an important component of the dynamic parameter optimization model. Its function is to detect abnormal patterns in the parameter feature vector set. The parameter feature vector set contains information such as temperature fluctuation feature values, electric field intensity gradient feature values, and magnetic field intensity deviation feature values, which reflect the state of the processing environment.

[0136] Anomalies may manifest as sudden changes in temperature fluctuation characteristics, abnormal increases or decreases in electric field intensity gradient characteristics, or magnetic field intensity deviations exceeding the normal range. These anomaly patterns can be detected promptly by analyzing the parameter feature vector set using an anomaly detection module.

[0137] Step S700: If a discontinuous jump is detected in the temperature fluctuation characteristic value or the electric field intensity gradient characteristic value, it is determined that the sensor is abnormal and a data re-acquisition command is triggered.

[0138] When the anomaly detection module detects discontinuous jumps in temperature fluctuation characteristic values ​​or electric field intensity gradient characteristic values, it indicates that these data may be abnormal. A discontinuous jump refers to a significant change in data within a short period of time, and this change does not conform to normal physical laws.

[0139] This situation could be caused by sensor malfunction, external interference, or other reasons. Therefore, it is determined to be a sensor malfunction, triggering a data re-acquisition command. Re-acquiring data can obtain more accurate processing environment information, providing a reliable data foundation for subsequent parameter optimization.

[0140] Step S800: If the abnormal mode still exists after the data is reacquired, switch to the backup sensor channel and generate an alarm log.

[0141] If abnormal patterns still exist in the parameter feature vector set after data re-acquisition, it indicates a possible fault in the primary sensor channel. In this case, switch to the backup sensor channel for data acquisition. The backup sensor channel is a pre-set backup channel that can provide reliable data when the primary sensor channel fails.

[0142] Simultaneously, an alarm log is generated, recording information such as the time of occurrence of the anomaly and the specific manifestation of the anomaly mode. The alarm log can be used for subsequent troubleshooting and analysis, helping technicians to promptly identify and resolve problems.

[0143] Step S900: Update the parameter feature vector set based on the data collected by the backup sensor channel, and re-execute the parameter adjustment instruction generation step of the dynamic parameter optimization model.

[0144] The parameter feature vector set is updated based on the data collected from the backup sensor channel. The newly collected data is then processed using the previous feature extraction method to generate a new parameter feature vector set.

[0145] The parameter adjustment instruction generation step of the dynamic parameter optimization model is re-executed, and the updated set of parameter feature vectors is input into the dynamic parameter optimization model to generate a new set of optimized parameter adjustment instructions. In this way, the processing parameters of the FIB equipment can still be optimized even in the event of sensor failure, improving processing stability and quality.

[0146] As one implementation method, the dynamic parameter optimization model also integrates a multi-objective optimization function, and the method provided in this embodiment of the invention may further include:

[0147] Step S1000: Extract the target processing accuracy priority label, material removal rate priority label and equipment energy consumption priority label based on the current processing task requirements, and generate a multi-target priority configuration set.

[0148] The current processing task requirements are determined based on specific processing requirements and objectives. Different processing tasks may have different priorities regarding processing accuracy, material removal rate, and equipment energy consumption. For example, for some high-precision processing tasks, processing accuracy may be a higher priority; for some large-scale processing tasks, material removal rate may be a higher priority; and for some energy-sensitive tasks, equipment energy consumption may be a higher priority.

[0149] Based on the current processing task requirements, priority tags for target processing accuracy, material removal rate, and equipment energy consumption are extracted. These tags indicate the relative importance of processing accuracy, material removal rate, and equipment energy consumption in the current processing task. These tags are then combined to generate a multi-target priority configuration set.

[0150] Step S1100: Based on the multi-objective priority configuration set, divide the processing accuracy error tolerance interval, material removal rate deviation tolerance interval, and energy consumption deviation tolerance interval in the three-dimensional target space, and generate a multi-objective constraint boundary condition set.

[0151] The three-dimensional target space is composed of three target dimensions: processing accuracy, material removal rate, and equipment energy consumption. Based on the multi-target priority configuration set, the three-dimensional target space is divided into processing accuracy error tolerance intervals, material removal rate deviation tolerance intervals, and energy consumption deviation tolerance intervals.

[0152] The tolerance range for machining accuracy error refers to the allowable range of machining accuracy error in the current machining task; the tolerance range for material removal rate deviation refers to the allowable deviation range between the material removal rate and the target rate; and the tolerance range for energy consumption deviation refers to the allowable deviation range between the equipment energy consumption and the target energy consumption.

[0153] These intervals are combined to generate a set of multi-objective constraint boundary conditions. The conditions in this set restrict the value range of each objective during the processing, ensuring that the processing result meets the requirements of multiple objectives.

[0154] Step S1200: Input the multi-objective constraint boundary condition set into the Pareto front solution generator, traverse the parameter adjustment instruction combinations that satisfy all constraint boundary conditions, and generate a candidate Pareto solution set.

[0155] The Pareto Front Solution Generator is a tool for generating Pareto optimal solutions. A Pareto optimal solution is a solution in a multi-objective optimization problem that cannot improve a particular objective without reducing the others. The Pareto Front Solution Generator takes the set of multi-objective constraints and boundary conditions as input, iterates through all possible combinations of parameter adjustments, and finds the combination that satisfies all the constraints and boundary conditions.

