Scintillator material performance prediction method and system based on artificial intelligence

Abstract

The invention relates to a scintillator material performance prediction method and system based on artificial intelligence. The method comprises the following steps: firstly, collecting basic characterization parameters and corresponding performance data of a scintillator material, establishing one-to-one mapping after first and second preprocessing, and constructing a mapping data set; dividing the data set into a training set, a verification set and a test set, and constructing a mixed type artificial neural network model containing submodules such as matrix material dynamic clustering to capture a doped ion regulation rule; and then training, verifying and testing the model based on the training set, the verification set and the test set, and finally inputting basic characterization parameters of a to-be-predicted scintillator material into the trained model after first preprocessing so as to output performance data of the scintillator material. Compared with the prior art, the method has the advantages of accuracy, rapidness, wide applicable material range and the like.

Description

Technical Field

[0001] This invention relates to the field of material performance prediction technology, and in particular to a method and system for predicting the performance of scintillator materials based on artificial intelligence. Background Technology

[0002] Scintillator materials can convert high-energy particles and rays into visible and ultraviolet light. Based on scintillator materials, high-energy ray or particle detectors can be fabricated, which have important application prospects in national defense, military, and medical detection. Therefore, accelerating the research and development of new high-performance scintillator materials is extremely crucial.

[0003] The performance of scintillator materials is mainly determined by two indicators: light yield and decay time. Traditional methods for developing scintillator materials primarily involve designing the matrix, designing different dopant ions, synthesizing the materials, and then characterizing their properties to obtain scintillator materials with varying performance characteristics. However, the wide variety of matrix materials used in scintillators, the complex modulation effects of dopant ions, and the high costs of material synthesis and characterization result in long development cycles for scintillator materials, hindering the rapid development of novel scintillator materials.

[0004] In recent years, artificial intelligence technology has been widely applied in data-driven research and development of new materials. Because different scintillator materials and dopant ions have different mechanisms of influence on light yield and decay time, early machine learning algorithms were only relatively accurate for predicting the performance of a single type of scintillator material and were difficult to generalize to other types. Although deep learning technology based on multilayer neural networks can automatically extract deep features from input data and improve prediction performance across different material types, the physical mechanisms between different matrices and dopant ions in scintillator materials are drastically different. Even trace amounts of doping can cause changes in light yield and decay time by 5 to 6 orders of magnitude, making it difficult for simple multilayer fully connected neural networks to learn the extremely complex control laws and meet the requirements of both accurate and rapid prediction of scintillator material performance.

[0005] Therefore, how to propose a new method that is accurate, fast, and applicable to a wide range of materials for the key performance indicators of scintillator materials has become a technical challenge restricting the rapid development of high-performance scintillator materials. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology by providing an artificial intelligence-based method and system for predicting the performance of scintillator materials.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] According to one aspect of the present invention, an artificial intelligence-based method for predicting the properties of scintillator materials is provided, characterized in that the method includes the following steps:

[0009] S1. Collect basic characterization parameters of multiple sets of scintillator materials, use them as input data after the first preprocessing process, collect the corresponding scintillator material performance data, use them as output data after the second preprocessing process, establish a one-to-one mapping relationship between the preprocessed input data and the output data, and construct a scintillator material mapping dataset.

[0010] S2. Construct a hybrid artificial neural network model, which includes a matrix material dynamic clustering submodule, a dopant ion nonlinear conversion submodule, a matrix material and dopant ion interaction submodule, a feature fusion and attention mechanism submodule, and a fully connected network submodule. The submodules work together to capture the regulation law of dopant ions on the performance of scintillator materials.

[0011] S3. Based on the scintillator material mapping dataset, train and validate the constructed hybrid artificial neural network model to obtain the trained hybrid artificial neural network model;

[0012] S4. After the basic characterization parameters of the scintillator material to be predicted are processed through the first preprocessing step, they are input into the trained hybrid artificial neural network model, and the performance data of the scintillator material is output.

[0013] As a preferred technical solution, the basic characterization parameters in S1 include the material type, the chemical formula of the matrix material, and the chemical formula of the dopant ions.

