A material molecular property prediction method and related device

Abstract

The application belongs to the technical field of material science, and discloses a material molecular attribute prediction method and related equipment, which obtains deep-level attribute prediction pre-training model to obtain organic molecular deep-level characterization vector, utilizes attention mechanism to fuse molecular structure features and deep-level attribute information, accurately predicts target attributes, and analyzes significant structure features after prediction, so as to solve the problems that the existing technology cannot effectively capture the internal correlation between material structure and performance and lacks interpretability, has the advantages that the deep-level attribute information and the molecular structure features are fused, the attention mechanism is utilized to accurately predict the target attributes of the material, and the analysis of the significant structure features is provided, so as to improve the material research and development efficiency and the interpretability.

Description

Technical Field

[0001] This application relates to the field of materials science and technology, and more specifically, to a method and related equipment for predicting the molecular properties of materials. Background Technology

[0002] With the rapid development of artificial intelligence (AI) technology, the deep integration of AI and materials science has become an important way to improve the automation and intelligence of materials research and development. Over the past decade, data-driven AI methods, such as black-box models like graph neural networks and recurrent neural networks, have made some progress in areas such as predicting organic molecule synthesis pathways, catalyst design, and predicting fundamental physical properties. However, existing technologies have significant limitations when dealing with complex systems such as OLED luminescent materials, alloy materials, and polymer materials. The molecular structures of these materials involve multi-scale and multi-level interactions, including complex structural features at the atomic, group, and molecular levels, as well as deep-level physical properties such as electronic structure and energy states. Traditional AI models have failed to effectively integrate physical computation methods such as density functional theory with deep learning techniques, making it difficult for models to accurately capture the intrinsic correlation mechanisms between material structure and properties. Especially when predicting the target properties of novel materials, existing methods cannot fully utilize the known deep-level physical property information of materials, nor do they possess the ability to interpretably analyze the correlation between molecular structural features and target properties. This severely restricts the improvement of materials research and development efficiency and results in a significant waste of computational resources.

[0003] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0004] The purpose of this application is to provide a method and related equipment for predicting the molecular properties of materials. By integrating deep-level property information and molecular structure features, the method uses an attention mechanism to accurately predict the target properties of materials and provides analysis of structural features that have a significant impact, thereby improving the efficiency and interpretability of materials research and development.

[0005] In a first aspect, this application provides a method for predicting the molecular properties of materials, the method comprising the following steps:

[0006] A1. Based on the DFT algorithm, deep-level property prediction models for multiple types of materials are established respectively; the deep-level property prediction models are used to predict the deep-level property information of materials based on the molecular geometric structure information of the materials;

[0007] A2. For a set of novel materials for which at least some attribute information has been determined, obtain the molecular geometric structure information and attribute labels of each novel material, and determine a deep attribute prediction model that matches the set of novel materials as the matching model; the attribute labels are used to indicate whether the novel material has the target attribute, or to indicate the attribute value of the target attribute of the novel material.

[0008] A3. Construct a material molecular property prediction model based on the matching model, and train the material molecular geometric structure information and attribute labels of the new material set; the material molecular property prediction model fuses the molecular structure feature information extracted from the material molecular geometric structure information of the new material and the deep attribute information of the new material extracted by the matching model through an attention mechanism, and outputs the prediction result of whether the new material has the target attribute or the attribute value of the target attribute of the new material;

[0009] A4. Using the trained material molecular property prediction model, predict whether the target novel material has the target property or the property value of the target novel material;

[0010] A5. If the prediction results indicate that the target novel material has the target attribute or the attribute value of the target novel material is higher than the preset value threshold, then based on the attention weight distribution information of the trained material molecular attribute prediction model, the part of the molecular structure feature information that has a significant impact on the target attribute is analyzed, and the analysis results are output to provide reference information for optimizing the target novel material.

[0011] Secondly, this application provides an electronic device including a processor and a memory, the memory storing a computer program executable by the processor, wherein when the processor executes the computer program, it performs the steps in the material molecular property prediction method described above.

[0012] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it performs the steps of the material molecular property prediction method as described above.

[0013] Beneficial Effects: The material molecular property prediction method and related equipment provided in this application establish a deep property prediction model, utilize an attention mechanism to fuse molecular structure features and deep property information, accurately predict target properties, and analyze significantly influential structural features after prediction. This solves the problems of existing technologies that cannot effectively capture the intrinsic relationship between material structure and performance and lack interpretability. It has the advantages of accurately predicting target material properties by fusing deep property information and molecular structure features, utilizing an attention mechanism, and providing analysis of significantly influential structural features, thereby improving the efficiency and interpretability of material research and development. Attached Figure Description

[0014] Figure 1 A flowchart of a method for predicting the molecular properties of materials provided in this application.

[0015] Figure 2 A schematic diagram of the structure of the electronic device provided in this application.

[0016] Figure 3 This is a schematic diagram of a material molecular property prediction model.

[0017] Labeling explanations: 301, processor; 302, memory; 303, communication bus. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0019] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0020] Please refer to Figure 1 A method for predicting the molecular properties of materials, as described in some embodiments of this application, includes the following steps:

[0021] A1. Based on the DFT (Density Functional Theory) algorithm, deep-level property prediction models are established for various types of materials; the deep-level property prediction models are used to predict the deep-level property information of materials based on the molecular geometric structure information of the materials.

[0022] A2. For a set of novel materials for which at least some attribute information has been determined, obtain the molecular geometric structure information and attribute labels of each novel material, and determine a deep attribute prediction model that matches the set of novel materials as the matching model; the attribute labels are used to indicate whether the novel material has the target attribute, or to indicate the attribute value of the target attribute of the novel material.