[0156] These combinations that satisfy the conditions constitute the candidate Pareto solution set. Each solution in the candidate Pareto solution set represents an optimal parameter adjustment scheme under multi-objective constraints.

[0157] Step S1300: Extract the processing accuracy weight value, material removal rate weight value and energy consumption weight value corresponding to each solution from the candidate Pareto solution set, and generate a multi-objective weight vector set.

[0158] In the candidate Pareto solution set, each solution corresponds to a set of processing accuracy weight values, material removal rate weight values, and energy consumption weight values. These weight values ​​indicate the importance of processing accuracy, material removal rate, and equipment energy consumption in the overall evaluation under that solution.

[0159] The weight values ​​corresponding to each solution are extracted from the candidate Pareto solution set and combined to generate a multi-objective weight vector set. The vectors in this set can reflect the weight distribution of different solutions in multi-objective optimization.

[0160] As one implementation method, step S1300, extracting the processing accuracy weight value, material removal rate weight value, and energy consumption weight value corresponding to each solution from the candidate Pareto solution set to generate a multi-objective weight vector set, can be specifically implemented as follows:

[0161] Step S1310: Perform target term contribution decomposition on each solution in the candidate Pareto solution set to obtain the contribution value of the decrease in processing accuracy error, the contribution value of the increase in material removal rate, and the contribution value of the decrease in energy consumption.

[0162] The objective contribution decomposition of each solution in the candidate Pareto solution set is performed to analyze the contribution of each solution to the three objectives of machining accuracy, material removal rate, and equipment energy consumption. By comparing the solution with the machining accuracy, material removal rate, and equipment energy consumption under the initial conditions, the decrease in machining accuracy error, the increase in material removal rate, and the reduction in equipment energy consumption can be calculated.

[0163] These changes are used as contribution values ​​for the reduction in processing accuracy error, the improvement in material removal rate, and the reduction in energy consumption. These contribution values ​​can reflect the optimization effect of each solution on different objectives.

[0164] Step S1320: Divide the three contribution values ​​of each solution by the total optimization benefit value corresponding to that solution to generate normalized processing accuracy weight component, normalized material removal rate weight component and normalized energy consumption weight component.

[0165] The total optimization benefit value refers to the comprehensive optimization effect of each solution on the three objectives, which can be obtained by adding the contribution value of the reduction in processing accuracy error, the contribution value of the improvement in material removal rate, and the contribution value of the reduction in energy consumption.

[0166] Dividing the three contribution values ​​of each solution by the corresponding total optimization benefit value yields the normalized processing accuracy weight component, the normalized material removal rate weight component, and the normalized energy consumption weight component. These normalized weight components range from [0,1] and represent the relative importance of each solution to different objectives.

[0167] Step S1330: Arrange the three normalized weight components of each solution in descending order according to the processing accuracy weight component to generate a sorted multi-objective weight sequence set.

[0168] The three normalized weight components of each solution are arranged in descending order according to the processing accuracy weight component. That is, they are first sorted from largest to smallest according to the processing accuracy weight component. If the processing accuracy weight components are the same, they are sorted according to the material removal rate weight component or the energy consumption weight component.

[0169] The sorted weight components are combined to generate a sorted multi-objective weight sequence set. This set of sequences more clearly shows the weight distribution of each solution in the multi-objective optimization.

[0170] Step S1340: Based on the priority labels in the multi-objective priority configuration set, locate the solution interval that meets the threshold of the processing accuracy weight component, the threshold of the material removal rate weight component, and the threshold of the energy consumption weight component in the sorted multi-objective weight sequence set, and generate a filtered weight vector subset.

[0171] The priority labels in the multi-objective priority configuration set specify the importance of processing accuracy, material removal rate, and equipment energy consumption. Based on these priority labels, the weighted component thresholds for processing accuracy, material removal rate, and energy consumption are determined.

[0172] Within the sorted multi-objective weight sequence set, the solution intervals that meet these thresholds are located. The weight vectors corresponding to these solutions are combined to generate a filtered subset of weight vectors. The weight vectors in this subset better reflect the multi-objective priority requirements of the current processing task.

[0173] Step S1350: Calculate the Euclidean distance between each vector in the filtered weight vector subset and the weight vector of historical successful cases, and generate a similarity score sequence.

[0174] The historical success case weight vector refers to the weight vector that achieved good processing results in past processing tasks. The Euclidean distance between each vector in the filtered weight vector subset and the historical success case weight vector is calculated. Euclidean distance refers to the distance between two vectors in Euclidean space.

[0175] The calculated Euclidean distance is converted into a similarity score; the smaller the Euclidean distance, the higher the similarity score. The similarity scores of each vector are combined to generate a similarity score sequence. This sequence reflects the similarity between each vector in the selected weighted vector subset and historical successful cases.

[0176] Step S1360: Select the top N weight vectors with the highest similarity scores and perform a weighted average to generate the dynamic optimization weight set for the current task, where N≥1.

[0177] Select the top N weight vectors with the highest similarity scores from the similarity score sequence. For example, take a weighted average of the top five weight vectors, and distribute the weights according to the similarity scores, with higher similarity scores resulting in higher weights.

