A neural network-guided material modification precision control method

Abstract

A neural network-guided method for precise control of material modification, relating to the field of material modification technology, addresses the problem of uneven surface modification caused by the inability of existing plasma modification techniques to precisely control plasma parameters for nonlinear phase transitions in microstructures. This invention first obtains the initial state parameters and target modification parameters of the material to be modified; then, the obtained parameters are input into a pre-trained neural network model to generate preliminary and high-precision plasma control parameters; next, the operating parameters of the plasma device are controlled by a control unit to modify the material; finally, the state parameters of the material are collected in real time during the modification process, and these real-time state parameters are used as feedback input to the neural network model to dynamically update the plasma control parameters. This invention enables precise and adaptive control of plasma material modification to address nonlinear phase transitions in microstructures.

Description

Technical Field

[0001] This invention relates to the field of materials modification technology, specifically a neural network-guided method for precise control of materials modification. Background Technology

[0002] Generating plasma control parameters based on the initial state parameters of the material and preset modification target parameters using neural network models, and then dynamically adjusting them using real-time feedback, has become the mainstream development direction in the industry. However, although these methods have made some progress in macroscopic parameter control, they still have significant limitations in handling complex physical phenomena at the microscopic scale.

[0003] In existing technologies, plasma control parameters are typically generated by inputting the initial state parameters and modification target parameters of the material into a pre-trained neural network model, and closed-loop control is achieved based on sensor feedback. This method relies on neural networks trained with historical experimental data and can effectively handle common macroscopic modification problems, such as surface roughness optimization and chemical composition distribution adjustment. However, it does not fully consider the nonlinear phase transitions of the microstructure that may occur during plasma treatment, such as local lattice reconstruction, amorphization, or phase separation. These microscopic phase transitions are highly nonlinear and random, often influenced by initial material defects and local plasma energy distribution, and the training data of existing neural network models usually cannot fully cover these rare but critical microscopic events. Furthermore, existing technologies lack real-time monitoring and intelligent adaptive control mechanisms for microstructural changes, resulting in lag in parameter optimization and an inability to respond promptly to the dynamic evolution of the material during treatment. For example, existing systems typically use macroscopic sensor feedback but lack specialized sensors and data fusion methods for microscopic phase transitions, and fail to integrate physical models with data-driven models.

[0004] Because existing technologies lack specific handling for nonlinear phase transitions in microstructures, neural network models may be biased in predicting control parameters, leading to excessive or insufficient plasma processing parameters in localized areas. In practical applications, this bias can cause uneven surface modification, localized thermal stress concentration, or phase transition-induced crack propagation, resulting in material performance degradation, functional coating failure, or damage to equipment components. For example, in semiconductor wafer modification, uncontrolled microphase transitions can increase interface defects and reduce device reliability; in biomedical implant surface treatment, it can cause decreased biocompatibility or fluctuations in antibacterial properties. These consequences not only affect product quality but also increase production costs and R&D cycles. Therefore, there is an urgent need in this field for a precise control method that can effectively address the problem of nonlinear phase transitions in microstructures to improve the overall reliability and applicability of plasma modification technology. Summary of the Invention

[0005] The purpose of this invention is to provide a neural network-guided method for precise control of material modification, which solves the problem of uneven material surface modification caused by the inability of existing plasma modification technology to precisely control plasma parameters for nonlinear phase transitions of microstructures.

[0006] The technical solution adopted by this invention to solve its technical problem is: a method for precise control of material modification guided by a neural network, comprising the following steps:

[0007] S1. Data Acquisition and Preprocessing.

[0008] The initial state parameters and preset modification target parameters of the material to be modified are obtained. The initial state parameters include the material surface morphology, chemical composition, physical property data, and microstructural feature parameters acquired by high-resolution microscopy. The modification target parameters include the target surface roughness, target chemical composition distribution, target mechanical property index, and target microstructural stability. All data are subjected to validity checks, unit consistency conversion, and normalization.

[0009] S2, Neural Network Model Processing.

[0010] The initial state parameter set processed in step S1 With the set of target parameters for modification The input is a pre-trained neural network model, which includes a multi-scale feature fusion module and a nonlinear dynamics compensator. The multi-scale feature fusion module extracts and fuses macroscopic and microscopic features of the input parameters to generate preliminary plasma control parameters. The nonlinear dynamics compensator fine-tunes and corrects the preliminary parameters based on the Avrami phase transition dynamics equation, and outputs high-precision plasma control parameters, including plasma power, gas flow rate, processing time and electrode spacing.

[0011] S3, Plasma Equipment Control.

[0012] Based on the generated plasma control parameters, the operating parameters of the plasma equipment are adjusted by the control unit to perform plasma modification treatment on the materials. The control unit calculates the desired voltage and current output of the plasma generator based on the control parameters and corrects deviations in real time through a closed-loop feedback mechanism. Simultaneously, the gas flow rate, processing time, and electrode spacing of the gas supply system are adjusted to ensure the stability of the discharge process and the controllability of the material modification reaction.

[0013] S4. Real-time feedback and dynamic updates.

[0014] During the plasma modification process, real-time state parameters of the material are collected in real time through a sensor array, including surface temperature, plasma emission spectrum data, and microstructure change data collected by an in-situ X-ray diffractometer. These data are used as feedback inputs to a neural network model to drive the multi-scale feature fusion module and nonlinear dynamic compensator to dynamically update the plasma control parameters.

[0015] Furthermore, the acquisition of initial state parameters includes the following steps: S1.1.1, obtaining high-resolution surface morphology image data by scanning the material surface with a scanning electron microscope, in the form of a grayscale matrix. This indicates that it is used to characterize the surface roughness distribution and microstructure undulations of materials; S1.1.2, X-ray photoelectron spectroscopy is used for component analysis to obtain the chemical elemental composition and bonding energy level distribution data of the material, forming a chemical composition vector. .in, The mole fraction of the i-th element is represented, i = 1, 2, 3…n; and the characteristic parameters of the chemical bonding state are calculated by the energy level peak position and peak intensity. This reflects the distribution characteristics of elemental bonding energies in the material; S1.1.3, Mechanical properties are tested using a universal testing machine, and physical property data are output, including hardness H, strength, etc. With elastic modulus E; where, F represents the applied load, and A represents the cross-sectional area subjected to force; S1.1.4. Microstructural observation data are provided through atomic force microscopy and transmission electron microscopy, and the lattice defect density is calculated through atomic resolution image sequences. .in, This indicates the number of defects per unit volume. The sampling volume is used; the phase transition tendency index is described by the rate of change of the energy potential well. ,in, This is the free energy difference during the phase transition. This represents the width of the phase transition temperature range.

