Explosion shock wave parameter prediction method and system based on physical information neural network

Abstract

The invention discloses an explosion shock wave parameter prediction method and system based on a physical information neural network, and the method comprises the steps: obtaining an explosion shock wave overpressure data set of a thermobaric explosive in an open space through refined numerical simulation, and dividing the overpressure data set into a training data set and a verification data set; a specific scene is used for solving an N-S equation as a physical problem, and a mass conservation equation and momentum conservation equations in the X direction and the Y direction are involved. A neural network with no less than four hidden layers and no less than 100 neurons in each layer is constructed, the hidden layers adopt a Tanh function, an output layer uses a Softplus function to ensure non-negativity for shock wave pressure P and density rho, and speeds U and V are unconstrained. The loss function comprises a data error term and a physical information error term, the total loss function is formed by adding the data error term and the physical information error term according to weights, and specific scene boundary conditions are set. A prediction model is established based on a training result, and especially in a complex fluid scene, the method can rapidly and accurately predict an explosion shock wave overpressure field and a distribution rule.

Description

Technical Field

[0001] The present invention belongs to the field of explosion shock calculation and evaluation, and more specifically, relates to a method and system for predicting explosion shock wave parameters based on a physical information neural network. Background Art

[0002] The prediction of explosion shock wave parameters plays a key role in explosion safety assessment and weapon effectiveness analysis, and is an indispensable part of related research and practice. Currently, traditional methods for predicting explosion shock wave parameters mainly rely on experimental measurements and computational fluid dynamics (CFD) numerical simulations.

[0003] While experimental measurement methods can obtain relatively accurate data, they have significant limitations. For one thing, they are expensive, requiring significant investment in both the purchase of experimental equipment, the preparation of the experimental site, and the various expenses involved in the experiment. Furthermore, when faced with complex scenarios, experimental measurements struggle to replicate specific explosion conditions. This results in the inability to obtain valid data under certain special or extreme conditions, hindering in-depth research into the characteristics of explosion shock waves in complex environments.

[0004] CFD methods based on the Navier-Stokes (NS) equations, such as the finite volume method and the finite element method, can theoretically provide high-precision solutions. However, this method faces the problem of extremely low computational efficiency in practical applications. In particular, when dealing with transient explosion flow fields, complex meshing is required, a process that is not only tedious but also extremely time-consuming. At the same time, the amount of computation required to solve the equations is also very large, making the entire calculation process time-consuming and difficult to meet the demand for real-time prediction in practical applications. In scenarios where prediction results need to be obtained quickly, the practicality of this method is greatly reduced.

[0005] In recent years, with the development of deep learning technology, purely data-driven deep learning models, such as CNNs (convolutional neural networks) and RNNs (recurrent neural networks), have been explored for application in shock wave prediction. By learning from large amounts of data, these models can predict shock wave parameters to a certain extent. However, they rely solely on data fitting and lack physical constraints. This leads to physical inconsistencies in the prediction results, such as mass / momentum non-conservation, under complex boundary conditions or in scenarios not covered by the training data. These physical inconsistencies significantly limit the model's generalization capabilities, making it impossible to reliably provide accurate predictions in actual engineering applications, faced with diverse operating conditions and complex environments.

[0006] In summary, the existing explosion shock wave prediction methods are mainly divided into two categories: traditional numerical simulation and pure data-driven model, but they all have significant limitations in computational efficiency, physical rationality and engineering applicability. Taking the adaptive mesh CFD method as an example, this method solves the three-dimensional N-S equation by coupling the Level-Set interface tracking technology, although it can ensure the strictness of the physical law, but the single simulation takes several hours to several days, and the parameters such as explosive equivalent and position need to be fixed in advance, it is difficult to realize the rapid prediction and dynamic visualization of multiple working conditions, and it cannot meet the demand of rapid analysis of multiple working conditions in actual engineering. While the pure data-driven model using convolution LSTM network, although the prediction time is compressed to minutes, it is heavily dependent on massive training data (10 4 ~10 5 groups of data), and cannot avoid physical contradictions, and is prone to problems such as abnormal peak overpressure, wave speed distortion, etc. under complex boundary conditions or untrained working conditions, affecting the accuracy and reliability of the prediction results. In addition, these two methods also have limitations in output form. Traditional CFD can only output the pressure field of the preset working condition, while the data-driven model focuses on the time history prediction of the overpressure at fixed measuring points, and neither of them can meet the joint demand of time history curve and full-field pressure contour at any position in the engineering scene, making it difficult to fully and accurately reflect the characteristics of explosion shock wave at different positions and times. SUMMARY

