Propellant agglomerate distribution prediction calculation method based on graph neural network
Through the graph neural network-based method, the problem of low computational efficiency of propellant agglomerate distribution is solved, and a rapid and accurate prediction of agglomerate distribution is achieved, adapting to the calculation needs of different environments and new types of propellants.
Patent Information
- Application Number
- CN202510523814.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-24
- Publication Date
- 2025-07-29
AI Technical Summary
The prior art cannot accurately predict the combustion process and particle size distribution of propellants in different environments, especially the intrinsic physical relationship between the agglomerates of aluminum particles and the properties of the condensate combustion products, resulting in low computational efficiency and limited types.
A graph neural network-based method is adopted to generate a mixed structure model of the internal particles of propellant through a random loading algorithm. Combined with graph convolutional neural network and self-attention neural network, the local and global characteristics of the particle infographic are captured, and the agglomeration distribution results are output, including the chemical composition ratio and critical distance of the multi-component propellant.
It realizes rapid calculation of multiple propellant structures in large-scale parallel computing, with fast calculation speed and high accuracy. When adapting to new types of propellants, it can be updated through transfer learning to simulate the propellant spatial structure to adjust the liquid bridge force, avoiding the problem of insufficient information capture of graph neural networks.
Smart Images
Figure CN120388652A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of methods for predicting the distribution of propellant aggregates, and particularly relates to a calculation method for predicting the distribution of propellant aggregates based on a graph neural network. Background Art
[0002] Aluminum powder is a commonly used metal additive in current solid propellants. Adding aluminum powder can suppress high-frequency unstable combustion and increase the specific impulse. The reaction process of aluminum can be divided into three stages.
[0003] First, during the combustion flow process of the propellant, as the propellant burning surface retreats, aluminum particles overflow from the interior of the propellant. Among them, a part of the aluminum particles stay on the burning surface and move with the propellant burning surface, aggregating with other nearby particles to form large-sized aggregates. Another part of the aluminum particles directly overflow from the propellant burning surface without agglomeration. The second stage is the combustion process of aluminum in the combustion chamber, including the combustion of aggregates and directly overflowed aluminum particles. The third stage is the process of forming condensed-phase combustion products after aluminum combustion. Generally, it is considered that the products of condensed-phase combustion include two substances: large particle aggregates and alumina smoke particles.
[0004] Aggregates are large-sized aluminum particles formed by the melting and polymerization of aluminum particles during the combustion process, or composed of large-sized alumina particles formed by the oxidation of aluminum vapor micron-sized alumina smoke particles. The combustion process of aluminum and the products of condensed-phase combustion have a significant impact, mainly manifested in specific impulse, ablation, combustion stability, slag accumulation, and energy release characteristics.
[0005] Therefore, studying the combustion of aluminum and the products of condensed-phase combustion is of great significance for the design and simulation of engines. Among them, the agglomeration of aluminum on the propellant surface is a key factor affecting its combustion and the products of condensed-phase combustion. Agglomeration will affect the combustion process of particles, resulting in significant changes in the physical and chemical properties of the products of condensed-phase combustion, especially the particle size distribution of aggregates.
[0006] Although some properties of the combustion products of the propellant are currently known, the internal physical relationship between the properties of aggregates and the products of condensed combustion is poorly understood, and it is impossible to accurately predict the combustion process and particle size distribution of the propellant under different environments. Summary of the Invention
[0007] The purpose of the present invention is to provide a calculation method for predicting the distribution of propellant aggregates based on a graph neural network, which is used to quickly predict the aggregate distribution of propellants with different components with high throughput, ensuring both calculation accuracy and saving calculation time, and solving the problems of slow calculation efficiency of aggregate distribution or limited types of propellants that can be calculated.
[0008] The present invention adopts the following technical solutions: A calculation method for predicting the distribution of propellant aggregates based on a graph neural network, including:
[0009] Step 1: Using the random packing algorithm, according to the chemical composition ratio of the multi-component propellant, generate a random distribution model representing different components in a virtual or physical space to simulate the complex particle mixing structure inside the actual propellant;
[0010] Step 2: Based on the random distribution model, calculate the internal particle agglomeration distribution of the propellant at a predetermined critical distance;
[0011] Step 3: Using the positions and volumes of the internal particles of the propellant as node information, connect the particles with a distance less than the predetermined critical distance to obtain a particle information graph, and use a neural network to process the particle information graph to output the agglomeration distribution result;
[0012] The neural network is used to capture the local and global features of the data of the particle information graph and sequentially output the proportion of particle agglomerates in each size interval after pooling, recapturing and pooling, attention learning, and dimensionality reduction;
[0013] The neural network includes: a first graph convolutional neural network module, a first self-attention graph pooling module, a second graph convolutional neural network module, a second self-attention graph pooling module, a first self-attention neural network module, a second self-attention neural network module, a global average pooling module, and a standard CNN layer connected in sequence.