[0178] A weighted average is used to generate a dynamic optimization weight set for the current task. This weight set comprehensively considers historical success cases and the multi-objective priority requirements of the current processing task, providing more suitable weights for the dynamic parameter optimization model.

[0179] Step S1370: Perform element-wise multiplication of the dynamically optimized weight set with the weight matrix of the fully connected layer of the dynamic parameter optimization model to generate a new weight matrix that adapts to the priority of multiple objectives.

[0180] The dynamic optimization weight set is multiplied element-wise with the weight matrix of the fully connected layer of the dynamic parameter optimization model, that is, each element in the dynamic optimization weight set is multiplied with the corresponding element in the fully connected layer weight matrix.

[0181] In this way, a new weight matrix is ​​generated that adapts to the priority of multiple objectives. The new weight matrix enables the dynamic parameter optimization model to pay more attention to the multi-objective priority requirements of the current processing task when outputting parameter adjustment instructions.

[0182] Step S1380: Each time a task priority reconfiguration instruction is received, the filtered weight vector subset is cleared and the weight component threshold positioning and similarity score calculation process is re-executed based on the new priority label to update the dynamically optimized weight set to drive model output adjustment.

[0183] When a task priority reconfiguration instruction is received, it indicates that the multi-objective priority of the current processing task has changed. At this time, the filtered weight vector subset is cleared, and the weight component threshold localization and similarity score calculation process is re-executed based on the new priority labels.

[0184] The dynamic optimization weight set is updated by recalculation. The updated dynamic optimization weight set is then input into the dynamic parameter optimization model, driving adjustments to the model's output. This allows the dynamic parameter optimization model to adapt promptly to changes in the processing task, generating parameter adjustment instructions that better meet the requirements of the new task.

[0185] Step S1400: Perform similarity matching between the multi-objective weight vector set and the weight distribution of successful cases in historical processing tasks, and select the target weight vector with the highest matching degree as the dynamic optimization weight set for the current task.

[0186] Matching the set of multi-objective weight vectors with the weight distributions of successful cases in historical processing tasks can be achieved using various methods, such as Euclidean distance and cosine similarity. The similarity score between each vector in the multi-objective weight vector set and the weight vectors of historical successful cases is calculated.

[0187] The target weight vector with the highest matching degree is selected and used as the dynamic optimization weight set for the current task. This weight set can combine historical success experience and the multi-objective priority requirements of the current processing task to provide more suitable weights for the dynamic parameter optimization model, thereby making the parameter adjustment instructions generated by the model more in line with actual needs.

[0188] Step S1500: Input the dynamically optimized weight set into the fully connected layer of the dynamic parameter optimization model, replace the original fixed weight matrix, and generate an adaptive multi-objective optimization network.

[0189] The dynamically optimized weight set is input into the fully connected layer of the dynamically parameterized model, replacing the original fixed weight matrix. The original fixed weight matrix is ​​determined during model training and may not be suitable for the multi-objective priority requirements of different processing tasks.

[0190] By replacing the weight matrix, an adaptive multi-objective optimization network is generated. This network can dynamically adjust the weights according to the priority of the multiple objectives of the current processing task, making the model pay more attention to the balance and optimization of each objective when outputting parameter adjustment instructions.

[0191] Step S1600: The parameter feature vector set is reweighted through an adaptive multi-objective optimization network to generate a set of optimization parameter adjustment instructions that emphasize different objective terms.

[0192] Adaptive multi-objective optimization networks reweight the parameter feature vector set based on a dynamically optimized weight set. This reweighting process adjusts the importance of different objective terms according to the dynamically optimized weight set when calculating parameter adjustment instructions.

[0193] For example, if the weight of machining accuracy is higher in the dynamic optimization weight set, the network will focus more on improving machining accuracy when generating parameter adjustment instructions; if the weight of material removal rate is higher, the network will tend to increase the material removal rate. Through reweighting, a set of optimization parameter adjustment instructions with different objectives is generated. The instructions in this set can better meet the multi-objective requirements of the current machining task.

[0194] Step S1700: During the processing, continuously monitor the external input processing task change instructions. When the task priority is reconfigured, extract the multi-objective priority configuration set of the new task and trigger the update process of the multi-objective constraint boundary condition set. Regenerate the dynamic optimization weight set adapted to the new task to adjust the output instructions of the dynamic parameter optimization model.

[0195] During processing, it is necessary to continuously monitor externally inputted processing task change commands. These commands may alter the multi-objective priorities of the current processing task. When a task priority reconfiguration is detected, the new multi-objective priority configuration set for the task is extracted.

[0196] Based on the new task's multi-objective priority configuration set, trigger the update process for the multi-objective constraint boundary condition set. Update the multi-objective constraint boundary condition set to meet the requirements of the new task.

[0197] Repeat the previous steps to generate a dynamically optimized weight set adapted to the new task. Input the new dynamically optimized weight set into the dynamic parameter optimization model and adjust the model's output instructions. This allows the dynamic parameter optimization model to adapt to changes in the processing task in a timely manner, ensuring high efficiency and quality in the processing.