[0007] The purpose of the present application is to provide a machine learning model combining deep learning and physical knowledge for predicting the overpressure distribution of open space explosive explosion shock wave, which should have the characteristics of high accuracy and high efficiency. Through this model and method, the significant limitations of existing explosion shock wave prediction methods in computational efficiency, physical rationality and engineering applicability can be effectively solved, and the joint demand of time history curve and full-field pressure contour at any position in the engineering scene can be met

[0008] In view of the above defects or improvement needs of the prior art, as a first aspect of the present application, the present application provides an explosion shock wave parameter prediction method based on physical information neural network, comprising:

[0009] S1. Obtain the overpressure data set of open space temperature and pressure explosive explosion shock wave through numerical simulation results, and divide the data set into two parts: training data set and verification data set;

[0010] S2. Solve the N-S equation as a physical problem in a specific scenario, involving the mass conservation equation, the momentum conservation equation in X direction and Y direction;

[0011] S3. According to the input parameters and output parameter characteristics, a neural network with not less than 4 layers of hidden layers and not less than 100 neurons in each layer is constructed; the hidden layer uses a Tanh function as an activation function, which is beneficial to automatic differentiation calculation of high-order derivatives of the N-S equation; the output layer uses a Softplus function to ensure non-negativity for the shock wave pressure P and density p, and has no constraint for the speed U and V;

[0012] S4. The loss function includes a data error term and a physical information error term; the data error term adopts a square error form to ensure that the neural network output is consistent with the real observation or measurement data; the physical information error term covers the momentum residual equation error, the X-direction mass residual equation and the Y-direction mass residual equation, so that the neural network meets the physical equation constraint; the total loss function is composed of the data error term and the physical information error term by adding them according to certain weights; and specific scene boundary conditions are set;

[0013] S5. The training result obtained based on the training model is used for an explosion shock wave prediction model to predict the P-t variation curve of any point in the open space explosion range and the two-dimensional pressure cloud diagram of the explosion range at any time.

[0014] Further, the S1 obtains the open space temperature and pressure explosive explosion shock wave overpressure data set through numerical simulation results, and the specific process is:

[0015] Through numerical simulation, the explosion process of temperature and pressure explosive in open space is simulated, and the explosion shock wave flow field change is recorded, including: explosion shock wave overpressure-time curve (ΔP-t), shock wave speed-time curve (U-t, V-t) and shock wave air density-time curve (p-t) at different distances from the explosion center; and the data set in the recorded database is normalized.

[0016] Further, the S2 solves the N-S equation as a physical problem in a specific scene, which involves the mass conservation equation, the X-direction and Y-direction momentum conservation equations, and specifically:

[0017] The physical problem is defined as solving the N-S equation in a specific scene:

[0018]

[0019] The continuity equation, i.e. the mass conservation equation:

[0020]

[0021] The X-direction momentum conservation equation:

[0022]

[0023] The Y-direction momentum conservation equation:

[0024]

[0025] where p represents the fluid density of the shock wave; u and v are the components of the fluid velocity of the shock wave in the x and y directions of the rectangular coordinate system, respectively; and P is the fluid pressure.

[0026] Further, the input parameters in S3 include:

[0027] two-dimensional coordinates (x, y), time t, and explosive mass m; and the input parameters are subjected to maximum-minimum normalization (range scaling [-1, 1]) to align with the output range of Tanh.

[0028] Further, the output parameters in S3 include:

[0029] explosive density p, overpressure p, X-direction wave speed u, and Y-direction wave speed v.

[0030] Further, the specific calculation method of the data error term in S4 is:

[0031] When there are real observation data or measurement data, the data error term is used to measure the difference between the neural network output and the real data; the data error term adopts a square error form:

[0032]

[0033] Mathematical expansion form of the data error term:

[0034]

[0035] where N: the number of training data points, p pred , p pred , u pred , v pred : density, pressure, x / y direction velocity predicted by the neural network; p true , p true , u true , v true : real values provided by experiments or CFD simulation.