[0014] As a preferred technical solution, the first preprocessing process in S1 specifically includes: a vectorization encoding process and a dimensionality reduction process;

[0015] The vectorization coding process includes the vectorization coding process for material categories, the vectorization coding process for the chemical formula of the matrix material, and the vectorization coding process for the chemical formula of the dopant ions;

[0016] The specific process of vectorizing material categories is as follows:

[0017] According to the preset mapping rules, the material category field in each data entry is converted into a unique thermal vector of the material category;

[0018] The vectorization encoding process for the basic material chemical formula specifically includes:

[0019] First, based on all elements in the periodic table, an ordered vector is constructed as the initial embedding vector. Then, natural language processing is used to parse the chemical formula of the base material, and regular expressions are used to identify the element symbols and their corresponding molar ratios in the chemical formula. The extracted molar ratios are then normalized and logarithmically or exponentially processed to obtain the first weighted ratio value. Based on this first weighted ratio value, values ​​are assigned to the corresponding positions of the initial embedding vector, and the remaining non-corresponding positions are assigned zero, thus obtaining the high-dimensional sparse vector of the matrix material.

[0020] The vectorization encoding process for the chemical formula of doped ions specifically includes:

[0021] First, based on all elements in the periodic table, an ordered vector is constructed as the initial embedding vector. Then, natural language processing is used to parse the chemical formula of the doped ions into strings. Regular expressions are used to identify the element symbols and their corresponding molar ratios in the chemical formula of the doped ions. The extracted molar ratios are then transformed by a power function to obtain a second weighted ratio value. Based on this first weighted ratio value, values ​​are assigned to the corresponding positions in the initial embedding vector, and the remaining non-corresponding positions are assigned zero, thus obtaining a high-dimensional sparse vector of the doped ions.

[0022] The dimensionality reduction process specifically includes:

[0023] Using a dimensionality reduction algorithm, the high-dimensional sparse vectors of the matrix material and the dopant ions are reduced in dimensionality. The high-dimensional sparse vectors are then mapped to a new set of feature vectors through linear or nonlinear transformations, thereby outputting the dense vectors of the matrix material and the dense vectors of the dopant ions.

[0024] The second preprocessing step in S1 specifically includes: adding a minimal constant term for bias, logarithmic transformation, and normalization or standardization.

[0025] As a preferred technical solution, in the hybrid artificial neural network model constructed in S2, the matrix material dynamic clustering submodule is used to structurally enhance the dense vectors of the matrix material. Its specific processing includes:

[0026] The dense vector of the matrix material is input into the dynamic clustering submodule of the matrix material. Then, multiple initial cluster centers are introduced into this submodule as reference points. By calculating the similarity between the current input vector and each cluster center, the clustering feature vector representing the material is generated and output.

[0027] In the training process of hybrid artificial neural networks, the initial cluster centers are used as weights and dynamically optimized through gradient backpropagation to realize the learnable function of the cluster centers and dynamically adjust the cluster centers to which different input vectors belong; the initial cluster centers are selected by random normal distribution or linear distribution method.

[0028] As a preferred technical solution, in the hybrid artificial neural network model constructed in S2, the dopant ion nonlinear conversion submodule consists of one or more fully connected networks, used to extract the features of the dopant ion chemical formula. Its specific processing includes:

[0029] The dense vector of doped ions is input into the doped ion nonlinear transformation submodule, and after activation function, batch normalization and discarding, the doped ion nonlinear transformation vector is output.

[0030] As a preferred technical solution, in the hybrid artificial neural network model constructed in S2, the matrix material and dopant ion interaction submodule is used to perform dimension alignment operations on the matrix material dense vector and to perform interactive processing on the matrix material dense vector and the dopant ion dense vector. The specific processing includes:

[0031] Input the matrix material density vector and the dopant ion density vector into the matrix material and dopant ion interaction submodule;

[0032] By using a fully connected neural network, the matrix material dense vector is mapped to the same dimensional space as the dopant ion dense vector, thus obtaining the dimension-aligned matrix material dense vector;

[0033] Element-wise multiplication is used to multiply the dimension-aligned matrix material dense vector and the dopant ion dense vector element-wise to form a linear interaction vector;

[0034] One fully connected network is used to extract the nonlinear interaction information between the dimension-aligned matrix material dense vector and the doped ion dense vector to form a nonlinear interaction vector; or a multi-layer fully connected network is used to convert the linear interaction vector into a nonlinear interaction vector and output the nonlinear interaction vector.