[0023] A3. Construct a material molecular property prediction model based on the matching model, and train the material molecular geometric structure information and attribute labels of the new material set; the material molecular property prediction model fuses the molecular structure feature information extracted from the material molecular geometric structure information of the new material and the deep attribute information of the new material extracted by the matching model through an attention mechanism, and outputs the prediction result of whether the new material has the target attribute or the attribute value of the target attribute of the new material;

[0024] A4. Using the trained material molecular property prediction model, predict whether the target novel material has the target property or the property value of the target novel material;

[0025] A5. If the prediction results indicate that the target novel material has the target attribute or the attribute value of the target novel material is higher than the preset value threshold, then based on the attention weight distribution information of the trained material molecular attribute prediction model, the part of the molecular structure feature information that has a significant impact on the target attribute is analyzed, and the analysis results are output to provide reference information for optimizing the target novel material.

[0026] This application proposes a method for predicting the molecular properties of materials. This method aims to address the problem that existing AI models fail to fully utilize the inherent logic of physical mechanisms to construct the correlation between material structure and performance when dealing with complex systems such as OLED luminescent materials, alloy materials, and polymer materials, resulting in low efficiency in material exploration and waste of computing resources.

[0027] Specifically: First, in step A1, deep-level property prediction models are established for multiple material categories based on the DFT algorithm. These models predict the deep-level properties of materials based on their molecular geometry. Different material categories can be selected, and a new model can be built for each category. For example, one model can be built for OLED luminescent materials, another for alloy materials, and yet another for polymer materials. These models can be based on different machine learning algorithms, such as Support Vector Machines (SVM) or decision trees, using the molecular structure information of their respective materials as input to predict their deep-level properties. In practice, each material category can be further classified into multiple subcategories (e.g., based on the similarity of molecular structure information), and a new deep-level property prediction model can be built for each subcategory.

[0028] Secondly, in step A2, for a set of novel materials with at least some attribute information already determined, the molecular geometry information and attribute labels of each novel material are acquired, and a deep-level attribute prediction model matching the set of novel materials is determined as the matching model. Attribute labels are used to indicate whether the novel material possesses a target attribute (a target attribute refers to a specific performance or characteristic that the novel material expects or needs to predict), or to indicate the attribute value of the target attribute of the novel material (the attribute value of a target attribute refers to an index value used to measure the level of the target attribute; for example, when the target attribute is luminescence, its attribute value can be luminance, luminous efficiency, or luminous chromaticity, etc.). Attribute labels can be determined experimentally. In one implementation, the known attribute information of the novel material set can be manually reviewed, and based on experience, it can be determined which pre-established deep-level attribute prediction model is most suitable for the novel material set. For example, if the novel material set mainly contains OLED luminescent materials, then the deep-level attribute prediction model established for OLED luminescent materials is manually selected as the matching model.

[0029] Next, in step A3, a material molecular property prediction model is constructed based on the matching model, and trained using the material molecular geometry information and attribute labels of the novel material set. The material molecular property prediction model fuses molecular structure feature information extracted from the material molecular geometry information of the novel materials and deep attribute information extracted using the matching model through an attention mechanism, outputting a prediction result of whether the novel material possesses the target attribute or the attribute value of the target attribute. In one implementation, a neural network structure can be designed, containing a feature extraction layer for extracting features from the molecular structure information, a deep attribute information input layer for receiving the output of the matching model, and a simple fusion layer to concatenate the two. Then, the neural network is trained using the molecular structure information and attribute labels of the novel material set to enable it to output a prediction result of whether the novel material possesses the target attribute or the attribute value of the target attribute.

[0030] Next, in step A4, the trained material molecular property prediction model is used to predict whether the target novel material possesses the target property or the property value of the target novel material's target property. In one implementation, the molecular structure information of the target novel material can be input into the trained material molecular property prediction model, and the prediction result output by the model can be directly read.

[0031] Finally, in step A5, if the prediction results indicate that the target novel material possesses the target property or the attribute value of the target novel material is higher than a preset numerical threshold, then based on the attention weight distribution information of the trained material molecular property prediction model, the parts of the molecular structure feature information that significantly affect the target property are analyzed, and the analysis results are output to provide reference information for optimizing the target novel material. In one implementation, the attention weight matrix within the model can be examined to identify molecular structure feature terms with high weight values. For example, if a specific atomic group or chemical bond dominates the attention weights, then that part can be considered to have a significant impact on the target property. Then, a report is generated explaining these significantly influential parts and provided to materials scientists as a reference for optimizing novel materials.

[0032] This application's method for predicting material molecular properties effectively addresses the inefficiencies and computational waste caused by insufficient utilization of physical logic in complex material systems by integrating a deep property prediction model based on physical mechanisms with attention-driven AI technology. Compared to traditional, purely data-driven "black box" AI models, in step A1, this application establishes a deep property prediction model for multiple types of materials based on the DFT algorithm. Utilizing the physical foundation of density functional theory, it directly predicts deep property information from the molecular geometric structure information of the materials. This ensures that the model is built on the inherent logic of physical mechanisms, avoiding the defects of purely data-driven black box models, thus providing reliable theoretical support for subsequent steps. In step A3, the attention mechanism fuses features extracted from molecular structure information with deep property information extracted using a matching model, achieving dynamic focusing of multi-scale features. This effectively handles complex interactions and enhances the accuracy and generalization ability of the prediction. Furthermore, in step A5, if the predicted value of the target attribute or the target attribute of the novel material is higher than a preset value threshold, the part of the molecular structure features that significantly affects the target attribute is analyzed based on the attention weight distribution information, and the analysis results are output. This provides interpretable optimization guidance, helps reduce the trial-and-error process, and further improves R&D efficiency. The technical solution of this application not only improves the efficiency and accuracy of material R&D but also provides interpretable guidance for material design, making a significant technical contribution.

[0033] Preferably, the multiple types of materials include OLED luminescent materials, alloy materials, and polymer materials. OLED luminescent materials typically refer to organic electroluminescent materials, which have complex molecular structures involving conjugated systems and luminescent groups. Alloy materials are materials with metallic properties formed by fusing two or more metals or metals with non-metals, characterized by their lattice arrangement and interatomic interactions. Polymer materials are macromolecules composed of numerous repeating structural units linked by covalent bonds, exhibiting multi-scale structural features, ranging from atomic bonding to molecular chain conformation to macroscopic aggregated structures. By clearly defining these material types, this application can customize the processing based on the characteristics of different material systems, improving the applicability and prediction accuracy of the model.