[0036] Further, the specific calculation method of the physical information error term in S4 is:

[0037] The physical information error term includes a mass residual equation, an X-direction momentum residual equation, and a Y-direction momentum residual equation.

[0038] Mass residual equation:

[0039]

[0040] X-direction momentum residual equation:

[0041]

[0042] Y direction momentum residual equation:

[0043]

[0044] The total physical information error term is:

[0045]

[0046] By minimizing the total physical information error term, the output of the neural network is closer to the solution of the physical equation.

[0047] Further, the specific calculation method of the total loss function in S4 is:

[0048] L total = λ data · L data + λ phys · L phys

[0049] In the formula, λ data , λ phys are the weights of the data error term and the physical information error term, respectively.

[0050] Further, the specific scene boundary condition in S4 is:

[0051] Since the ground is a solid reflective boundary, the normal velocity is 0, the tangential velocity is consistent with the wall velocity, the shock wave pressure and the density do not need to be constrained, and the N-S continuity equation is naturally satisfied, so the boundary condition is set as:

[0052]

[0053] In the formula, u and v are the components of the shock wave fluid velocity in the x and y directions of the rectangular coordinate system, respectively; ρ is the density of the shock wave fluid; and p is the fluid pressure.

[0054] As a second aspect of the application, a physical information neural network-based explosion shock wave parameter prediction system is provided, comprising:

[0055] The explosion overpressure data set unit is used to obtain the open space temperature and pressure explosive explosion shock wave overpressure data set through numerical simulation results, and the data set is divided into two parts: a training data set and a verification data set;

[0056] The physical problem and equation unit is used to solve the N-S equation as a physical problem in a specific scene, involving the mass conservation equation, the X direction and Y direction momentum conservation equation;

[0057] A special neural network unit is built to construct a neural network with no less than 4 hidden layers and no less than 100 neurons in each layer according to input parameters and output parameter characteristics; the hidden layer adopts a Tanh function as an activation function, which is beneficial to automatic differentiation calculation of high-order derivatives of N-S equations; the output layer uses a Softplus function to ensure non-negativity for shock wave pressure P and density p, and has no constraints for speed U and V;

[0058] A loss function and boundary condition unit is designed, wherein the loss function includes a data error term and a physical information error term; the data error term adopts a square error form to ensure that the neural network output is consistent with the real observation or measurement data; the physical information error term covers momentum residual equation error, X-direction mass residual equation and Y-direction mass residual equation, so that the neural network meets the physical equation constraint; the total loss function is composed of the data error term and the physical information error term added according to certain weights; and specific scene boundary conditions are set;

[0059] An explosion parameter prediction unit is used to obtain training results based on a trained model, and is used for an explosion shock wave prediction model to predict P-t variation curves of any point in an open space explosion range and two-dimensional pressure cloud diagrams of the explosion range at any time.

[0060] Overall, compared with the prior art, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0061] 1. The explosion shock wave parameter prediction method based on the physical information neural network of the present application effectively improves the prediction accuracy and physical rationality by integrating physical laws into the neural network to construct a loss function. In traditional methods, pure data-driven models lack physical constraints, resulting in physical contradictions in the prediction results under complex scenarios. However, the present application combines the mass conservation equation and the momentum conservation equation to construct a physically constrained loss function, which not only makes the neural network fit the data, but also makes it follow the physical laws during model training. In this way, when facing complex fluid scenarios, this method can accurately predict the explosion shock wave overpressure distribution, greatly improving the accuracy and reliability of the prediction solution, and providing a more reliable basis for explosion safety evaluation and weapon effectiveness analysis.

[0062] 2. The physical information neural network-based explosion shock wave parameter prediction method of the present application realizes an efficient calculation process and ensures the rationality of the output through the design of a specific structure of neural network. A neural network with not less than 4 hidden layers and not less than 100 neurons in each layer is constructed, the hidden layer adopts a Tanh function, the derivative of which is smooth, which is beneficial to automatic differentiation calculation of high-order derivatives of N-S equation and accelerates model training. Different processing methods are adopted for different physical quantities in the output layer, such as using a Softplus function to ensure non-negativity for shock wave pressure P and density p, and unconstrained for speed U and V. Compared with the finite element / finite volume simulation model, the method only needs to input the coordinate points and the mass of the explosive, and the shock wave pressure-time variation curve can be quickly obtained, which greatly reduces the calculation time while ensuring the calculation accuracy, and significantly improves the calculation efficiency.