[0016] Furthermore, the modification target parameters are obtained by inputting through a graphical user interface or by loading from a modification target database.

[0017] Furthermore, the training process of the neural network model includes steps S2.1, data collection, and S2.2, data preprocessing; S2.1, data collection, refers to collecting historical data from multiple plasma material modification experiments, including the initial state parameter set of the material. Modification target parameter set Plasma control parameter set Modified material performance parameters Microscopic phase transition characteristic data obtained from in-situ monitoring ;in, Indicates plasma power. Indicates gas flow rate, Indicates processing time. Indicates the electrode spacing.

[0018] Further, step S2.2, data preprocessing includes: S2.2.1, performing outlier removal and noise point removal operations, using the mean square error filtering function. Outliers and noise points are removed to retain valid samples within the statistical distribution interval; among which, This represents the i-th original sample data value in a certain dimension; This represents the mean of the feature in historical samples; This represents the standard deviation of the characteristic; Indicates the absolute distance of a sample from the mean; The threshold representing the statistically valid fluctuation range; This is an indicator function; it takes a value of 1 when the condition is true, indicating that the sample is retained, and a value of 0 when the condition is false, indicating that the sample is removed; S2.2.2, Perform normalization transformation on all feature parameters. Where x represents the original feature value, This represents the minimum value of the feature dimension. This represents the maximum value of the feature dimension; S2.2.3, subsequently, feature enhancement operations are performed, applying rotation and scaling transformation matrices to the image data. Gaussian noise is superimposed on numerical sequence data. ,in, This represents the random perturbation noise term superimposed on the original feature. Indicates a Gaussian distribution. This represents the noise variance.

[0019] Furthermore, the pre-trained neural network model structure includes an input layer, three hidden modules, and an output layer; the input layer receives the standardized parameter set from step S1. and The input dimension is The hidden module includes a convolutional neural network module, a long short-term memory network module, and a multi-scale pyramid network module. The convolutional neural network module consists of four convolutional layers with kernel sizes of 3×3, 5×5, 5×5, and 3×3, respectively. The convolution operation is defined as... ;in, This represents the output of the l-th convolutional layer. One feature point, For convolution kernel weights, For bias terms, activation function The ReLU function is used for nonlinear feature mapping; the convolutional neural network module extracts the surface morphology of the material. and chemical composition vector Macroscopic spatial characteristics, outputting a three-dimensional feature tensor The Long Short-Term Memory (LSTM) network module consists of a single-layer LSTM, containing three control units: a forget gate, an input gate, and an output gate. The computation process is as follows: ; ; Where t represents the time step index, This represents the feature vector input to the Long Short-Term Memory network at time step t. This represents the hidden state vector at time step t-1. W represents the joint input after concatenating the hidden state from the previous time step with the current input feature vector. f W represents the weight matrix of the forget gate. i W represents the weight matrix of the input gate. o Let b represent the weight matrix of the output gate. f The bias vector b represents the forget gate. i b represents the bias vector of the input gate. o This represents the weight matrix of the output gate. This represents the Sigmoid activation function. This represents the output vector of the forget gate. This represents the output vector of the input gate. This represents the output vector of the output gate.

[0020] Furthermore, the multi-scale pyramid network module consists of three parallel convolutional paths with kernel sizes of 3×3, 5×5, and 7×7. Each path independently extracts feature responses at different spatial scales, defined as... ;in, Representing paths at different scales; the three feature maps are concatenated through feature stitching operations. After fusion, the dimension reduction mapping function Generate a unified feature representation.

[0021] Furthermore, the output layer consists of a fully connected layer and a linear activation function, and the mapping function is defined as follows: ;in, This represents the predicted plasma parameters output by the output layer. and These are the trainable parameters for the output layer.

[0022] Furthermore, the control unit is based on the plasma control parameters from step S2. High-precision dynamic control of key operating parameters of the plasma equipment ensures the stability of the discharge process and the controllability of the material modification reaction. S3.1 The control unit first calculates the desired voltage output based on the power component in the control parameters. With current output And its deviation is corrected in real time through a closed-loop feedback mechanism. The discharge stability constraint expression is: .in, As a discharge error indicator, For voltage weighting coefficients, For the current, For the voltage monitored in real time, The current value is monitored in real time. The output signal is adjusted via proportional-integral control to ensure discharge power. The power target range predicted by the neural network is stabilized, thus maintaining the consistency of plasma energy distribution. S3.2 The control unit then analyzes the gas flow rate component in the control parameters, adjusts the input rate of the reactant gas by driving the mass flow meter, achieving uniform gas distribution within the reaction chamber, and dynamically corrects the flow control threshold based on real-time gas pressure feedback signals to maintain a matching relationship between the local gas density and the electric field distribution of the discharge region. The processing time component is managed by the digital timer module; the control unit processes the time parameters... Input timer, combined with status monitoring signals during the modification process The time is adjusted, and the adjustment function is defined as follows: .in, The actual processing time is corrected, where λ is the dynamic time correction coefficient; the electrode spacing component is executed by the stepper motor control module, and the control signal is provided by... The spatial adjustment parameters are derived, and linear displacement adjustment is achieved through the conversion function between the motor stepping angle θ and the pitch.

