A plasma electromagnetic field detection method, device, terminal and storage medium

Abstract

The present application provides a plasma electromagnetic field detection method, device, terminal and storage medium, which relate to the field of plasma diagnostic technology. The method includes: constructing a plasma electromagnetic field physical model, collecting electromagnetic field data, the electromagnetic field data including the electric field intensity and magnetic induction intensity at each time and space point in the plasma; constructing a plasma time-space domain physical information neural network architecture, a multi-objective loss function and an adaptive sampling strategy, and training the plasma time-space domain physical information neural network architecture based on the multi-objective loss function, the adaptive sampling strategy and the electromagnetic field data; using the trained plasma time-space domain physical information neural network architecture to detect the electromagnetic field distribution of the plasma to be tested. The present application can improve the accuracy and reliability of detection.

Description

Technical Field

[0001] The present application relates to the field of plasma diagnosis technology, and in particular to a plasma electromagnetic field detection method, device, terminal and storage medium. Background Art

[0002] Plasma, as the fourth state of matter, is widely present in fusion energy devices, plasma thrusters, material surface treatment equipment, and astrophysical environments. Accurate detection and analysis of the electromagnetic field distribution in plasma is of great significance for understanding plasma dynamics, optimizing the performance of plasma devices, and achieving precise plasma control. Traditional plasma electromagnetic field detection methods mainly include Langmuir probes, magnetic probes, microwave interferometers, Thomson scattering, and other technologies. However, these methods have the following limitations: (1) Probe-based methods are invasive measurements that will cause disturbances to the plasma and affect measurement accuracy; (2) Although optical diagnostic methods are non-invasive, they can usually only obtain information along the line of sight, making it difficult to reconstruct three-dimensional spatial distributions; (3) The time resolution of traditional methods is limited by the hardware sampling rate, making it difficult to capture rapidly changing plasma phenomena; (4) Under complex geometric structures and strong nonlinear conditions, the computational cost of traditional numerical simulation methods is extremely high.

[0003] In recent years, machine learning, especially deep neural networks, has shown great potential in the field of scientific computing. As an emerging scientific machine learning method, physics-informed neural networks (PINNs) can accurately solve and predict the behavior of physical systems under limited observation data conditions by embedding physical laws (such as partial differential equations) in the loss function of the neural network. The advantages of the PINNs method are: (1) it can integrate multi-source heterogeneous data, including experimental measurements, numerical simulations, and theoretical models; (2) it has strong nonlinear fitting capabilities and is suitable for complex plasma systems; (3) after training, it has fast inference speed and can achieve real-time detection; (4) it can handle incomplete data and noisy data and has good robustness.

[0004] However, the direct application of PINNs to plasma electromagnetic field detection still faces many challenges: the multi-scale nature of plasma systems makes it difficult for neural networks to simultaneously capture physical phenomena at different scales; the strong nonlinearity and instability of plasmas complicate network training; the vectorial nature of electromagnetic fields and the complexity of Maxwell's equations complicate network design; and the spatial inhomogeneity and temporal evolution of plasma parameters require networks with good generalization capabilities. Therefore, it is urgent to develop PINNs specifically for plasma electromagnetic field detection to overcome these technical difficulties. Summary of the Invention

[0005] The present application provides a plasma electromagnetic field detection method, device, terminal and storage medium to solve the problems of intrusive interference, low spatial resolution, slow time response and high computational cost existing in traditional plasma electromagnetic field detection technology.

[0006] In a first aspect, the present application provides a plasma electromagnetic field detection method, comprising:

[0007] Constructing a plasma electromagnetic field physical model and collecting electromagnetic field data, wherein the electromagnetic field data includes the electric field intensity and magnetic induction intensity at each time and space point in the plasma;

[0008] Constructing a plasma space-time domain physical information neural network architecture, a multi-objective loss function, and an adaptive sampling strategy, and training the plasma space-time domain physical information neural network architecture based on the multi-objective loss function, the adaptive sampling strategy, and the electromagnetic field data;

[0009] The trained plasma spatiotemporal physical information neural network architecture is used to detect the electromagnetic field distribution of the plasma to be tested.

[0010] In a second aspect, the present application provides a plasma electromagnetic field detection device, comprising:

[0011] A model building module is used to build a plasma electromagnetic field physical model and collect electromagnetic field data, wherein the electromagnetic field data includes the electric field intensity and magnetic induction intensity at each time and space point in the plasma;

[0012] An architecture construction and training module is used to construct a plasma space-time domain physical information neural network architecture, a multi-objective loss function and an adaptive sampling strategy, and train the plasma space-time domain physical information neural network architecture based on the multi-objective loss function, the adaptive sampling strategy and the electromagnetic field data;

[0013] The electromagnetic field detection module is used to detect the electromagnetic field distribution of the plasma to be tested using the trained plasma spatiotemporal physical information neural network architecture.

[0014] In a third aspect, the present application provides a terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the method described in the first aspect or any possible implementation of the first aspect are implemented.

[0015] In a fourth aspect, the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the method described in the first aspect or any possible implementation of the first aspect.

[0016] The present application provides a plasma electromagnetic field detection method, device, terminal and storage medium. By constructing a plasma electromagnetic field physical model, electromagnetic field data is collected, and the electromagnetic field data includes the electric field intensity and magnetic induction intensity at each time and space point in the plasma; a plasma time-space domain physical information neural network architecture, a multi-objective loss function and an adaptive sampling strategy are constructed, and the plasma time-space domain physical information neural network architecture is trained based on the multi-objective loss function, the adaptive sampling strategy and the electromagnetic field data; and the trained plasma time-space domain physical information neural network architecture is used to detect the electromagnetic field distribution of the plasma to be tested. This application constructs a plasma spatiotemporal physical information neural network architecture to deeply integrate physical information with neural networks, so that the model can better understand and simulate the complex spatiotemporal changes of the plasma electromagnetic field, thereby improving the accuracy and reliability of detection; and, by introducing a multi-objective loss function and an adaptive sampling strategy, the neural network can simultaneously optimize multiple objectives during the training process and adaptively adjust the sampling strategy according to the characteristics of the plasma, thereby more effectively utilizing data resources and improving the training efficiency and generalization ability of the model; at the same time, by using the trained plasma spatiotemporal physical information neural network architecture, the electromagnetic field distribution of the plasma to be tested can be quickly and accurately detected, which not only improves the detection efficiency, but also reduces human intervention and errors, making the detection results more objective and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0018] Figure 1 This is a flow chart of the implementation of the plasma electromagnetic field detection method provided in the embodiment of the present application;

[0019] Figure 2 Schematic diagram of the structure of the plasma electromagnetic field detection device provided in an embodiment of the present application;

[0020] Figure 3 It is a schematic diagram of a terminal provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0022] In order to make the purpose, technical solutions and advantages of this application clearer, specific embodiments will be described below with reference to the accompanying drawings.

[0023] Figure 1 The following is a flowchart of the plasma electromagnetic field detection method provided in the embodiment of the present application:

[0024] In step 101, a plasma electromagnetic field physical model is constructed and electromagnetic field data are collected. The electromagnetic field data includes the electric field intensity and magnetic induction intensity at each time and space point in the plasma.

[0025] In the embodiments of the present application, a set of governing equations describing the evolution of the electromagnetic field in a plasma is established as a plasma electromagnetic field physics model, and based on the plasma electromagnetic field physics model, corresponding electromagnetic field data is obtained. The electromagnetic field data may include the electromagnetic field intensity and magnetic induction intensity at each spacetime point and time in the plasma.

[0026] The embodiment of the present application utilizes a plasma electromagnetic field physical model to obtain electromagnetic field data, which can comprehensively and accurately describe the electromagnetic field characteristics of the plasma and provide a reliable basis for subsequent detection and analysis.