[0063] 3. The physical information neural network-based explosion shock wave parameter prediction method of the present application balances data fitting and physical constraint loss through a hybrid training strategy, and enhances the generalization ability of the model. In the training process, the total loss function is composed of data error term and physical information error term added according to a certain weight. This way dynamically balances the data fitting loss and the physical constraint loss, so that the model can not only learn the characteristics of the existing data, but also meet the constraints of the physical equation. The trained model has strong generalization ability and can quickly learn different equivalent, different distances and other working conditions, realize efficient and high-precision prediction of two-dimensional explosion overpressure distribution in open space, and the calculation results strictly meet the physical law, effectively expanding the application range of the method in different explosion scenes. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 A physical information neural network-based explosion shock wave parameter prediction method flowchart of an embodiment of the present application;

[0065] Figure 2 A calculation step schematic diagram of an embodiment of the present application;

[0066] Figure 3 A neural network structure schematic diagram of an embodiment of the present application;

[0067] Figure 4 A comparison schematic diagram of shock wave overpressure time history curve of any point of an embodiment of the present application;

[0068] Figure 5 A shock wave pressure cloud chart in the explosion range at any time of an embodiment of the present application;

[0069] Figure 6 A system unit diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0070] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not used to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0071] Embodiment 1

[0072] Please refer to Figure 1 The embodiment 1 provides a physical information neural network based explosion shock wave parameter prediction method, comprising

[0073] S1. Obtain the open space temperature and pressure explosive explosion shock wave overpressure data set through numerical simulation results, and divide the data set into two parts: training data set and validation data set;

[0074] S2. The specific scene solving N-S equation is used as a physical problem, which involves the mass conservation equation, the momentum conservation equation in X direction and Y direction;

[0075] S3. According to the characteristics of input parameters and output parameters, a neural network with not less than 4 layers of hidden layers and not less than 100 neurons in each layer is constructed; the hidden layer uses Tanh function as the activation function, which is beneficial to automatic differentiation calculation of high order derivative of N-S equation; the output layer uses Softplus function to ensure non-negativity for shock wave pressure P and density p, and has no constraint for speed U and V;

[0076] S4. The loss function includes data error term and physical information error term; the data error term adopts square error form to ensure that the neural network output is consistent with the real observation or measurement data; the physical information error term covers the momentum residual equation error, X direction mass residual equation and Y direction mass residual equation, so that the neural network meets the physical equation constraint; the total loss function is composed of the data error term and the physical information error term by adding a certain weight; and the specific scene boundary conditions are set;

[0077] S5. Based on the training results obtained by the training model, the explosion shock wave prediction model is used to predict the P-t variation curve of any point in the open space explosion range and the two-dimensional pressure cloud picture of the explosion range at any time.

[0078] Please refer to Figure 2 The embodiment 1 further expands the description of the above method steps.

[0079] (1) Construction of explosion shock wave overpressure data set

[0080] First, the numerical simulation technology is used to simulate the explosion process of warm-pressed explosive in an open space. During the simulation, the changes of the explosion shock wave flow field are recorded comprehensively, and the overpressure-time curve (ΔP-t), the shock wave speed-time curve (U-t, V-t), and the shock wave air density-time curve (p-t) at different distances from the explosion center are obtained in detail. After completing the data recording, the processed data is divided into training data set and verification data set according to a certain proportion, the training data set is used for subsequent neural network model training, and the verification data set is used for accuracy and reliability verification of the trained model.

[0081] (2) Definition of physical problem and related equations

[0082] In order to accurately describe the physical process of explosion shock wave, the embodiment takes solving N-S equation in a specific scenario as the core physical problem. N-S equation is a motion equation for describing momentum conservation of viscous Newtonian fluid, which is of great significance in the study of explosion shock wave. The equation involves several important sub-equations, including continuity equation (i.e. mass conservation equation) and momentum conservation equations in X and Y directions.