[0014] Further, the input of the first graph convolutional neural network module is a feature matrix representing the information of each aluminum particle and an adjacency matrix representing the connection relationship between aluminum particles, and the output is a first feature matrix with a high feature dimension representing the information of aluminum particles and an adjacency matrix representing the connection relationship between aluminum particles; it is used to extract the local high-dimensional feature information of each aluminum particle.
[0015] Further, the input of the first self-attention graph pooling module is the first feature matrix and the adjacency matrix, and the output is a second feature matrix with a low node dimension and a high feature dimension and a first adjacency matrix with a low dimension, which is used to reduce the node dimension of aluminum particles.
[0016] Further, the input of the second graph convolutional neural network module is the second feature matrix and the first adjacency matrix, and the output is a third feature matrix with a high feature dimension representing the information of aluminum particles and the first adjacency matrix, which is used to extract the local high-dimensional feature information of each aluminum particle.
[0017] Further, the input of the second self-attention graph pooling module is the third feature matrix and the first adjacency matrix, and the output is a fourth feature matrix with a low node dimension and a high feature dimension and a second adjacency matrix with a low dimension; it is used to reduce the node dimension of aluminum particles.
[0018] Further, the input of the first self-attention neural network module is the fourth feature matrix, and the output is the fifth feature matrix, which is used to capture the high-dimensional global features between aluminum particles.
[0019] Further, the input of the second self-attention neural network module is the fifth feature matrix, and the output is the sixth feature matrix; it is used to capture the high-dimensional global features among aluminum particles.
[0020] Further, the input of the global average pooling module is the sixth feature matrix, and the output is a feature row vector; the input of the standard CNN layer is the feature row vector, and the output is a feature row vector representing the proportion of particle agglomerates in each size interval.
[0021] Further, the formulation of the multi-component propellant is HTPB:AP:AN:Al = 16.5:15.5:40:18; the predetermined critical distance includes a predetermined horizontal critical distance and a predetermined vertical critical distance, the predetermined horizontal critical distance is 30 microns, and the predetermined vertical critical distance is 120 microns.
[0022] The beneficial effects of the present invention are as follows:
[0023] The present invention can perform large-scale parallel operations on multiple propellant structures, and the operation speed is faster than the normal calculation method for propellant agglomeration distribution based on random packing, and the calculation accuracy is basically the same.
[0024] For new types of propellants that may appear in the future, the present invention does not need to reconstruct the calculation method, and can save the consumption of updating the calculation method through transfer learning and reinforcement learning;
[0025] Since the data volume of the propellant agglomeration distribution is not abundant, it is difficult to directly use real data to achieve port-to-port prediction. The present invention can be based on the calculation method of the propellant agglomeration model of random packing, and perform agglomeration calculations using the same parameters (propellant spatial structure and agglomeration critical distance); the neural network of the present invention can be perfectly combined with the random packing algorithm, embed the critical distance into the adjacency matrix, and can adjust the critical distance according to requirements to change the adjacency matrix, simulating the adjustment of the liquid bridge force size by adjusting the binder in the propellant. To avoid the problem of insufficient global information capture ability of the graph neural network, a self-attention neural network is used for supplementation. Description of the Drawings
[0026] Figure 1 The propellant spatial structure generated by the random packing model in the prior art;
[0027] Figure 2 The internal particle agglomeration distribution of the propellant generated by the present invention;
[0028] Figure 3 The agglomerate distribution calculated based on the agglomeration model of random packing;
[0029] Figure 4This is the framework diagram of the neural network of the present invention;
[0030] Figure 5 This is the structure of the graph convolutional neural network module of the present invention;
[0031] Figure 6 This is the structure of the self-attention graph pooling neural network module of the present invention;
[0032] Figure 7 This is the structure of the self-attention neural network module of the present invention. Detailed implementation manners
[0033] The present invention will be described in detail below with reference to the drawings and specific implementation manners.