[0023] Furthermore, the real-time state parameter acquisition and feedback process is based on a multi-source sensor fusion mechanism, achieving dynamic monitoring and data feedback of the material modification state through high-frequency sampling; the sensor array includes an infrared thermal imager, a spectrometer, and an in-situ X-ray diffractometer, with the infrared thermal imager outputting surface temperature field data. The spectrometer outputs plasma emission spectral signals. The in-situ X-ray diffractometer outputs a sequence of microstructural changes; the acquired analog signal undergoes noise suppression and filtering operations through a signal conditioning circuit, and the output signal is converted into digital form by an analog-to-digital converter to form a real-time state parameter vector.

[0024] Furthermore, all sensor data undergoes scale unification and unit conversion through a standardization function to ensure that the input format conforms to the numerical requirements of the multi-scale feature fusion module. After the real-time state parameters are input into the neural network model in step S2, the system calculates the feature weights based on the multi-scale feature fusion module. During the update phase, phase transition dynamic constraints are introduced through a nonlinear dynamic compensator, and the phase transition volume fraction is corrected according to the Avrami model. Its time derivative Characterize the phase transition rate; calculate the correction factor systematically. The plasma control parameters are updated based on feedback, using the formula... ;in, For the updated plasma control parameters, The fine-tuning parameters output in step S2 are used; the update frequency is set between 10 and 100 times per second. The feedback system prioritizes analyzing the lattice constant change rate and amorphization rate signals provided by the in-situ X-ray diffractometer, and uses a dynamic weighting function. The degree to which microstructure data dominates the overall feedback is determined, and the updated plasma control parameters are then re-input into the control unit of step S3 to form a real-time closed-loop control circuit. The control unit then... The changes in voltage, current, gas flow rate, and electrode spacing are adjusted synchronously to maintain a dynamic balance between energy density and material reaction kinetics in the multidimensional parameter space.

[0025] The beneficial effects of this invention are: (1) Through the multi-scale feature fusion module, a unified characterization of macroscopic and microscopic features is achieved, generating plasma control parameters that can take into account both the overall state of the material and the local microscopic characteristics, thereby improving the comprehensiveness and accuracy of parameter prediction from the source. (2) Through the nonlinear dynamic compensator, the physical model of materials science and the data-driven neural network model are coordinated, significantly enhancing the adaptability and suppression capability of the control system to the nonlinear phase transition of the material microstructure, effectively avoiding material damage caused by local overheating or insufficient treatment. (3) Through the dynamic closed-loop control mechanism based on real-time microscopic feedback, high-frequency and precise control of the modification process is achieved, ensuring the consistency and stability of the material modification effect in time and space, and solving the problem of control lag caused by the lack of in-situ microscopic monitoring in traditional methods. (4) A complete technology chain was constructed through the overall scheme, from multi-source data acquisition, intelligent parameter prediction, physical model correction to precise equipment execution and real-time feedback. This achieved unprecedented precision and adaptive control of the plasma material modification process. It can improve the functionality and durability of semiconductors, biomedical materials, etc., while actively suppressing the degradation and damage of material properties caused by nonlinear phase transitions in their microstructure. Attached Figure Description

[0026] Figure 1 A diagram illustrating the neural network-guided precise control architecture for plasma material modification.

[0027] Figure 2 This is a schematic diagram of the internal structure of a neural network model;

[0028] Figure 3 Here is a flowchart of the feature weighting and splicing mechanism based on the attention mechanism;

[0029] Figure 4 This is a flowchart of the parameter correction process for a nonlinear dynamic compensator. Detailed Implementation

[0030] This invention first obtains the initial state parameters and preset modification target parameters of the material to be modified. Then, the obtained parameters are input into a pre-trained neural network model. The multi-scale feature fusion module and nonlinear dynamics compensator in the neural network model respectively generate preliminary plasma control parameters and fine-tuning correction parameters, outputting high-precision control parameters. Then, the operating parameters of the plasma device are adjusted by the control unit to modify the material. Finally, the state parameters of the material are collected in real time during the modification process, and these real-time state parameters are used as feedback input to the neural network model to dynamically update the plasma control parameters. This invention can achieve precise and adaptive control of plasma material modification to address nonlinear phase transition problems in microstructures, improving the overall reliability and applicability of plasma modification technology. The following is a related description... Figures 1 to 4 The present invention provides a detailed description of a neural network-guided method for precise control of material modification, which includes the following steps.

[0031] S1. Data Acquisition and Preprocessing.

[0032] Obtain the initial state parameters and preset modification target parameters of the material to be modified.

[0033] Initial state parameters include material surface morphology, chemical composition, physical property data, and microstructural characteristic parameters acquired through high-resolution microscopy, such as lattice defect density and phase transition tendency index.

[0034] The target parameters for modification include the target surface roughness, the target chemical composition distribution, the target mechanical properties, and the target microstructure stability.

[0035] Data was collected through a multi-source, high-precision testing approach, including scanning electron microscopy, X-ray photoelectron spectroscopy, atomic force microscopy, and transmission electron microscopy. All data underwent validity checks, unit consistency conversion, and normalization to ensure that the parameters input to the neural network model were standardized and dimensionless.

[0036] S1.1 Obtaining the initial state parameters.

[0037] S1.1.1 High-resolution surface morphology image data is obtained by scanning the material surface using a scanning electron microscope, and is presented in grayscale matrix form. It indicates that it is used to characterize the surface roughness distribution and microstructure undulations of materials.

[0038] S1.1.2. Use X-ray photoelectron spectroscopy to perform compositional analysis, obtain data on the chemical elemental composition and bonding energy level distribution of the material, and form a chemical composition vector. .in, The mole fraction of the i-th element is represented, i = 1, 2, 3…n; and the characteristic parameters of the chemical bonding state are calculated by the energy level peak position and peak intensity. It reflects the distribution characteristics of elemental bonding energy in the material.

[0039] S1.1.3. Mechanical properties are tested using a universal testing machine, and the output physical property data includes hardness (H) and strength. And the elastic modulus E. Wherein, F is the applied load, and A is the cross-sectional area under force.

[0040] S1.1.4. Microstructure observation data are provided using atomic force microscopy and transmission electron microscopy, and lattice defect density is calculated using atomic resolution image sequences. .in, This indicates the number of defects per unit volume. The sampling volume is used; the phase transition tendency index is described by the rate of change of the energy potential well. ,in, This is the free energy difference during the phase transition. This represents the width of the phase transition temperature range.