[0083] Through these equations, the internal relationship between the parameters such as shock wave fluid density p, fluid velocity components u and v in x and y directions of rectangular coordinate system, and fluid pressure P is determined, which provides a solid physical theoretical basis for subsequent model construction.

[0084] In a preferred embodiment, the physical problem is defined as solving N-S equation in a specific scenario, specifically:

[0085]

[0086] Continuity equation, i.e. mass conservation equation:

[0087]

[0088] It reflects that in the process of explosion shock wave propagation, the fluid mass neither appears nor disappears, and the rate of change of mass in any small volume is balanced with the mass flux through the boundary of the volume. This means that under the action of explosion shock wave, although the density, velocity and other parameters of the fluid will change with time and space, the total mass remains constant.

[0089] Momentum conservation equation in X direction:

[0090]

[0091] This equation describes the rate of change of momentum of a unit volume of fluid in the X direction over time, equal to the combined force of pressure gradient, viscous force, and other external forces acting on that volume. It embodies the law of momentum change in the X direction during the propagation of the blast shock wave, which is crucial for understanding the propagation and mechanism of the shock wave in the horizontal direction.

[0092] Momentum conservation equation in Y direction:

[0093]

[0094] Similar to the momentum conservation equation in the X direction, it describes the change of momentum in the Y direction, taking into account the influence of pressure, viscous force, etc. in the Y direction on the fluid momentum, providing a theoretical basis for studying the characteristics of the blast shock wave in the vertical direction.

[0095] Through these equations, the internal relationship between the fluid density ρ, the components of fluid velocity u and v in the x and y directions of the Cartesian coordinate system, and the fluid pressure P, etc. is clearly defined. These equations provide a solid physical and theoretical basis for the construction of subsequent neural network models, ensuring that the model can reflect the real physical process of the blast shock wave.

[0096] (3) Build a special neural network

[0097] Please refer to Figure 3 , according to the determined input parameters and output parameter characteristics to build a neural network.

[0098] The shock wave overpressure data set of the training set contains multiple important parameters, among which the two-dimensional coordinates (x, y), time t, and explosive mass m are input parameters, representing the spatial location, time of explosion, and basic properties of the explosive. These parameters are substituted into the above N-S equation to provide physical background information for the model. The explosive density ρ, overpressure p, X direction wave speed u, and Y direction wave speed v are output parameters, which are the target quantities that the model needs to predict.

[0099] Input layer: normalization processing, unified to international units to avoid dimensional differences affecting training, and maximum-minimum normalization (range scaling [-1, 1]) to align with the output range of Tanh and accelerate convergence.

[0100] Before the data enters the neural network, the input parameters need to be preprocessed. First, all parameters are unified to international units to eliminate the interference of dimensional differences on training. Then, the maximum-minimum normalization method is used to process the input, scaling its range to [-1, 1], which matches the output range of the Tanh function used in the subsequent hidden layer, helping to accelerate the convergence speed of the neural network and improve the training efficiency.

[0101] Hidden layer: The neural network constructed in this embodiment has 4 hidden layers, with 100 neurons in each layer. The hidden layer uses the Tanh function as the activation function, which has good characteristics. Its output range is [-1, 1], and the function graph is symmetrical about the origin. This symmetry makes it perform well when processing smooth physical fields. In addition, the derivative of the Tanh function is smooth, and its derivative formula is Tanh'(z) = 1-Tanh 2 (z), this feature facilitates automatic differentiation to calculate higher-order derivatives of the NS equations. During neural network training, automatic differentiation can efficiently calculate gradients, helping the model converge to the optimal solution more quickly, thereby better learning the complex relationship between input and output parameters.

[0102]

[0103] Output layer: Considering the physical properties of shock wave pressure P and density ρ, it is necessary to ensure that their values ​​are non-negative. Therefore, the Softplus function is used to process these two parameters in the output layer. The expression of the Softplus function is Softplus(z)=ln(1+e z ), which can map any real number to the non-negative real number domain. As for the velocities U and V, since they have no non-negative restrictions in the physical sense, no special constraints are imposed and the calculation results of the neural network are directly output.