[0034] Specifically, for OLED luminescent materials and polymer materials, the molecular geometric structure information includes atomic-level, group-level, and molecular-level structural diagrams. Atomic-level structural diagrams typically represent the types and positions of individual atoms in the molecule, as well as the chemical bond relationships between them, for example, through adjacency matrices or graph structures. Group-level structural diagrams focus on the representation of chemical groups with specific functions or structural features (such as benzene rings, carboxyl groups, luminescent groups, etc.) and their interconnections, which helps capture local interactions within the molecule. Molecular-level structural diagrams provide information on the topological structure and spatial conformation of the entire molecule, for example, through molecular diagrams or three-dimensional coordinates. This multi-level structural representation can comprehensively capture the complex structural features of OLED and polymer materials from the microscopic to the macroscopic level, providing richer and more refined input information for the model.

[0035] For alloy materials, the molecular geometry information includes a microcrystalline lattice network diagram. This diagram describes the arrangement of atoms within the crystal structure, cell parameters, atom types and their positions in the lattice, as well as the nearest neighbor relationships between atoms. This representation effectively captures the periodic structure and long-range order of alloy materials, which is crucial for understanding and predicting their physicochemical properties.

[0036] Specifically, the deep-level attribute information includes at least one of electronic structure, energy state, energy level distribution, charge distribution, and vibrational modes. Electronic structure refers to the arrangement and motion of electrons in a material, such as band structure, density of states, and HOMO-LUMO band gap; energy state refers to the total energy, formation energy, and binding energy of the material system; energy level distribution describes the specific arrangement of electronic energy levels in the material; charge distribution reflects the non-uniformity of charge within atoms or molecules; and vibrational modes describe the vibrational behavior of atoms in a lattice or molecule, such as phonon spectra and infrared / Raman active modes. This deep-level attribute information forms the microscopic physical basis of the macroscopic properties of materials. Obtaining this information through the DFT algorithm can provide the model with first-principles-based physical insights, thereby establishing a more accurate structure-property correlation.

[0037] This application's approach involves clearly classifying multiple types of materials and employing multi-scale molecular geometric structure information representations that match the structural characteristics of different material types. This is combined with deep-level attribute information based on physical mechanisms to construct a more universal and accurate material molecular property prediction model. Specifically, when establishing deep-level attribute prediction models for multiple material types, they are first classified into categories such as OLED luminescent materials, alloy materials, and polymer materials based on their inherent characteristics. For OLED luminescent materials and polymer materials, due to their complex molecular structures and multi-scale features, the model is designed to simultaneously process atomic-level, group-level, and molecular-level structural diagrams to comprehensively capture their structural information (for example, the model can utilize a graph neural network module to process atomic-level structural diagrams, a Transformer module with a graph attention perceptron to process group-level structural diagrams, and a neural network module to process molecular-level structural diagrams, and then fuse the feature information extracted by the three modules). This multi-level representation allows the model to understand molecular structure at different granularities. For example, the atomic level focuses on local bonding, the group level on functional units, and the molecular level on overall topology, thus avoiding information loss that might occur with a single-scale representation. For alloy materials, the structural characteristic lies in their periodic lattice arrangement. Therefore, using microscopic lattice network diagrams as information on the material's molecular geometry can effectively reflect its crystal structure and interatomic interactions (e.g., using graph neural network models to construct deep-level property prediction models for alloy materials). This refined molecular geometry information, as input to the deep-level property prediction model, can more accurately characterize the material's essential features. Based on this, deep-level property information such as electronic structure, energy state, energy level distribution, charge distribution, and vibrational modes obtained through the DFT algorithm serve as prediction targets or auxiliary features for the model, providing a solid physical foundation. This deep-level property information directly reflects the quantum mechanical behavior of the material and is a determining factor in its macroscopic properties. Therefore, when constructing molecular property prediction models for novel material sets, these physically enhanced deep-level property prediction models can be used to extract deep-level property information of the novel materials and fuse it with molecular structural feature information extracted from the molecular geometry information of the novel materials. This fusion not only considers the apparent structural features of the materials but also incorporates their inherent physical essence, making the final prediction results of whether the novel material possesses the target property or the property value of the target property more reliable and interpretable. In this way, the solution proposed in this application can effectively overcome the problems of inaccurate predictions and low efficiency caused by the failure of traditional "black box" models to fully utilize physical mechanisms when dealing with complex material systems, significantly improving the efficiency and accuracy of materials exploration.

[0038] In some implementations, step A1 includes:

[0039] A101. Obtain material molecular geometric structure information for multiple types of materials;

[0040] A102. Based on the DFT algorithm, obtain the deep-level property information of the multiple types of materials;

[0041] A103. Using material molecular geometric structure information as sample data and corresponding deep-level attribute information as label data, construct a sample dataset for each type of material;

[0042] A104. Based on deep learning algorithms, construct deep property prediction models for each type of material, and train the deep property prediction models using the corresponding sample datasets.

[0043] First, in step A101, it is necessary to obtain the molecular geometry information of multiple types of materials. This molecular geometry information is the basic input data for subsequent model training, and its comprehensiveness and accuracy directly affect the model's generalization ability. Acquisition methods may include, but are not limited to: collecting data on multiple types of materials from existing large-scale materials databases, such as the Materials Project, OQMD (Open Quantum Materials Database), or other publicly available materials science databases; characterizing the actually synthesized materials through experimental methods, such as X-ray diffraction (XRD) and nuclear magnetic resonance (NMR), to obtain their molecular structure information; or using computational chemistry software, such as Gaussian and VASP, to optimize the structure of theoretically designed materials and obtain their molecular structure information.