[0027] In one possible implementation, constructing a plasma electromagnetic field physical model may include:

[0028] The plasma electromagnetic field physics model is constructed using Maxwell's equations and plasma fluid equations. Maxwell's equations are used to describe the behavior of the electromagnetic field, and the plasma fluid equations are used to describe the macroscopic motion of the plasma.

[0029] Optionally, the plasma electromagnetic field physics model in the embodiment of the present application is constructed by a group of control equations for the evolution of the electromagnetic field in the plasma, which couples the Maxwell equations describing the behavior of the electromagnetic field and the plasma fluid equations describing the macroscopic motion of the plasma, thereby achieving a self-consistent description of the electromagnetic field and plasma dynamics.

[0030] Among them, the plasma electromagnetic field physical model is:

[0031]

[0032]

[0033]

[0034]

[0035] in, For any spacetime point and time in the plasma The electric field strength at , (vector), For any spacetime point and time in the plasma The magnetic induction intensity (vector) at , both of which are key physical quantities to be solved in the plasma electromagnetic field physics model. is the current density (vector) generated by the motion of charged particles in the plasma, is the net charge density (scalar), is the vacuum permeability (constant), is the dielectric constant of vacuum (constant).

[0036] The above set of equations together constitute the eigenconstraints of the electromagnetic field.

[0037] Among them, the current density generated by the movement of charged particles in the plasma As the core coupling term, this embodiment uses the generalized Ohm's law, which takes into account the response of electron motion in the plasma under the action of electric fields, magnetic fields, and pressure gradients, and more accurately reflects the conductive properties of the plasma, namely:

[0038]

[0039] in, The conductivity tensor of the plasma depends on the temperature, density, magnetic field and other parameters of the plasma, reflecting the ability of the plasma to conduct current. is the plasma macroscopic velocity (vector), which describes the overall motion of the plasma; is the electron charge (constant); is the electron density (scalar), which reflects the number of electrons per unit volume; is the electron pressure (scalar), which describes the random thermal motion of electrons.

[0040] The generalized Ohm's law used in the embodiments of this application not only includes the classic Ohm's term , the Hall effect term is also introduced and electron inertia and electron pressure gradient effects , which enables the plasma electromagnetic field physics model to more accurately describe the current generation mechanism in complex magnetohydrodynamic plasmas.

[0041] In step 102, a plasma space-time domain physical information neural network architecture, a multi-objective loss function and an adaptive sampling strategy are constructed, and the plasma space-time domain physical information neural network architecture is trained based on the multi-objective loss function, the adaptive sampling strategy and the electromagnetic field data.

[0042] In an embodiment of the present application, an end-to-end learning framework is designed, namely a plasma space-time domain physical information neural network architecture. The plasma space-time domain physical information neural network architecture adopts a multi-branch structure, and each branch is specifically responsible for learning the specific physical characteristics of the plasma electromagnetic field at different spatial frequencies and time scales, aiming to maximize the expression ability and compliance with the laws of physics of the plasma space-time domain physical information neural network architecture. Then a plasma region adaptive multi-objective loss function is constructed for multi-objective optimization. In addition, an adaptive sampling strategy for plasma structural features needs to be constructed to improve training efficiency and plasma space-time domain physical information neural network architecture accuracy. Then, the plasma space-time domain physical information neural network architecture is trained using the constructed multi-objective loss function and adaptive sampling strategy and the electromagnetic field data collected in step 101.

[0043] In the embodiment of the present application, a plasma spatiotemporal domain physical information neural network architecture is constructed, and physical information is deeply integrated with the neural network, so that the plasma spatiotemporal domain physical information neural network architecture can better understand and simulate the complex spatiotemporal changes of the plasma electromagnetic field, thereby improving the accuracy and reliability of detection. In addition, the introduction of a multi-objective loss function and an adaptive sampling strategy enables the neural network to optimize multiple objectives simultaneously during the training process and adaptively adjust the sampling strategy according to the characteristics of the plasma, thereby more effectively utilizing data resources and improving the training efficiency and generalization ability of the plasma spatiotemporal domain physical information neural network architecture.

[0044] In one possible implementation, building a plasma spatiotemporal physics information neural network architecture may include:

[0045] Constructing a multi-scale feature extraction module, wherein the multi-scale feature extraction module includes a multi-scale convolution kernel and a Fourier feature map. The multi-scale convolution kernel is used to extract local patterns of spatiotemporal coordinates, and the Fourier feature map is used to map the spatiotemporal coordinates into a high-dimensional space. The spatiotemporal coordinates are composed of spatiotemporal points and time.

[0046] Using automatic differentiation technology, a physical constraint coding module is constructed, which is used to calculate the partial derivatives of electromagnetic field data with respect to space-time coordinates;

[0047] Constructing adaptive activation functions;

[0048] Residual connections and attention mechanisms are introduced to construct a plasma spatiotemporal domain physical information neural network architecture using a multi-scale feature extraction module, a physical constraint encoding module, and an adaptive activation function.

[0049] Optionally, the core input of the plasma space-time domain physical information neural network architecture is a four-dimensional vector consisting of three-dimensional space coordinates and time coordinates , whose output is the electric field intensity at any time and space point in the plasma and magnetic induction intensity All components of The plasma space-time domain physics information neural network architecture includes:

[0050] (1) The construction of multi-scale feature extraction module aims to efficiently capture the fine features of plasma electromagnetic field at different frequency and time scales. First, a multi-scale convolution kernel is used, i.e., different sizes (3 3, 5 5, 7 7) Convolution kernels extract local patterns of input spatiotemporal coordinates in parallel. Larger receptive field convolution kernels can capture macroscopic, low-frequency global features, while smaller receptive field convolution kernels focus on microscopic, high-frequency local details. Secondly, the original input spatiotemporal coordinates are transformed into Mapping to a high-dimensional space addresses the "spectral bias" problem of plasma spatiotemporal physics neural network architectures when processing high-frequency information. Fourier eigenmapping enhances the network's ability to learn and represent rapidly changing electromagnetic field components, such as plasma waves.

[0051] Among them, the Fourier eigenmap is defined as:

[0052]

[0053] in, is the space-time coordinate vector , It is a fixed matrix randomly sampled from a Gaussian distribution (such as the standard normal distribution), whose elements control the range and direction of the characteristic frequency after mapping. By introducing high-frequency components, the neural network can better fit the electromagnetic field with complex oscillation characteristics.

[0054] (2) Construction of the physical constraint coding module. The physical constraint coding module is the core of the plasma space-time domain physical information neural network architecture. By directly embedding the physical laws into the loss function of the neural network, the network output is forced to meet the known physical laws. Specifically, the automatic differentiation technology can be used to efficiently calculate the output of the plasma space-time domain physical information neural network architecture (i.e., the electric field intensity). and magnetic induction intensity ) with respect to the input spacetime coordinates, thus directly constructing the residual of Max's equations. The physical residual is calculated as:

[0055] The residual of Faraday's law is expected to approach zero, that is:

[0056]

[0057] Ampere's law residual, calculated in conjunction with generalized Ohm's law ,Right now:

[0058]

[0059] The residual of Gauss's electric field law, if the charge density is known or can be derived, is:

[0060]

[0061] The residual of Gauss's magnetic field law is always expected to be zero, that is:

[0062]

[0063] The above residual terms will be incorporated into the total loss function to guide the plasma spatiotemporal physical information neural network architecture to automatically learn solutions that satisfy these physical equations during the training process.