[0104] (4) Design loss function and boundary conditions

[0105] The design of the loss function directly affects the training effect and prediction accuracy of the neural network. The loss function constructed in this embodiment consists of a data error term and a physical information error term.

[0106] Data error term: When there is real observation data or measurement data, the data error term is used to measure the difference between the neural network output and the real data. This embodiment uses the square error form to calculate the data error term:

[0107]

[0108] The mathematical expansion form of the data error term is:

[0109]

[0110] Where: N: the number of training data points, ρ pred 、p pred 、u pred 、v pred : density, pressure, and x / y direction velocity predicted by the neural network; ρ true 、p true 、u true 、vtrue : Real values ​​provided by experiments or CFD simulations.

[0111] Physical information error terms: Ideally, the solutions output by a neural network should fully satisfy the physical equations. However, in actual training, errors often exist due to factors such as model complexity and data noise. To ensure that the output of the neural network is more consistent with physical laws, this embodiment introduces physical information error terms. These include the mass residual equation, the X-direction momentum residual equation, and the Y-direction momentum residual equation.

[0112] Mass residual equation:

[0113]

[0114] The residual equation of momentum in the X direction is:

[0115]

[0116] The residual equation of momentum in the Y direction is:

[0117]

[0118] The total physical information error term is:

[0119]

[0120] By minimizing the total physical information error term, the output of the neural network can be made closer to the solution of the physical equation.

[0121] The physical information error term is added to the total loss function of PINN and minimized together with the data error or boundary conditions, initial conditions, etc. Its purpose is to make the neural network not only fit the data but also satisfy the constraints of the physical equations. The total loss function term is:

[0122] L total =λ data ·L data +λ phys ·L phys

[0123] Where λ data ,λ phys are the weights of the data error term and the physical information error term respectively.

[0124] Scenario-Specific Boundary Conditions: In actual explosion scenarios, the ground acts as a solid reflective boundary with unique physical properties. The normal velocity is 0, meaning that the fluid velocity perpendicular to the ground is zero at the boundary; the tangential velocity is consistent with the wall velocity, ensuring fluid continuity at the boundary. The shock wave pressure and density require no additional constraints, as they naturally satisfy the NS continuity equation. Therefore, the boundary conditions are set as follows:

[0125]

[0126] where u and v are the components of the shock wave fluid velocity in the x and y directions of the Cartesian coordinate system, respectively; p is the fluid density; and p is the fluid pressure.

[0127] The setting of these boundary conditions enables the model to more accurately simulate actual explosion scenarios, improving the accuracy of the predictions.

[0128] (5) Performing explosion parameter prediction

[0129] Based on the training results obtained from the previous model training, an explosion shock wave prediction model is constructed. This prediction model integrates the relationships between explosion shock wave parameters and physical laws learned during the training process. It can predict the pressure-time curve (P-t curve) of any point in the open space explosion range over time. By inputting the coordinates (x, y) and time t of the point, the model can output the pressure value of the point at different times, thereby drawing a complete P-t curve.

[0130] At the same time, the model can also generate a two-dimensional pressure cloud map of the explosion range at any time. By calculating the pressure of each point in the explosion range, the pressure values are presented in the form of an image, which intuitively shows the pressure distribution of the explosion shock wave at a certain time. The pressure cloud map can clearly show the high and low distribution areas of the pressure, as well as the propagation range and intensity changes of the shock wave, providing comprehensive and intuitive data support for the analysis of explosion shock wave parameters.

[0131] Please refer to Figure 4 and Figure 5 , Figure 4 for the comparison of the shock wave overpressure time curve for any point, Figure 5 for the shock wave pressure cloud map in the explosion range at any time. Comparing the prediction results with the validation data set, it can be found that the prediction curve is highly consistent with the true data curve, and the pressure cloud map can accurately reflect the pressure distribution characteristics of the explosion shock wave. This indicates that the explosion shock wave parameter prediction method based on physical information neural network of the present application has high accuracy and can provide reliable technical support for explosion safety evaluation, weapon effectiveness analysis, and other fields.

[0132] Example 2

[0133] Please refer to Figure 6 , this embodiment 2 provides a kind of explosion shock wave parameter prediction system based on physical information neural network, comprising:

[0134] A unit of constructing an explosion overpressure dataset is configured to obtain an open space temperature-pressure explosive explosion shock wave overpressure dataset through numerical simulation results, and divide the dataset into two parts: a training dataset and a verification dataset.

