A method for predicting multi-task properties of energetic materials

Abstract

The application relates to the technical field of energetic material design and performance prediction, and relates to an energetic material multitask property prediction method. Standardized data sets are constructed by collecting energetic material sample data and performing molecular graph data enhancement processing on a training subset; molecular SMILES strings are respectively parsed into molecular topological graphs and word sequences, three-dimensional graph neural network is used to encode atomic space position information to extract molecular local topological features, a Transformer encoder is used to extract molecular global long-range correlation features, and a dynamic weighting mechanism is used to realize adaptive fusion of the two types of features, so that the depth and comprehensiveness of feature mining are improved; a multitask prediction model with a shared encoding layer, multiple independent decoding layers and a task decoupling constraint mechanism is built, an end-to-end training is completed by using a multitask loss function, the problems that existing multitask methods are difficult to balance the internal conflicts between energy performance and safety performance and are difficult to simultaneously predict multiple key indicators are solved, and the simultaneous prediction of multiple performance indicators is realized.

Description

Technical Field

[0001] This application relates to the field of energetic materials design and performance prediction technology, and more specifically, to a method for predicting the multi-task properties of energetic materials. Background Technology

[0002] Energetic materials are core functional materials in the fields of national defense, aerospace propulsion, and civilian blasting. Their detonation properties (such as detonation velocity and detonation pressure), energy properties (such as detonation heat and specific impulse), safety properties (such as impact, friction, and electrostatic sensitivity), and thermal stability properties (such as thermal decomposition temperature) directly determine the reliability of the formulation and the safety of its application. Traditionally, obtaining the above properties has been highly dependent on experimental synthesis, loading, explosion testing, and sensitivity testing, which has bottleneck problems such as long cycle, high cost, high risk, and difficulty in achieving high-throughput screening, making it difficult to meet the needs of modern energetic materials research and development for high efficiency, safety, and high throughput.

[0003] In recent years, with the rapid development of artificial intelligence technology, data-driven machine learning methods have provided a new paradigm for molecular performance prediction. Especially in the field of drug discovery, deep learning methods such as Transformer and graph neural networks have been widely used in molecular performance prediction. For example, chemical language models based on Transformer can achieve high-precision prediction of drug solubility, toxicity and binding affinity by pre-training on large-scale SMILES data. Hybrid architectures that integrate sequence information and graph structure information have shown excellent performance on multiple drug benchmark datasets.

[0004] However, despite significant progress in drug molecule prediction, research on the prediction of energetic material properties remains relatively immature; existing methods for predicting energetic materials can be mainly divided into three categories: The first category is traditional machine learning methods based on quantitative structure-property relationships (QSPR), which establish regression models using predefined molecular descriptors (such as molecular weight, enthalpy of formation, oxygen balance, etc.). For example, Chinese invention patent CN116721726A discloses a method and system for predicting the properties of energetic materials based on automatic machine learning. It uses the AutoML framework to train the model and realizes the automated prediction of the properties of energetic materials. However, this type of method relies on manually designed descriptors, which makes it difficult to fully explore the chemical bond connections, functional group spatial arrangement, and overall molecular topology information within the molecule. For energetic materials, their key properties such as sensitivity and stability are highly correlated with the microstructure within the molecule. A single descriptor input limits the upper limit of the model's prediction accuracy for such structure-sensitive properties.

[0005] The second category is based on graph neural networks, which represent molecules as topological graphs with atoms as nodes and chemical bonds as edges, and learn local chemical environment features through message passing mechanisms. This type of method has achieved good results in predicting energy properties such as detonation velocity and detonation pressure. For example, Chinese invention patent CN117747014A discloses an artificial intelligence design method for energetic materials based on multi-objective optimization, which combines deep learning and multi-objective optimization for the generation and screening of energetic molecules. However, pure GNN methods extract features by aggregating neighborhoods layer by layer, and their effective information range is limited by the number of network layers and the problem of oversmoothing, making it difficult to efficiently capture long-range dependencies such as synergistic effects and conjugation effects between distal groups in the molecule. Studies have shown that the thermal stability and sensitivity of energetic materials are not only affected by local triggering bonds, but also closely related to the overall conjugated system of the molecule and the electronic effects of distal substituents. It is difficult to fully model these properties by relying solely on local graph features. In recent years, advanced GNN architectures such as DimeNet++ have demonstrated excellent geometric modeling capabilities in the field of molecular property prediction by jointly encoding interatomic distance and bond angle information through spherical basis functions, but they have not yet been applied to the multi-task performance prediction of energetic materials.