[0064] (3) Construction of adaptive activation function: Traditional fixed activation functions (such as ReLU, Sigmoid) may not be able to fully capture the nonlinear mapping relationship of complex physical systems. The embodiment of the present application adopts a learnable activation function in the form of a linear combination of multiple basic activation functions, allowing the plasma spatiotemporal domain physical information neural network architecture to adaptively adjust the nonlinear characteristics during the training process, namely:

[0065]

[0066] in, 、 、 、 、 、 These parameters are all learnable parameters. These parameters are optimized together with other network weights during the training process of the plasma space-time domain physical information neural network architecture, so that the activation function can dynamically adjust its shape, slope and range, thereby more flexibly fitting the complex nonlinear relationship of the electromagnetic field and improving the expression ability and convergence speed of the plasma space-time domain physical information neural network architecture. For example, and Provides nonlinear saturation characteristics, while Introducing periodicity helps capture fluctuation phenomena.

[0067] (4) Introduction of residual connection and attention mechanism: In order to build a deeper network and solve the problem of gradient disappearance during training, the plasma spatiotemporal domain physical information neural network architecture introduces residual connection, which allows information to directly skip several layers, alleviating the difficulty of deep network training and accelerating convergence. In addition, in order to capture the relationship between different components of the electromagnetic field (such as and The self-attention mechanism is introduced to address the complex coupling relationships and non-local dependencies between features. The self-attention mechanism can dynamically learn the correlation strength between different features, providing a more global perception capability for the plasma spatiotemporal physics information neural network architecture, especially when dealing with non-uniform or local effects that may exist in plasma. The self-attention calculation formula is:

[0068]

[0069] in, 、 、 They are respectively the query matrix, key matrix, and value matrix obtained from the input features (electromagnetic field components) through linear transformation. The key dimension is used for scaling to prevent the inner product from being too large. Through this mechanism, the network can focus on the electromagnetic field components or regions most relevant to the current prediction, enhancing the ability of the plasma spatiotemporal physics information neural network architecture to model complex electromagnetic field structures.

[0070] In one possible implementation, constructing a multi-objective loss function may include:

[0071] Obtain data fitting loss, physical constraint loss, boundary condition loss and regularization term, and obtain a first weight, a second weight, a third weight and a fourth weight, wherein the first weight is the weight corresponding to the data fitting loss, the second weight is the weight corresponding to the physical constraint loss, the third weight is the weight corresponding to the boundary condition loss, and the fourth weight is the weight corresponding to the regularization term;

[0072] The product of the data fitting loss and the first weight, the product of the physical constraint loss and the second weight, the product of the boundary condition loss and the third weight, and the product of the regularization term and the fourth weight are added together to construct a plasma region adaptive multi-objective loss function.

[0073] Optionally, the multi-objective loss function in the embodiment of the present application is composed of data fitting loss, physical constraint loss, boundary condition loss and regularization term weighted by different weights to balance the requirements of the plasma spatiotemporal domain physical information neural network architecture for observation data, physical laws and solution smoothness. The multi-objective loss function is:

[0074]

[0075] in, is a multi-objective loss function, is the data fitting loss, is the physical constraint loss, is the boundary condition loss, is the regularization term, is the first weight, is the second weight, is the third weight, is the fourth weight. The above weights are all adjustable weight coefficients used to balance the importance of various losses.

[0076] Among them, the data fitting loss The deviation between the neural network prediction results and the limited measured (or simulated) data is measured to ensure that the plasma spatiotemporal physics information neural network architecture can accurately fit the known observations. The mean square error (MSE) is used for calculation, that is:

[0077]

[0078] in, is the total number of observation data points, For the neural network The electromagnetic field component values ​​predicted at each observation point (including , , , , , ), is the corresponding actual observed value.

[0079] Physical constraint loss Ensure that the prediction results of the neural network strictly satisfy the Maxwell equations in the entire computational domain. In order to solve the problem that the universal loss function cannot reflect the regional differences in the physical processes inside the plasma, the points in the computational domain are first dynamically divided into different physical regions according to the local physical parameters of the plasma (such as normalized magnetic pressure, collision frequency, Hall parameters, etc.), such as the core high temperature region, edge collision region, sheath strong electric field region, etc. Then different weights are applied to the physical residuals in different regions, so that the network training focuses on the dominant physical processes in each region. Specifically, the physical constraint loss is defined as the weighted sum of all physical silent losses, that is:

[0080]

[0081] in, is the total number of physical areas divided, For the physical area, For the The number of points in a physical area, 、 etc. is for the A specific weight factor for each physical area. For example:

[0082] 1) In the core region of magnetic confinement fusion plasma (i.e., ideal magnetic fluid region), the weight of the relevant terms of Ampere's law describing ideal conductivity can be increased, such as .

[0083] 2) In the sheath region near the wall, the charge separation effect is significant, which can increase the residual error of Gaussian electric field law. weights to accurately capture the sheath electric field.

[0084] 3) In the current sheet region where magnetic field reconnection exists, the plasma resistance effect cannot be ignored and can increase the weight of the resistance term in the generalized Ohm's law.

[0085] In the embodiment of the present application, through a regionally adaptive weighting strategy, the plasma space-time domain physical information neural network architecture is forced to learn solutions that obey the local dominant physical laws, thereby significantly improving the expression ability, generalization ability and physical rationality of the plasma space-time domain physical information neural network architecture for complex plasma physical phenomena, especially in areas with sparse data.

[0086] Boundary condition loss Ensure that the predicted electromagnetic fields of the neural network meet specific physical conditions at the boundary of the computational domain. For example, for the boundary of an ideal conductor, the tangential electric field is zero and the normal magnetic field is zero. This loss term ensures the boundary consistency of the neural network architecture solution of the plasma space-time physics information, namely:

[0087]

[0088] in, is the total number of boundary sampling points, is the unit external normal vector on the boundary, is the tangential component of the electric field, is the normal component of the magnetic field.

[0089] Regularization term It is used to prevent the plasma spatiotemporal physics information neural network architecture from overfitting the training data, and to improve the smoothness and stability of the prediction solution, avoiding physically unreasonable violent oscillations or discontinuities, namely:

[0090]

[0091] in, is L2 weight regularization, For the neural network The weight matrix of the layer, is the Frobenius norm, which limits the complexity of the plasma spatiotemporal physical information neural network architecture by penalizing the weight value. Regularization for high-order derivatives encourages predictive solutions over the entire computational domain by minimizing the squared norm of the second-order derivatives of the electric and magnetic fields. Maintain smoothness internally to prevent noise and oscillation in the solution. 、 is the regularization coefficient, which is used to control the regularization strength of each item.

[0092] In one possible implementation, building an adaptive sampling strategy may include:

[0093] Using a uniform random sampling method, N initial collocation points are selected from all the time and space domains in the plasma electromagnetic field, where N is a positive integer greater than or equal to 1;

[0094] Calculate the sampling weight of each initial collocation point, and use the sampling weight of each initial collocation point to calculate the sampling probability of the corresponding initial collocation point. The sampling weight includes the physical residual and the physical field gradient.

[0095] The sampling probability of each initial matching point is used to determine the matching point set for training the plasma spatiotemporal physical information neural network architecture.

[0096] Optionally, in order to significantly improve the training efficiency and accuracy of the plasma space-time domain physical information neural network architecture, the embodiment of the present application uses physical residuals and physical field gradients to construct an adaptive sampling strategy for plasma structural characteristics, so that the training process can more effectively concentrate computing resources on areas where the plasma space-time domain physical information neural network architecture is difficult to learn and where the physical structure is complex (such as high-gradient areas such as sheaths, double layers, and current sheets).