[0198] As one implementation method, the method provided in this embodiment of the invention may further include:

[0199] Step S1800: During real-time processing based on the adjusted processing parameter set, surface morphology data and ion beam residence time distribution data of the processing area are collected in real time to generate a processing surface quality monitoring dataset.

[0200] When performing real-time machining based on an adjusted set of machining parameters, it is necessary to acquire surface morphology data and ion beam residence time distribution data of the machining area in real time. Surface morphology data reflects the microstructure and shape characteristics of the surface of the machining area and is crucial for evaluating machining accuracy and surface quality. Surface morphology data can be acquired using equipment such as optical microscopes and scanning electron microscopes, which can capture detailed information about the machined surface at high resolution. Ion beam residence time distribution data records the dwell time of the ion beam at different locations in the machining area, directly affecting the amount of material removed and the machining effect. A dedicated ion beam monitoring system can accurately measure the ion beam residence time distribution.

[0201] The collected surface morphology data and ion beam residence time distribution data are integrated to generate a processed surface quality monitoring dataset. This dataset provides comprehensive and detailed information for subsequent analysis of processed surface quality, helping to promptly identify potential problems during processing, such as surface roughness not meeting requirements or processing accuracy deviations.

[0202] Step S1900: Compare the surface quality monitoring dataset with the theoretical morphology contour data in the target machining design drawing pixel by pixel to generate a surface morphology error mapping map and a residence time deviation distribution map.

[0203] The theoretical topographic contour data in the target machining design drawing is a pre-defined ideal machining result, which specifies the shape and size that the surface of the machining area should achieve. Comparing the machining surface quality monitoring dataset with the theoretical topographic contour data pixel by pixel means accurately comparing each pixel of the machining surface with the corresponding pixel of the theoretical contour.

[0204] During the comparison process, the difference between the actual morphology and the theoretical morphology at each pixel is calculated. These difference values ​​constitute a surface morphology error map. The surface morphology error map visually displays the error status of the processed surface at various locations. Areas with larger errors may indicate processing defects or insufficient precision.

[0205] Simultaneously, by comparing the ion beam residence time distribution data with the theoretical residence time, the residence time deviation at each location is calculated, thereby generating a residence time deviation distribution map. Residence time deviations can lead to uneven material removal, thus affecting the quality of the processed surface. The residence time deviation distribution map provides a clear visual indication of the deviations in ion beam residence time within the processing area.

[0206] Step S2000: Extract the coordinate set of abnormal regions that exceed the tolerance threshold from the surface morphology error map, and match the abnormal values ​​of ion beam scanning speed corresponding to the coordinates according to the residence time deviation distribution map to generate a set of scanning speed correction values.

[0207] The tolerance threshold is the allowable error range determined based on processing requirements and process standards. Pixels whose error values ​​exceed the tolerance threshold are selected from the surface topography error map; the regions containing these pixels are considered abnormal areas. The coordinates of these abnormal areas are extracted to form an abnormal area coordinate set.

[0208] Based on the coordinate set of the abnormal regions, the corresponding ion beam residence time deviation value is found in the residence time deviation distribution map. Since ion beam scanning speed is closely related to residence time, a deviation in residence time often indicates an abnormality in scanning speed. Through a preset mapping relationship (the specific mapping is not limited; for example, it can be based on the basic relationship that ion beam residence time is inversely proportional to scanning speed. If the residence time deviation is Δt and the original scanning speed is v0, the new scanning speed v can be calculated using the formula v = v0 × t0 / (t0 + Δt), where t0 is the theoretical residence time, thus obtaining the abnormal scanning speed value and correction amount), the residence time deviation value is converted into an abnormal ion beam scanning speed value. The abnormal scanning speed value for each abnormal region is analyzed and calculated to determine the required scanning speed correction amount. The scanning speed correction amounts for all abnormal regions are combined to generate a set of scanning speed correction amounts. This set provides specific parameter basis for subsequent adjustments to the ion beam scanning speed.

[0209] Step S2100: Input the abnormal region coordinate set and the scan speed correction amount set into the historical data cache queue of the dynamic parameter optimization model, trigger the online incremental learning process of the dynamic parameter optimization model, and generate updated focus compensation coefficient correction values ​​and scan path offset compensation amounts.

[0210] The historical data cache queue of the dynamic parameter optimization model stores historical data related to the processing procedure, which can provide a reference for model learning and optimization. The coordinate set of abnormal regions and the set of scanning speed correction values ​​are input into the historical data cache queue, and the model integrates this new data with the previous historical data.

[0211] This triggers an online incremental learning process for the dynamic parameter optimization model. Online incremental learning refers to the model continuously learning and adjusting based on new data during operation to adapt to changes in the processing. The model uses new input data to update and optimize its parameters, especially those related to focus compensation coefficients and scan path offset.

[0212] During the learning process, the model analyzes the characteristics of abnormal areas and the impact of scanning speed correction on the processing effect. Through continuous iteration and optimization, it generates updated focus compensation coefficient correction values ​​and scanning path offset compensation values. These correction values ​​and compensation values ​​can be precisely adjusted to address problems in abnormal areas, improving processing accuracy and quality.