[0006] The third category is multi-task learning methods, which simultaneously predict multiple related performance characteristics by sharing representations. For example, Chinese invention patent CN117747014A considers multi-objective optimization problems in energetic material design, attempting to balance energy and safety requirements. However, existing multi-task methods mainly adopt simple shared underlying or fixed-weight loss functions, lacking an effective balancing mechanism for the correlation and conflict between tasks, making it difficult to simultaneously and accurately predict related indicators with inherent trade-offs, such as detonation velocity, sensitivity, and decomposition temperature.

[0007] It is important to note that although there have been explorations of integrating GNNs and Transformers in drug molecule prediction, there are fundamental differences between energetic material prediction and drug prediction: First, drug properties (such as solubility and toxicity) mainly depend on the characteristics of local functional groups in the molecule, while the sensitivity of energetic materials is highly dependent on the global structural characteristics of the molecule, such as the electron distribution of the overall conjugated system, the molecular packing mode, and the synergistic effect between distal groups; Second, drug datasets can reach the scale of millions, while experimental data for energetic materials are only around the scale of thousands, highlighting the scarcity of data (Chinese invention patent CN114861532A discloses a small-sample energetic material performance prediction method using meta-learning, attempting to alleviate the small-sample problem through meta-learning, but this method targets mechanical properties and does not solve the structural sensitivity modeling problem); Third, energetic materials require simultaneous consideration of energy performance and safety performance, which are inherently conflicting, placing higher demands on multi-task collaborative modeling; therefore, successful experiences in the pharmaceutical field cannot be simply extrapolated to the field of energetic materials.

[0008] In summary, although existing technologies, represented by CN116721726A and CN109411029A, have achieved automated prediction based on machine learning, and advancements, represented by CN117747014A, have explored multi-objective collaborative design, significant shortcomings remain in areas such as in-depth feature mining (fusion of graph structure information and global dependencies), collaborative processing of complex tasks (dynamic balance between energy and safety), and generalization ability under small sample conditions. There is an urgent need for a multi-task property prediction method for energetic materials that can deeply integrate local features of molecular graphs with global features of Transformers, coordinate energy performance and safety performance in collaborative modeling, and possess excellent generalization ability under small sample conditions. This method would address the deficiencies of existing technologies and meet the high-throughput, high-efficiency, and high-safety research and development needs of modern energetic materials. Summary of the Invention

[0009] This invention provides a multi-task property prediction method for energetic materials, which aims to solve the technical problems of low prediction accuracy, limited task coverage, insufficient feature extraction, and difficulty in simultaneously predicting multiple key performance indicators in existing energetic material performance prediction methods.

[0010] This invention provides a method for predicting the multi-task properties of energetic materials, comprising the following steps: Collect energetic material sample data, which includes the molecular SMILES string and corresponding multiple performance labels. Construct a standardized dataset based on the sample data. Perform data cleaning, numerical normalization, and dataset partitioning on the standardized dataset in sequence. Then perform molecular graph data augmentation on the partitioned training subset to expand the sample size. The SMILES string is parsed into a molecular topology graph, which includes atomic node features, chemical bond edge features, and an atomic adjacency matrix. A three-dimensional graph neural network is used to extract features from the molecular topology graph, encode the spatial position information between atoms, and obtain the local topological feature vector of the molecule. The SMILES string of molecules is converted into a word segment sequence and position encoding is added. The word segment sequence is input into the Transformer encoder, and the global long-range correlation feature vector of molecules is extracted through the self-attention mechanism. An adaptive fusion of the molecular local topological feature vector and the molecular global long-range correlation feature vector is performed using a dynamic weighting mechanism to obtain the target fused feature. A multi-task prediction model is constructed, which is configured with a shared encoding layer, multiple independent decoding layers, and a task decoupling constraint mechanism. The target fusion features are input into the multi-task prediction model, and end-to-end training is completed using a multi-task loss function. After extracting the target fusion features from the SMILES string of the energetic material to be predicted, the target fusion features are input into the trained multi-task prediction model, and the corresponding multiple performance indicators are output.

[0011] This invention alleviates the scarcity of energetic material data by collecting sample data of energetic materials and constructing a standardized dataset. Molecular graph data augmentation is then performed on the training subset to provide data support for high-precision prediction. The SMILES string is parsed into a molecular topology graph and a word segmentation sequence. A 3D graph neural network is used to encode atomic spatial location information to extract local molecular topological features. A Transformer encoder extracts global long-range correlation features of the molecules. A dynamic weighting mechanism is then used to adaptively fuse the two types of features, effectively overcoming the limitations of traditional QSPR methods that rely on manual descriptors and where pure graph neural networks struggle to capture long-range dependencies. This approach addresses the shortcomings of insufficient single feature extraction by improving the depth and comprehensiveness of feature mining. It constructs a multi-task prediction model with a shared encoding layer, multiple independent decoding layers, and a task decoupling constraint mechanism. End-to-end training is completed using a multi-task loss function, resolving the existing multi-task methods' difficulty in balancing the inherent conflict between energy performance and safety performance, and the challenge of simultaneously predicting multiple key indicators. This enables simultaneous prediction of multiple performance indicators, thereby improving prediction accuracy, expanding task coverage, and meeting the research and development needs of energetic materials for high throughput, high efficiency, and high safety. Furthermore, it boasts high data utilization, comprehensive feature extraction, high prediction accuracy, and the ability to simultaneously predict multiple key performance indicators.