[0034] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "center", "longitudinal", "transverse", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise specified, the meaning of "a plurality" is two or more.
[0035] The present invention discloses a calculation method for predicting the distribution of propellant agglomerates based on a graph neural network, including:
[0036] Step 1: Using a random packing algorithm, according to the chemical composition ratio of the multi-component propellant, generate a random distribution model representing different components in a virtual or physical space to simulate the complex particle mixing structure inside the actual propellant;
[0037] Step 2: Calculate the internal particle agglomeration distribution of the propellant at a predetermined critical distance based on the random distribution model;
[0038] Step 3: Using the position and volume of the internal particles of the propellant as node information, connect the particles with a distance less than the predetermined critical distance to obtain a particle information graph, and use a neural network to process the particle information graph to output the agglomeration distribution result.
[0039] The neural network is used to capture the local and global features of the data of the particle information graph and sequentially output the proportion of the particle agglomerates in each size interval after pooling, recapturing and pooling again, attention learning, and dimensionality reduction.
[0040] As Figure 4 shown, the neural network includes: a first graph convolutional neural network module, a first self-attention graph pooling module, a second graph convolutional neural network module, a second self-attention graph pooling module, a first self-attention neural network module, a second self-attention neural network module, a global average pooling module, and a standard CNN layer connected in sequence.
[0041] The input of the first graph convolutional neural network module is a feature matrix representing the information of each aluminum particle and an adjacency matrix representing the connection relationship between aluminum particles, and the output is a first feature matrix with a high feature dimension representing the information of aluminum particles and an adjacency matrix representing the connection relationship between aluminum particles; it is used to extract the local high-dimensional feature information of each aluminum particle.
[0042] The input of the first self-attention graph pooling module is the first feature matrix and the adjacency matrix, and the output is a second feature matrix with a low node dimension and a high feature dimension and a first adjacency matrix with a low dimension, which is used to reduce the node dimension of aluminum particles.
[0043] The input of the second graph convolutional neural network module is the second feature matrix and the first adjacency matrix, and the output is a third feature matrix with a high feature dimension representing the information of aluminum particles and the first adjacency matrix, which is used to extract the local high-dimensional feature information of each aluminum particle.
[0044] The input of the second self-attention graph pooling module is the third feature matrix and the first adjacency matrix, and the output is a fourth feature matrix with a low node dimension and a high feature dimension and a second adjacency matrix with a low dimension; it is used to reduce the node dimension of aluminum particles.
[0045] The input of the first self-attention neural network module is the fourth feature matrix, and the output is the fifth feature matrix, which is used to capture the high-dimensional global features between each aluminum particle.
[0046] The input of the second self-attention neural network module is the fifth feature matrix, and the output is the sixth feature matrix; it is used to capture the high-dimensional global features between each aluminum particle.
[0047] The input of the global average pooling module is the sixth feature matrix, and the output is a feature row vector; the input of the standard CNN layer is the feature row vector, and the output is a feature row vector representing the proportion of particle aggregates in each size interval.
[0048] The formula of the multi-component propellant is HTPB: AP: AN: Al = 16.5: 15.5: 40: 18; the predetermined critical distance includes a predetermined horizontal critical distance and a predetermined vertical critical distance, the predetermined horizontal critical distance is 30 microns, and the predetermined vertical critical distance is 120 microns.
[0049] Example 1
[0050] The internal particle agglomeration distribution of the propellant obtained in this embodiment is as follows Figure 2 shown. It can be seen from Figure 2 that the positions of the aluminum particles forming the same agglomerate are basically close, and there will be no situation where completely unrelated aluminum particles agglomerate with each other, which proves the robustness of the algorithm in this embodiment.
[0051] The spatial structure of the propellant generated by the random packing model in the prior art is as follows Figure 1 shown. It can be seen from Figure 1 the spatial distribution and size of AP, AN and Al particles. The binder, as the continuous phase, is not shown. The generated particles have various gradations, meeting the requirements of existing propellants.
[0052] In the prior art, the agglomerate distribution calculated by the agglomeration model based on random packing is as follows Figure 3 shown. It can be seen from Figure 3 that the agglomerate distribution shows a log-normal distribution, which conforms to the general distribution of propellant agglomerates, proving that the conventional algorithm conforms to the objective law. Only when the conventional algorithm conforms to the objective law can the data generated enable the deep learning model to conform to the objective law.