[0041] S1.2 Obtaining the target modification parameters. The target modification parameters are obtained by inputting them through the graphical user interface or by loading them from the target modification database.

[0042] S2, Neural Network Model Processing.

[0043] The initial state parameter set processed in step S1 With the set of target parameters for modification Input as follows Figure 2 The pre-trained neural network model shown ensures that the multi-scale feature fusion module receives standardized and dimensionless input in the parameter space, thus providing a stable feature parameter basis for the subsequent generation of high-precision plasma control parameters. For example... Figure 2 As shown, the neural network model includes a multi-scale feature fusion module and a nonlinear dynamics compensator. Figure 3 As shown, the feature weighting and splicing mechanism module based on the attention mechanism extracts and fuses macroscopic and microscopic features of the input parameters to generate preliminary plasma control parameters. Figure 4 As shown, the nonlinear dynamic compensator fine-tunes the initial parameters based on the Avrami phase transition dynamics equation, and outputs high-precision plasma control parameters, including plasma power, gas flow rate, processing time, and electrode spacing.

[0044] Specifically, during the training of the neural network model, historical data from multiple plasma material modification experiments are first collected. The data includes the initial state parameter set of the material. Modification target parameter set Plasma control parameter set Modified material performance parameters Microscopic phase transition characteristic data obtained from in-situ monitoring .in, Indicates plasma power. Indicates gas flow rate, Indicates processing time. Indicates the electrode spacing.

[0045] Specifically, in the data preprocessing stage: first, outlier removal and noise point removal operations are performed, using the mean squared error filtering function. ,in, This represents the value of the i-th original sample data in a certain dimension. This represents the mean of the feature across historical samples, used to characterize the central tendency of the data. The standard deviation, representing this characteristic, is used to measure the magnitude of data fluctuation. This represents the absolute distance of a sample from the mean. This represents the statistically significant effective fluctuation range threshold. This is an indicator function; it takes a value of 1 when the condition is true, indicating that the sample is retained, and a value of 0 when the condition is false, indicating that the sample is removed. This function is used to remove outliers and noise points to retain valid samples within the statistical distribution interval. Then, normalization transformation is performed on all feature parameters. Where x represents the original feature value, This represents the minimum value of the feature dimension. This represents the maximum value of the feature dimension. This process achieves parameter scale independence, allowing feature parameters of different dimensions to be input into the network within a unified range. Subsequently, feature enhancement operations are performed, applying rotation and scaling transformation matrices to the image-like data. Gaussian noise is superimposed on numerical sequence data. ,in, This represents the random perturbation noise term superimposed on the original feature. This indicates a Gaussian distribution, with a mean of 0 indicating that noise does not introduce a systematic bias. This represents the noise variance, used to control the noise intensity. This noise is used in the feature enhancement stage to improve the model's robustness to experimental perturbations and measurement uncertainties, generating an expanded sample set to balance the sample distribution. Alignment of micro-features is achieved through a time index mapping function. This method synchronizes time-series data with spatial distribution characteristics, enabling dynamic phase transition signals and microscopic structure data to correspond within the same time frame. The training architecture employs a hybrid structure of convolutional neural networks and long short-term memory networks, embedding a multi-scale pyramid network. The convolutional neural network utilizes multiple layers of convolutional kernels. Extracting macroscopic spatial features ,in, This represents the macroscopic spatial feature representation vector extracted by the convolutional neural network. This represents a nonlinear mapping function consisting of multiple convolutional and pooling layers. This is the set of initial state parameters for the material. Given the set of target parameters for modification, this function extracts the spatial correlation features between material surface morphology, composition distribution, and target constraints; Long Short-Term Memory Network uses hidden state units. ,in, This represents the hidden state vector at time step t, used to store historical dynamic information about the modification process. This represents the input feature vector corresponding to time step t, which contains real-time plasma state or microstructure change features. This indicates the hidden state of the previous time step. This represents the state update function of the Long Short-Term Memory (LSTM) network. This function adjusts the fusion ratio of historical information and current input through a gating structure, capturing time-dependent information during plasma processing. The multi-scale pyramid network uses a set of multi-scale convolutional kernels... ,in, Let represent the set of convolution kernel parameters at the k-th scale, s represent the scale level index, and n represent the number of scales. In the multi-scale pyramid network, this set simultaneously extracts feature information of the microstructure at different spatial scales, thereby achieving the fusion expression of multi-scale features such as lattice defect distribution and phase transition region evolution; output layer generation parameters During training, the backpropagation algorithm is used to calculate the gradient of the loss function. .in, Represents the i-th component of the actual plasma control parameters. These are the model's predicted values. Let be the j-th weight parameter of the model, and λ be the regularization coefficient used to constrain model complexity and prevent overfitting. The model is optimized using a stochastic gradient descent optimizer and based on the weight update rule. Iterative optimization of parameters. Here, η is the learning rate. The training process continues until the root mean square error of the model's predictions on the independent validation set is reached. Below the threshold of 0.01, the output shows a stable and convergent pre-trained neural network model. This is used to generate the preliminary plasma control parameters in step S2, so as to realize intelligent prediction and dynamic guidance of the material modification process.

[0046] Specifically, the pre-trained neural network model structure includes an input layer, three hidden layers, and an output layer. The input layer receives the standardized parameter set from step S1. and The input dimension is The hidden layers include a convolutional neural network module, a long short-term memory network module, and a multi-scale pyramid network module. The convolutional neural network module consists of four convolutional layers with kernel sizes of 3×3, 5×5, 5×5, and 3×3, respectively. The convolution operation is defined as... .in, This represents the output of the l-th convolutional layer. One feature point, For convolution kernel weights, For bias terms, activation function This is the ReLU function, used for nonlinear feature mapping.

