Crystal material property prediction system and method based on graph network and denoising pre-training

Abstract

The invention discloses a crystal material property prediction system and method based on a graph network and de-noising pre-training, and belongs to the crossing field of biological materials and machine learning, the crystal material property prediction system comprises a structural feature extraction module, a de-noising pre-training module and a crystal material property prediction module, the method comprises the following steps: collecting a large amount of unlabeled crystal material structure data, including atomic types, atomic positions and lattice parameters, and preprocessing; the preprocessed data modeling graph structure is used for extracting local features, and the local features are fused with the extracted global features to obtain an original crystal material structure; and de-noising pre-training is carried out, the crystal material structure is reconstructed through the mask atom type, the disturbing atom position and the disturbing lattice parameter, and the property of the crystal material is predicted. According to the crystalline material property prediction system and method provided by the invention, the problem of insufficient precision of crystalline material property prediction caused by limited labeled data in practical application can be effectively solved, so that the discovery process of new materials in the medical field is accelerated.

Description

Technical Field

[0001] The present invention belongs to the cross - field of biomaterials and machine learning, and particularly relates to a crystal material property prediction system and method based on graph network and denoising pre - training. Background Art

[0002] Crystal materials play a crucial role in the sustainable development of society, especially in fields such as advanced manufacturing, nanotechnology, and the medical field. In the medical field, crystal materials play a vital role. For example, magnetic nanocrystal materials are used in magnetic resonance imaging; in terms of treatment applications, such as magnetic nanocrystal material hyperthermia for treating tumors and magnetic targeting drug carrier technology, and in vitro applications include magnetic separation, cell separation, purification, immunoassay, and magnetic infection. Another example is that scintillation crystals are applied in nuclear medicine imaging technology for the early diagnosis of diseases such as tumors and cardiovascular and cerebrovascular diseases; and molecularly imprinted photonic crystal hydrogel sensors prepared by combining photonic crystal materials and molecular imprinting technology can play a great application value in fields such as drug analysis, clinical diagnosis, biomedicine, toxicant analysis, and virus detection.

[0003] With the continuous growth of human demand for high - performance materials, accurately predicting the properties of crystal materials is crucial for the discovery and optimization of new materials, which can significantly shorten the R & D cycle and reduce experimental and computational costs.

[0004] Traditional machine learning methods learn patterns and rules from a large amount of data to predict crystal properties at a relatively low computational cost. The accuracy of predicting material properties is close to that of ab - initio methods, but the computational speed is several orders of magnitude faster. However, the effectiveness of machine learning methods applied to materials science depends to a large extent on the rational representation of materials and molecules. This representation must be encoded as a fixed - length vector to be compatible with machine learning algorithms, and the variability of crystal structure sizes makes this requirement more complex.

[0005] In recent years, deep learning methods have made remarkable progress in the field of materials science. For example, the Crystal Graph Convolutional Neural Network (CGCNN) proposed in Reference 1 (Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties, Phys. Rev. Lett, 2018, 120, 145301) encodes the crystal structure as a graph, with nodes representing atoms and the edges connecting the nodes as the bonding keys between atoms. Property prediction is performed based on the graph representation of the crystal material structure, which not only provides material property predictions with DFT accuracy but also provides atomic-level chemical insights. However, the number of element types and the complex interactions are more complex than the encoded graph. The extraction of feature information focuses on the constituent atoms and the overall structure, and there is insufficient learning of local structural features such as bonding keys. When predicting certain material properties that are highly sensitive to structural features (such as electronic features like band gaps), there are significant prediction biases.

[0006] Again, for example, the ALIGNN model enhances the modeling ability of inter-atomic interactions by introducing angular information; Matformer adopts a Transformer-based periodic feature extraction method. Although both further improve the accuracy of crystal material property prediction and perform well in prediction tasks for basic properties such as formation energy and band gap, these methods usually rely on a large amount of labeled data, which is mainly obtained through physical experiments or simulations based on first-principles calculations (DFT). Such labeled data has significant limitations, specifically manifested as limited data volume and uneven distribution among different crystal material properties. For example, in the JARVIS dataset, although it contains 55,714 labeled crystal material data, the labeled data for predicting shear modulus is only 4,664, far less than the massive datasets commonly seen in fields such as computer vision (such as ImageNet). In addition, the correlation between different crystal material properties is weak, making it difficult for traditional supervised learning methods to effectively extract meaningful features.