[0053] In this embodiment, the structures of the first graph convolutional neural network module and the second graph convolutional neural network module (GCN) are the same. Their input is the feature matrix X and the adjacency matrix A, as follows Figure 5 shown. The feature matrix X is an N×a matrix, where N represents the number of nodes in the graph; a represents the features of each node. A is an N×N matrix representing the connection situation between nodes, with the connection positions being 1 and otherwise 0. D is an N×N diagonal matrix generated based on A, and the values on the main diagonal are the number of 1s in the corresponding rows of matrix A.
[0054] The operator in the form of the matrix for graph convolution operation is:
[0055]
[0056] W is the parameter matrix obtained by training the output of the graph convolutional neural network. a is the feature dimension of matrix X, and b is the output feature dimension. The calculation process of graph convolution is:
[0057]
[0058] Relu is an activation function. X ′ is the output of the GCN, which is an N×a matrix, and the adjacency matrix A is directly propagated downward. The graph convolutional neural network can capture the local features between each node.
[0059] As follows Figure 6As shown, the structures of the first self-attention graph pooling module and the second self-attention graph pooling module (SAGpooling) are the same. By combining the self-attention operator and the activation function, the self-attention scores can be calculated:
[0060]
[0061] Then, sort according to the calculated attention scores Z and take the indices idx of the top kN nodes:
[0062] idx = Top_Rank(Z, kN) (4)
[0063] The feature matrix X finally output by SAGpooling ′ is a kN×a matrix, calculated by formula (5):
[0064] X ′ = Z·X idx (5)
[0065] X idx represents the rows of X indexed by idx. The adjacency matrix A finally output by SAGpooling ′ is a kN×kN adjacency matrix, calculated by formula (6):
[0066] A ′ = A idx，idx (6)
[0067] A idx，idx represents the rows and columns of A indexed by idx. SAGpooling can reduce the node dimension and fuse node information.
[0068] As Figure 7 shown, the structures of the first self-attention neural network module and the second self-attention neural network module (self-Attention) are the same. The self-attention neural network only requires the feature matrix X as the input. The model requires three parameter matrices W1, W2, W3 of a×b. By multiplying the feature vector with the parameter matrices, the query vector, key vector, and value vector can be obtained:
[0069] K = XW1 (7)
[0070] Q = XW2 (8)
[0071] V = XW3 (9)
[0072] The score matrix Z is obtained by formula (10):
[0073]
[0074] The output X of the network′ is an N×b matrix, obtained from Equation (11):
[0075] X ′ = Z·V (11)
[0076] Self-Attention can capture the global information among all features.
[0077] The input of the global average pooling module (GAP) is the feature matrix X of size N×a, and the output is a 1×a row vector. The specific calculation method is to calculate the average value of each column of the feature matrix:
[0078] X ′ = mean(X, axis = 1) (11)
[0079] GAP can integrate the global feature information, which is convenient for integrating inputs of different lengths into a fixed-length output.
[0080] The fully connected neural network is the most basic neural network module and can be used as the output of the entire neural network at the end of the neural network. For an input feature matrix X of size N×a, the output is:
[0081] X ′ = XW (11)
[0082] Since this module is used as the output of the neural network, there is no need to use an activation function anymore.
[0083] The initial input of the network is the feature matrix X1 of size N×4 and the adjacency matrix A of size N×N. The four features of the feature matrix are the X, Y, Z coordinates and volume of the aluminum particles, and the adjacency matrix A. The generation of the adjacency matrix A is based on the distance between pairs of aluminum particles. For example, if two particles are within the critical distance of agglomeration, the position is set to 1, otherwise 0. The specific process is as follows:
[0084] (1) After passing through a GCN with a parameter matrix of 4×16, the N×16 feature matrix X2 is obtained.
[0085] (2) After passing through a SAGpooling with k = 0.6, the 0.6N×16 feature matrix X3 and the 0.6N×0.6N adjacency matrix A2 are obtained
[0086] (3) After passing through a GCN with a parameter matrix of 16×32, the N×32 feature matrix X4 is obtained.
[0087] (4) After passing through a SAGpooling with k = 0.6, the 0.36N×32 feature matrix X5 and the 0.36N×0.36N adjacency matrix A2 are obtained
[0088] After discarding the adjacency matrix, a 0.36N×64 feature matrix X6 is obtained after one layer of self-attention with parameters of three 32×64 matrices.