[0035] As a preferred technical solution, in the hybrid artificial neural network model constructed in S2, the feature fusion and attention mechanism submodule is used to fuse vectors and extract attention feature vectors. Its specific processing includes:

[0036] The material category unique thermal vector, cluster feature vector, doped ion nonlinear transformation vector, and nonlinear interaction vector are input into the feature fusion and attention mechanism submodule and fused to obtain the initial fusion vector.

[0037] The Reshape algorithm is used to convert the initial fusion vector into the input format of the attention mechanism, thus obtaining the fusion input vector of the attention mechanism.

[0038] By utilizing a multi-head attention mechanism, we extract the attention feature vector from the fused input vector.

[0039] The attention feature vector is concatenated with the initial fusion vector again to output the final fusion vector.

[0040] As a preferred technical solution, in the hybrid artificial neural network model constructed in S2, the fully connected network sub-module consists of multiple layers of fully connected artificial neural networks. The first layer of the artificial neural network receives the final fusion vector from the attention module, and subsequent layers adopt fully connected layers or residual connection structures, and are optimized by combining regularization, normalization and activation functions.

[0041] As a preferred technical solution, the specific process of training, validating, and testing the constructed hybrid artificial neural network model in S3 includes:

[0042] S31. Divide the scintillator material mapping dataset into a training set, a validation set, and a test set;

[0043] S32. Define the set of hyperparameter values ​​to be searched, and use the network search method, with mean squared error as the loss function, to select the hyperparameter combination with the smallest error based on the training set.

[0044] S33. Using the selected optimal hyperparameter combination, the model is validated and trained based on the validation set using a multi-fold cross-validation method.

[0045] The test set is not involved in the training and validation process; it is only used to evaluate the model after training is completed.

[0046] According to another aspect of the present invention, an artificial intelligence-based scintillator material performance prediction system is provided, the system comprising a data acquisition module, a data input module, a data preprocessing module, a hybrid artificial neural network model, a network optimization module, and a visualization module;

[0047] The data acquisition module is used to collect basic characterization parameters and corresponding scintillator material performance data of multiple sets of scintillator materials. In conjunction with the data preprocessing module, the basic characterization parameters are used as input data after a first preprocessing process, and the material performance data are used as output data after a second preprocessing process. A one-to-one mapping relationship is established between the preprocessed input data and the output data to construct a scintillator material mapping dataset.

[0048] The data input module is used to input the basic characterization parameters of the scintillator material to be predicted into the trained hybrid artificial neural network model;

[0049] The hybrid artificial neural network model is used to capture the regulatory mechanism of dopant ions on the performance of scintillator materials. It includes a matrix material dynamic clustering submodule, a dopant ion nonlinear conversion submodule, a matrix material and dopant ion interaction submodule, a feature fusion and attention mechanism submodule, and a fully connected network submodule.

[0050] The network optimization module trains and validates the constructed hybrid artificial neural network model based on the scintillator material mapping dataset, and obtains the trained hybrid artificial neural network model.

[0051] The visualization module is connected to the trained hybrid artificial neural network model and is used to visualize the output of the trained hybrid artificial neural network model.

[0052] The data preprocessing module is a dual-input processing logic. It performs preprocessing on the dataset from the data acquisition module and outputs it to the dataset partitioning module. For the parameters to be predicted in the data input module, it performs the same preprocessing process as the training data and outputs it to the optimized model to ensure data distribution consistency.