[0007] Therefore, how to make full use of a large amount of unlabeled crystal material structure data has become a key technical challenge in the current field of crystal material property prediction. Summary of the Invention

[0008] To solve the above technical problems, the present invention provides a crystal material property prediction system and method based on graph network and denoising pre-training. Based on a large amount of unlabeled crystal material structure data, through denoising pre-training technology, combined with graph network representation, it effectively learns the crystal material structure features, significantly improves the accuracy of crystal material property prediction, and accelerates the discovery of new materials.

[0009] To achieve the above invention object, an embodiment provides a crystal material property prediction system based on graph network and denoising pre-training, including: A structural feature extraction module, which is used to collect crystal material structure data and perform preprocessing. The crystal material structure data includes atomic type, atomic position and lattice parameters. Based on the preprocessed atomic type and atomic position, it is modeled into a graph structure, and the local features of the crystal material structure are extracted. At the same time, the global features of the crystal material structure are extracted based on the lattice parameters, and the local features and global features are fused to obtain the original crystal material structure; A denoising pre-training module, which is used to randomly select some atoms from the original crystal material structure and perform masking processing, predict the atomic type of the masked atoms. At the same time, random noise is added to the atomic position and lattice parameters in the original crystal material structure for denoising pre-training, and the reconstructed crystal material structure is obtained according to the predicted atomic type, the atomic position and lattice parameters after denoising pre-training; A crystal material property prediction module, which is used to input the reconstructed crystal material structure into a multi-layer fully connected network, complete the construction of the crystal material property model, and predict the crystal material properties.

[0010] In one embodiment, the crystal material structure data is collected by means of a crystal material database or computational simulation; The crystal material database includes the ICSD database, the Materials Project database or the JARVIS database; The computational simulation method is to generate crystal material structure data based on first-principles quantum calculations.

[0011] In one embodiment, the collection of crystal material structure data and preprocessing includes: cleaning and deduplication of the collected atomic type, atomic position and lattice parameters; Normalizing the atomic position in each crystal material structure; At the same time, the lattice parameters are standardized by dividing the lattice parameters by the maximum size of the crystal material structure, where the lattice parameters include lattice constants and lattice angles; And integer encoding or one-hot encoding is used for different types of atoms to convert the atomic type into a numerical form that can be processed by a machine learning model.

[0012] By normalizing the atomic positions in each crystal material structure, all atomic positions are located within a unified standardized range, where the standardized range is 0 to 1, to eliminate the deviations in the size and position of the crystal material structure and ensure that the atomic positions of different crystal material structures can be compared on the same scale; normalizing the lattice parameters can ensure that crystal materials of different sizes are uniformly processed, thereby standardizing the representation of the geometric characteristics of crystal materials; by encoding different types of atoms and converting them into a numerical form that can be processed by a machine learning model, the atomic types can be effectively represented and calculations can be performed. Through the above-mentioned preprocessing steps, the crystal material structure data can be normalized into a unified format, providing high-quality input for subsequent model training and prediction of crystal material properties.

[0013] In one embodiment, the preprocessed atomic types and atomic positions are modeled as a graph structure to extract the local features of the crystal material structure, including: constructing atomic nodes based on the preprocessed atomic types and atomic positions, and defining edges based on the Euclidean distance between atoms and periodic boundary conditions to form the graph structure of the crystal material, and using a graph convolutional neural network or a periodic graph neural network to extract the features of the graph structure as the local features of the crystal material structure.

[0014] Modeling the preprocessed atomic types and atomic positions as a graph structure and using a graph convolutional neural network or a periodic graph neural network to extract the local features between atoms can learn the influence of short-range interactions on the properties of crystal materials.

[0015] In one embodiment, extracting the global features of the crystal material structure based on the lattice parameters includes: extracting the global features of the crystal material structure through a multi-layer perceptron based on the lattice parameters.

[0016] In one embodiment, the comprehensive crystal material structure features include the spatial and geometric features of the crystal material.

[0017] In one embodiment, randomly selecting some atoms from the original crystal material structure and masking them, and predicting the types of the masked atoms, including: randomly selecting some atoms from the original crystal material structure, masking the types of the selected atoms, and using the remaining atomic information, the positional relationship of neighboring atoms, and the lattice parameters to predict the types of the masked atoms.