[0135] A unit of defining a physical problem and an equation is configured to solve the N-S equation as a physical problem for a specific scenario, involving a mass conservation equation, X-direction and Y-direction momentum conservation equations.

[0136] A unit of building a special neural network is configured to construct a neural network with no less than 4 layers of hidden layers and no less than 100 neurons in each layer according to input parameters and output parameter characteristics; the hidden layer uses a Tanh function as an activation function, which is beneficial to automatic differentiation calculation of high-order derivatives of the N-S equation; the output layer uses a Softplus function to ensure non-negativity for shock wave pressure P and density p, and has no constraints for speed U and V.

[0137] A unit of designing a loss function and boundary conditions is configured to include a data error term and a physical information error term in the loss function; the data error term uses a squared error form to ensure that the neural network output is consistent with the true observation or measurement data; the physical information error term includes momentum residual equation error, X-direction mass residual equation, and Y-direction mass residual equation to make the neural network meet the physical equation constraints; the total loss function is composed of the data error term and the physical information error term added according to certain weights; and specific scenario boundary conditions are set.

[0138] An explosion parameter prediction unit is configured to obtain training results based on a trained model, and is used for an explosion shock wave prediction model to predict P-t variation curves at any point in an open space explosion range and two-dimensional pressure cloud maps at any time in the explosion range.

[0139] Embodiment 3

[0140] Embodiment 3 also provides a computer-readable storage medium having a computer program stored thereon, and the computer program is executable by a processor to implement any step of a method for predicting explosion shock wave parameters based on physical information neural networks.

[0141] The computer-readable storage medium can include a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes. For the computer-readable storage medium provided in the present application, please refer to the above method embodiments, which will not be repeated here.

[0142] Those skilled in the art can understand that the above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for predicting explosion shock wave parameters based on physical information neural network, characterized by ,include: S1. Obtain a dataset of overpressure from the shock wave of thermobaric explosives in open space through numerical simulation results, and divide the dataset into two parts: a training dataset and a validation dataset. S2. Solve the NS equations for a specific scenario as a physical problem, involving the mass conservation equation and the momentum conservation equations in the X and Y directions; S3. Based on the characteristics of the input and output parameters, construct a neural network with at least four hidden layers and at least 100 neurons in each layer. The hidden layers use the Tanh function as the activation function to facilitate automatic differentiation to calculate high-order derivatives of the NS equation. The output layer uses the Softplus function to ensure the non-negativity of the shock wave pressure P and density ρ, and the velocities U and V are unconstrained. S4. The loss function includes data error terms and physical information error terms. The data error terms take the form of squared errors to ensure that the neural network output is consistent with the actual observed or measured data. The physical information error terms include the momentum residual equation error, the X-direction mass residual equation, and the Y-direction mass residual equation, ensuring that the neural network satisfies the physical equation constraints. The total loss function is constructed by adding the data error terms and the physical information error terms according to certain weights. Specific scenario boundary conditions are also set. S5. The training results obtained based on the training model are used in the explosion shock wave prediction model to predict the Pt change curve of any point in the explosion range of the open space and the two-dimensional pressure cloud map of the explosion range at any time.

2. The method for predicting explosion shock wave parameters based on physical information neural network according to claim 1, characterized in that: In S1, the overpressure data set of the shock wave of the open space thermobaric explosive explosion is obtained through the numerical simulation results. The specific process is as follows: Through numerical simulation, the explosion process of thermobaric explosives in open space is simulated, and the changes in the explosion shock wave flow field are recorded, including: the explosion shock wave overpressure-time curve (ΔP-t), shock wave velocity-time curve (Ut, Vt) and shock wave air density-time curve (ρ-t) at different distances from the explosion center; and the data sets in the recorded database are normalized.

3. The method for predicting explosion shock wave parameters based on physical information neural network according to claim 1, characterized in that: In S2, solving the NS equations in a specific scenario is considered a physical problem, involving the mass conservation equation and the momentum conservation equations in the X and Y directions, specifically: Define the physical problem as solving the NS equations for a specific scenario: The continuity equation, also known as the mass conservation equation: The momentum conservation equation in the X direction is: The momentum conservation equation in the Y direction is: Where ρ represents the shock wave fluid density; u and v are the components of the shock wave fluid velocity in the x and y directions of the rectangular coordinate system, respectively; and P is the fluid pressure.