[0097] The specific implementation steps of the adaptive sampling strategy are as follows:

[0098] (1) Initialization sampling: At the beginning of training, in order to cover the entire computational domain, a uniform random sampling method is used to select N initial points in all time and space domains of the plasma electromagnetic field. These initial points are used to preliminarily calculate various losses and guide the initial learning direction of the plasma time and space domain physical information neural network architecture.

[0099] (2) Calculate sampling weights: After each training cycle (or at regular intervals), for the current N initial points, calculate the sampling weight of each initial point. The sampling weight consists of two parts: the physical residual and the physical field gradient. The physical residual is used to identify the area where the prediction of the plasma spatiotemporal physical information neural network architecture deviates the most from the physical laws, while the physical field gradient is used to identify the structural areas in the plasma where the physical quantities change dramatically.

[0100] Correspondingly, the sampling weight The calculation formula is:

[0101]

[0102] in, For the Initial points, For the The total physical residual of the initial collocation points, and They are the gradients of the electric field tensor and the magnetic field tensor (such as the Frobenius norm of the Jacobian matrix), which are used to measure the intensity of spatial changes in field quantities. is a hyperparameter that balances the importance of the residual term and the gradient term. is a very small integer (such as 10 -8 ), which is used to prevent division by zero errors when the weights are too small or even zero, and to ensure that all points have the probability of being sampled.

[0103] (3) Probability sampling: The sampling weight of each initial distribution point is obtained based on the calculation , using the importance sampling method to select the points for the next round of training. The larger the weight, the higher the probability that the point will be selected, thereby allocating more computing resources to the areas of the plasma spatiotemporal physical information neural network architecture that are "difficult to learn (i.e., high residual)" or "complex structure (i.e., high gradient)", namely:

[0104]

[0105] in, For the The probability of a point being selected. is the temperature parameter, the value is greater than 0, when When , the sampling probability is linearly related to the sampling weight; when When it increases, sampling will be more concentrated in high-weight areas, thereby controlling the concentration of sampling and the balance between exploration and exploitation.

[0106] (4) Dynamic update: In order to continuously optimize the training process, the adaptive sampling strategy is dynamically updated. Every T training cycles, the sampling weights are recalculated and a new set of points is generated based on the latest physical residual distribution. Specifically, a hybrid strategy is adopted: a portion of the current good performance (low residual) points is retained, and a portion of new points is generated based on the new probability distribution. In particular, the old points that have converged in the learning process but have low contribution are replaced, and new points are introduced in the high residual area to ensure that the plasma time-space domain physical information neural network architecture can continue to focus on the non-convergence area, further improving the training efficiency and the accuracy of the plasma time-space domain physical information neural network architecture.

[0107] In one possible implementation, training a plasma spatiotemporal physics information neural network architecture based on a multi-objective loss function, an adaptive sampling strategy, and electromagnetic field data may include:

[0108] Sequentially pre-train and fine-tune the plasma spatiotemporal physics information neural network architecture;

[0109] Based on the multi-objective loss function, the multi-objective loss function of the finely tuned plasma time-space domain physical information neural network architecture is calculated, and the multi-objective loss function of the finely tuned plasma time-space domain physical information neural network architecture is used as a joint loss function;

[0110] Using the adaptive sampling strategy, the weights of each loss term in the joint loss function are calculated, and the total loss weight is calculated using the weights of each loss term;

[0111] Determine whether the total loss weight converges;

[0112] If the total loss weight converges, it is determined that the training of the plasma spatiotemporal domain physical information neural network architecture is completed;

[0113] If the total loss weight does not converge, the process returns to the steps of pre-training and fine-tuning the plasma spatiotemporal physical information neural network architecture.

[0114] Optionally, in the embodiment of the present application, after constructing the plasma spatiotemporal domain physical information neural network architecture, a two-stage training is first performed, namely a pre-training stage and a fine-tuning stage.

[0115] The pre-training phase aims to provide a good initial weight for the plasma spatiotemporal physical information neural network architecture, enabling preliminary learning of the basic characteristics and physical laws of the plasma electromagnetic field. The specific training process includes:

[0116] Data source training: Mainly uses high-precision physical simulation data (such as numerical simulation results based on magnetohydrodynamics equations). These data usually cover a wider range of physical parameter spaces and have higher accuracy.

[0117] Optimizer training: The training process primarily uses the Adam optimizer. The Adam optimizer combines the advantages of Adagrad and RMSprop, adaptively adjusting the learning rate and demonstrating fast and stable convergence in the early and middle stages of training.

[0118] Learning rate adjustment: The learning rate adopts the cosine annealing strategy. This strategy makes the learning rate start from a higher initial value. Smoothly decreases to a lower minimum , whose variation curve is in the form of a cosine function. This annealing method helps the plasma space-time domain physical information neural network architecture to quickly explore the solution space in the early stages of training and make fine adjustments in the later stages to achieve better convergence. The calculation formula of the cosine annealing strategy is as follows:

[0119]

[0120] in, For the current training cycle The learning rate, is the total number of training cycles.

[0121] After the plasma space-time physics information neural network architecture has achieved a good initial state through pre-training, the embodiment of the present application enters the fine-tuning stage. This stage will further optimize the plasma space-time physics information neural network architecture based on actual measurement data so that it can accurately fit real-world observations. The specific training process includes:

[0122] Data source training: actual measurement data, which are usually sparse and may be noisy, but contain unique information about the real physical system.

[0123] Optimizer Training: This phase primarily uses the L-BFGS optimizer. The L-BFGS optimizer is a quasi-Newton method that uses the second-order derivative of the loss function (an approximate Hessian matrix) to guide optimization. It typically converges faster and achieves higher accuracy near the optimal solution than first-order optimizers (such as the Adam optimizer). It is particularly well-suited for optimization problems involving complex loss functions, such as those found in plasma spatiotemporal physics information neural network architectures.

[0124] Then, the multi-objective loss function constructed as described above is used to calculate the finely tuned multi-objective loss function of the plasma spatiotemporal physical information neural network architecture, and use it as a joint loss function. And under the premise that the joint loss function satisfies the adaptive sampling strategy, the weights of each loss term in the joint loss function are calculated, and the total loss weight is calculated using the weights of each loss term. Specifically, in order to solve the problem of weight imbalance caused by differences in numerical values ​​or different convergence speeds of each loss term in the multi-objective loss function, the embodiment of the present application adopts an adaptive sampling strategy to dynamically adjust the weights of each loss term of the joint loss function, that is, dynamically adjust the weights of each loss term in the total loss function according to the relative size of each loss term, that is:

[0125]

[0126] in, For the The first iteration cycle The weight of the loss, is the corresponding loss value, To adjust the rate parameter, usually a positive value.

[0127] Through the above-mentioned exponential decay or growth strategy, the embodiment of the present application can dynamically give higher weights to the loss terms that are currently relatively large or difficult to converge, thereby balancing the contributions of various loss terms and avoiding a certain loss term from dominating the training process too early.

[0128] Then determine whether the calculated total loss weight converges. If it converges, it indicates that the plasma space-time domain physical information neural network architecture training is completed; if it does not converge, return to the pre-training and fine-tuning stage for retraining.

[0129] In addition, the trained plasma spatiotemporal physics information neural network architecture can be used as an efficient proxy model for real-time electromagnetic field detection and reconstruction. In order to evaluate the reliability of the prediction and provide a basis for subsequent decision-making, the embodiment of the present application also introduces an uncertainty quantification mechanism. Specifically, it includes:

[0130] (1) Training multiple plasma spatiotemporal physics information neural network architectures: In order to capture the inherent epistemic uncertainty of the plasma spatiotemporal physics information neural network architecture (i.e., the uncertainty caused by the model structure and parameter selection), instead of training only one plasma spatiotemporal physics information neural network architecture, an integrated model consisting of multiple (e.g., K = 5 to 10) plasma spatiotemporal physics information neural network architectures with different randomly initialized weights, or slight architectural differences, or different training data subsets is trained. That is:

[0131]

[0132] Among them, each plasma space-time domain physical information neural network framework Both can independently predict the electromagnetic field.