[0012] Preferably, the three-dimensional graph neural network is a DimeNet++ network; the DimeNet++ network has a built-in spherical basis function layer to perform feature mapping on the inter-atomic distance data and angle data respectively, and then aggregates the mapped distance features and angle features through a message passing layer to generate a molecular local topological feature vector.

[0013] Preferably, the DimeNet++ network is further configured with a global average pooling layer, a Dropout layer, and a batch normalization layer; the features aggregated by the message passing layer are then aggregated by the global average pooling layer to generate a molecular local topological feature vector. The Dropout layer and batch normalization layer are used for regularization during the model training phase. The DimeNet++ network is configured with at least three message passing layers and a global average pooling layer.

[0014] Preferably, the Transformer encoder adopts a stacked structure of at least two coding layers, each coding layer is configured with a set of multi-head self-attention sub-layers and a set of feedforward neural network sub-layers, the encoder does not add extra molecular modules to simplify the network architecture; learnable positional encoding vectors and word segmentation sequences are used for feature concatenation to associate molecular topological distance information; the number of heads in the multi-head self-attention sub-layers is at least 8.

[0015] Preferably, the number of independent decoding layers is eight; the eight task-independent decoding layers are eight parallel single-layer fully connected networks, realizing multi-task decoupled prediction.

[0016] Preferably, the task decoupling constraint mechanism includes a triple constraint of feature subspace partitioning, orthogonality constraint, and gradient projection; the shared coding layer output features are split into energy performance subspace features and safety performance subspace features by feature subspace partitioning, the two types of subspace feature vectors are forced to remain orthogonal by orthogonality constraint, and the gradient projection is used to resolve conflicts in the backpropagation gradient of the multi-task.

[0017] Preferably, the dynamic weighting mechanism includes the following steps: A vector concatenation operation is performed on the local topological feature vector of the molecule and the global long-range correlation feature vector of the molecule. The result of the concatenation operation is used to calculate a dynamic weight vector that matches the feature dimension through an activation function. Based on the dynamic weight vector, the two types of features are fused element by element to obtain the target fused feature.

[0018] Preferably, the multi-task loss function is a weighted sum of weighted loss based on different variance uncertainties and orthogonal constraint loss, and the loss weights of each task are dynamically adjusted by learnable noise parameters.

[0019] Preferably, the performance indicators include eight performance indicators: detonation velocity, detonation pressure, detonation heat, specific impulse, impact sensitivity, friction sensitivity, electrostatic spark sensitivity, and thermal decomposition temperature.

[0020] Preferably, the molecular graph data augmentation process includes at least one of the following augmentation operations: atomic masking, bond perturbation, and subgraph clipping.

[0021] The beneficial effects of this invention include: This invention addresses the three major characteristics of energetic materials: scarce data, global structural sensitivity, and energy-security conflict. It designs molecular graph data enhancement and lightweight regularization strategies, a Transformer-specific design for long-range dependencies, and a task decoupling mechanism based on subspace partitioning and orthogonal constraints. These interconnected designs form a systematic technical solution that ensures the model's prediction accuracy and generalization ability for structurally sensitive energy under small sample conditions.

[0022] This invention innovatively integrates DimeNet++ with Transformer. DimeNet++ uses spherical basis functions to jointly encode interatomic distances and bond angles, enabling precise capture of molecular geometry and electronic interactions, and fully exploiting local topological features such as functional groups and chemical bonds in molecular structures. Simultaneously, it utilizes Transformer's self-attention mechanism to capture long-range dependencies and synergistic effects between distal molecular groups. This dual-modal feature extraction mechanism effectively solves the prediction accuracy bottleneck caused by single feature extraction in traditional methods, significantly improving the prediction reliability and accuracy of structural sensitivity properties such as impact sensitivity, friction sensitivity, electrostatic sensitivity, and thermal decomposition temperature.

[0023] This invention achieves data-driven and intelligent molecular structure calculation, eliminating the need for high-risk experiments such as physical synthesis, loading, explosion testing, and sensitivity testing. It avoids safety risks at the source and significantly reduces R&D costs and time. It is a green, safe, efficient, and economical new method for evaluating the performance of energetic materials. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 The flowchart below shows the overall process of the multi-task performance prediction method for energetic materials provided in this embodiment of the invention.