[0053] As a preferred technical solution, the system also includes a dataset partitioning module, which divides the preprocessed dataset of basic characterization parameters of the scintillator material into a training set, a validation set, and a test set. The training set is used to train the network model, the validation set is used to validate the trained network model, and the test set is used for the final evaluation of the trained network model.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. In this invention, the scintillator material performance prediction method is based on a hybrid artificial neural network model. Through the model's dynamic clustering of matrix materials, nonlinear conversion of dopant ions, interaction between matrix materials and dopant ions, feature fusion and attention mechanisms, and fully connected networks, it achieves accurate prediction of the light yield and decay time of different types of scintillator materials. Because it only requires training and inputting the basic characterization parameters of the scintillator material to be predicted into the model, it is applicable to a wide variety of matrix materials and dopant ions. Furthermore, after training with a large amount of data, the model can accurately capture the control patterns, resulting in high prediction accuracy. This method is accurate, fast, and applicable to a wide range of materials for scintillator material performance prediction.

[0056] 2. In this invention, a first preprocessing and a second preprocessing are employed. The first preprocessing process transforms various parameters into a vector form that the model can process through vectorization encoding, reducing data redundancy and improving overall computational efficiency through dimensionality reduction. The second preprocessing process optimizes the distribution characteristics of performance data through bias, logarithmic transformation, and normalization, avoiding the impact of differences in data volume on model training accuracy. The overall preprocessing process ensures both the integrity and validity of the data and improves the training efficiency and prediction accuracy of the subsequent model.

[0057] 3. In this invention, the matrix material dynamic clustering submodule overcomes the limitations of traditional fixed clustering by optimizing initial cluster centers and gradient backpropagation. During training, it dynamically optimizes cluster centers, adaptively capturing structural differences in different matrix materials and generating more discriminative cluster feature vectors. This design makes the model more adaptable to novel or rare matrix materials, avoiding insufficient feature extraction due to special matrix types, and further improving the model's compatibility with different matrix materials.

[0058] 4. In this invention, the doped ion nonlinear transformation submodule employs a multilayer fully connected network, activation functions, batch normalization, and discarding techniques. On one hand, it accurately extracts the complex features of doped ions through nonlinear transformation; on the other hand, batch normalization accelerates model convergence, and discarding techniques suppress overfitting. Compared to simple linear transformation, this submodule can more realistically reproduce the regulatory law of doped ions on scintillator performance and reduce prediction errors caused by insufficient feature extraction.

[0059] 5. In this invention, the matrix material and dopant ion interaction submodule achieves dimensional alignment and interactive processing between the two. It can handle matrix materials and dopant ion combinations with different dimensional characteristics, is applicable to a wider range of material combinations, and increases the variety of applicable materials. It accurately captures the linear and nonlinear interaction relationships between the two, enabling the model to gain a deeper understanding of the material performance regulation mechanism and improve prediction accuracy. Furthermore, by introducing dynamic clustering of matrix materials and the dopant ion-matrix material interaction mechanism, it effectively enhances the model's ability to learn deep-seated implicit features, achieving rapid prediction of light yield and decay time. Compared to existing methods, this method can reduce the total number of parameters in the artificial neural network model, improve training and prediction speed, and quickly predict the performance indicators of different types of scintillator materials.

[0060] 6. In this invention, the feature fusion and attention mechanism submodule integrates multi-source information such as material category, clustering features, doped ion conversion features, and interaction features to ensure feature integrity. Secondly, by concatenating attention features with the initial fusion vector, key information is further enhanced, providing high-value fusion features for the subsequent accurate prediction performance of the fully connected network and improving the model's sensitivity to core influencing factors.

[0061] 7. The fully connected network submodule of this invention employs a multi-layer structure, residual connections, and regularization or normalization operations. The multi-layer network can handle the complex mapping relationships of fused features, while residual connections solve the gradient vanishing problem in deep network training, ensuring the model can effectively learn the correlation between features and performance. Regularization and normalization further suppress overfitting and accelerate convergence. This design gives the model a stronger feature fitting ability, enabling it to stably and accurately map fused features to performance results, thus improving the stability and accuracy of predictions.