[0044] Secondly, in step A102, deep-level attribute information of the various materials is obtained based on the DFT algorithm. This deep-level attribute information serves as the model's label data, and its high reliability is crucial for the model to effectively utilize the inherent logic of the physical mechanisms. Specifically, DFT calculation software (such as VASP, Quantum ESPRESSO, Gaussian, etc.) can be used to perform first-principles calculations on the molecular structures of the materials obtained in step A101 to obtain deep-level attribute information such as electronic structure, energy state, energy level distribution, charge distribution, and vibrational modes. Alternatively, a high-throughput computing platform can be used to automatically execute the DFT calculation process, batch-acquire deep-level attribute information of multiple materials, and perform data cleaning and processing.

[0045] Next, in step A103, sample datasets for each material class are constructed using the material's molecular geometry information as sample data and the corresponding deep-level attribute information as label data. This step aims to achieve a precise structure-attribute correspondence, providing a standardized data foundation for supervised learning. Specifically, the material's molecular geometry information obtained in step A101 (such as SMILES strings, graph representations, crystal structure files, etc.) can be used as input features, and the deep-level attribute information obtained in step A102 (such as band gap values, formation energies, density of states curves, etc.) can be used as corresponding output labels. Data alignment and format conversion are then performed to form structured sample data pairs. Subsequently, according to the material category (e.g., OLED luminescent materials, alloy materials, polymer materials), these sample data pairs are organized into independent sample datasets to ensure the consistency of material properties within each dataset.

[0046] Finally, in step A104, a deep-level property prediction model for each type of material is constructed based on deep learning algorithms, and the model is trained using the corresponding sample dataset. This step aims to combine the specificity of the sample dataset, capture complex patterns, and optimize the training process to improve the model's prediction accuracy and efficiency. Specifically, appropriate deep learning model architectures can be selected for different types of materials. For example, for molecular structure diagrams, graph neural networks (GNNs), such as graph convolutional networks (GCN), GraphSAGE, and message passing neural networks (MPNN), can be used; for sequential molecular representations, recurrent neural networks (RNNs) or Transformer models can be used. Then, using the corresponding sample dataset constructed in step A103, the deep learning model is trained through backpropagation algorithms and optimizers (such as Adam and SGD), adjusting the model parameters to enable it to accurately predict deep-level property information from the molecular geometry of the material. Cross-validation and early stopping strategies can be used during training to prevent overfitting.

[0047] This application's solution provides a systematic and highly reliable method for establishing deep-level property prediction models through the synergistic effect of the aforementioned steps. First, step A101 acquires diverse molecular geometric structure information of materials, providing a comprehensive input foundation for the model. Next, step A102 utilizes the physical rigor of the DFT algorithm to generate high-precision deep-level property labels for these structures, thus integrating the inherent logic of the physical mechanism into the data. Based on this, step A103 precisely pairs structural and property information to construct a structured sample dataset, laying the data foundation for the effective training of the deep learning model. Finally, step A104 trains the model based on deep learning algorithms, enabling it to learn complex structure-property mapping relationships from this high-quality data. This interconnected process transforms the establishment of deep-level property prediction models in material molecular property prediction methods from a vague "black box" process into a transparent, efficient, and accurate systems engineering approach based on a combination of physical principles and data-driven approaches. This not only solves the ambiguity in data acquisition, dataset construction, and model training steps in traditional methods, but also ensures the sufficiency of model training, significantly improving the efficiency of materials exploration and the reliability of predictions.

[0048] In some implementations, in step A2, a deep property prediction model that matches the new material set can be determined based on the determined property information, and used as the matching model.

[0049] The identified attribute information refers to material characteristic data in the new material set that has been clearly defined or obtained through experiments, literature, databases, etc. This information may include, but is not limited to, the material's physical properties (e.g., melting point, density), chemical properties (e.g., reactivity, stability), biological activities (e.g., toxicity, biocompatibility), or specific functional properties (e.g., conductivity, luminescence efficiency). Its role is to provide a direct and reliable basis for subsequently selecting an appropriate prediction model. Specifically, the identified attribute information can be obtained in various ways. For example, through laboratory testing, such as spectral analysis, X-ray diffraction, thermogravimetric analysis, etc., to directly measure certain key properties of the new material; or by consulting existing materials science databases or literature to obtain attribute information of known materials with similar structures or the same composition as the new material and using them as a reference; or by utilizing expert experience or domain knowledge to conduct a preliminary assessment and determination of the potential properties of the new material.

[0050] Determining the deep-level attribute prediction model that best matches the new material set refers to selecting the model that most accurately predicts the deep-level attributes of the current new material set from a pool of pre-established deep-level attribute prediction models for various material categories. This process aims to ensure that the selected model is highly correlated with the characteristics of the new materials, thereby improving the accuracy and reliability of subsequent predictions. This determination process can be based on comparing the established attribute information with the applicability of the pre-trained model. For example, if the new material is identified as an OLED luminescent material, and only one deep-level attribute prediction model is built for OLED luminescent materials, then that deep-level attribute prediction model is used as the matching model. If corresponding deep-level attribute prediction models are built for multiple subcategories of OLED luminescent materials, then the matching model can be determined based on the similarity between the established attribute information and the attribute information of each subcategory of materials.

[0051] Through the above technical solution, this application can selectively choose the most suitable deep-level property prediction model based on the determined attribute information of the novel material set. This precise matching mechanism based on actual attribute information ensures that the selected model is highly correlated with the characteristics of the novel material, thereby avoiding prediction bias caused by model mismatch. In the subsequent construction and training of the material molecular property prediction model, molecular structure feature information and deep-level attribute information can be more reliably integrated, significantly improving the accuracy and reliability of the prediction results. Furthermore, by selecting the most suitable model, unnecessary computational resource consumption can be effectively reduced, prediction efficiency can be improved, and more accurate and efficient guidance can be provided for the research and optimization of novel materials.

[0052] In some implementations, step A3 includes:

[0053] A301. Replace the last output layer of the matching model with a fully connected layer to obtain the replaced matching model;

[0054] A302. Connect the output of the replaced matching model and the output of a molecular structure feature extraction module to the same attention fusion module to construct the material molecular property prediction model; the molecular structure feature extraction module is used to extract molecular structure feature vectors from the material molecular geometry information of the novel material; the fully connected layer is used to convert the deep attribute information extracted by the matching model from the material molecular geometry information of the novel material into deep attribute feature vectors with the same dimension as the molecular structure feature vectors; the attention fusion module is used to fuse the molecular structure feature vectors and the deep attribute feature vectors and output the prediction result of whether the novel material has the target attribute or the attribute value of the target attribute of the novel material;

[0055] A303. The molecular property prediction model of the material is trained using the molecular geometric structure information and the attribute labels of the novel material set.