[0047] Specifically, the convolutional neural network module extracts the surface morphology of the material. and chemical composition vector Macroscopic spatial characteristics, outputting a three-dimensional feature tensor It contains multi-scale structural response information. The Long Short-Term Memory (LSTM) network module consists of a single-layer LSTM, containing three control units: a forget gate, an input gate, and an output gate. Its computation process is as follows: ; ; Where t represents the time step index, corresponding to the processing time of discrete sampling during plasma modification; This represents the feature vector input to the Long Short-Term Memory network at time step t. This feature is output by the preceding convolutional neural network module and contains temporal feature information of the material's surface morphology, chemical composition, and physical properties. The hidden state vector at time step t-1 records the historical information of the material state evolution during previous plasma interactions; This represents the joint input after concatenating the hidden state from the previous time step with the current input feature vector, which is used to simultaneously consider historical information and the current state. These represent the weight matrices corresponding to the forget gate, input gate, and output gate, respectively, which are used to perform linear mapping on the concatenated input features. These weights are updated during training using the backpropagation algorithm. These represent the bias vectors of the forget gate, input gate, and output gate, respectively, which are used to adjust the activation threshold of each gate unit and improve the model's ability to express data under different operating conditions. This represents the Sigmoid activation function, whose output range is from 0 to 1. It is used to constrain the gating coefficients within a continuously adjustable range, thereby controlling the proportion of information retained, introduced, and output in the time dimension. The output vector of the forget gate determines which information from the cell state at the previous time step is retained or decayed, and is used to suppress historical states that are irrelevant to the current plasma modification stage. The output vector of the input gate is represented by the value of which determines the proportion of new feature information written into the cell state at the current time step, and is used to introduce effective information related to the current phase transition behavior of the material. The output vector of the output gate determines which information in the cell state is mapped to the hidden state output at the current time step, thus affecting the subsequent fusion and modeling of dynamically modified features by the multi-scale pyramid network.

[0048] Specifically, the cell state is updated to The output in the hidden state is .in, Indicates the forgetting gate control factor. Indicates the input gate adjustment factor. This represents the output gate adjustment factor. Represents the Hadamard product, parameter matrix With bias term All variables are trainable. This module receives the temporal unfolded sequence of the output feature maps from the convolutional neural network module, capturing the physical performance parameters during the modification process. The dynamic dependencies are output as time-dependent feature representations. .

[0049] Specifically, the multi-scale pyramid network module consists of three parallel convolutional paths with kernel sizes of 3×3, 5×5, and 7×7. Each path independently extracts feature responses at different spatial scales, defined as... .in, Representing paths at different scales. The three feature maps are concatenated using a feature stitching operation. After fusion, the dimension reduction mapping function A unified feature representation is generated, which integrates macroscopic morphology, chemical composition, time dependence, and microscopic structural scale features.

[0050] Specifically, the output layer consists of a fully connected layer and a linear activation function, and the mapping function is defined as follows: .in, This represents the predicted plasma parameters output by the output layer. and These are the trainable parameters for the output layer, which can be compared with the preliminary plasma control parameter prediction results output by the attention-based feature weighting and splicing mechanism module.

[0051] Specifically, the attention-based feature weighting and concatenation mechanism module works based on the hierarchical structure of the feature pyramid network. It constructs a composite feature representation that combines global perception and local sensitivity by weightedly fusing macroscopic and microscopic features in the feature space. The input to this module includes the macroscopic feature vector output from the convolutional neural network module. and the micro-feature vectors output from the multi-scale pyramid network modules .in, This represents the high-dimensional spatial feature dimension of the convolutional layer output. This represents the low-dimensional feature dimension after fusion through multi-scale convolutional paths. Macroscopic feature vector. The global structural distribution characterizing the surface morphology and chemical composition of materials reflects the statistical mapping of macroscopic information obtained through scanning electron microscopy and X-ray photoelectron spectroscopy in the feature space; microscopic eigenvectors This describes the local structural features of the material, including the density of lattice defects. Phase transition tendency index The variation patterns in the spatial resolution dimension are used to reflect the nonlinear response characteristics of the material's microstructure under plasma interaction. The fusion process first calculates feature weights through an attention mechanism; the attention allocation function is defined as follows: .in, This represents the concatenation operation between macroscopic and microscopic eigenvectors. For learnable weight matrix, For bias vectors, This is used to normalize the feature weights, ensuring they are distributed within the [0,1] interval and sum to 1. This process automatically adjusts the relative contributions of macroscopic and microscopic features during the fusion process through gradient learning, thereby suppressing redundant features and enhancing key feature dimensions. The weighted feature vectors are then concatenated to form a fused feature map. , where ⊙ represents element-wise multiplication. and These are sub-vectors assigned from the weight vector to the corresponding feature dimensions. The fused feature map undergoes principal component analysis after dimensionality reduction, and the covariance matrix is ​​calculated. Extract the eigenvectors corresponding to the k largest eigenvalues ​​to construct the projection matrix. Perform linear transformation To retain key information and reduce feature dimensions to a preset size This dimensionality reduction process preserves high-variance features while reducing noise interference, achieving an optimal balance between information density and discriminability in the fusion result. The final dimensionality-reduced feature representation... As input passed to the fully connected layer, the mapping function is defined as follows: .in, and These are the trainable parameters for the output layer. The preliminary plasma control parameter prediction results serve as input to a nonlinear dynamics compensator, used for parameter fine-tuning based on the material phase transition dynamics equations. This ensures a precise dynamic correspondence between the plasma control process and the material's microscopic evolution behavior. Through this fusion mechanism, a unified characterization of the material's macroscopic structural evolution features and microscopic defect dynamics information is achieved at the feature level. This ensures a dynamic balance between global consistency and local accuracy in the predicted plasma parameters, significantly improving the prediction accuracy and controllability of the material modification process.