[0213] As one implementation method, in step S2100, triggering the online incremental learning process of the dynamic parameter optimization model to generate updated focusing compensation coefficient correction values ​​and scan path offset compensation amounts can be specifically implemented as follows:

[0214] Step S2110: Extract the coordinate set of the abnormal area, the set of scanning speed corrections, and the corresponding surface roughness verification results of the most recent five processing cycles from the historical data cache queue to generate an incremental training sample set.

[0215] The historical data cache queue stores relevant data from multiple processing cycles. To ensure the timeliness and relevance of the data, the coordinate sets of abnormal areas, the set of scanning speed correction values, and the corresponding surface roughness verification results for the most recent five processing cycles are extracted from the queue. The surface roughness verification results are obtained by measuring the roughness of the processed surface, reflecting the quality status of the processed surface.

[0216] These data are combined to generate an incremental training sample set. This set contains the latest processing information and feedback, providing valuable training data for the online incremental learning of dynamic parameter optimization models.

[0217] Step S2120: Convert the set of coordinates of abnormal regions in the incremental training sample set into a spatial location encoding vector, and associate the set of scanning speed correction quantities with the surface roughness verification results as a parameter adjustment label vector to generate spatiotemporal correlation training data pairs.

[0218] To ensure that the coordinate set of anomaly regions can be effectively processed by the dynamic parameter optimization model, it is converted into a spatial location encoding vector. Various encoding methods can be used for spatial location encoding vectors, such as one-hot encoding and binary encoding, to convert the coordinate information of the anomaly regions into a vector form that the model can understand.

[0219] The scan speed correction set is correlated with the surface roughness verification results to form a parameter adjustment label vector. This vector represents the target surface roughness to be achieved under the current anomaly region and scan speed correction conditions. The spatial location encoding vector and the parameter adjustment label vector are combined to generate a spatiotemporally correlated training data pair. This data pair effectively correlates spatial location information with the parameter adjustment target, providing more comprehensive information for model learning.

[0220] Step S2130: Use the spatiotemporal correlation training data to perform short-term temporal dependency reinforcement training on the LSTM layer of the dynamic parameter optimization model, and only update the weight matrix in the fully connected layer that is related to the scan path offset and focus compensation coefficient, to generate a dynamic parameter optimization sub-model containing the updated weight matrix.

[0221] The LSTM layer is a crucial component in dynamic parameter optimization models for processing time series data, enabling it to capture short-term temporal dependencies within the data. When spatiotemporally correlated training data pairs are input into the LSTM layer, it processes and analyzes the data, extracting its temporal features.

[0222] During training, only the weight matrices in the fully connected layers related to the scan path offset and focus compensation coefficients are updated. This is done to selectively adjust parameters in the model relevant to solving the current anomaly, avoiding large-scale updates to the entire model and improving training efficiency.

[0223] By using short-term temporal dependency reinforcement training, a dynamic parameter optimization sub-model is generated, which includes updated weight matrices. This sub-model can better adapt to newly emerging anomalies, providing support for generating more accurate focus compensation coefficient correction values ​​and scan path offset compensation amounts.

[0224] Step S2140: The dynamic parameter optimization sub-model performs forward calculations on the current processing parameter set and outputs a preliminary focus compensation coefficient correction sequence and a scan path offset compensation sequence.

[0225] The current set of processing parameters is input into the dynamic parameter optimization sub-model for forward computation. Forward computation refers to the process where data travels from the model's input layer through various hidden layers to the output layer. During this process, the dynamic parameter optimization sub-model processes and transforms the input processing parameters based on the updated weight matrix.

[0226] The system outputs a preliminary focus compensation coefficient correction sequence and a scan path offset compensation sequence. The preliminary focus compensation coefficient correction sequence contains suggested corrections to the focus compensation coefficients for the current machining parameter set, while the scan path offset compensation sequence provides a compensation scheme for the scan path offset. These sequences provide a preliminary reference for subsequent parameter adjustments.

[0227] Step S2150: Perform historical adjustment amplitude filtering on the preliminary focus compensation coefficient correction sequence to remove abnormal correction values ​​that exceed the preset multiple of historical fluctuation range, and generate updated focus compensation coefficient correction values.

[0228] Historical adjustment amplitude filtering is performed to ensure that the correction value of the focus compensation coefficient is within a reasonable range. The historical fluctuation range of the preset multiple is a safety interval determined based on past processing experience and equipment performance. Each correction value in the initial focus compensation coefficient correction sequence is checked. If a correction value exceeds the historical fluctuation range of the preset multiple, it is considered abnormal and is discarded.

[0229] After filtering, updated focus compensation coefficient correction values ​​are obtained. These correction values ​​take into account the current abnormal conditions while ensuring that the adjustment of the focus compensation coefficient is not too drastic, thereby ensuring the stability and reliability of the processing.

[0230] Step S2160: Perform path continuity verification in adjacent regions on the scan path offset compensation sequence, eliminate path jump conflict points, and generate the updated scan path offset compensation amount.

[0231] Each compensation amount in the scan path offset compensation sequence corresponds to a position in the machining area. In actual machining, the scan path needs to maintain continuity to avoid path jumps and conflict points; otherwise, discontinuities or defects may appear on the machined surface.