[0026] Figure 2 A scatter plot comparing the predicted and experimental values ​​of eight performance indicators—detonation velocity, detonation pressure, detonation heat, specific impulse, impact sensitivity, friction sensitivity, electrostatic sensitivity, and thermal decomposition temperature—provided for embodiments of the present invention.

[0027] Figure 3 The diagram shows the weight distribution of local and global features in the adaptive dynamic weighted fusion mechanism provided in this embodiment of the invention. Detailed Implementation

[0028] To make the technical problems, solutions, and beneficial effects of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. Example

[0029] This embodiment provides a multi-task property prediction method for energetic materials. Since the properties of energetic materials (such as detonation velocity, sensitivity, thermal decomposition temperature, etc.) are closely related to their microscopic molecular structure, traditional descriptor-based machine learning methods are difficult to fully explore the chemical bond connections, functional group spatial arrangement and synergistic effects between distal groups within the molecule, resulting in limited prediction accuracy, especially poor prediction effect for structure-sensitive safety performance indicators.

[0030] This implementation encodes energetic material molecules into graph structures (preserving atomic connections and local topology) and sequence structures (preserving global order through SMILES token sequences), extracts local chemical environment features using DimeNet++ (using spherical basis functions to jointly encode distance and angle information), and extracts global long-range interaction features using Transformer.

[0031] By fusing local and global features through a dynamic weighting mechanism based on input features, complementary feature enhancement can be achieved.

[0032] A shared encoding layer and a task-independent decoding layer are constructed. The energy-security conflict is resolved by using feature subspace partitioning, orthogonal constraints and gradient projection mechanisms. A multi-task loss function based on homoscedastic uncertainty is used to dynamically balance the learning progress of different tasks.

[0033] Based on the above, this invention can achieve high-precision synchronous prediction of eight key properties of energetic materials (detonation velocity, detonation pressure, detonation heat, specific impulse, impact sensitivity, friction sensitivity, electrostatic sensitivity, and thermal decomposition temperature), providing a powerful tool for the design of new energetic materials.

[0034] To better understand the present invention, the specific steps of the embodiments are further described below, but the content of the present invention is not limited to the following embodiments: See appendix Figure 1 The method for predicting the multi-task properties of energetic materials includes the following steps: S1. Energetic molecule data were collected from public databases (such as the Cambridge Structural Database (CSD) and energetic materials databases) and literature, resulting in 1860 energetic molecule samples, including five major categories: nitramines, nitrate esters, azides, high-nitrogen compounds, and aromatic nitro compounds. Each data point included basic characteristics such as SMILES, density, enthalpy of formation, molecular weight, oxygen balance, and eight performance labels: detonation velocity, detonation pressure, detonation heat, specific impulse, impact sensitivity, friction sensitivity, electrostatic spark sensitivity, and thermal decomposition temperature. Missing values, outliers, and duplicate samples were removed from the data. Numerical features were normalized using min-max normalization and divided into training, validation, and test sets in an 8:1:1 ratio. In this embodiment, in view of the scarcity of experimental data for energetic materials, a molecular graph data enhancement strategy based on chemical knowledge is introduced before training. Specifically, it includes: (1) Atom masking: randomly setting 5%-15% of the atomic node features in the molecule to zero, forcing the model to learn robust local structural representations; (2) Bond perturbation: randomly deleting or changing the type of chemical bonds with a probability of 10%, simulating the effect of molecular conformational changes on performance; (3) Subgraph trimming: randomly retaining the main molecular structure and trimming the side chains to generate variant samples with similar structures. Through the above enhancement strategy, the effective training sample size is expanded to 3-5 times that of the original data, effectively addressing the problem of overfitting with small samples.

[0035] In this embodiment, data cleaning was performed on the 1860 energetic molecule samples obtained, removing 23 obvious outliers (such as detonation velocities exceeding physical limits, negative sensitivity values, etc.). The remaining 1837 samples underwent min-max normalization for their three numerical features: density, enthalpy of formation, and oxygen balance, scaling them to the [0,1] interval. The dataset was randomly divided in an 8:1:1 ratio, resulting in a training set of 1469 samples, a validation set of 184 samples, and a test set of 184 samples.

[0036] For the training set, before training, a molecular graph data augmentation strategy is performed: for each molecule, 5% to 15% of the atomic node features are randomly masked with a 10% probability; chemical bond types are randomly deleted or changed with a 10% probability; and the main structure is randomly retained while side chains are pruned. Through the above augmentation, the training set is expanded to approximately 5000 valid samples.

[0037] In the data construction phase of this embodiment, the system covers the main categories of energetic materials such as nitramines, nitrate esters, azides, high-nitrogen compounds, and aromatic nitro compounds. The subsequent model learns the common molecular structure characteristics and performance correlation rules of different categories of materials, and has excellent cross-category generalization ability. For newly designed energetic molecules, there is no need to model or retrain for specific categories to achieve high-precision performance prediction, which effectively supports the rapid screening and design of new energetic materials across categories.