[0062] 8. The scintillator material performance prediction system of this invention, after training, directly generates the performance indicators of the scintillator material by only inputting the type of the material to be predicted, the chemical formula of the matrix material, and the chemical formula of the dopant ions. It does not rely on any manually selected feature engineering, nor does it require experimental synthesis and characterization. Compared with existing methods, the cost of obtaining scintillator material performance using this system is significantly reduced. Attached Figure Description

[0063] Figure 1 This is a schematic diagram illustrating the steps of an artificial intelligence-based scintillator material performance prediction method in this invention.

[0064] Figure 2 This is a flowchart illustrating the application of an artificial intelligence-based method for predicting the properties of scintillator materials in this invention.

[0065] Figure 3 This is a schematic diagram of the hybrid artificial neural network model structure used to predict the properties of scintillator materials in this invention.

[0066] Figure 4 This is a graph showing the coefficient of determination between the predicted and actual values ​​of the light yield index of the scintillator material in the embodiment.

[0067] Figure 5 For the example, the predicted Ce-doped Lu3Al5O 12 A comparison chart of the light yield index of single crystal materials and experimental data;

[0068] Figure 6 This is a graph showing the coefficient of determination between the predicted and actual values ​​of the decay time index of the scintillator material in the embodiment.

[0069] Figure 7 This is a comparison chart of the predicted decay time index of Ce-doped LaBr3 single crystal material in the example and the experimental data.

[0070] Figure 8 This is a schematic diagram of the physical structure of an artificial intelligence-based scintillator material performance prediction electronic device in an embodiment. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0072] Example 1

[0073] In this embodiment, an artificial intelligence-based method for predicting the properties of scintillator materials is employed. The steps of this method are as follows: Figure 1 As shown, the specific steps include:

[0074] S1. Collect basic characterization parameters of multiple sets of scintillator materials, use them as input data after the first preprocessing process, collect the corresponding scintillator material performance data, use them as output data after the second preprocessing process, establish a one-to-one mapping relationship between the preprocessed input data and the output data, and construct a scintillator material mapping dataset.

[0075] S2. Construct a hybrid artificial neural network model, which includes a matrix material dynamic clustering submodule, a dopant ion nonlinear conversion submodule, a matrix material and dopant ion interaction submodule, a feature fusion and attention mechanism submodule, and a fully connected network submodule. The submodules work together to capture the regulation law of dopant ions on the performance of scintillator materials.

[0076] S3. Based on the scintillator material mapping dataset, train and validate the constructed hybrid artificial neural network model to obtain the trained hybrid artificial neural network model;

[0077] S4. After the basic characterization parameters of the scintillator material to be predicted are processed through the first preprocessing step, they are input into the trained hybrid artificial neural network model, and the performance data of the scintillator material is output.

[0078] In this embodiment, the method is used to predict the light yield of scintillator materials, and the specific process is as follows: Figure 2 As shown. In this embodiment, the material type is selected as a single crystal, and the chemical formula of the matrix material is Lu3Al5O. 12 The dopant ion is Ce ion, and the dopant ion varies from 0.00015 to 0.0012, with a step size of 0.000002.

[0079] The specific implementation steps are as follows:

[0080] S101. First, obtain publicly available research literature data related to the performance of scintillator materials through scientific literature databases, focusing on collecting material names, chemical formulas of matrix materials, chemical formulas of dopant ions, and light yield indicators. Conduct keyword searches. Keyword combinations such as "scintillator AND light yield" are used to filter literature by year.

[0081] S102. Perform natural language processing on the collected literature, and use regular expressions and structured rules to extract key fields from the literature, such as matrix material category, chemical formula of matrix material, chemical formula of dopant ion, and light yield index. Combined with a large language model based on prompt word engineering, extract the required data and establish a structured scintillator material light yield dataset.

[0082] S201. Vectorize the scintillator material category field. Map the material type in each sample record of the original dataset, where "ceramic" represents ceramic and "crystal" represents single crystal, to a fixed-length two-dimensional vector to characterize the material category features. Specifically, if the material type is ceramic, it is represented as [1.0, 0.0]; if the material type is crystal, it is represented as [0.0, 1.0].