[0232] The scan path offset compensation sequence undergoes a continuity check between adjacent regions to examine whether the scan path offset compensation at adjacent positions would cause path discontinuities. If path jump conflict points are found, the compensation amounts at adjacent positions are adjusted to maintain scan path continuity. After eliminating path jump conflict points, updated scan path offset compensation amounts are generated. These compensation amounts ensure a smooth transition of the ion beam scan path within the processing area, improving processing quality and accuracy.

[0233] Step S2200: Generate a secondary calibration parameter set based on the focus compensation coefficient correction value and the scan path offset compensation amount, and superimpose the secondary calibration parameter set onto the adjusted processing parameter set to generate an iteratively optimized processing parameter configuration file.

[0234] The focus compensation coefficient correction value and the scan path offset compensation amount are key parameters for adjusting for abnormal situations that occur during the processing. Combining the focus compensation coefficient correction value and the scan path offset compensation amount generates a secondary calibration parameter set.

[0235] The secondary calibration parameter set is superimposed onto the adjusted processing parameter set. The superposition process involves adding the parameter values ​​from the secondary calibration parameter set to the corresponding parameter values ​​from the adjusted processing parameter set, or performing other appropriate operations, thereby updating the processing parameters.

[0236] After the overlay operation, an iteratively optimized machining parameter configuration file is generated. This configuration file contains machining parameters that have undergone secondary calibration, enabling it to more accurately adapt to the actual conditions during the machining process and further improve machining precision and quality.

[0237] Step S2300: In subsequent processing cycles, the iteratively optimized processing parameter configuration file is applied first to perform reprocessing of local areas. At the same time, surface roughness data of the reprocessed area is continuously collected to verify the effect of parameter correction. If the roughness data does not meet the preset standard, the online incremental learning process is triggered again.

[0238] In subsequent processing cycles, the iteratively optimized processing parameter configuration file is used first to reprocess local areas. Local areas refer to regions where anomalies were previously detected; reprocessing can improve the processing quality of these areas.

[0239] During the reprocessing process, surface roughness data of the reprocessed area is continuously collected. Surface roughness is an important indicator for measuring the quality of the processed surface. By monitoring surface roughness data in real time, the effectiveness of parameter corrections can be evaluated promptly.

[0240] The collected surface roughness data is compared with a preset standard. The preset standard is a target surface roughness value determined according to processing requirements and process specifications. If the roughness data does not meet the preset standard, it indicates that the current parameter correction is not ideal and further optimization is needed. At this point, the online incremental learning process of the dynamic parameter optimization model is triggered again. The model is trained and adjusted using the newly collected data to generate new focus compensation coefficient correction values ​​and scan path offset compensation amounts, so as to achieve continuous optimization of processing parameters until the surface roughness of the reprocessed area reaches the preset standard.

[0241] Through the above steps, this real-time monitoring and feedback-based FIB equipment processing parameter optimization method enables comprehensive monitoring and precise optimization of the FIB equipment processing process. From real-time acquisition of processing environment and effect data, to feature extraction, model training, and optimization command generation, and then to parameter adjustment and iterative optimization based on feedback, the entire process forms a closed-loop optimization system. This system can respond promptly to changes and anomalies during the processing, continuously adjust processing parameters, improve processing accuracy, efficiency, and quality, and adapt to the multi-objective requirements of different processing tasks, exhibiting strong flexibility and adaptability. In practical applications, this optimization method can be widely used in semiconductor manufacturing, micro-nano fabrication, and other fields, providing strong support for improving product quality and production efficiency.

[0242] Please see Figure 2 , Figure 2 This is a schematic diagram of a computer system provided in an embodiment of the present invention. The computer system can be embedded in or connected to a FIB device, and includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include standard wired interfaces or wireless interfaces (such as Wi-Fi, mobile communication interfaces, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a memory device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system; the system architecture is not limited in this invention.