[0018] The method of modeling by masking atomic types can enhance the model's understanding of the chemical composition of crystal materials and effectively improve its prediction ability for the properties of crystal materials in low-data scenarios.

[0019] In one embodiment, adding random noise to the atomic positions and lattice parameters in the original crystal material structure for denoising pre-training includes: Add Gaussian-distributed random noise to the atomic positions in the structure of the original crystal material, and denoise and pre-train the atomic positions of the original crystal material structure through the atomic positions after adding random noise: , wherein, represents the atomic positions after adding random noise, is the scaling factor of the random noise, represents the random noise conforms to the Gaussian distribution, represents the position coordinates of the th atom in the structure of the original crystal material; Meanwhile, add Gaussian-distributed random noise to the lattice parameters in the structure of the original crystal material, and denoise and pre-train the lattice parameters of the original crystal material structure through the lattice parameters after adding random noise: , wherein, represents the lattice parameters after adding random noise, is the scaling factor of the random noise, is used to represent that the random noise conforms to the Gaussian distribution, represents the lattice parameters in the structure of the original crystal material.

[0020] By adding random noise conforming to the Gaussian distribution to the atomic coordinates, denoising pre-training is performed to restore the noisy atomic coordinates from the perturbed state to the original state, improving the adaptability to changes in the crystal material structure, helping to capture the influence of the microscopic vibrations of atoms in different materials on the overall structural stability, and thus improving the robustness of the atomic position denoising pre-training model to atomic position perturbations.

[0021] By adding random noise conforming to the Gaussian distribution to the lattice parameters, denoising pre-training is performed to restore the noisy lattice parameters to the true lattice parameters, which can improve the perception ability of the periodicity and structural distortion of the crystal material, and help the prediction model of the crystal material better adapt to the lattice characteristics of different materials, where the lattice parameters include the lengths and angles of three lattice vectors.

[0022] In one embodiment, the multi-layer fully connected network includes a hidden layer and an output layer; The hidden layer is used to perform non-linear transformation on the features of the reconstructed crystal material structure input by using a multi-layer perceptron, and combine batch normalization and Dropout to obtain complex feature representations.

[0023] The output layer is used to predict the properties of crystal materials through linear transformation based on complex feature representations, where the properties of crystal materials include one or more of formation energy, band gap, bulk modulus, and shear modulus.

[0024] Furthermore, the ReLU or GELU activation function is used in the hidden layer for non-linear transformation.

[0025] On the other hand, the present invention also provides a method for predicting the properties of crystal materials based on graph networks and denoising pre-training. The method for predicting the properties of crystal materials applies the system for predicting the properties of crystal materials based on graph networks and denoising pre-training, and includes the following steps: Collect crystal material structure data and perform preprocessing. Model it as a graph structure based on the preprocessed atomic types and atomic positions, extract the local features of the crystal material structure, and at the same time extract the global features of the crystal material structure based on the lattice parameters. Fusion of the local features and the global features gives the original crystal material structure; Randomly select some atoms from the original crystal material structure and perform masking processing to predict the atomic types of the masked atoms. At the same time, add random noise to the atomic positions and lattice parameters in the original crystal material structure for denoising pre-training. Obtain the reconstructed crystal material structure according to the predicted atomic types, atomic positions and lattice parameters after denoising pre-training, and form a pre-training model of the crystal material structure; Use the mean square error as the loss function, and combine ADAMW to iteratively optimize the parameters of the pre-training model of the crystal material structure. At the same time, use the K-fold cross-validation method and combine L1 regularization to optimize the hyperparameters of the pre-training model of the crystal material structure, and form a prediction model for the properties of crystal materials to predict the properties of crystal materials.

[0026] In one embodiment, the use of the K-fold cross-validation method includes: using the preprocessed crystal material structure data as the input training set, dividing the input training set into several subsets, selecting one subset as the validation subset, and the remaining subsets as the training subsets. Train with the training subsets and calculate the mean square error on the validation subset, and iterate multiple times to optimize the hyperparameters of the pre-training model of the crystal material structure.

[0027] In one embodiment, the evaluation index of the prediction model for the properties of crystal materials is the mean square error deviation between the predicted value and the true value of the properties of crystal materials : , where is the predicted value of the properties of crystal materials, is the true value of the properties of crystal materials, is the number of samples of the properties of crystal materials.