[0083] S202. Perform vectorization encoding and dimensionality reduction on the chemical formula of the matrix material. The dimension of the dense vector after dimensionality reduction is 20.

[0084] S203. Vectorize and encode the chemical formulas of the dopant ions, and perform dimensionality reduction. The resulting dense vector has a dimension of 30.

[0085] S301. In this embodiment, the scintillator material performance to be predicted is the light yield. Therefore, 10 is added to the original light yield data. -5 The bias is then applied, followed by a base-10 logarithmic transformation, and the original light yield is normalized to the range of 0 to 1 using the max-min normalization method.

[0086] S401. The dataset is divided using the quantile binning method. The model is trained on the training set and validated on the validation set. The test set does not participate in the training and validation process and is only used to evaluate the model after the final training is completed.

[0087] S501. Design a hybrid artificial neural network with the following structure: Figure 3 As shown, it includes a matrix material dynamic clustering submodule, a doped ion nonlinear conversion submodule, a matrix material and doped ion interaction submodule, a feature fusion and attention mechanism submodule, and a fully connected network submodule.

[0088] In the S502 matrix material dynamic clustering submodule, 10 initial cluster centers are set using a random normal distribution. A cosine similarity algorithm is employed to calculate the similarity between the input matrix material vector and each cluster center. The most similar cluster center is used as the category to which the matrix material is classified. After training, the cluster center parameters are continuously optimized through backpropagation during model training. Therefore, the cluster center positions will change after training, and the final similarity vector will be output.

[0089] S503. By splicing the similarity vector and the matrix material density vector, the matrix material similarity enhancement vector is obtained.

[0090] S504. Construct a doped ion nonlinear conversion submodule. In this embodiment, the doped ion nonlinear conversion vector is input into a 1-layer fully connected neural network with a ReLU activation function (Rectified Linear Unit). The number of neurons is 512. Batch normalization and dropout processing are used, with a dropout rate of 10%. -5 This yields the nonlinear transformation vector of the doped ions.

[0091] S505. The matrix material density vector is converted to the same dimension as the dopant density vector, and then element-wise multiplied with the dopant ion density vector to generate a linear interaction feature vector. A fully connected network is then added after the interaction feature vector, employing batch normalization and dropout processing. The activation function is ReLU, the number of neurons is 32, and the dropout rate is 10%. -5 The nonlinear interaction vector between the matrix material and the doped ions is obtained.

[0092] S506. By splicing the material category vector, matrix material similarity enhancement vector, dopant ion nonlinear transformation vector, and matrix material and dopant ion nonlinear interaction vector, an initial fusion vector is obtained. In this embodiment, the dimension is 576.

[0093] S507. The Reshape algorithm (Reshape Operation) is used to convert the fused feature vector into a tensor form that can be input into a multi-head attention mechanism.

[0094] S508. Input the transformed tensor into the multi-head attention module. In this embodiment, the number of attention heads is 4 and the key dimension is 32.

[0095] S509. The output of multi-head attention and the original fusion features are flattened and then concatenated to obtain the attention-enhanced fusion vector.

[0096] S510. Construct a fully connected network submodule. The first layer of the fully connected neural network module receives the concatenated vector after multi-head attention output. Multiple fully connected layers are set within this module, with batch normalization and dropout processing performed after each layer. In this embodiment, the number of layers is 5, the number of neurons in each layer is 256, and the dropout rate is 10%. -5 .

[0097] S511. The last layer of the fully connected neural network backbone is the output layer, with a linear activation function. The output value is the light yield after minimum-maximum normalization. According to the transformation steps during preprocessing, an inverse transformation is performed to restore the light yield value at the original scale.

[0098] S601. Construct a hyperparameter searcher. This searcher defines the hyperparameter space to be searched at the model level. Hyperparameters include L2 regularization coefficient, learning rate, etc., and calculates the loss function of the model on the validation set under different hyperparameter conditions. In this embodiment, an efficient hyperparameter optimization method based on the bandwidth allocation principle is used for the search, and the loss function is the mean squared error. Based on the calculation results, the optimal hyperparameter combination is selected.