[0056] First, in step A301, the matching model is a model that predicts the deep-level properties of materials based on their molecular geometry. Its original final output layer may be designed for a specific deep-level property prediction task (e.g., regression prediction of energy values ​​or classification prediction of electronic states), and its output dimension and format may be inconsistent. To effectively fuse the deep-level property information extracted by this model with molecular structure features, its output needs to be standardized. This is achieved by replacing the original final output layer with a fully connected layer. This fully connected layer can map the high-dimensional features learned within the matching model to a pre-defined, uniform-dimensional feature space. For example, this fully connected layer could be a linear layer with a specific number of neurons, whose output dimension is designed to match the dimension of the molecular structure feature vector; or it could be a fully connected network with multiple hidden layers, transforming layer by layer to finally output a feature vector of the required dimension. This replacement operation allows the deep-level property information obtained from the matching model to participate in subsequent feature fusion in the form of a standardized "deep-level property feature vector."

[0057] Secondly, in step A302, the output of the replaced matching model provides a deep-level attribute feature vector with unified dimensions. The molecular structure feature extraction module is an independent neural network component whose function is to extract molecular structure feature vectors representing the molecular structure characteristics from the molecular geometry information of the novel material. For example, this module can be a graph neural network (such as GCN, GAT, or MPNN) capable of directly processing molecular structure diagrams, capturing atomic, bond, and topological information, and encoding it into fixed-dimensional vectors; or, this module can be an encoder based on sequence models (such as RNN or Transformer) to process serialized molecular structure representations such as SMILES strings. These two feature vectors (deep-level attribute feature vector and molecular structure feature vector) are then jointly input into an attention fusion module. The attention fusion module is an advanced feature fusion mechanism that dynamically assigns weights based on the importance of the input features, thereby achieving more effective feature integration. For example, the attention fusion module can employ a self-attention mechanism to calculate the correlation between different feature dimensions and generate a weighted combination; or, it can employ a cross-attention mechanism, allowing one feature vector to act as a query and the other as a key and value, thereby achieving biased information extraction and fusion. In this way, the constructed material molecular property prediction model can simultaneously utilize deep property information guided by physical mechanisms and data-driven molecular structural feature information. The resulting material molecular property prediction model is as follows: Figure 3As shown in the figure, the deep attribute feature extraction module is the replaced matching model. The deep attribute extraction network layer in this deep attribute feature extraction module refers to the part remaining after the last output layer of the matching model is removed.

[0058] After replacing the output layer of the matching model, the core function of the fully connected layer is to perform a dimensional alignment and feature transformation task. It receives deep-level attribute information (e.g., electronic structure, energy state) obtained by the matching model after processing the molecular geometry information of novel materials, and maps it to a deep-level attribute feature vector with the same dimension as the molecular structure feature vector output by the molecular structure feature extraction module. This dimensional unification is a key prerequisite for subsequent attention fusion, ensuring that features from two different sources can be effectively compared and interacted with within the attention mechanism.

[0059] The attention fusion module is the core fusion unit of the entire model. It receives molecular structure feature vectors from the molecular structure feature extraction module and deep attribute feature vectors from the replaced matching model. Through its internal attention mechanism, this module can learn and identify which feature (or which part of the feature) is more important when predicting whether a novel material possesses the target attribute, and assign different weights accordingly. For example, when predicting the fluorescence properties of a material, the attention mechanism may assign higher weights to deep attribute features related to electronic structure; while when predicting the mechanical strength of a material, specific groups or connection methods in the molecular structure may receive higher weights. Finally, the attention fusion module outputs the fused feature vectors through the corresponding output layer according to the specific prediction task. If the prediction task is to determine whether a novel material possesses the target attribute (i.e., a classification task), the output layer can be a classifier (e.g., a sigmoid layer or a softmax layer), and the output result is usually a binary label indicating whether the novel material has or does not possess the target attribute. If the prediction task is to determine the attribute value of the target attribute of a novel material (i.e., a regression task), the output layer can be a regression layer (e.g., a linear layer), and the output result is a specific attribute value.

[0060] Finally, in step A303, after constructing the aforementioned material molecular property prediction model, it needs to be trained to achieve accurate prediction capabilities. The training process involves using the molecular geometric structure information of the novel material set as input and its corresponding attribute labels (i.e., the true markers of whether the material possesses the target attribute) as supervision signals. During training, the model iteratively adjusts all trainable parameters within the model (including the parameters of the molecular structure feature extraction module, fully connected layers, and attention fusion module) using optimization algorithms (such as Adam and SGD) based on the difference between the predicted results and the true labels (measured by a loss function). Through extensive training data and iterative optimization, the model can learn the complex nonlinear relationships between molecular structure feature information, deep-level attribute information, and the target attribute, thereby improving its prediction accuracy and generalization ability.