[0052] Specifically, the implementation of the nonlinear dynamic compensator is based on the material phase transformation dynamics equation. A physical constraint mechanism is introduced to dynamically correct the neural network output, thereby ensuring that the generated plasma control parameters conform to the actual material evolution. The input data of the compensator includes preliminary plasma control parameters from the multi-scale feature fusion module. The real-time microstructure change data from step S4 reflects the temporal evolution trends of physical quantities such as lattice distortion, phase boundary migration, and defect density changes. The compensator is internally loaded with a nonlinear phase transition kinetic equation based on the Avrami model, the expression of which is: .in, The phase transition volume fraction represents the degree of phase transformation of the material at time t; k is the phase transition rate constant, which determines the nucleation and growth rates; and n is the Avrami exponent, reflecting the geometric dimension of the phase transition and the characteristics of the growth mechanism. Based on real-time microstructure change data, the system calculates the derivative of the instantaneous phase transition rate. And by introducing a correction factor This achieves nonlinear feedback correction of the neural network output. Specifically, This is an adaptive coefficient, whose value is adjusted in real time based on the neural network prediction error and the physical measurement deviation. It is used to control the dynamic correction amplitude and ensure the stability and physical consistency of the feedback process. Subsequently, the compensator performs a weighted adjustment of the initial control parameters based on the correction factor. The correction formula is as follows: .in, These are the high-precision plasma control parameters corrected by physical constraints; β is the learning rate parameter, which adjusts the correction step size to maintain parameter convergence and system response balance. The correction process establishes a closed-loop feedback between neural network prediction and dynamic equation constraints, achieving dual correction of parameters in both physical and feature spaces, thereby avoiding overresponse caused by high-dimensional feature fitting errors. The compensated parameters are... The output to the control unit in step S3 enables the control variables such as discharge power, gas flow rate, and voltage of the plasma device to evolve synchronously with the material phase transformation dynamics in the time domain, thereby achieving high-precision control of the plasma modification process.

[0053] S3, Plasma Equipment Control.

[0054] Based on the generated plasma control parameters, the operating parameters of the plasma equipment are adjusted by the control unit to perform plasma modification treatment on the materials. The control unit calculates the desired voltage and current output of the plasma generator based on the control parameters and corrects deviations in real time through a closed-loop feedback mechanism. Simultaneously, the gas flow rate, processing time, and electrode spacing of the gas supply system are adjusted to ensure the stability of the discharge process and the controllability of the material modification reaction. All control processes are executed under the drive of real-time data streams, achieving coordinated convergence of parameters in multi-dimensional space.

[0055] Specifically, the control unit is based on the plasma control parameters from step S2. High-precision dynamic control of key operating parameters of plasma equipment ensures the stability of the discharge process and the controllability of material modification reactions.

[0056] S3.1 The control unit first calculates the desired voltage output based on the power component in the control parameters. With current output And its deviation is corrected in real time through a closed-loop feedback mechanism. The discharge stability constraint expression is: .in, As a discharge error indicator, For voltage weighting coefficients, For the current, For the voltage monitored in real time, The current value is monitored in real time. The output signal is adjusted via proportional-integral control to ensure discharge power. The power is stabilized within the target power range predicted by the neural network, thereby maintaining the consistency of plasma energy distribution. The closed-loop adjustment range of the power parameters is a voltage range of 100V-1000V and a current range of 1A-10A, and all adjustment processes are performed with millisecond-level time resolution.

[0057] S3.2 The control unit then analyzes the gas flow rate component in the control parameters and adjusts the input rate of the reactant gas by driving the mass flow meter to achieve uniform gas distribution within the reaction chamber. The flow rate adjustment range is 0.1 L / min to 10 L / min, and the flow control threshold is dynamically corrected based on the real-time gas pressure feedback signal to maintain a matching relationship between the local gas density and the electric field distribution of the discharge area. The processing time component is managed by the digital timer module, and the control unit processes the time parameters. Input timer, combined with status monitoring signals during the modification process The time is adjusted, and the adjustment function is defined as follows: .in, The actual processing time is corrected, where λ is the dynamic time correction factor, reflecting the proportional relationship between the material response rate and the energy deposition rate. The electrode spacing component is executed by the stepper motor control module, and the control signal is provided by... The spatial adjustment parameters are derived and linear displacement adjustment is achieved through a conversion function between the motor stepping angle θ and the screw pitch. The displacement adjustment range is from 1mm to 50mm, ensuring optimal overlap between the discharge spatial field distribution and the energy absorption region on the material surface. All control processes are executed under the drive of real-time data streams. The system continuously iteratively updates parameters based on feedforward prediction and feedback correction mechanisms, ensuring that voltage, current, gas flow rate, processing time, and electrode spacing converge synergistically in the multi-dimensional parameter space. This maintains physical consistency between plasma energy density and material surface reaction kinetics, thereby achieving precise control and stable output of the material modification process.

[0058] S4. Real-time feedback and dynamic updates.

[0059] During plasma modification, real-time state parameters of the material, such as surface temperature, plasma emission spectrum data, and microstructural change data acquired through in-situ X-ray diffraction, including lattice constant change rate and amorphization rate, are collected in real time using a sensor array. These data are fed back into a neural network model at a high frequency of 10 to 100 times per second, driving a multi-scale feature fusion module and a nonlinear dynamics compensator to dynamically update plasma control parameters, forming a process such as... Figure 1 The real-time closed-loop control loop shown effectively avoids local damage caused by nonlinear phase transitions in the microstructure of the material surface.

[0060] Specifically, the real-time state parameter acquisition and feedback process is based on a multi-source sensor fusion mechanism, using high-frequency sampling to achieve dynamic monitoring and data feedback of the material modification state. The sensor array includes an infrared thermal imager, a spectrometer, and an in-situ X-ray diffractometer, wherein the infrared thermal imager outputs surface temperature field data. It is used to characterize the energy absorption distribution and time-varying properties of the material surface; the spectrometer outputs plasma emission spectrum signals. This reflects the concentration of active particles and their energy level distribution in the discharge region; the in-situ X-ray diffractometer outputs sequence data of microstructure changes, including the rate of change of lattice constant. With amorphization rate 'a' represents the lattice constant, and 'Ф' represents the proportion of the amorphous phase. The acquired analog signal undergoes noise suppression and filtering operations through a signal conditioning circuit, and the output signal is then converted to a digital signal at the sampling frequency by an analog-to-digital converter. Converted into digital form, forming a real-time state parameter vector. .