[0038] S2. Use RDKit to convert the SMILES string into a molecular graph object (molecular topology graph); the molecular topology graph contains atomic node features and chemical bond edge features; the atomic node features include: atomic number (one-hot encoded, dimension 10), hybridization type (sp, sp², sp³, etc.), whether it is an aromatic atom, formal charge, hydrogen bond donor / acceptor identifier, with a total feature dimension of 10; the chemical bond edge features include: bond type (single bond, double bond, triple bond, aromatic bond), whether it is a conjugated bond, whether it is a cyclic bond, bond length (normalized), with a total feature dimension of 6; the adjacency matrix dimension is N×N, where N is the total number of atoms in the molecule.

[0039] S3. The three-dimensional graph neural network model DimeNet++ is used to perform graph convolution operation on the molecular topology graph. DimeNet++ uses spherical 2D Fourier-Bessel basis to jointly encode interatomic distance and bond angle information, which can accurately capture molecular geometry and electronic interactions. Secondly, in this embodiment, considering the limited data scale of energetic materials, DimeNet++ is designed to be lightweight: the number of message passing layers is set to 3 (to avoid oversmoothing), the hidden dimension is set to 128 (to achieve a balance between representation ability and parameter quantity), and Dropout (drop rate 0.2) and batch normalization are introduced as regularization methods. After global pooling, the local topological feature vector of the molecular structure is obtained.

[0040] S4. Convert the SMILES string into a token sequence, where each token corresponds to an atomic symbol or a special structure mark. Map each token to a high-dimensional vector through a learnable embedding layer, add positional encoding, and then input it into the Transformer encoder. In this embodiment, considering the influence of the overall molecular conjugation system and the synergistic effect of distal groups on the mechanical sensitivity and other properties of energetic materials, the Transformer is designed in the following way: Global self-attention mechanism: Transformer's multi-head self-attention allows the direct calculation of interaction weights between tokens at any two positions in a molecule (no matter how far apart they are), which is highly consistent with the intrinsic mechanism of the sensitivity of energetic materials being affected by the electronic effects of distant substituents and the long-range conjugation synergy. Compared with GNN, which requires multiple layers to aggregate distant information (and faces the problem of oversmoothing), Transformer can capture global dependencies in one step.

[0041] Lightweight encoder: Considering the scale of energetic material data, the number of encoder layers is set to 2 and the number of attention heads is set to 8, so as to achieve a balance between capturing long-range dependencies and preventing overfitting; the output sequence is obtained by global average pooling to obtain the global sequence feature vector.

[0042] Specifically, the exemplary parameter setting and execution logic for step S4 in this embodiment is as follows: DimeNet++ is used as the backbone of the graph neural network. DimeNet++ integrates geometric structure information into the message passing process by jointly encoding interatomic distances and bond angles using spherical 2D Fourier-Bessel basis functions. In this embodiment, three message passing layers are set, with each layer having a hidden dimension of 128 and an embedding dimension of 64 for the radial basis functions (RBF) and spherical basis functions (SBF). After three layers of message passing, global average pooling is performed on the node features to obtain a 128-dimensional molecular local topological feature vector. The model incorporates Dropout (with a dropout rate of 0.2) and batch normalization to prevent overfitting. The SMILES string is split into a token sequence character by character, with the sequence length uniformly padded / truncated to 128. Each token is mapped to a 128-dimensional vector through a learnable embedding layer, and learnable positional encoding is added. The Transformer encoder adopts a 2-layer structure with an 8-head self-attention mechanism, a feedforward network hidden layer dimension of 256, and a Dropout rate of 0.1. The output sequence is then subjected to global average pooling to obtain a 128-dimensional global sequence feature vector. .

[0043] S5. A dynamic weighting mechanism based on input features is used to fuse the local topological features extracted by DimeNet++ with the global sequence features extracted by Transformer. The calculation formula is as follows:

[0044]

[0045] In the formula: and These are local features and global features, respectively. This represents vector concatenation. and For learnable parameters, For activation function, This indicates element-wise multiplication. A dynamic weight vector with the same dimension as the features; the fused feature vector. It is a high-dimensional compact representation with 256 dimensions.

[0046] S6. Construct a multi-task prediction model containing a shared encoding layer and eight independent task decoding layers; the shared encoding layer consists of multiple fully connected networks, used to further abstract and reduce the dimensionality of the fused features. For example, the shared encoding layer uses two fully connected networks with dimensions of 256→128→64 respectively, and the activation function is ReLU. A Dropout layer (dropout rate set to 0.2) is added after each layer to suppress overfitting; the independent decoding layer has eight parallel task branches, each branch is a single fully connected network (input dimension 64, output dimension 1), and no activation function is added because it is a regression task.