[0061] This application's solution addresses the issues of dimensionality mismatch and training instability in feature fusion by optimizing the model construction process, thereby improving prediction accuracy and efficiency. Specifically, when constructing the material molecular property prediction model, the last output layer of the matching model is first replaced with a fully connected layer. This converts the deep-level attribute information extracted from the molecular geometry of the novel material by the matching model into deep-level attribute feature vectors with the same dimension as the molecular structure feature vectors. This dimension-unified conversion effectively solves the problem of inconsistent feature dimensions from different sources, laying the foundation for subsequent feature fusion. Based on this, the output of the replaced matching model and the output of a molecular structure feature extraction module are connected to the same attention fusion module. This attention fusion module dynamically fuses the molecular structure feature vectors and deep-level attribute feature vectors through an attention mechanism, outputting a prediction result indicating whether the novel material possesses the target attribute or the attribute value of the target attribute of the novel material. This attention-based fusion method allows the model to adaptively adjust weights according to the importance of features, achieving efficient and intelligent feature integration and avoiding the inefficiency and information loss problems that may exist in traditional fusion methods. Meanwhile, the molecular structure feature extraction module is dedicated to extracting structural features, the fully connected layer is dedicated to attribute information transformation, and the attention fusion module is dedicated to intelligent fusion and prediction output. These synergistic effects enhance the model's ability to handle complex material structures. By training the entire material molecular property prediction model using the material molecular geometry information and attribute labels of a novel material set, effective learning based on the fused features is ensured, further stabilizing the training process and optimizing the prediction results. This method not only fully utilizes the deep attribute information guided by the physical mechanism obtained based on the DFT algorithm but also combines data-driven molecular structure feature information, enabling the model to understand material properties from multiple dimensions and levels, thereby significantly improving the accuracy and reliability of material molecular property prediction.

[0062] Preferably, in step A303, during the training of the material molecular property prediction model, the portion of the replaced matching model other than the fully connected layer is frozen.

[0063] Specifically, the part of the replaced matching model excluding the fully connected layer refers to the network layer used to extract deep-level attribute information that remains in the original matching model after replacing the last output layer of the matching model with a fully connected layer in step A301 (i.e., Figure 3 The deep attribute extraction network layers in the model are pre-trained on a large amount of data and possess the ability to extract high-dimensional, abstract features from the molecular geometry of materials. These layers have already learned rich physicochemical laws and material structure-property relationships during the pre-training phase (i.e., step A1). "Freezing" refers to setting the parameters (including weights and biases) of specific network layers to a non-updatable state during model training. This means that the parameters of these frozen layers will not be adjusted by the optimizer during backpropagation gradient calculation.

[0064] The proposed solution freezes all but the fully connected layers in the replaced matching model during the training of the material molecular property prediction model, ensuring the preservation and stability of the deep property extraction capabilities learned by the pre-trained model. In step A301, the output layer of the matching model is replaced with a fully connected layer. This fully connected layer, along with the output of the molecular structure feature extraction module, is connected to the attention fusion module to construct the material molecular property prediction model. The freezing operation keeps the underlying network structure responsible for extracting deep property information in the matching model unchanged, and its weights and biases are not modified during training. In this way, the model can continuously utilize its profound understanding of deep material properties obtained through pre-training on a large amount of multi-class material data. At the same time, the parameters of the newly added fully connected layer and the attention fusion module are updated according to the material molecular geometry information and attribute labels of the novel material set, enabling the model to effectively fuse the pre-trained deep property information with molecular structure feature information and make accurate predictions for specific target properties. This mechanism avoids unnecessary adjustments to the already learned feature extraction layer, thereby significantly improving training efficiency, reducing computational resource consumption, and helping to prevent overfitting of the model on limited new material data, allowing the model to focus more on learning the mapping relationship between target attributes and fused features.

[0065] In some implementations, step A4 includes:

[0066] A401. Obtain the molecular geometric structure information of the target novel material;

[0067] A402. Input the molecular geometric structure information of the target novel material into the trained molecular property prediction model to obtain the prediction result output by the trained molecular property prediction model. The prediction result indicates whether the target novel material has the target property or the attribute value of the target novel material.

[0068] Obtaining the molecular geometry information of the target novel material refers to acquiring molecular structure data of the material to be predicted through various means. This information is the foundation for predicting material properties, and its accuracy and completeness directly affect the reliability of the prediction results. Acquisition methods may include, but are not limited to: directly determining the molecular structure of the target novel material through experimental analysis, such as using X-ray diffraction, nuclear magnetic resonance spectroscopy, and mass spectrometry, and then digitally representing it; or using computational chemistry methods, such as using density functional theory (DFT) software to theoretically simulate the target novel material and generate its molecular structure data; or retrieving similar or related molecular structure information from existing publicly available material databases; or manually constructing or editing the molecular structure of the target novel material using a chemical structure editor.

[0069] The molecular geometry information of the target novel material is input into a trained molecular property prediction model. The model outputs a prediction result indicating whether the target novel material possesses the target property. This involves using the acquired molecular geometry information as input to the trained model for inference and calculation. The model processes and analyzes the input structural information based on its internal learning mechanism, ultimately outputting a prediction result. This prediction result is typically presented as a binary classification label or attribute value. The binary classification result indicates whether the target novel material possesses the target property, while the attribute value indicates the level of the target property.

[0070] This application's solution ensures high accuracy and standardization of input data for the prediction process by explicitly acquiring the molecular geometric structure information of the target novel material, thus avoiding prediction bias caused by missing, incomplete, or inconsistent data formats. Based on this, the precise structural information is input into a pre-trained material molecular property prediction model, enabling the model to perform efficient inference calculations based on reliable input. During training, the prediction model integrates molecular structural feature information extracted from the molecular geometric structure information of the novel material and deep-level property information extracted using a matching model through an attention mechanism. Therefore, when receiving structural information of the target novel material, it can fully utilize its inherent physical mechanisms and structure-property correlations for prediction. This explicit input acquisition and model execution process standardizes and automates the entire prediction process, greatly improving prediction efficiency and reliability, and effectively solving the problems of high ambiguity and poor reliability in prediction steps.

[0071] In some implementations, step A5 includes:

[0072] A501. If the prediction result indicates that the target novel material has the target attribute or the attribute value of the target novel material is higher than the preset value threshold, then the attention weights assigned by the attention fusion module to each structural feature item of the molecular structure feature vector are extracted to obtain the attention weight distribution information used as a reference.

[0073] A502. Based on the attention weight magnitude of each structural feature item in the attention weight distribution information used as a reference, determine the structural feature items that have a significant impact on the target attribute, and obtain the analysis results;

[0074] A503. Output an analysis report containing the analysis results to provide reference information for optimizing the target novel material.