[0028] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention constructs the crystal material characterization based on the graph neural network. By modeling the crystal structure as graph data, where nodes represent atoms and edges represent the interactions between atoms, it can effectively utilize a large amount of unlabeled crystal material structure data for characterization learning, significantly improving the prediction accuracy and generalization ability of the crystal material property prediction model.

[0029] (2) Based on the graph network, denoising pre-training for masked atom type recovery, perturbed atom position reconstruction, and lattice parameter perturbation correction is respectively performed, which can learn the essential features from the crystal material structure and effectively address the problems of data missing and noise in crystal material data. Thereby enhancing the modeling ability of the crystal material property prediction model for the complex relationships between different properties and improving the stability and accuracy of the crystal material property prediction results.

[0030] (3) The crystal material property prediction system and method provided by the present invention have strong compatibility with downstream tasks and can be applied to various crystal material structure modeling methods, such as CGCNN, ALIGNN, Matformer, CATGNN, or CrysDiff, and can be widely applied to fields such as drug analysis, clinical diagnosis, biomedicine, toxicant analysis, and virus detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic structural diagram of the crystal material property prediction system based on the graph network and denoising pre-training provided by the present invention.

[0032] Figure 2 It is a schematic diagram of the crystal material unit structure provided by the present invention.

[0033] Figure 3 It is a schematic flowchart of the crystal material property prediction method based on the graph network and denoising pre-training provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] The present invention will be further described in detail below with reference to the drawings and embodiments. It should be noted that the following embodiments are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0035] A crystal material property prediction system based on a denoising pre-training framework, as Figure 1 shown, includes a structural feature extraction module, a denoising pre-training module, and a crystal material property prediction module.

[0036] The structural feature extraction module is used to collect crystal material structure data and perform preprocessing. The crystal material structure data includes atomic types, atomic positions, and lattice parameters. Based on the preprocessed atomic types and atomic positions, it is modeled as a graph structure to extract local features of the crystal material structure. At the same time, based on the lattice parameters, global features of the crystal material structure are extracted, and the local features and global features are fused to obtain comprehensive crystal material structure features, forming the original crystal material structure.

[0037] In the embodiments, the crystal material structure data usually comes from publicly available crystal material databases such as ICSD (Inorganic Crystal Structure Database), Materials Project, or JARVIS, or is generated by computational simulation methods (such as quantum calculations based on first principles).

[0038] To ensure the quality and uniformity of the data, preprocessing of the crystal material structure data is required, such as Figure 2 The following shows a schematic diagram of the unit structure of the crystal material, which specifically includes the following steps: (1) Standardization of atomic positions: Normalize the atomic positions in each crystal material structure so that all atomic positions are within a unified standard range (specifically 0 to 1) to eliminate the deviation in the size and position of the crystal material structure and ensure that the atomic positions of different crystal material structures can be compared on the same scale.

[0039] (2) Standardization of lattice parameters: Standardize the lattice parameters (including lattice constants and lattice angles) by dividing the lattice parameters by the maximum size of the crystal material structure to ensure that different-sized crystal materials are uniformly processed in the model, thereby standardizing the representation of the geometric features of the crystal material.

[0040] (3) Encoding of atomic types: Use integer encoding or one-hot encoding for different types of atoms, encoded into 95 types, and convert them into a numerical form that can be processed by a machine learning model to effectively represent atomic types and perform calculations in the model.

[0041] Through the above preprocessing steps, the crystal material structure data can be standardized into a unified format, providing high-quality input for subsequent model training and prediction.

[0042] Next, based on the preprocessed crystal material structure data, a model based on graph neural network (GNN) is used to extract the features of the crystal material structure. Specifically: Based on the preprocessed atomic types and atomic positions as node feature inputs, atomic nodes are constructed, and edges are defined based on the Euclidean distance between atoms and periodic boundary conditions. Then, the crystal material structure is modeled as a graph, and the interactions between nodes are represented by edges. A graph convolutional neural network (GCN) or a periodic graph neural network, such as ALIGNN, is used to extract the features of the graph structure as the local features of the crystal material structure to learn the influence of short-range interactions on the properties of the crystal material.

[0043] At the same time, based on the lattice parameters, a multi-layer perceptron (MLP) is used to extract the global features of the crystal material structure. The lattice parameters reflect the macroscopic geometric features of the crystal material. The local features and global features are fused to obtain the comprehensive crystal material structure features. The comprehensive crystal material structure features include the spatial and geometric features of the crystal material, which can effectively support the subsequent crystal material property prediction tasks. Then, the original crystal material structure is formed for subsequent model use. Through the structure feature extraction module, the local and global features of the crystal material structure can be comprehensively captured, providing a high-quality feature representation for the crystal material property prediction.