[0047] In this embodiment, considering the inherent conflict between the energy performance and safety performance of energetic materials, an explicit task decoupling mechanism is introduced into the multi-task model: Feature Subspace Partitioning: The feature vector (64 dimensions) output from the shared encoding layer is further divided into an energy task subspace (32 dimensions) and a safety task subspace (32 dimensions). The energy task subspace connects to the decoding layers for four energy tasks: detonation velocity, detonation pressure, detonation heat, and specific impulse; the safety task subspace connects to the decoding layers for four safety tasks: impact sensitivity, friction sensitivity, electrostatic sensitivity, and thermal decomposition temperature. This hard partitioning forces the model to separately encode energy-related and safety-related information. Orthogonal constraint loss: During training, subspace orthogonal constraints are introduced to force the inner product of the energy subspace feature vector and the safety subspace feature vector to approach zero.

[0048] In the formula: is the orthogonality constraint loss value. The smaller this value, the closer the feature vectors of the two subspaces are to orthogonality; B represents the batch size, which is the total number of samples involved in the calculation in one training iteration; This is represented as the summation of the loss for each sample in the batch; Represents the normalization factor, where It is the dimension of the feature vector; This represents summing over each dimension of the eigenvectors. It involves summing the element-wise multiplications of two vectors across all dimensions, i.e., calculating the inner product. For the first The sample, the first The eigenvector components of the "energy subspace" of the dimension, where the subscript "energy" indicates that they belong to energy-related properties. No. The sample, the first The eigenvector components of the "safety subspace" of the dimension. The subscript "safety" indicates that it belongs to the safety-related characteristics.

[0049] Gradient projection mechanism: When the gradient directions of the energy task and the safety task are opposite, the gradient of one task is projected onto the orthogonal direction of the gradient of the other task.

[0050] In the formula: This is the projected gradient for the security task. This gradient is obtained by transforming the original security gradient. Subtract its energy gradient The component in the direction is obtained, making and Orthogonal (i.e., the inner product is zero); The square of the magnitude of the energy gradient vector, i.e. As the denominator, it is used to normalize the step size of the projection, ensuring that the projection operation is an orthogonal projection.

[0051] The model loss function is a weighted sum of the multi-task loss based on homoscedastic uncertainty and the orthogonal constraint loss:

[0052] In the formula, This represents the total loss value, used for backpropagation optimization during model training; This represents the summation over 8 different tasks. The 8 here indicates that there are a total of 8 sub-tasks (the specific tasks depend on the model design, such as energy-related tasks, security-related tasks, etc.). Indicates the first The weight coefficients of each task are used to achieve automatic task weighting based on homoscedastic uncertainty; For the first Mean squared error loss for each task: is the true value of the k-th task, and pred represents the predicted value; It is the first The logarithmic regularization term for each task. As a logarithmic regularization term for each task. Punishment to prevent Excessive increases lead to insufficient task weight. Together they constitute a multi-task loss based on uncertainty, the theoretical basis of which is maximizing Gaussian likelihood; To balance the hyperparameter (set to 0.1 in this embodiment); This is the orthogonal constraint loss.

[0053] The optimizer used is Adam, and the initial learning rate is set to 1×10⁻ 4 The training phase uses a batch size of 32 and a maximum number of training rounds of 150, and introduces an early stopping mechanism: if the loss on the validation set does not decrease for 10 consecutive rounds, the training is terminated early; during the training process, the loss changes of the training set and the validation set are monitored simultaneously, and the model with the smallest loss on the validation set is finally selected as the final prediction model.

[0054] S7. After extracting the target fusion features from the SMILES string of the energetic material to be predicted, input the target fusion features into the trained multi-task prediction model and output the corresponding multiple performance indicators, that is, output the prediction results of eight key performance indicators in parallel at one time: detonation velocity, detonation pressure, detonation heat, specific impulse, impact sensitivity, friction sensitivity, electrostatic spark sensitivity, and thermal decomposition temperature.

[0055] Example 2 In this embodiment, the model trained in Example 1 is applied to the test set (184 samples) to predict 8 performance indicators, and the coefficient of determination between the predicted values ​​and the experimental values ​​is calculated. R ²), root mean square error (RMSE), and other evaluation indicators; the scatter plot of the prediction results is shown below. Figure 2 As shown: It can be seen that the method of this invention has extremely high prediction accuracy for detonation performance indicators (detonation velocity, detonation pressure, detonation heat, specific impulse). R ² All values ​​reached above 0.95; the prediction accuracy for safety performance indicators (sensitivity, thermal decomposition temperature) was good. R ² is between 0.60 and 0.74; it should be noted that sensitivity and thermal decomposition temperature are affected by various factors such as test environment, crystal form, particle size, and purity, and the experimental measurement itself has a large degree of uncertainty (according to literature reports, the coefficient of variation of the measurement results of the impact sensitivity of the same sample by different laboratories can reach more than 15%). Therefore, the prediction accuracy of the present invention on the above indicators can meet the needs of actual engineering screening and preliminary evaluation.