[0075] Specifically, subsequent analysis steps are only executed when the material molecular property prediction model predicts a positive result for the target novel material, indicating that it possesses the target property, or when the value of the target property of the target novel material is higher than a preset numerical threshold. This ensures that resources are concentrated on analyzing materials that have been preliminarily identified as having potential, avoiding invalid analysis of materials that do not possess the target property or whose target property level is too low. Under this premise, this application obtains the weight values ​​assigned to each component (i.e., each structural feature term) of the molecular structure feature vector by the attention fusion module in the material molecular property prediction model when fusing the molecular structure feature vector and the deep attribute feature vector. These weight values ​​quantify the degree of attention or importance that the model pays to each structural feature term when making predictions. The attention weights can be extracted directly by accessing the weight matrix or attention score calculation results within the attention fusion module. For example, if the attention mechanism uses dot product attention, the weights obtained by normalizing the dot product of the query vector and the key vector using the softmax function can be extracted. Alternatively, the output weights of the attention layer can be obtained by calling a specific function after the model inference is completed through the application programming interface (API) or interface provided by the model. The extracted attention weights represent the attention weight value corresponding to each structural feature item in the molecular structure feature vector in a structured form (e.g., a vector, list, or dictionary), forming attention weight distribution information used as a reference. This information forms the basis for subsequent analysis, intuitively reflecting which molecular structural features the model believes contribute more to the prediction of the target attribute. For example, the extracted attention weights can be associated with corresponding structural feature item identifiers (such as atom index, bond type, group ID, etc.) to form a list of key-value pairs or a dictionary; or the attention weights can be directly stored as a numerical array, combined with a predefined feature item order, and indexed to correspond to each structural feature item.

[0076] Based on this, this application identifies molecular structural features that have a significant impact on the prediction of target properties by comparing the weight values ​​of different structural feature items, using the obtained attention weight distribution information. The larger the weight value, the more important the role of the structural feature item in the model prediction, and therefore the more significant its impact on the target property. Methods for determining significantly influential structural features include: setting a preset threshold and marking structural features with attention weights exceeding that threshold as significantly influential features; or ranking the attention weights and selecting the top N (N being a preset number) structural features as significantly influential features; or using statistical methods, such as calculating the mean and standard deviation of the weights, and identifying features corresponding to weights exceeding a certain standard deviation range as significant features. The identified molecular structural features that significantly influence the target property constitute the analysis results, which are key information providing specific guidance for material optimization. The analysis results can be a list containing the identifiers of the identified significant structural features and their corresponding attention weights, or a structured data object that, in addition to the identifiers and weights of the significant features, may also include a brief description of these features or their location information in the molecular structure. Finally, this application presents the analytical results obtained above in an easily understandable and viewable document format, outputting an analysis report containing the results. The analysis report not only includes a list of significant structural features, but may also include explanations of these features, visualizations (such as highlighting key parts on a molecular structure diagram), and suggestions for material optimization. For example, a PDF or HTML report can be generated, containing text descriptions, charts, and visualizations of the molecular structure, highlighting significantly influential parts; or a JSON or XML data file can be generated, which structurally contains the analytical results, facilitating further processing and presentation by other systems or tools. This analysis report provides reference information for optimizing the target novel material; that is, by identifying molecular structural features that significantly affect the target properties, it provides specific and actionable guidance for materials scientists or engineers to improve or design novel materials with superior performance. This reference information can guide the material's synthetic pathways, structural modifications, or component adjustments, for example, by providing suggestions directly in the analysis report, or by passing it as input to automated materials design or optimization platforms.

[0077] This application's solution, through the aforementioned steps, provides a precise and interpretable analysis mechanism based on a material molecular property prediction model. When the prediction model (e.g., a material molecular property prediction model constructed and trained according to methods A1 to A3 above) predicts that a target novel material possesses the target property or that the target property level is sufficiently high (above a preset numerical threshold), this method can delve into the attention fusion module within the model, extracting the attention weights assigned to each structural feature term of the molecular structure feature vector. These weights directly reflect the degree of attention the model pays to different parts of the molecular structure when making predictions. By analyzing this attention weight distribution information, for example, by identifying structural feature terms with higher weight values, this application can determine the molecular structural parts that significantly influence the target property. This method of directly obtaining interpretive information from within the model avoids the inaccuracy of external empirical judgments and ensures the reliability of the analysis results. Finally, these analysis results are compiled into an analysis report and output, providing materials scientists with clear and actionable reference information to guide them in targeted material structure optimization.

[0078] Please refer to Figure 2This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device includes a processor 301 and a memory 302. The processor 301 and the memory 302 are interconnected and communicate with each other via a communication bus 303 and / or other forms of connection mechanisms (not shown). The memory 302 stores a computer program executable by the processor 301. When the electronic device is running, the processor 301 executes the computer program to perform the material molecular property prediction method in any optional implementation of the above embodiments, to achieve the following functions: establishing deep-level property prediction models for multiple types of materials based on the DFT (Density Functional Theory) algorithm; the deep-level property prediction model is used to predict the deep-level property information of materials based on the material molecular geometric structure information; for a set of novel materials whose at least some property information has been determined, obtaining the material molecular geometric structure information and attribute labels of each novel material, and determining the deep-level property prediction model matching the set of novel materials as a matching model. A matching model is used; the attribute labels are used to indicate whether the novel material has the target attribute; a material molecular attribute prediction model is constructed based on the matching model, and the material molecular geometric structure information and attribute labels of the novel material set are used to train the material molecular attribute prediction model; the material molecular attribute prediction model fuses the molecular structure feature information extracted from the material molecular geometric structure information of the novel material and the deep attribute information of the novel material extracted by the matching model through an attention mechanism, and outputs the prediction result of whether the novel material has the target attribute or the attribute value of the target attribute of the novel material; using the trained material molecular attribute prediction model, the target novel material is predicted to have the target attribute; if the prediction result shows that the target novel material has the target attribute, the part of the molecular structure feature information that has a significant impact on the target attribute is analyzed according to the attention weight distribution information of the trained material molecular attribute prediction model, and the analysis result is output to provide reference information for optimizing the target novel material.