[0061] Specifically, all sensor data undergoes scale unification and unit conversion using a standardization function to ensure the input format conforms to the numerical requirements of the multi-scale feature fusion module. After the real-time state parameters are input into the neural network model in step S2, the system calculates the feature weights based on the multi-scale feature fusion module. During the update phase, phase transition dynamic constraints are introduced through a nonlinear dynamic compensator, and the phase transition volume fraction is corrected according to the Avrami model. Its time derivative Characterize the phase transition rate. Calculate the correction factor systematically. The plasma control parameters are updated based on feedback, using the formula... .in, For the updated plasma control parameters, This refers to the fine-tuning parameters output from step S2. The update frequency is set between 10 and 100 times per second to ensure that the parameter correction frequency matches the time scale of discharge energy fluctuations. The feedback system prioritizes analyzing the lattice constant change rate and amorphization rate signals provided by the in-situ X-ray diffractometer, using a dynamic weighting function. The degree to which microstructure data dominates the overall feedback is determined; the closer ω is to 1, the higher the priority of the amorphization process in the control. The finally updated plasma control parameters are re-inputted into the control unit in step S3, forming a real-time closed-loop control circuit. The control unit then... The system synchronously adjusts voltage, current, gas flow rate, and electrode spacing to maintain a dynamic balance between energy density and material reaction kinetics in a multidimensional parameter space. Through continuous iterative high-frequency feedback correction, the system achieves suppression of nonlinear phase transitions in the microstructure and adaptive coordination of energy input, thereby maintaining the stability and predictability of the plasma modification process.

Claims

1. A method for precise control of material modification guided by a neural network, characterized in that, Includes the following steps: S1. Data acquisition and preprocessing; The initial state parameters and preset modification target parameters of the material to be modified are obtained. The initial state parameters include the material surface morphology, chemical composition, physical property data, and microstructural feature parameters acquired by high-resolution microscopy. The target parameters for modification include the target surface roughness, the target chemical composition distribution, the target mechanical properties, and the target microstructure stability. All data has undergone validity checks, unit consistency conversion, and normalization. S2, Neural Network Model Processing; The initial state parameter set processed in step S1 With the set of target parameters for modification The input is a pre-trained neural network model, which includes a multi-scale feature fusion module and a nonlinear dynamics compensator. The multi-scale feature fusion module extracts and fuses macroscopic and microscopic features of the input parameters to generate preliminary plasma control parameters. The nonlinear dynamics compensator fine-tunes and corrects the preliminary parameters based on the Avrami phase transition dynamics equation, and outputs high-precision plasma control parameters, including plasma power, gas flow rate, processing time and electrode spacing. S3, Plasma equipment control; Based on the generated plasma control parameters, the operating parameters of the plasma device are adjusted by the control unit to perform plasma modification treatment on the material; The control unit calculates the desired voltage and current output of the plasma generator based on the control parameters and corrects the deviation in real time through a closed-loop feedback mechanism; at the same time, it adjusts the gas flow rate, processing time and electrode spacing of the gas supply system to ensure the stability of the discharge process and the controllability of the material modification reaction. S4. Real-time feedback and dynamic updates; During the plasma modification process, real-time state parameters of the material are collected in real time through a sensor array, including surface temperature, plasma emission spectrum data, and microstructure change data collected by an in-situ X-ray diffractometer. These data are used as feedback inputs to a neural network model to drive the multi-scale feature fusion module and nonlinear dynamic compensator to dynamically update the plasma control parameters.

2. The method for precise control of material modification guided by a neural network according to claim 1, characterized in that, The initial state parameters are obtained through the following steps: S1.1.1, high-resolution surface morphology image data is obtained by scanning the material surface with a scanning electron microscope, in the form of a grayscale matrix. This indicates that it is used to characterize the surface roughness distribution and microstructure undulations of materials; S1.1.2, X-ray photoelectron spectroscopy is used for component analysis to obtain the chemical elemental composition and bonding energy level distribution data of the material, forming a chemical composition vector. ;in, The mole fraction of the i-th element is represented, i = 1, 2, 3…n; and the characteristic parameters of the chemical bonding state are calculated by the energy level peak position and peak intensity. This reflects the distribution characteristics of elemental bonding energies in the material; S1.1.3, Mechanical properties are tested using a universal testing machine, and physical property data are output, including hardness H, strength, etc. With elastic modulus E; where, F represents the applied load, and A represents the cross-sectional area subjected to force; S1.1.

4. Microstructural observation data are provided through atomic force microscopy and transmission electron microscopy, and the lattice defect density is calculated through atomic resolution image sequences. ;in, This indicates the number of defects per unit volume. The sampling volume is used; the phase transition tendency index is described by the rate of change of the energy potential well. ,in, This is the free energy difference during the phase transition. This represents the width of the phase transition temperature range.

3. The method for precise control of material modification guided by a neural network according to claim 2, characterized in that, The training process of the neural network model includes steps S2.1, data collection, and S2.2, data preprocessing. S2.1, data collection, refers to collecting historical data from multiple plasma material modification experiments. The data includes the initial state parameter set of the material. Modification target parameter set Plasma control parameter set Modified material performance parameters Microscopic phase transition characteristic data obtained from in-situ monitoring ;in, Indicates plasma power. Indicates gas flow rate, Indicates processing time. Indicates the electrode spacing.

4. The method for precise control of material modification guided by a neural network according to claim 3, characterized in that, Step S2.2, data preprocessing includes: S2.2.1, performing outlier removal and noise point removal operations, using the mean square error filtering function. Outliers and noise points are removed to retain valid samples within the statistical distribution interval; among which, This represents the i-th original sample data value in a certain dimension; This represents the mean of the feature in historical samples; This represents the standard deviation of the characteristic; Indicates the absolute distance of a sample from the mean; The threshold representing the statistically valid fluctuation range; This is an indicator function; it takes a value of 1 when the condition is true, indicating that the sample is retained, and a value of 0 when the condition is false, indicating that the sample is removed; S2.2.2, Perform normalization transformation on all feature parameters. Where x represents the original feature value, This represents the minimum value of the feature dimension. This represents the maximum value of the feature dimension; S2.2.3, subsequently, feature enhancement operations are performed, applying rotation and scaling transformation matrices to the image data. Gaussian noise is superimposed on numerical sequence data. ,in, This represents the random perturbation noise term superimposed on the original feature. Indicates a Gaussian distribution. This represents the noise variance.