[0099] S701. Based on the found optimal hyperparameter combination, perform multi-fold cross-validation training on the model. In this embodiment, 10 folds are used.

[0100] S702. After training is complete, save the trained model and preprocessor files as keras and pkl files respectively.

[0101] S801. Call the saved model and preprocessor file, and input the Ce-doped Lu3Al5O to be predicted in this embodiment. 12 Single crystal, performing prediction;

[0102] S901. Call the visualization module to plot the relationship curve between the predicted light yield and the doped ion ratio.

[0103] In this embodiment, as Figure 4 As shown, the artificial intelligence model used to predict the light yield index of scintillator materials has a determination coefficient of 0.92 on the test set.

[0104] In this embodiment, as Figure 5 As shown, the predicted Ce-doped Lu3Al5O3 was obtained. 12 The variation curves of the light yield index of single crystal materials. In the literature, Ce ion doping of Lu3Al5O3 at concentrations of 0.018%, 0.06%, and 0.1% were experimentally characterized. 12 The light yield indicators were 10200, 13000, and 14100 Ph / MeV, respectively. Comparing the results in this embodiment with those in the literature, it can be seen that the method accurately predicts the light yield indicators, and the regulation of Ce ion doping is consistent.

[0105] In this embodiment, the total number of parameters of the hybrid artificial neural network is 1.0M, and the time consumed to perform the prediction is 5.63 seconds.

[0106] In summary, this method, based on a hybrid artificial neural network model, utilizes dynamic clustering of matrix materials, nonlinear transformation of dopant ions, interaction between matrix materials and dopant ions, feature fusion and attention mechanisms, and a fully connected network to accurately predict the light yield and decay time of different types of scintillator materials. Because it only requires training and inputting the basic characterization parameters of the scintillator material to be predicted into the model, it is applicable to a wide variety of matrix materials and dopant ions. Furthermore, after training with a large amount of data, the model can accurately capture the modulation rules, resulting in high prediction accuracy. This method is accurate, fast, and applicable to a wide range of materials for predicting the performance of scintillator materials.

[0107] Example 2

[0108] In this embodiment, an artificial intelligence-based scintillator material performance prediction system is employed. This system first collects and preprocesses data, constructing a hybrid artificial neural network that includes dynamic clustering of matrix materials, nonlinear conversion of dopant ions, interaction between matrix materials and dopant ions, feature fusion and attention mechanisms, and a fully connected network. The preprocessed input data is used to train this hybrid artificial neural network, enhancing the model's ability to extract features related to matrix materials, dopant ions, and their interaction. Ultimately, the model learns the regulatory mechanisms by which different dopant ions affect the performance of scintillator materials, enabling accurate and rapid prediction of the performance of different types of scintillator materials using the trained model.

[0109] The system includes a data acquisition module, a data input module, a data preprocessing module, a hybrid artificial neural network model, a network optimization module, and a visualization module;

[0110] The data acquisition module is used to collect basic characterization parameters and corresponding scintillator material performance data of multiple sets of scintillator materials. It works with the data preprocessing module to take the basic characterization parameters as input data after the first preprocessing process and the material performance data as output data after the second preprocessing process. A one-to-one mapping relationship is established between the preprocessed input data and the output data to construct a scintillator material mapping dataset.

[0111] The data input module is used to input the basic characterization parameters of the scintillator material to be predicted into the trained hybrid artificial neural network model;

[0112] A hybrid artificial neural network model is used to capture the regulatory mechanism of dopant ions on the performance of scintillator materials. It includes a matrix material dynamic clustering submodule, a dopant ion nonlinear conversion submodule, a matrix material and dopant ion interaction submodule, a feature fusion and attention mechanism submodule, and a fully connected network submodule.

[0113] The network optimization module trains and validates the constructed hybrid artificial neural network model based on the scintillator material mapping dataset, resulting in the trained hybrid artificial neural network model.