[0089] After one layer of self-attention with parameters of three 64×128 matrices, a 0.36N×128 feature matrix X7 is obtained.
[0090] After one layer of GAP, a 1×128 feature matrix X8 is obtained.
[0091] After one layer of CNN with parameters of three 128×100 matrices, a 1×100 feature matrix Y is obtained.
[0092] Each value in the finally obtained feature matrix Y correspondingly represents the proportion of aggregates in each size interval.
[0093] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A calculation method for predicting the distribution of propellant aggregates based on a graph neural network, characterized in that, Including: Step 1: Using the random packing algorithm, according to the chemical composition ratio of the multi-component propellant, generate a random distribution model representing different components in a virtual or physical space to simulate the complex particle mixing structure inside the actual propellant; Step 2: Calculate the internal particle agglomeration distribution of the propellant at a predetermined critical distance based on the random distribution model; Step 3: Using the positions and volumes of the internal particles of the propellant as node information, connect the particles with a distance less than the predetermined critical distance to obtain a particle information graph, and use a neural network to process the particle information graph to output the agglomeration distribution result; The neural network is used to capture the local and global features of the data of the particle information graph and output the proportion of particle agglomerates in each size interval after pooling, recapturing and pooling, attention learning, and dimensionality reduction in sequence; The neural network includes: a first graph convolutional neural network module, a first self-attention graph pooling module, a second graph convolutional neural network module, a second self-attention graph pooling module, a first self-attention neural network module, a second self-attention neural network module, a global average pooling module, and a standard CNN layer connected in sequence.
2. The calculation method for predicting the distribution of propellant agglomerates based on a graph neural network according to claim 1, characterized in that, The input of the first graph convolutional neural network module is a feature matrix representing the information of each aluminum particle and an adjacency matrix representing the connection relationship between aluminum particles, and the output is a first feature matrix with a high feature dimension representing the information of aluminum particles and an adjacency matrix representing the connection relationship between aluminum particles; it is used to extract the local high-dimensional feature information of each aluminum particle.
3. The calculation method for predicting the distribution of propellant aggregates based on a graph neural network according to claim 2, wherein The input of the first self-attention graph pooling module is the first feature matrix and the adjacency matrix, and the output is a second feature matrix with a low node dimension and a high feature dimension and a first adjacency matrix with a low dimension, which is used to reduce the node dimension of aluminum particles.
4. The calculation method for predicting the distribution of propellant aggregates based on a graph neural network according to claim 3, wherein The input of the second graph convolutional neural network module is the second feature matrix and the first adjacency matrix, and the output is a third feature matrix with a high feature dimension representing the information of aluminum particles and the first adjacency matrix, which is used to extract the local high-dimensional feature information of each aluminum particle.
5. The calculation method for predicting the distribution of propellant agglomerates based on a graph neural network according to claim 4, characterized in that, The input of the second self-attention graph pooling module is the third feature matrix and the first adjacency matrix, and the output is a fourth feature matrix with a low node dimension and a high feature dimension and a second adjacency matrix with a low dimension; it is used to reduce the node dimension of aluminum particles.
6. The calculation method for predicting the distribution of propellant aggregates based on a graph neural network according to claim 5, characterized in that The input of the first self-attention neural network module is the fourth feature matrix, and the output is a fifth feature matrix, which is used to capture the high-dimensional global features between aluminum particles.
7. A calculation method for predicting the distribution of propellant aggregates based on a graph neural network according to claim 6, characterized in that, The input of the second self-attention neural network module is the fifth feature matrix, and the output is a sixth feature matrix; it is used to capture the high-dimensional global features between aluminum particles.
8. A calculation method for predicting the distribution of propellant agglomerates based on a graph neural network according to claim 7, characterized in that, The input of the global average pooling module is the sixth feature matrix, and the output is a feature row vector; the input of the standard CNN layer is the feature row vector, and the output is a feature row vector representing the proportion of particle agglomerates in each size interval.
9. The calculation method for predicting the distribution of propellant agglomerates based on a graph neural network according to claim 1, wherein, The formula of the multi-component propellant is HTPB: AP: AN: Al = 16.5: 15.5: 40: 18; the predetermined critical distance includes a predetermined horizontal critical distance and a predetermined vertical critical distance, the predetermined horizontal critical distance is 30 microns, and the predetermined vertical critical distance is 120 microns.