4. The method for predicting explosion shock wave parameters based on physical information neural network according to claim 1, characterized in that: The input parameters in S3 include: Two-dimensional coordinates (x, y), time t, and explosive mass m; and the input parameters are normalized to the maximum and minimum, and the range is scaled to [-1, 1] to align with the output range of Tanh.

5. The method for predicting explosion shock wave parameters based on physical information neural network according to claim 1, characterized in that: The output parameters in S3 include: Explosive density ρ, overpressure p, wave velocity u in the X direction, and wave velocity v in the Y direction.

6. The method for predicting explosion shock wave parameters based on physical information neural network according to claim 1, characterized in that: The specific calculation method of the data error term in S4 is: When there is real observation data or measurement data, the data error term is used to measure the difference between the neural network output and the real data; the data error term takes the form of square error: The mathematical expansion form of the data error term is: Where: N: the number of training data points, ρ pred 、p pred 、u pred 、v pred : density, pressure, and x / y direction velocity predicted by the neural network; ρ true 、p true 、u true 、v true : Real values ​​provided by experiments or CFD simulations.

7. The method for predicting explosion shock wave parameters based on physical information neural network according to claim 1, characterized in that: The specific calculation method of the physical information error term in S4 is as follows: the physical information error term includes a mass residual equation, an X-direction momentum residual equation, and a Y-direction momentum residual equation; Mass residual equation: The residual equation of momentum in the X direction is: The residual equation of momentum in the Y direction is: The total physical information error term is: By minimizing the total physical information error term, the output of the neural network is made closer to the solution of the physical equation.

8. The method for predicting explosion shock wave parameters based on physical information neural network according to claim 1 is characterized in that: The specific calculation method of the total loss function in S4 is: L total =λ data ·L data +λ phys ·L phys Where λ data ,λ phys are the weights of the data error term and the physical information error term respectively.

9. The method for predicting explosion shock wave parameters based on physical information neural network according to claim 1, characterized in that: The specific scenario boundary conditions in S4 are specifically: Since the ground is a solid reflective boundary, the normal velocity is 0, the tangential velocity is consistent with the wall velocity, and the shock wave pressure and density do not need to be constrained, naturally satisfying the NS continuity equation. The boundary conditions are set as follows: Where u and v are the components of the shock wave fluid velocity in the x and y directions of the rectangular coordinate system, respectively; ρ is the shock wave fluid density; and p is the fluid pressure.

10. An explosion shock wave parameter prediction system based on physical information neural network, characterized in that: include: An explosion overpressure dataset unit is constructed to obtain the overpressure dataset of the shock wave of thermobaric explosive explosion in open space through numerical simulation results. The dataset is divided into two parts: a training dataset and a validation dataset. Define the physical problem and equation unit, which is used to solve the NS equations in a specific scenario as a physical problem, involving the mass conservation equation and the momentum conservation equations in the X and Y directions; A dedicated neural network unit is built to construct a neural network with no fewer than four hidden layers and no fewer than 100 neurons per layer based on the characteristics of the input and output parameters. The hidden layer uses the Tanh function as the activation function to facilitate automatic differentiation to calculate high-order derivatives of the NS equation. The output layer uses the Softplus function to ensure the non-negativity of the shock wave pressure P and density ρ, and the velocities U and V are unconstrained. Design loss function and boundary condition units, so that the loss function includes data error terms and physical information error terms. The data error term uses the squared error form to ensure that the neural network output is consistent with the actual observation or measurement data. The physical information error term includes the momentum residual equation error, the X-direction mass residual equation, and the Y-direction mass residual equation, so that the neural network satisfies the physical equation constraints. The total loss function is composed of the data error term and the physical information error term added according to certain weights. Specific scene boundary conditions are also set. The explosion parameter prediction unit is used to predict the explosion shock wave prediction model based on the training results obtained by the training model, and predict the Pt change curve of any point in the explosion range of the open space and the two-dimensional pressure cloud map of the explosion range at any time.

Images