[0133] (2) Calculate the mean and variance of the prediction of the integrated model: For any space-time point x to be queried, each plasma space-time domain physical information neural network architecture in the integration will give a prediction value The uncertainty is quantified by computing a statistic about these predictions:

[0134] 1) Predicted mean : represents the optimal estimated value of the electromagnetic field at this point by the integrated model, which serves as the final real-time detection result. That is:

[0135]

[0136] 2) Prediction Variance : represents the degree of dispersion between the prediction values ​​of the integrated model, which directly reflects the uncertainty of the prediction of the integrated model at this point. The larger the variance, the greater the difference in predictions of different plasma spatiotemporal physical information neural network architectures at this point, and the higher the uncertainty. That is:

[0137]

[0138] in, is the Euclidean norm, which is used to calculate the difference in the predicted vectors of the electromagnetic field components.

[0139] (3) Constructing confidence intervals and identifying high uncertainty areas: Based on the predicted mean and variance, the confidence intervals for the electromagnetic field prediction can be constructed (95% confidence intervals are .96 ), provides users with a prediction range rather than a single value to assess the reliability of the prediction. At the same time, by visualizing the variance The spatial distribution of the electromagnetic field can clearly identify areas of high uncertainty. These areas may be due to data sparseness, conflicting physical constraints, or insufficient model learning. This uncertainty information can further guide the placement of subsequent measurement points (for example, adding sensors in areas of high uncertainty) or prompt operators to more carefully assess the electromagnetic field in that area.

[0140] At the same time, in order to achieve millisecond-level real-time detection, the embodiment of the present application may also have the following operations:

[0141] (1) Model quantization: quantize the weights of the trained plasma spatiotemporal physics information neural network architecture from 32-bit floating point numbers to 8-bit integers to reduce computational and storage overheads.

[0142]

[0143] in, is the weight corresponding to the 8-bit integer, is the weight corresponding to the 32-bit floating point, is zero point, is the scaling factor.

[0144] (2) Graph optimization: Optimize the computational graph through techniques such as operator fusion and constant folding to reduce the number of memory accesses.

[0145] (3) Parallel reasoning: Utilizing the parallel computing capability of the GPU, the reasoning tasks of spatial grid points are assigned to different CUDA cores.

[0146] (4) Pipeline processing: Design an inference pipeline to achieve parallel execution of data preprocessing, neural network inference, and post-processing.

[0147] In step 103, the trained plasma spatiotemporal physical information neural network architecture is used to detect the electromagnetic field distribution of the plasma to be tested.

[0148] In an embodiment of the present application, the space-time coordinates of each space-time point and time component in the plasma to be measured are input into the trained plasma space-time domain physical information neural network architecture to obtain the corresponding electromagnetic field distribution.

[0149] The embodiment of the present application utilizes a trained plasma spatiotemporal physical information neural network architecture to quickly and accurately detect the electromagnetic field distribution of the plasma to be tested. This method not only improves detection efficiency, but also reduces human intervention and errors, making the detection results more objective and accurate.

[0150] In one possible implementation, using a trained plasma spatiotemporal physics information neural network architecture to detect the electromagnetic field distribution of the plasma to be tested may include:

[0151] Each space-time point and time in the plasma to be tested is input into the trained plasma space-time domain physical information neural network architecture, and the electromagnetic field data corresponding to each space-time point and time in the plasma to be tested is output.

[0152] Optionally, the plasma space-time domain physical information neural network architecture trained in step 102 is used to input the space-time coordinates of each space-time point and time component in the plasma to be measured into the architecture, and the electromagnetic field data corresponding to each space-time coordinate is output.

[0153] Furthermore, the embodiments of the present application demonstrate excellent scalability and can be easily extended to plasma systems involving more complex physical processes, thereby enabling simultaneous, self-consistent detection of multiple key physical parameters of the plasma, far exceeding the capabilities of detecting only electromagnetic fields.

[0154] To achieve a comprehensive description of the plasma dynamics, in addition to the governing equations describing the electromagnetic field, the following plasma fluid equations can be introduced as additional physical constraints into the plasma space-time domain physical information neural network architecture, namely:

[0155] 1) Particle continuity equation: This describes the conservation of particle number density in plasma, taking into account particle convection, diffusion, and creation / annihilation processes. The formula is:

[0156]

[0157] in, is the particle number density, is the average particle velocity, It is a particle source term used to describe particle generation or loss, such as ionization and recombination. are different types of particles (for example, electrons e or ions i).

[0158] 2) Momentum equation: This describes the conservation of momentum of particles in plasma, taking into account the electromagnetic force (i.e., the Lorentz force), the pressure gradient force, and the collision resistance between particles. The formula is:

[0159]

[0160] in, is the particle mass, is the particle charge, is the particle pressure, is the collision term between different particle types, used to describe the momentum exchange.

[0161] Then, by adding the residual term of the above equation to the multi-objective loss function, the plasma spatiotemporal domain physics information neural network architecture will simultaneously learn solutions that satisfy all these coupled equations. This means that the embodiments of the present application can not only detect electromagnetic fields, but also simultaneously reconstruct and predict core parameters such as plasma density, temperature, and flow rate. This multi-physics field coupling detection capability is crucial for a deep understanding of plasma behavior, optimizing plasma devices, and achieving precise control of complex plasma processes.

[0162] The embodiments of the present application can have the following important values ​​in multiple plasma application fields, namely:

[0163] In the field of magnetic confinement fusion, this technology can be used to address key issues such as plasma equilibrium reconstruction, rupture prediction, and transport analysis in tokamaks and stellarators. By monitoring the plasma current distribution and magnetic field configuration in real time, it provides accurate feedback signals for plasma configuration control. Combined with machine learning predictive models, it can identify rupture precursors in advance and implement mitigation measures to prevent device damage.

[0164] In the field of plasma thrusters, this technology can be used to optimize the performance of devices such as Hall thrusters and ion thrusters. By detecting the electric field and plasma density distribution within the thruster channel, the magnetic field configuration design can be optimized to improve propulsion efficiency and specific impulse. Real-time monitoring of plasma oscillations and instabilities ensures the stable operation of the thruster.

[0165] In the field of plasma material processing, it can be used for process monitoring of plasma etching, deposition, surface modification, etc. By detecting the plasma sheath electric field and ion energy distribution, the material processing process can be precisely controlled to improve process consistency and yield.

[0166] In space plasma physics research, it can be used to analyze the plasma environment around satellites and spacecraft, reconstruct large-scale electromagnetic field structures using limited onboard detector data, and study important physical processes such as magnetosphere-ionosphere coupling and solar wind-magnetosphere interaction.

[0167] For example, two different processes are tested using the embodiments of the present application, as follows:

[0168] (1) Tokamak plasma equilibrium reconstruction

[0169] For example, a medium-sized tokamak has a maximum radius R = 1.5 m, a minimum radius a = 0.5 m, a toroidal magnetic field Bt = 2 T, and a plasma current Ip = 500 kA. The diagnostic system includes 96 magnetic probes distributed along the outer wall of the vacuum chamber to measure the poloidal magnetic field; 16 flux rings to measure the poloidal magnetic flux; and eight far-infrared interferometers to measure the electron density line integral.