[0079] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it executes the material molecular property prediction method in any optional implementation of the above embodiments to achieve the following functions: Based on the DFT (Density Functional Theory) algorithm, it establishes deep-level property prediction models for multiple types of materials; the deep-level property prediction model is used to predict the deep-level property information of materials based on the material molecular geometric structure information; for a set of novel materials whose at least some property information has been determined, it obtains the material molecular geometric structure information and attribute tags of each novel material, and determines the deep-level property prediction model matching the novel material set as the matching model; the attribute tags are used to indicate whether the novel material has the target property; based on the matching model, it constructs a material molecular property prediction model and utilizes the... The molecular geometry information and attribute labels of the novel materials are used to train a molecular property prediction model. This model integrates molecular structure features extracted from the molecular geometry information of the novel materials with deep-level attribute information extracted using the matching model via an attention mechanism. The model outputs a prediction result indicating whether the novel material possesses a target attribute or the attribute value of that target attribute. Using the trained molecular property prediction model, the model predicts whether the target novel material possesses the target attribute. If the prediction result indicates that the target novel material possesses the target attribute, the model analyzes the significant portion of the molecular structure features that significantly influences the target attribute based on the attention weight distribution information of the trained molecular property prediction model, and outputs the analysis results to provide reference information for optimizing the target novel material. The computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0080] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for predicting the molecular properties of materials, characterized in that, The method includes the following steps: A1. Based on the DFT algorithm, deep-level property prediction models for multiple types of materials are established respectively; the deep-level property prediction models are used to predict the deep-level property information of materials based on the molecular geometric structure information of the materials; A2. For a set of novel materials for which at least some attribute information has been determined, obtain the molecular geometric structure information and attribute labels of each novel material, and determine a deep attribute prediction model that matches the set of novel materials as the matching model; the attribute labels are used to indicate whether the novel material has the target attribute, or to indicate the attribute value of the target attribute of the novel material. A3. Construct a material molecular property prediction model based on the matching model, and train the material molecular geometric structure information and attribute labels of the new material set; the material molecular property prediction model fuses the molecular structure feature information extracted from the material molecular geometric structure information of the new material and the deep attribute information of the new material extracted by the matching model through an attention mechanism, and outputs the prediction result of whether the new material has the target attribute or the attribute value of the target attribute of the new material; A4. Using the trained material molecular property prediction model, predict whether the target novel material has the target property or the property value of the target novel material; A5. If the prediction results indicate that the target novel material has the target attribute or the attribute value of the target novel material is higher than the preset value threshold, then based on the attention weight distribution information of the trained material molecular attribute prediction model, the part of the molecular structure feature information that has a significant impact on the target attribute is analyzed, and the analysis results are output to provide reference information for optimizing the target novel material. Step A3 includes: A301. Replace the last output layer of the matching model with a fully connected layer to obtain the replaced matching model; A302. Connect the output of the replaced matching model and the output of a molecular structure feature extraction module to the same attention fusion module to construct the material molecular property prediction model; the molecular structure feature extraction module is used to extract molecular structure feature vectors from the material molecular geometry information of the novel material; the fully connected layer is used to convert the deep attribute information extracted by the matching model from the material molecular geometry information of the novel material into deep attribute feature vectors with the same dimension as the molecular structure feature vectors; the attention fusion module is used to fuse the molecular structure feature vectors and the deep attribute feature vectors and output the prediction result of whether the novel material has the target attribute or the attribute value of the target attribute of the novel material; A303. The molecular property prediction model of the material is trained using the molecular geometric structure information and the attribute labels of the novel material set.

2. The method for predicting the molecular properties of materials according to claim 1, characterized in that, The aforementioned materials include OLED luminescent materials, alloy materials, and polymer materials; For OLED luminescent materials and polymer materials, the molecular geometric structure information includes atomic-level structure diagrams, group-level structure diagrams, and molecular-level structure diagrams; for alloy materials, the molecular geometric structure information includes microcrystalline lattice network diagrams. The deep-level attribute information includes at least one of electronic structure, energy state, energy level distribution, charge distribution, and vibrational mode.

3. The method for predicting the molecular properties of materials according to claim 1, characterized in that, Step A1 includes: A101. Obtain material molecular geometric structure information for multiple types of materials; A102. Based on the DFT algorithm, obtain the deep-level property information of the multiple types of materials; A103. Using material molecular geometric structure information as sample data and corresponding deep-level attribute information as label data, construct a sample dataset for each type of material; A104. Based on deep learning algorithms, construct deep property prediction models for each type of material, and train the deep property prediction models using the corresponding sample datasets.

4. The method for predicting the molecular properties of materials according to claim 1, characterized in that, In step A2, based on the determined attribute information, a deep attribute prediction model that matches the new material set is determined as the matching model.

5. The method for predicting the molecular properties of materials according to claim 1, characterized in that, In step A303, during the training of the material molecular property prediction model, the part of the replaced matching model except for the fully connected layer is frozen.

6. The method for predicting molecular properties of materials according to claim 1, characterized in that, Step A4 includes: A401. Obtain the molecular geometric structure information of the target novel material; A402. Input the molecular geometric structure information of the target novel material into the trained molecular property prediction model to obtain the prediction result output by the trained molecular property prediction model. The prediction result indicates whether the target novel material has the target property or the attribute value of the target novel material.

7. The method for predicting molecular properties of materials according to claim 1, characterized in that, Step A5 includes: A501. If the prediction result indicates that the target novel material has the target attribute or the attribute value of the target novel material is higher than the preset value threshold, then the attention weights assigned by the attention fusion module to each structural feature item of the molecular structure feature vector are extracted to obtain the attention weight distribution information used as a reference. A502. Based on the attention weight magnitude of each structural feature item in the attention weight distribution information used as a reference, determine the structural feature items that have a significant impact on the target attribute, and obtain the analysis results; A503. Output an analysis report containing the analysis results to provide reference information for optimizing the target novel material.

8. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program executable by the processor, which, when executing the computer program, performs the steps in the material molecular property prediction method as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it performs the steps of the material molecular property prediction method as described in any one of claims 1-7.

Images