[0056] Example 3 To verify the superiority of the method of the present invention, three sets of comparative experiments were set up, using the exact same dataset partitioning and preprocessing method as in Example 1, and the following models were constructed respectively: Comparative Example 1: GCN Single-Task Model Only A 3-layer Graph Convolutional Network (GCN) is used to extract molecular graph features with 128 hidden dimensions. After global pooling, the features are fed into 8 independent single-task output layers. Each task is trained separately with MSE as the loss function, Adam as the optimizer, and a learning rate of 1×10⁻. 4 150 training rounds.

[0057] Comparative Example 2: MLP Multi-Task Model Only Without using graph structure information, only pre-computed molecular descriptors (200 2D descriptors computed from RDKit) are used as input, passed through a 2-layer fully connected network (200→128→64), and then fed into an 8-task multi-task output layer. Multi-task training is performed, and the loss function is the weighted sum of the MSEs of the 8 tasks.

[0058] Comparative Example 3: XGBoost Model Only XGBoost (eXtreme Gradient Boosting) was used as the base model, with the 200 molecular descriptors used in Comparative Example 2 as input features. Eight independent XGBoost regression models were trained for each of the eight tasks, and the hyperparameters were optimized using grid search.

[0059] The comparison results are shown in Table 1 below (average). R² (The arithmetic mean of 8 indicators) Table 1 Model Average² Explosive speed Explosive pressure Extremely popular Bichong Impact Sensitivity friction sensitivity electrostatic sensitivity Thermal decomposition temperature Method of the present invention 0.812 0.972 0.968 0.969 0.959 0.601 0.655 0.735 0.636 Comparative Example 1 0.672 0.823 0.815 0.802 0.791 0.512 0.503 0.587 0.543 Comparative Example 2 0.738 0.912 0.908 0.897 0.886 0.602 0.595 0.578 0.524 Comparative Example 3 0.746 0.894 0.886 0.875 0.867 0.589 0.581 0.662 0.612 Experimental results show that the method of the present invention achieves an average performance improvement across eight performance indicators. R The result reached 0.812, which is better than all comparative examples, verifying the effectiveness of the present invention.

[0060] Example 4 To verify the generalization ability of the method of the present invention, the 184 samples in the test set were divided into categories of energetic materials, and the prediction performance (represented by detonation velocity and electrostatic spark sensitivity) of each category was statistically analyzed. The results are shown in Table 2 below: Table 2 Material Category Sample size Explosive speed² Electrostatic spark sensitivity² Nitroamines 52 0.975 0.751 Nitrates 38 0.968 0.729 Azides 31 0.973 0.702 High nitrogen compounds 29 0.966 0.718 Aromatic Nitros 34 0.969 0.735 The results show that the method of the present invention exhibits stable predictive performance across different types of energetic materials, with varying detonation velocities. R ² All are above 0.96, electrostatic spark sensitivity R The values ​​were all above 0.70, and there was no significant decrease in the predictive performance for any category, indicating that the model has good generalization ability.

[0061] To verify the effectiveness of the adaptive fusion mechanism, 10 samples were randomly selected from the test set, and the mean of their corresponding dynamic weight vector w was calculated and plotted as follows. Figure 3 The weight distribution diagram is shown below; in the diagram, the horizontal axis represents the sample number, the vertical axis represents the fusion weight, the dark part represents the local feature weight w, and the light part represents the global feature weight 1-w.

[0062] from Figure 3 It can be seen that there are significant differences in the fusion weights among different samples. Samples S5 and S7 have higher local feature weights (>0.6), indicating that the local functional groups of these molecules play a dominant role in performance prediction. In contrast, samples S4, S5, and S8 have higher global feature weights (>0.6), indicating that the synergistic effect of the distal groups of these molecules is more critical. This result verifies that the adaptive fusion mechanism proposed in this invention can dynamically adjust the contribution ratio of local and global features according to the molecular structure characteristics, thereby achieving accurate modeling of the molecular structures of different energetic materials.

[0063] Experimental conditions and test methods All experiments in this invention were conducted in the following environment: Hardware: Intel Xeon Gold 6242R CPU @ 3.10GHz, 256GB RAM, NVIDIA Tesla V100 GPU (32GB VRAM) Software: Ubuntu 20.04 LTS operating system, Python 3.8, PyTorch 1.10.0, RDKit 2020.09.1, DGL 0.7.2 Evaluation indicator: Coefficient of determination ( R ²), Root Mean Square Error (RMSE), Mean Absolute Error (MAE).