[0170] Step 1: Data preprocessing. Digitally filter the magnetic probe signal to remove high-frequency noise. Normalize the measured data to the [-1, 1] interval:

[0171]

[0172] in, is the normalized magnetic induction intensity, is the minimum value of magnetic induction intensity, is the maximum value of magnetic induction intensity.

[0173] Step 2: Construct the plasma space-time domain physical information neural network architecture. The input is the cylindrical coordinates , the output is the three components of the magnetic vector potential The magnetic field passes The network uses 8 hidden layers, 256 neurons in each layer, and the Swish activation function.

[0174] Step 3: Define the region-adaptive physical constraints. Under the axisymmetric assumption (i.e. ), the plasma equilibrium satisfies the magnetohydrodynamic equation:

[0175]

[0176] in, is the poloidal flux function. Tokamak plasma has significant regional characteristics, such as the core, the susceptor region, and the scraping layer. To address this problem, the present embodiment constructs a regionally adaptive loss function.

[0177] First, during training, the poloidal flux function is predicted by the neural network itself. , dynamically divide the points into different areas. For example, the magnetic flux value of the last closed magnetic surface For the boundary, The points are divided into the core confinement area, The point is divided into scraping layer / vacuum area.

[0178] Then, the residuals of the magnetohydrodynamic equations in different regions are Assign different weights. In the core confinement area, the magnetohydrodynamic equations should be strictly satisfied, so high weights are given. ; In the scraping layer, different or lower weights can be given .Right now:

[0179]

[0180] in, is the total number of higher weights, is the total number of different or lower weights.

[0181] Step 4: Training process using gradient-aware adaptive sampling. Training is divided into two stages:

[0182] Phase 1: Pre-training for 1000 epochs using the balanced analytical solution with a learning rate of 1e-3.

[0183] Phase 2: Add measured data, start the gradient-aware adaptive sampling strategy, and train for 5000 rounds.

[0184] In each round or several rounds of training, recalculate the sampling weights of the points :

[0185]

[0186] The weight formula here automatically concentrates the sampling points in areas with drastic magnetic flux changes (such as the base area and near the magnetic axis). Importance sampling is then performed based on this weight, and the cosine annealing learning rate is reduced from 1e-3 to 1e-5.

[0187] Step 5: Verify the results. The trained plasma spatiotemporal physics neural network architecture can accurately reconstruct the magnetic field distribution across the entire poloidal cross-section from sparse magnetic probe data. Compared to traditional EFIT reconstruction results, this embodiment achieves higher reconstruction accuracy due to its enhanced resolution of high-gradient regions such as the base area, with a magnetic axis position deviation of less than 3 mm and a final closed magnetic surface deviation of less than 0.8 cm. A single reconstruction time of 15 ms meets real-time control requirements.

[0188] (2) Industrial plasma etching process monitoring

[0189] Wafer etching in an inductively coupled plasma reactor used in semiconductor manufacturing. Process parameters: RF power 500W, bias power 100W, operating pressure 20mTorr, CF4 / O2 gas mixture.

[0190] Step 1: Construct a plasma spatiotemporal physics information neural network architecture for etching physics. The etching process involves complex plasma physics and surface chemistry. This embodiment of the application constructs a multi-physics field plasma spatiotemporal physics information neural network architecture, with the input being the spatiotemporal coordinates. , the output is the key plasma parameters (electron density , electron temperature ,electric field ) and the density of various particles (ions, free radicals) Physical constraints Including particle balance equation, energy balance equation and Poisson equation, etc.

[0191] Step 2: Use regional adaptive loss and gradient-aware sampling for training. This step is used to solve the problem of physical differences between the sheath and the body in inductively coupled plasma. First, a regional adaptive loss function is constructed. According to the potential and electron density predicted by the model, the computational domain is dynamically divided into the sheath and the body. In the sheath, the residual weights of the Poisson equation are strengthened to accurately capture the electric field structure of the sheath. In the body, the residual weights of the particle balance and energy balance equations are strengthened. Secondly, gradient-aware adaptive sampling is used. Use the following sampling weight formula to tilt computing resources toward the sheath where the physical changes are most drastic:

[0192]

[0193] This strategy ensures that the plasma spatiotemporal physics information neural network architecture can resolve the sheath with extremely high resolution, which is crucial for accurately predicting the energy and angular distribution of ions bombarding the wafer.

[0194] Step 3: Deploy a digital twin system based on the plasma space-time physics information neural network architecture. Integrate the trained, high-precision plasma space-time physics information neural network architecture into the manufacturing execution system to achieve the following functions:

[0195] 1) Real-time state monitoring: By inputting real-time sensor data, the plasma spatiotemporal physical information neural network architecture can output a three-dimensional image of the plasma state of the entire chamber within milliseconds.

[0196] 2) Virtual Metrology: Predicting etch depth and profiles thanks to high-precision analysis of the sheath layer. A plasma spatiotemporal physics neural network architecture accurately predicts the ion flux and energy distribution reaching the wafer surface. Combined with a surface chemical reaction model, the system can non-invasively predict key dimensions such as etch depth and sidewall angle at any location on the wafer in real time, significantly reducing or replacing expensive offline SEM measurements.

[0197] 3) Predictive Maintenance: Predicting equipment maintenance needs based on plasma parameter trends. By continuously monitoring subtle drifts in plasma uniformity and abnormal changes in sheath thickness, and by leveraging deep physics information provided by the plasma spatiotemporal physics neural network architecture, the system can proactively predict the aging of equipment components or contamination of the cavity walls. When a predicted parameter trend exceeds a healthy baseline, the system automatically triggers a maintenance work order, transitioning from reactive maintenance to proactive predictive maintenance.

[0198] 4) Process optimization: Through this digital twin, the impact of different process parameter combinations on etching rate, uniformity and selectivity can be quickly evaluated, greatly shortening the R&D cycle of new processes.

[0199] The present application provides a plasma electromagnetic field detection method, which constructs a plasma electromagnetic field physical model and collects electromagnetic field data, wherein the electromagnetic field data includes the electric field intensity and magnetic induction intensity at each spacetime point and time in the plasma; constructs a plasma spacetime domain physical information neural network architecture, a multi-objective loss function and an adaptive sampling strategy, and trains the plasma spacetime domain physical information neural network architecture based on the multi-objective loss function, the adaptive sampling strategy and the electromagnetic field data; and uses the trained plasma spacetime domain physical information neural network architecture to detect the electromagnetic field distribution of the plasma to be tested. This application constructs a plasma spatiotemporal physical information neural network architecture to deeply integrate physical information with neural networks, so that the model can better understand and simulate the complex spatiotemporal changes of the plasma electromagnetic field, thereby improving the accuracy and reliability of detection; and, by introducing a multi-objective loss function and an adaptive sampling strategy, the neural network can simultaneously optimize multiple objectives during the training process and adaptively adjust the sampling strategy according to the characteristics of the plasma, thereby more effectively utilizing data resources and improving the training efficiency and generalization ability of the model; at the same time, by using the trained plasma spatiotemporal physical information neural network architecture, the electromagnetic field distribution of the plasma to be tested can be quickly and accurately detected, which not only improves the detection efficiency, but also reduces human intervention and errors, making the detection results more objective and accurate.

[0200] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0201] The following are device embodiments of the present application. For details not fully described therein, please refer to the corresponding method embodiments described above.