[0243] In one embodiment, the processor 101 executes the real-time monitoring and feedback method for optimizing FIB equipment processing parameters provided in the above embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A method for optimizing FIB device processing parameters with real-time monitoring feedback, comprising: The method comprises: Real-time acquisition of processing environment monitoring data and processing effect monitoring data of the FIB device to generate a real-time monitoring data set; the processing environment monitoring data comprises temperature distribution data, electric field intensity distribution data and magnetic field intensity distribution data, and the processing effect monitoring data comprises ion beam focusing accuracy data and material removal rate data; According to the processing environment characteristic parameter set and the processing effect characteristic parameter set in the real-time monitoring data set, parameter characteristic extraction processing is performed on the real-time monitoring data set to generate a parameter characteristic vector set; each characteristic vector in the parameter characteristic vector set comprises a combination of temperature fluctuation characteristic values, electric field intensity gradient characteristic values and magnetic field intensity deviation characteristic values; The parameter characteristic vector set is input into a pre-trained dynamic parameter optimization model to generate an optimization parameter adjustment instruction set; the dynamic parameter optimization model is generated by training the mapping relationship between environment parameters and processing parameters in historical processing data, and the optimization parameter adjustment instruction set comprises ion beam intensity adjustment values, scanning path offset amounts and focusing compensation coefficients; According to the optimization parameter adjustment instruction set, the current processing parameters of the FIB device are adjusted to generate an adjusted processing parameter set; Based on the adjusted processing parameter set, real-time processing is performed, and the process of real-time acquisition is cyclically executed to update the real-time monitoring data set, specifically comprising: in the process of executing the adjusted processing parameter set, real-time thermal imaging data and ion beam trajectory data of the processing area are synchronously acquired; According to the real-time thermal imaging data, a thermal effect accumulation amount is calculated, and when the thermal effect accumulation amount exceeds a preset safety threshold, a processing interruption instruction is triggered; During the processing interruption, according to the deviation amount of the ion beam trajectory data and the target processing path, path calibration parameters are generated and updated to the adjusted processing parameter set; After restarting the processing, the updated adjusted processing parameter set is input into the dynamic parameter optimization model for secondary optimization to generate a corrected optimization parameter adjustment instruction set; According to the corrected optimization parameter adjustment instruction set, the processing is continued until the entire processing task is completed; In the process of executing real-time processing based on the adjusted processing parameter set, surface topography data and ion beam residence time distribution data of the processing area are acquired in real time to generate a processing surface quality monitoring data set; The processing surface quality monitoring data set and theoretical topography profile data in the target processing design diagram are compared pixel by pixel to generate a surface topography error mapping diagram and a residence time deviation distribution diagram; From the surface topography error mapping diagram, an abnormal region coordinate set exceeding a tolerance threshold is extracted, and according to the residence time deviation distribution diagram, ion beam scanning speed abnormal values corresponding to the coordinates are matched to generate a scanning speed correction amount set; The abnormal region coordinate set and the scanning speed correction amount set are input into the historical data buffer queue of the dynamic parameter optimization model to trigger an online incremental learning process of the dynamic parameter optimization model to generate an updated focusing compensation coefficient correction value and a scanning path offset compensation amount; According to the focus compensation coefficient correction value and the scanning path offset compensation amount, a secondary calibration parameter set is generated, and the secondary calibration parameter set is superimposed to the adjusted machining parameter set to generate an iteratively optimized machining parameter configuration file; In a subsequent machining cycle, the iteratively optimized machining parameter configuration file is preferentially applied to execute re-machining of a local area, and surface roughness data of the re-machining area is continuously collected to verify the parameter correction effect, and if the roughness data does not reach a preset standard, the online incremental learning process is triggered again.

2. The method of claim 1, wherein, The parameter feature extraction processing on the real-time monitoring data set according to the machining environment feature parameter set and the machining effect feature parameter set in the real-time monitoring data set generates a parameter feature vector set, including: The time series fluctuation feature of the temperature distribution data is extracted from the machining environment monitoring data to generate a temperature fluctuation feature value sequence; The spatial gradient change feature is extracted from the electric field intensity distribution data to generate an electric field intensity gradient feature value matrix; The deviation feature relative to a preset reference magnetic field is extracted from the magnetic field intensity distribution data to generate a magnetic field intensity deviation feature value set; The temperature fluctuation feature value sequence, the electric field intensity gradient feature value matrix and the magnetic field intensity deviation feature value set are normalized to generate a standardized environment feature data set; According to the correlation between the ion beam focusing accuracy data and the material removal rate data in the machining effect monitoring data, a machining effect weight coefficient is generated, and the standardized environment feature data set and the machining effect weight coefficient are weighted and fused to generate the parameter feature vector set.

3. The method of claim 2, wherein, The parameter feature vector set is input into a pre-trained dynamic parameter optimization model to generate an optimization parameter adjustment instruction set, including: A multi-layer time series analysis network in the dynamic parameter optimization model is obtained; the multi-layer time series analysis network includes an LSTM layer for extracting environmental parameter time series dependence and a fully connected layer for generating parameter adjustment instructions; The parameter feature vector set is divided into multiple subsequences according to a time window, and each subsequence is input into the LSTM layer to generate a corresponding hidden state vector set; According to the multiplication operation of each vector in the hidden state vector set and the weight matrix of the fully connected layer, a preliminary parameter adjustment value set is generated; The preliminary parameter adjustment value set is subjected to constraint condition filtering processing to generate a candidate parameter adjustment instruction set that meets the physical limitations of the FIB device; the constraint conditions include an ion beam intensity maximum threshold and a focus compensation coefficient change rate threshold; According to the predicted machining effect scores of each instruction in the candidate parameter adjustment instruction set, the instruction with the highest score is selected as the optimization parameter adjustment instruction set.

4. The method of claim 1, wherein, The training process of the dynamic parameter optimization model includes: acquire a historical processing environment data set and a historical processing parameter adjustment record set; the historical processing environment data set includes multiple groups of temperature distribution sequences, electric field intensity distribution sequences and magnetic field intensity distribution sequences arranged in time sequence, and the historical processing parameter adjustment record set includes ion beam intensity adjustment sequences, scanning path offset sequences and focusing compensation coefficient sequences corresponding to each group of historical processing environment data; perform time domain filtering processing on the temperature distribution sequences in the historical processing environment data set to remove high-frequency noise components, generate a smooth temperature distribution sequence set, perform spatial alignment correction on the electric field intensity distribution sequences to generate a corrected electric field intensity distribution set, and perform difference calculation on the magnetic field intensity distribution sequences and a device reference magnetic field configuration to generate a magnetic field intensity deviation sequence set; align and combine the smooth temperature distribution sequence set, the corrected electric field intensity distribution set and the magnetic field intensity deviation sequence set according to time stamps to generate a historical environment feature matrix set, and convert the ion beam intensity adjustment sequences, the scanning path offset sequences and the focusing compensation coefficient sequences in the historical processing parameter adjustment record set into a parameter adjustment vector set; use the historical environment feature matrix set as an input sample set and the parameter adjustment vector set as an output label set to form a supervised training data set; perform multiple rounds of iterative training on an initial neural network model through the supervised training data set, calculate the mean square error of the model predicted parameter adjustment vector and the real parameter adjustment vector in each iteration, and update the model weight according to the error back propagation until the prediction error is stable within a preset convergence interval to generate a trained dynamic parameter optimization model.