[0064] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting the multi-task properties of energetic materials, characterized in that, Includes the following steps: Collect energetic material sample data, which includes the molecular SMILES string and corresponding multiple performance labels. Construct a standardized dataset based on the sample data. Perform data cleaning, numerical normalization, and dataset partitioning on the standardized dataset in sequence. Then perform molecular graph data augmentation on the partitioned training subset to expand the sample size. The SMILES string is parsed into a molecular topology graph, which includes atomic node features, chemical bond edge features, and an atomic adjacency matrix. A three-dimensional graph neural network is used to extract features from the molecular topology graph, encode the spatial position information between atoms, and obtain the local topological feature vector of the molecule. The SMILES string of molecules is converted into a word segment sequence and position encoding is added. The word segment sequence is input into the Transformer encoder, and the global long-range correlation feature vector of molecules is extracted through the self-attention mechanism. An adaptive fusion of the molecular local topological feature vector and the molecular global long-range correlation feature vector is performed using a dynamic weighting mechanism to obtain the target fused feature. A multi-task prediction model is constructed, which is configured with a shared encoding layer, multiple independent decoding layers, and a task decoupling constraint mechanism. The target fusion features are input into the multi-task prediction model, and end-to-end training is completed using a multi-task loss function. After extracting the target fusion features from the SMILES string of the energetic material to be predicted, the target fusion features are input into the trained multi-task prediction model, and the corresponding multiple performance indicators are output.

2. The method for predicting the multi-task properties of energetic materials according to claim 1, characterized in that, The three-dimensional graph neural network is a DimeNet++ network; the DimeNet++ network has a built-in spherical basis function layer to perform feature mapping on the distance data and angle data between atoms respectively, and then aggregates the mapped distance features and angle features through a message passing layer to generate a molecular local topological feature vector.

3. The method for predicting the multi-task properties of energetic materials according to claim 2, characterized in that, The DimeNet++ network is also configured with a global average pooling layer, a Dropout layer, and a batch normalization layer; the features aggregated by the message passing layer are further aggregated by the global average pooling layer to generate a molecular local topological feature vector. The Dropout layer and batch normalization layer are used for regularization during the model training phase. The DimeNet++ network is configured with at least three message passing layers and a global average pooling layer.

4. The method for predicting the multi-task properties of energetic materials according to claim 1, characterized in that, The Transformer encoder employs a stacked structure of at least two coding layers. Each coding layer is configured with a set of multi-head self-attention sub-layers and a set of feedforward neural network sub-layers. The encoder does not add extra molecular modules to simplify the network architecture. It uses learnable positional encoding vectors and word segmentation sequences for feature concatenation to associate molecular topological distance information. The multi-head self-attention sub-layers have at least 8 heads.

5. The method for predicting the multi-task properties of energetic materials according to claim 1, characterized in that, The number of independent decoding layers is 8; the 8 task-independent decoding layers are 8 parallel single-layer fully connected networks, realizing multi-task decoupled prediction.

6. The method for predicting the multi-task properties of energetic materials according to claim 1, characterized in that, The task decoupling constraint mechanism includes a triple constraint of feature subspace partitioning, orthogonality constraint, and gradient projection. The feature subspace partitioning divides the output features of the shared coding layer into energy performance subspace features and safety performance subspace features. The orthogonality constraint forces the feature vectors of the two types of subspaces to remain orthogonal. The gradient projection is used to resolve conflicts in the backpropagation gradients of the multi-task system.

7. The method for predicting the multi-task properties of energetic materials according to claim 1, characterized in that, The dynamic weighting mechanism includes the following steps: A vector concatenation operation is performed on the local topological feature vector of the molecule and the global long-range correlation feature vector of the molecule. The result of the concatenation operation is used to calculate a dynamic weight vector that matches the feature dimension through an activation function. Based on the dynamic weight vector, the two types of features are fused element by element to obtain the target fused feature.

8. The method for predicting the multi-task properties of energetic materials according to claim 1, characterized in that, The multi-task loss function is a weighted sum of weighted losses based on different variance uncertainties and orthogonal constraint losses, and the loss weights of each task are dynamically adjusted through learnable noise parameters.

9. The method for predicting the multi-task properties of energetic materials according to claim 1, characterized in that, The performance indicators include eight items: detonation velocity, detonation pressure, detonation heat, specific impulse, impact sensitivity, friction sensitivity, electrostatic spark sensitivity, and thermal decomposition temperature.

10. The method for predicting the multi-task properties of energetic materials according to claim 1, characterized in that, The molecular graph data augmentation process includes at least one of the following enhancement operations: atomic masking, bond perturbation, and subgraph pruning.

Images