[0044] The denoising pre-training module is used to randomly select some atoms from the original crystal material structure and perform masking processing, predict the types of the masked atoms. At the same time, random noise is added to the atomic positions and lattice parameters in the original crystal material structure for denoising pre-training, and the reconstructed crystal material structure is obtained according to the predicted atomic types, the atomic positions and lattice parameters after denoising pre-training.

[0045] The main task of the denoising pre-training module is to pre-train the model through various denoising tasks so that it can learn the essential features of the crystal material structure. In the embodiment, the denoising tasks include masked atomic type modeling, atomic position denoising, and lattice parameter denoising. Specifically, (1) Masked atomic type modeling: Randomly select a part of the atoms in the original crystal material structure, mask their types as "unknown", and then use the remaining atomic information, the positional relationship of neighboring atoms, and the lattice parameters to predict the types of the masked atoms : , where, represents the atomic number; is an indicator. When the atom is masked, otherwise it is 0.

[0046] This method can enhance the model's understanding of the chemical composition of the crystal material and effectively improve its prediction ability in low-data scenarios.

[0047] (2) Atomic position denoising: Add Gaussian-distributed random noise to the atomic positions in the original crystal material structure for denoising pre-training, so that the atomic coordinates with added noise are restored from the perturbed state to the original state, thereby improving the robustness of the atomic position denoising pre-training model to atomic position perturbations. The mathematical formula is expressed as: , where represents the atomic position after adding random noise, is the scaling factor of the random noise, represents the random noise conforms to the Gaussian distribution, represents the position coordinate of the -th atom in the original crystal material structure.

[0048] This method helps to capture the influence of the microscopic vibrations of atoms in different materials on the overall structural stability.

[0049] (3) Lattice parameter denoising: Add random noise to the lattice parameters, and predict the lattice parameters of the original crystal material structure through the lattice parameter denoising pre-training with the added random noise, thereby improving the recovery ability of the lattice parameter denoising pre-training model for lattice geometric features. The mathematical formula is expressed as: , where represents the lattice parameters after adding random noise, is the scaling factor of the random noise, represents the random noise conforms to the Gaussian distribution, represents the lattice parameters in the original crystal material structure, where the lattice parameters include the lengths and angles of three lattice vectors.

[0050] This method can improve the model's perception ability of the periodicity and structural distortion of crystal materials and help it better adapt to the lattice characteristics of different materials. Through the above three denoising tasks, the essential features of the crystal material structure can be effectively learned, providing high-quality feature representations for subsequent crystal material property prediction tasks.

[0051] The crystal material property prediction module is used to input the reconstructed crystal material structure into a multi-layer fully connected network to complete the construction of the crystal material property model and predict the crystal material properties.

[0052] In the embodiment, in the crystal material property prediction module, the reconstructed crystal material structure is input into a multi-layer fully connected network to predict the properties of the crystal material structure features after denoising pre-training. Feature transformation is performed through a non-linear activation function (such as RELU or GELU), and finally the properties of the crystal material are output, specifically including the following steps: Input the reconstructed crystal material structure features output by the denoising pre-training module into a multi-layer fully connected network, and perform feature transformation layer by layer. The mapping function of the multi-layer fully connected network is a multi-layer perceptron (MLP), and each layer is mapped through a non-linear activation function (such as RELU or GELU) to obtain a more complex feature representation; Prediction output: At the output layer, the properties of the crystal material are predicted through a linear transformation, such as formation energy, band gap, bulk modulus, shear modulus, etc. The mathematical formula is expressed as: , where, is the reconstructed crystal material structure features output by the denoising pre-training module, is the mapping function of the multi-layer fully connected network, is the predicted crystal material property.