5. The method of claim 4, wherein, The multiple rounds of iterative training on the initial neural network model through the supervised training data set include: extract a historical environment feature matrix sub-set and a corresponding parameter adjustment vector sub-set from the supervised training data set in batches, input the historical environment feature matrix sub-set into the feature encoding layer of the initial neural network model to generate an environment feature encoding vector set; input the environment feature encoding vector set into the LSTM layer of the initial neural network model to extract the time-varying dependence mode of the environment parameters and generate a time sequence feature vector set; input the time sequence feature vector set into the full connection layer of the initial neural network model to generate a predicted parameter adjustment vector set, and calculate the error distribution between the predicted parameter adjustment vector set and the real parameter adjustment vector sub-set; calculate the gradient value of the weight of each layer of the initial neural network model according to the error distribution, scale the gradient value by using an adaptive learning rate algorithm to generate an adjusted weight update amount; iteratively correct the weight parameters of the initial neural network model according to the weight update amount, and verify the prediction accuracy of the model on an independent verification data set after the end of each iteration, terminate the training when the verification accuracy improvement amplitude of five consecutive iterations is less than a preset threshold, and generate a final optimized dynamic parameter optimization model.

6. The method of claim 1, wherein, The calculation of the thermal effect accumulation amount from the real-time thermal imaging data includes: dividing the real-time thermal imaging data into multiple sub-regions and calculating the temperature rise rate and spatial temperature gradient of each sub-region; According to the product of the temperature rise rate of each sub-region and the thermal expansion coefficient of the material, a local thermal deformation prediction value is generated; The local thermal deformation prediction values of all sub-regions are weighted and summed to generate an overall thermal effect accumulation; The overall thermal effect accumulation is compared with the safety threshold corresponding to different material types, and if the threshold is exceeded, a processing interruption instruction and corresponding cooling parameter adjustment suggestion are generated.

7. The method of claim 1, wherein, The dynamic parameter optimization model further includes an anomaly detection module; the method further includes: After generating the optimization parameter adjustment instruction set, the anomaly detection module is used to analyze the abnormal patterns in the parameter feature vector set; If the temperature fluctuation feature value or the electric field intensity gradient feature value shows a non-continuous jump, it is determined that the sensor is abnormal and a data re-acquisition instruction is triggered; If the abnormal pattern still exists after data re-acquisition, a backup sensor channel is switched to and an alarm log is generated; The parameter feature vector set is updated according to the data collected by the backup sensor channel, and the parameter adjustment instruction generation step of the dynamic parameter optimization model is re-executed.

8. The method of claim 1, wherein, The dynamic parameter optimization model further integrates a multi-objective optimization function, and the method further includes: Based on the current processing task requirements, target processing precision priority labels, material removal rate priority labels and equipment energy consumption priority labels are extracted to generate a multi-objective priority configuration set; According to the multi-objective priority configuration set, a processing precision error tolerance interval, a material removal rate deviation tolerance interval and an energy consumption deviation tolerance interval are divided in a three-dimensional target space to generate a multi-objective constraint boundary condition set; The multi-objective constraint boundary condition set is input into a Pareto frontier solution generator to traverse parameter adjustment instruction combinations that satisfy all constraint boundary conditions, and a candidate Pareto solution set is generated; From the candidate Pareto solution set, the processing precision weight value, the material removal rate weight value and the energy consumption weight value corresponding to each solution are extracted to generate a multi-objective weight vector set; The multi-objective weight vector set is similarity matched with the weight distribution of successful cases in historical processing tasks, and the target weight vector with the highest matching degree is selected as the dynamic optimization weight set of the current task; The dynamic optimization weight set is input into the fully connected layer of the dynamic parameter optimization model to replace the original fixed weight matrix, and an adaptive multi-objective optimization network is generated; The parameter feature vector set is re-weighted by the adaptive multi-objective optimization network to generate an optimization parameter adjustment instruction set that is biased towards different target items; During processing, the externally input processing task change instruction is continuously monitored, and when a task priority reconfiguration is detected, the multi-objective priority configuration set of the new task is extracted and the update process of the multi-objective constraint boundary condition set is triggered, and a dynamic optimization weight set that adapts to the new task is regenerated to adjust the output instruction of the dynamic parameter optimization model.

9. A computer system, characterized by It includes: a memory in which a computer program is stored; a processor for loading the computer program to implement the real-time monitoring feedback FIB equipment processing parameter optimization method of any one of claims 1-8.

Images