[0202] Figure 2 The following is a schematic diagram of the structure of the plasma electromagnetic field detection device provided in an embodiment of the present application. For ease of explanation, only the parts related to the embodiment of the present application are shown, which are detailed as follows:

[0203] like Figure 2 As shown, the plasma electromagnetic field detection device 2 includes:

[0204] Model building module 21, used to build a plasma electromagnetic field physical model and collect electromagnetic field data, the electromagnetic field data including the electric field intensity and magnetic induction intensity at each time and space point in the plasma;

[0205] An architecture construction and training module 22 is used to construct a plasma space-time domain physical information neural network architecture, a multi-objective loss function, and an adaptive sampling strategy, and to train the plasma space-time domain physical information neural network architecture based on the multi-objective loss function, the adaptive sampling strategy, and electromagnetic field data;

[0206] The electromagnetic field detection module 23 is used to detect the electromagnetic field distribution of the plasma to be tested using the trained plasma spatiotemporal physical information neural network architecture.

[0207] The present application provides a plasma electromagnetic field detection device, which collects electromagnetic field data by constructing a plasma electromagnetic field physical model, wherein the electromagnetic field data includes the electric field intensity and magnetic induction intensity at each spacetime point and time in the plasma; constructs a plasma spacetime domain physical information neural network architecture, a multi-objective loss function and an adaptive sampling strategy, and trains the plasma spacetime domain physical information neural network architecture based on the multi-objective loss function, the adaptive sampling strategy and the electromagnetic field data; and uses the trained plasma spacetime domain physical information neural network architecture to detect the electromagnetic field distribution of the plasma to be tested. This application constructs a plasma spatiotemporal physical information neural network architecture to deeply integrate physical information with neural networks, so that the model can better understand and simulate the complex spatiotemporal changes of the plasma electromagnetic field, thereby improving the accuracy and reliability of detection; and, by introducing a multi-objective loss function and an adaptive sampling strategy, the neural network can simultaneously optimize multiple objectives during the training process and adaptively adjust the sampling strategy according to the characteristics of the plasma, thereby more effectively utilizing data resources and improving the training efficiency and generalization ability of the model; at the same time, by using the trained plasma spatiotemporal physical information neural network architecture, the electromagnetic field distribution of the plasma to be tested can be quickly and accurately detected, which not only improves the detection efficiency, but also reduces human intervention and errors, making the detection results more objective and accurate.

[0208] In one possible implementation, the model building module can be used to:

[0209] The plasma electromagnetic field physics model is constructed using Maxwell's equations and plasma fluid equations. Maxwell's equations are used to describe the behavior of the electromagnetic field, and the plasma fluid equations are used to describe the macroscopic motion of the plasma.

[0210] In one possible implementation, the architecture building and training module can be used to:

[0211] Constructing a multi-scale feature extraction module, wherein the multi-scale feature extraction module includes a multi-scale convolution kernel and a Fourier feature map. The multi-scale convolution kernel is used to extract local patterns of spatiotemporal coordinates, and the Fourier feature map is used to map the spatiotemporal coordinates into a high-dimensional space. The spatiotemporal coordinates are composed of spatiotemporal points and time.

[0212] Using automatic differentiation technology, a physical constraint coding module is constructed, which is used to calculate the partial derivatives of electromagnetic field data with respect to space-time coordinates;

[0213] Constructing adaptive activation functions;

[0214] Residual connections and attention mechanisms are introduced to construct a plasma spatiotemporal domain physical information neural network architecture using a multi-scale feature extraction module, a physical constraint encoding module, and an adaptive activation function.

[0215] In one possible implementation, the architecture building and training module can also be used to:

[0216] Obtain data fitting loss, physical constraint loss, boundary condition loss and regularization term, and obtain a first weight, a second weight, a third weight and a fourth weight, wherein the first weight is the weight corresponding to the data fitting loss, the second weight is the weight corresponding to the physical constraint loss, the third weight is the weight corresponding to the boundary condition loss, and the fourth weight is the weight corresponding to the regularization term;

[0217] The product of the data fitting loss and the first weight, the product of the physical constraint loss and the second weight, the product of the boundary condition loss and the third weight, and the product of the regularization term and the fourth weight are added together to construct a plasma region adaptive multi-objective loss function.

[0218] In one possible implementation, the architecture building and training module can also be used to:

[0219] Using a uniform random sampling method, N initial collocation points are selected from all the time and space domains in the plasma electromagnetic field, where N is a positive integer greater than or equal to 1;

[0220] Calculate the sampling weight of each initial collocation point, and use the sampling weight of each initial collocation point to calculate the sampling probability of the corresponding initial collocation point. The sampling weight includes the physical residual and the physical field gradient.

[0221] The sampling probability of each initial matching point is used to determine the matching point set for training the plasma spatiotemporal physical information neural network architecture.

[0222] In one possible implementation, the architecture building and training module can also be used to:

[0223] Sequentially pre-train and fine-tune the plasma spatiotemporal physics information neural network architecture;

[0224] Based on the multi-objective loss function, the multi-objective loss function of the finely tuned plasma time-space domain physical information neural network architecture is calculated, and the multi-objective loss function of the finely tuned plasma time-space domain physical information neural network architecture is used as a joint loss function;

[0225] Using the adaptive sampling strategy, the weights of each loss term in the joint loss function are calculated, and the total loss weight is calculated using the weights of each loss term;

[0226] Determine whether the total loss weight converges;

[0227] If the total loss weight converges, it is determined that the training of the plasma spatiotemporal domain physical information neural network architecture is completed;

[0228] If the total loss weight does not converge, the process returns to the steps of pre-training and fine-tuning the plasma spatiotemporal physical information neural network architecture.

[0229] In one possible implementation, the electromagnetic field detection module may be used to:

[0230] Each space-time point and time in the plasma to be tested is input into the trained plasma space-time domain physical information neural network architecture, and the electromagnetic field data corresponding to each space-time point and time in the plasma to be tested is output.

[0231] Figure 3 Schematic diagram of the terminal provided in the embodiment of the present application. Figure 3 As shown, the terminal 3 of this embodiment includes: a processor 30, a memory 31, and a computer program 32 stored in the memory 31 and executable on the processor 30. When the processor 30 executes the computer program 32, the steps in the above-mentioned embodiments of the plasma electromagnetic field detection method are implemented, for example Figure 1 Alternatively, when the processor 30 executes the computer program 32, the functions of the modules / units in the above-mentioned device embodiments are realized, for example, Figure 2 The functions of each module are shown.

[0232] Exemplarily, the computer program 32 may be divided into one or more modules / units, which are stored in the memory 31 and executed by the processor 30 to complete the present application. The one or more modules / units may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program 32 in the terminal 3. For example, the computer program 32 may be divided into Figure 2 The modules shown.

[0233] The terminal 3 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The terminal 3 may include, but is not limited to, a processor 30 and a memory 31. Those skilled in the art will understand that Figure 3 It is only an example of terminal 3 and does not constitute a limitation on terminal 3. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the terminal may also include input and output devices, network access devices, buses, etc.

[0234] The processor 30 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0235] The memory 31 can be an internal storage unit of the terminal 3, such as a hard drive or memory of the terminal 3. The memory 31 can also be an external storage device of the terminal 3, such as a plug-in hard drive, a Smart Media Card (SMC), a Secure Digital (SD) card, a flash memory card, etc. equipped with the terminal 3. Furthermore, the memory 31 can include both the internal storage unit of the terminal 3 and an external storage device. The memory 31 is used to store the computer program and other programs and data required by the terminal. The memory 31 can also be used to temporarily store data that has been output or is about to be output.

[0236] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0237] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0238] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0239] In the embodiments provided in this application, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For example, the division of the modules or units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0240] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0241] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0242] If the integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above-mentioned various plasma electromagnetic field detection method embodiments. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practices in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practices, computer-readable media does not include electrical carrier signals and telecommunication signals.