[0053] In this embodiment, a crystal material property prediction method based on a graph network and denoising pre-training is also provided. The crystal material property prediction method applies the crystal material property prediction system based on the graph network and denoising pre-training, as Figure 3 shown, including the following steps: (1) Collect a large amount of crystal material structure data and perform preprocessing to ensure the availability of the data after normalization and encoding. Model the preprocessed atomic types and atomic positions as a graph structure, extract the local features of the crystal material structure, and at the same time extract the global features of the crystal material structure based on the lattice parameters, and fuse the local features and the global features to obtain the original crystal material structure; (2) Randomly select some atoms from the original crystal material structure and perform masking processing to predict the atomic types of the masks. At the same time, add random noise to the atomic positions and lattice parameters in the original crystal material structure for denoising pre-training, and obtain the reconstructed crystal material structure according to the predicted atomic types, the atomic positions and lattice parameters after denoising pre-training, and form a pre-training model of the crystal material structure.

[0054] (3) After completing the denoising pre-training, fine-tune the pre-training model of the crystal material structure so that it can be better applied to the crystal material property prediction task. Specifically, when fine-tuning, the mean square error (MSE) is used as the loss function: , where, is the predicted value of the crystal material property, is the true value of the crystal material property, The number of samples of the crystal material properties is used, and the parameters of the pre-trained model for optimizing the crystal material structure are iteratively optimized by ADAMW. At the same time, the K-fold cross-validation method is adopted and combined with L1 regularization to optimize the hyperparameters of the pre-trained model for the crystal material structure, forming a prediction model for the crystal material properties to predict the crystal material properties. In this embodiment, five-fold cross-validation is selected for evaluation to ensure that the model has good generalization ability.

[0055] To verify the effectiveness of the crystal material property prediction system and method proposed by the present invention, a series of experimental verifications were carried out on the Materials Project benchmark dataset in the embodiment. The quantitative results are shown in Table 1, with the mean absolute error (MAE) as the index. The best results are shown in bold, and the sub-optimal results are underlined. γ is the masking ratio of atomic types: Table 1

[0056] From the results in Table 1, it can be seen that the crystal material property prediction system provided by the present invention significantly outperforms the baseline model Matformer in three of the four subtasks of the Materials Project benchmark task by adopting different masking ratios of atomic types. It is worth noting that due to the fact that the Bulk Moduli task requires learning the key factors affecting this property from limited data, the improvement of all previous methods in this task has always been very limited. By pre-training on a large number of unlabeled crystal structures, the crystal material property prediction system and method provided by the present invention already have the ability to capture the crystal structure features and show stronger robustness.

[0057] Based on the above, the crystal material property prediction system and method based on graph network and denoising pre-training provided by the present invention can accurately predict various properties of crystal materials (such as formation energy, band gap, bulk modulus, etc.) by analyzing crystal material structure data (such as atomic type, atomic position, and lattice constant) in practical applications. Based on the prediction results, researchers can more efficiently screen and design new materials, and optimize the material properties accordingly, thus accelerating the process of discovering new materials and promoting technological innovation in fields such as advanced manufacturing, nanotechnology, and the medical field, especially in fields such as drug analysis, clinical diagnosis, biomedicine, toxicant analysis, and virus detection in the medical field, and playing a huge application value.

[0058] The above-described embodiments have detailed the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, supplements, and equivalent replacements made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A crystal material property prediction system based on graph network and denoising pre-training, characterized in that Including: A structural feature extraction module, which is used to collect crystal material structure data and perform preprocessing. The crystal material structure data includes atomic types, atomic positions, and lattice parameters. Based on the preprocessed atomic types and atomic positions, it is modeled into a graph structure, local features of the crystal material structure are extracted. At the same time, global features of the crystal material structure are extracted based on the lattice parameters, and the local features and global features are fused to obtain comprehensive crystal material structure features, forming the original crystal material structure; A denoising pre-training module, which is used to randomly select some atoms from the original crystal material structure and perform masking processing, predict the atomic types of the masked atoms. At the same time, random noise is added to the atomic positions and lattice parameters in the original crystal material structure for denoising pre-training, and the reconstructed crystal material structure is obtained according to the predicted atomic types, the atomic positions and lattice parameters after denoising pre-training; A crystal material property prediction module, which is used to input the reconstructed crystal material structure into a multi-layer fully connected network, complete the construction of the crystal material property model, and predict the crystal material properties.

2. The crystal material property prediction system according to claim 1, wherein The above-mentioned preprocessing includes: cleaning and de-duplicating the collected atomic types, atomic positions, and lattice parameters; Normalizing the atomic positions in each crystal material structure; At the same time, by dividing the lattice parameters by the maximum size of the crystal material structure, standardization processing of the lattice parameters is realized, where the lattice parameters include lattice constants and lattice angles; And integer encoding or one-hot encoding is adopted for different types of atoms to convert the atomic types into numerical forms that can be processed by machine learning models.