5. The method for precise control of material modification guided by a neural network according to claim 4, characterized in that, The pre-trained neural network model structure includes an input layer, three hidden layers, and an output layer; the input layer receives the standardized parameter set from step S1. and The input dimension is The hidden module includes a convolutional neural network module, a long short-term memory network module, and a multi-scale pyramid network module. The convolutional neural network module consists of four convolutional layers with kernel sizes of 3×3, 5×5, 5×5, and 3×3, respectively. The convolution operation is defined as... ;in, The output of the l-th convolutional layer represents the first... One feature point, For convolution kernel weights, For bias terms, activation function The ReLU function is used for nonlinear feature mapping; the convolutional neural network module extracts the surface morphology of the material. and chemical composition vector Macroscopic spatial characteristics, outputting a three-dimensional feature tensor The Long Short-Term Memory (LSTM) network module consists of a single-layer LSTM, containing three control units: a forget gate, an input gate, and an output gate. The computation process is as follows: ; ; Where t represents the time step index, This represents the feature vector input to the Long Short-Term Memory network at time step t. This represents the hidden state vector at time step t-1. W represents the joint input after concatenating the hidden state from the previous time step with the current input feature vector. f W represents the weight matrix of the forget gate. i W represents the weight matrix of the input gate. o Let b represent the weight matrix of the output gate. f b represents the bias vector of the forget gate. i b represents the bias vector of the input gate. o This represents the weight matrix of the output gate. This represents the Sigmoid activation function. This represents the output vector of the forget gate. This represents the output vector of the input gate. This represents the output vector of the output gate.

6. The method for precise control of material modification guided by a neural network according to claim 5, characterized in that, The multi-scale pyramid network module consists of three parallel convolutional paths with kernel sizes of 3×3, 5×5, and 7×7. Each path independently extracts feature responses at different spatial scales, defined as... ;in, Representing paths at different scales; the three feature maps are concatenated through feature stitching operations. After fusion, the dimension reduction mapping function Generate a unified feature representation.

7. The method for precise control of material modification guided by a neural network according to claim 6, characterized in that, The output layer consists of a fully connected layer and a linear activation function, and the mapping function is defined as follows: ;in, This represents the predicted plasma parameters output by the output layer. and These are the trainable parameters for the output layer.

8. The method for precise control of material modification guided by a neural network according to claim 7, characterized in that, The control unit is based on the plasma control parameters from step S2. High-precision dynamic control of key operating parameters of plasma equipment ensures the stability of the discharge process and the controllability of the material modification reaction; S3.1, the control unit first calculates the desired voltage output based on the power component in the control parameters. With current output And it corrects its deviation in real time through a closed-loop feedback mechanism; The expression for discharge stability constraint is as follows: ;in, As a discharge error indicator, For voltage weighting coefficients, For the current, For the voltage monitored in real time, The current value is monitored in real time; the output signal is adjusted through proportional-integral control to ensure discharge power. The power is stabilized within the target range predicted by the neural network, thus maintaining the consistency of plasma energy distribution. S3.2 The control unit then analyzes the gas flow component in the control parameters, adjusts the input rate of the reactant gas by driving the mass flow meter, achieving uniform gas distribution within the reaction chamber, and dynamically corrects the flow control threshold based on real-time gas pressure feedback signals to maintain a matching relationship between the local gas density and the electric field distribution of the discharge region. The processing time component is managed by the digital timer module, and the control unit processes the time parameters... Input timer, combined with status monitoring signals during the modification process The time is adjusted, and the adjustment function is defined as follows: ;in, The actual processing time is corrected, where λ is the dynamic time correction coefficient; the electrode spacing component is executed by the stepper motor control module, and the control signal is provided by... The spatial adjustment parameters are derived, and linear displacement adjustment is achieved through the conversion function between the motor stepping angle θ and the pitch.

9. The method for precise control of material modification guided by a neural network according to claim 8, characterized in that, The real-time state parameter acquisition and feedback process is based on a multi-source sensor fusion mechanism, which achieves dynamic monitoring and data feedback of the material modification state through high-frequency sampling. The sensor array includes an infrared thermal imager, a spectrometer, and an in-situ X-ray diffractometer. The infrared thermal imager outputs surface temperature field data. The spectrometer outputs plasma emission spectral signals. The in-situ X-ray diffractometer outputs a sequence of microstructural changes; the acquired analog signal undergoes noise suppression and filtering operations through a signal conditioning circuit, and the output signal is converted into digital form by an analog-to-digital converter to form a real-time state parameter vector.

10. The method for precise control of material modification guided by a neural network according to claim 9, characterized in that, All sensor data undergoes scaling and unit conversion using a standardization function to ensure the input format meets the numerical requirements of the multi-scale feature fusion module. After real-time state parameters are input into the neural network model in step S2, the system calculates feature weights based on the multi-scale feature fusion module. During the update phase, phase transition dynamic constraints are introduced through a nonlinear dynamic compensator, and the phase transition volume fraction is corrected according to the Avrami model. Its time derivative Characterize the phase transition rate; calculate the correction factor systematically. The plasma control parameters are updated based on feedback, using the formula... ;in, For the updated plasma control parameters, The fine-tuning parameters output in step S2 are used; the update frequency is set between 10 and 100 times per second. The feedback system prioritizes analyzing the lattice constant change rate and amorphization rate signals provided by the in-situ X-ray diffractometer, and uses a dynamic weighting function. The degree to which microstructure data dominates the overall feedback is determined, and the updated plasma control parameters are then re-input into the control unit of step S3 to form a real-time closed-loop control circuit. The control unit then... The changes in voltage, current, gas flow rate, and electrode spacing are adjusted synchronously to maintain a dynamic balance between energy density and material reaction kinetics in the multidimensional parameter space.

Images