[0243] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A plasma electromagnetic field detection method, characterized in that: include: Constructing a plasma electromagnetic field physical model and collecting electromagnetic field data, wherein the electromagnetic field data includes the electric field intensity and magnetic induction intensity at each time and space point in the plasma; Constructing a plasma space-time domain physical information neural network architecture, a multi-objective loss function, and an adaptive sampling strategy, and training the plasma space-time domain physical information neural network architecture based on the multi-objective loss function, the adaptive sampling strategy, and the electromagnetic field data; The electromagnetic field distribution of the plasma to be tested is detected using the trained plasma spatiotemporal physics information neural network architecture; The construction of a plasma spatiotemporal physical information neural network architecture includes: Constructing a multi-scale feature extraction module, wherein the multi-scale feature extraction module includes a multi-scale convolution kernel and a Fourier feature map, the multi-scale convolution kernel is used to extract local patterns of spatiotemporal coordinates, and the Fourier feature map is used to map the spatiotemporal coordinates to a high-dimensional space, the spatiotemporal coordinates consisting of spatiotemporal points and time; Using automatic differentiation technology, constructing a physical constraint coding module, wherein the physical constraint coding module is used to calculate the partial derivatives of the electromagnetic field data with respect to the space-time coordinates; Constructing adaptive activation functions; Introducing residual connections and an attention mechanism, the multi-scale feature extraction module, the physical constraint encoding module, and the adaptive activation function are constructed into the plasma spatiotemporal domain physical information neural network architecture; Among them, constructing a multi-objective loss function includes: Obtaining data fitting loss, physical constraint loss, boundary condition loss and regularization term, and obtaining a first weight, a second weight, a third weight and a fourth weight, wherein the first weight is the weight corresponding to the data fitting loss, the second weight is the weight corresponding to the physical constraint loss, the third weight is the weight corresponding to the boundary condition loss, and the fourth weight is the weight corresponding to the regularization term; the physical constraint loss is to dynamically divide the points in the calculation domain into a core high temperature region, an edge collision region and a sheath strong electric field region according to the normalized magnetic pressure, collision frequency and Hall parameter of the plasma, and apply different weights to the physical residuals of the core high temperature region, the edge collision region and the sheath strong electric field region respectively; Adding the product of the data fitting loss and the first weight, the product of the physical constraint loss and the second weight, the product of the boundary condition loss and the third weight, and the product of the regularization term and the fourth weight to construct the plasma region adaptive multi-objective loss function; Among them, building an adaptive sampling strategy includes: Using a uniform random sampling method, N initial collocation points are selected from all the time and space domains in the plasma electromagnetic field, where N is a positive integer greater than or equal to 1; Calculating a sampling weight for each initial collocation point, and using the sampling weight for each initial collocation point to calculate a sampling probability for the corresponding initial collocation point, wherein the sampling weight includes a physical residual and a physical field gradient; The sampling probability of each initial collocation point is used to determine the collocation point set for training the plasma spatiotemporal domain physical information neural network architecture.

2. The plasma electromagnetic field detection method according to claim 1, characterized in that: The construction of the plasma electromagnetic field physical model includes: The plasma electromagnetic field physical model is constructed using Maxwell's equations and plasma fluid equations. The Maxwell's equations are used to describe the electromagnetic field behavior, and the plasma fluid equations are used to describe the plasma macroscopic motion.

3. The plasma electromagnetic field detection method according to claim 1, characterized in that: The training of the plasma spatiotemporal domain physical information neural network architecture based on the multi-objective loss function, the adaptive sampling strategy, and the electromagnetic field data includes: Pre-training and fine-tuning the plasma spatiotemporal domain physical information neural network architecture in sequence; Based on the multi-objective loss function, calculating the multi-objective loss function of the finely adjusted plasma time-space domain physical information neural network architecture, and using the multi-objective loss function of the finely adjusted plasma time-space domain physical information neural network architecture as a joint loss function; Utilizing the adaptive sampling strategy, calculating the weight of each loss term in the joint loss function, and calculating the total loss weight using the weight of each loss term; Determining whether the total loss weight converges; If the total loss weight converges, it is determined that the plasma spatiotemporal domain physical information neural network architecture training is completed; If the total loss weight has not converged, the process returns to the steps of pre-training and fine-tuning the plasma spatiotemporal physical information neural network architecture in sequence and continues to execute.

4. The plasma electromagnetic field detection method according to claim 1, characterized in that: The method of detecting the electromagnetic field distribution of the plasma to be tested by using the trained plasma spatiotemporal physical information neural network architecture includes: Each time-space point and time in the plasma to be measured is input into the trained plasma time-space domain physical information neural network architecture, and the electromagnetic field data corresponding to each time-space point and time in the plasma to be measured is output.

5. A plasma electromagnetic field detection device, characterized in that: include: A model building module is used to build a plasma electromagnetic field physical model and collect electromagnetic field data, wherein the electromagnetic field data includes the electric field intensity and magnetic induction intensity at each time and space point in the plasma; An architecture construction and training module is used to construct a plasma space-time domain physical information neural network architecture, a multi-objective loss function and an adaptive sampling strategy, and train the plasma space-time domain physical information neural network architecture based on the multi-objective loss function, the adaptive sampling strategy and the electromagnetic field data; An electromagnetic field detection module is used to detect the electromagnetic field distribution of the plasma to be tested using a trained plasma spatiotemporal physical information neural network architecture; The architecture construction and training module is used to: Constructing a multi-scale feature extraction module, wherein the multi-scale feature extraction module includes a multi-scale convolution kernel and a Fourier feature map, the multi-scale convolution kernel is used to extract local patterns of spatiotemporal coordinates, and the Fourier feature map is used to map the spatiotemporal coordinates to a high-dimensional space, the spatiotemporal coordinates consisting of spatiotemporal points and time; Using automatic differentiation technology, constructing a physical constraint coding module, wherein the physical constraint coding module is used to calculate the partial derivatives of the electromagnetic field data with respect to the space-time coordinates; Constructing adaptive activation functions; Introducing residual connections and an attention mechanism, the multi-scale feature extraction module, the physical constraint encoding module, and the adaptive activation function are constructed into the plasma spatiotemporal domain physical information neural network architecture; The architecture construction and training module is used to: Obtaining data fitting loss, physical constraint loss, boundary condition loss and regularization term, and obtaining a first weight, a second weight, a third weight and a fourth weight, wherein the first weight is the weight corresponding to the data fitting loss, the second weight is the weight corresponding to the physical constraint loss, the third weight is the weight corresponding to the boundary condition loss, and the fourth weight is the weight corresponding to the regularization term; the physical constraint loss is to dynamically divide the points in the calculation domain into a core high temperature region, an edge collision region and a sheath strong electric field region according to the normalized magnetic pressure, collision frequency and Hall parameter of the plasma, and apply different weights to the physical residuals of the core high temperature region, the edge collision region and the sheath strong electric field region respectively; Adding the product of the data fitting loss and the first weight, the product of the physical constraint loss and the second weight, the product of the boundary condition loss and the third weight, and the product of the regularization term and the fourth weight to construct the plasma region adaptive multi-objective loss function; The architecture construction and training module is used to: Using a uniform random sampling method, N initial collocation points are selected from all the time and space domains in the plasma electromagnetic field, where N is a positive integer greater than or equal to 1; Calculating a sampling weight for each initial collocation point, and using the sampling weight for each initial collocation point to calculate a sampling probability for the corresponding initial collocation point, wherein the sampling weight includes a physical residual and a physical field gradient; The sampling probability of each initial collocation point is used to determine the collocation point set for training the plasma spatiotemporal domain physical information neural network architecture.

6. A terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the plasma electromagnetic field detection method according to any one of claims 1 to 4 are implemented.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the plasma electromagnetic field detection method according to any one of claims 1 to 4 are implemented.

Images