3. The crystal material property prediction system according to claim 1, characterized in that, The above-mentioned modeling into a graph structure based on the preprocessed atomic types and atomic positions to extract local features of the crystal material structure includes: constructing atomic nodes based on the preprocessed atomic types and atomic positions, and defining edges based on the Euclidean distance between atoms and periodic boundary conditions to form the graph structure of the crystal material, and using a graph convolutional neural network or a periodic graph neural network to extract the features of the graph structure as local features of the crystal material structure.

4. The crystal material property prediction system according to claim 1, wherein The above-mentioned randomly selecting some atoms from the original crystal material structure and performing masking processing and predicting the atomic types of the masked atoms includes: randomly selecting some atoms from the original crystal material structure, masking the selected atomic types, and using the remaining atomic information, the positional relationship of neighboring atoms, and the lattice parameters to predict the atomic types of the masked atoms.

5. The crystal material property prediction system according to claim 1, wherein The above-mentioned adding random noise to the atomic positions and lattice parameters in the original crystal material structure for denoising pre-training includes: Adding random noise with a Gaussian distribution to the atomic positions in the original crystal material structure, and predicting the atomic positions of the original crystal material structure through denoising pre-training with the atomic positions after adding random noise: ， Among them, represents the atomic position after adding random noise, is the scaling factor of the random noise, represents the random noise conforms to the Gaussian distribution, represents the position coordinates of the th atom in the original crystal material structure; At the same time, adding random noise with a Gaussian distribution to the lattice parameters in the original crystal material structure, and predicting the lattice parameters of the original crystal material structure through denoising pre-training with the lattice parameters after adding random noise: ， Among them, represents the lattice parameter after adding random noise, is the scaling factor of the random noise, represents the random noise conforms to a Gaussian distribution, represents the lattice parameter in the original crystal material structure.

6. The crystal material property prediction system according to claim 1, wherein The above-mentioned multi-layer fully connected network includes a hidden layer and an output layer; The hidden layer is used to perform non-linear transformation on the features of the input reconstructed crystal material structure by using a multi-layer perceptron, and combine batch normalization and Dropout to obtain complex feature representations; The output layer is used to predict crystal material properties through linear transformation based on complex feature representations, where the crystal material properties include one or more of formation energy, band gap, bulk modulus, and shear modulus.

7. The crystal material property prediction system according to claim 6, wherein The RELU or GELU activation function is used in the hidden layer for non-linear transformation.

8. A method for predicting the properties of crystal materials based on graph networks and denoising pre-training, characterized in that Applying the crystal material property prediction system based on graph network and denoising pre-training according to any one of claims 1 to 7, comprising the following steps: Collect crystal material structure data and perform preprocessing, model it into a graph structure based on the preprocessed atomic types and atomic positions, extract local features of the crystal material structure, and at the same time extract global features of the crystal material structure based on lattice parameters, and fuse the local features and global features to obtain the original crystal material structure; Randomly select some atoms from the original crystal material structure and perform masking processing, predict the types of the masked atoms, and at the same time add random noise to the atomic positions and lattice parameters in the original crystal material structure for denoising pre-training to obtain a reconstructed crystal material structure, forming a pre-training model of the crystal material structure; Use the mean square error as the loss function, and combine ADAMW to iteratively optimize the parameters of the pre-training model of the crystal material structure. At the same time, use the K-fold cross-validation method and combine L1 regularization to optimize the hyperparameters of the pre-training model of the crystal material structure, form a crystal material property prediction model, and predict crystal material properties.

9. The method for predicting the properties of a crystal material according to claim 8, wherein The said K-fold cross-validation method includes: using the preprocessed crystal material structure data as the input training set, dividing the input training set into several subsets, selecting one of the subsets as the validation subset, and the remaining subsets as the training subsets, training with the training subsets and calculating the mean square error on the validation subset, and iterating multiple times to optimize the hyperparameters of the pre-training model of the crystal material structure.

10. The method for predicting the properties of a crystal material according to claim 8, characterized in that, The evaluation index of the crystal material property prediction model is the mean square error between the predicted value and the true value of the crystal material property : ， Among them, is the predicted value of the properties of the crystal material, is the true value of the properties of the crystal material, is the number of samples of the properties of the crystal material.

Images