[0114] The visualization module is connected to the trained hybrid artificial neural network model and is used to visualize the output of the trained hybrid artificial neural network model.

[0115] The data preprocessing module is a dual-input processing logic. It performs preprocessing on the dataset from the data acquisition module and outputs it to the dataset partitioning module. For the parameters to be predicted in the data input module, it performs the same preprocessing process as the training data and outputs it to the optimized model to ensure data distribution consistency.

[0116] The system also includes a dataset partitioning module, which divides the preprocessed dataset of basic characterization parameters of the scintillator material into a training set, a validation set, and a test set. The training set is used to train the network model, the validation set is used to validate the trained network model, and the test set is used to perform the final evaluation of the trained network model.

[0117] The specific steps for predicting the material properties of scintillators using this system include:

[0118] Step 1: Using natural language processing methods, obtain data on scintillator materials from scientific literature and establish a dedicated dataset for scintillator materials that includes material category, chemical formula of matrix material, chemical formula of dopant ion, and material properties;

[0119] Step 2: Preprocess the material type, matrix material chemical formula, and dopant ion chemical formula as input data;

[0120] Step 3: Preprocess the performance indicators of the scintillator material to obtain the output data;

[0121] Step 4: Divide the input and output data into training set, validation set, and test set. Train the model with the training set, validate it with the validation set, and evaluate the trained model with the test set.

[0122] Step 5: Construct a hybrid artificial neural network, including a matrix material dynamic clustering submodule, a dopant ion nonlinear conversion submodule, a matrix material and dopant ion interaction submodule, a feature fusion and attention mechanism submodule, and a fully connected network submodule, to capture the regulation law of dopant ions on the performance of scintillator materials;

[0123] Step 6: Optimize the hyperparameters in the hybrid artificial neural network and select the hyperparameter set with the smallest error in the test set;

[0124] Step 7: Based on the optimal set of hyperparameters, train the hybrid artificial neural network using a multi-fold cross-validation method, and save the trained model and preprocessor files.

[0125] Step 8: Load the model and preprocessor, input the material category to be predicted, the chemical formula of the matrix material and the chemical formula of the dopant ions, predict the scintillation performance of the material, and generate the raw data file.

[0126] Step 9: Based on the original data file, plot the curves showing the change in the scintillator material properties as the doping ion ratio changes.

[0127] Step 1 involves using natural language processing methods to obtain data on scintillator materials from scientific literature and establishing a dedicated dataset for scintillator materials that includes material category, matrix material chemical formula, dopant ion chemical formula, and material properties. This specifically includes:

[0128] Data on scintillator material categories, matrix chemical formulas, dopant ion chemical formulas, and scintillation performance are automatically or manually obtained from scientific literature. Text mining methods such as natural language processing and regular expression matching are employed, while manual screening methods can also be used.

[0129] Specialized datasets are highly structured datasets that include material categories, element types and proportions in the matrix material, types and proportions of dopant ions, and performance data of scintillating materials. In specialized datasets, the element types and proportions in the matrix material and the types and proportions of dopant ions are represented by key-value pairs.

[0130] The preprocessing of input data in step 2 includes the following steps:

[0131] The vectorized encoding of material categories involves obtaining the material category field and converting it into a unique thermal vector form according to a preset mapping rule. The vectorized encoding of the matrix material chemical formula involves obtaining the matrix material chemical formula field from each data entry and converting it into a fixed-length element vector using a predefined matrix material chemical formula embedding function.

[0132] In step 2, the predefined matrix material chemical formula embedding function constructs an ordered vector with all elements of the periodic table. Natural language processing is used to parse the chemical formulas of the matrix materials in the dataset. Regular expressions and other methods are used to identify the element symbols and their corresponding molar ratios. The extracted ratios are normalized, and the weighted ratio values ​​are assigned to the corresponding positions in the vector, while the remaining positions are assigned zero, resulting in a fixed-dimensional sparse vector of matrix materials.

[0133] The predefined matrix material chemical formula embedding function in step 2 is used to take the logarithm or exponential function of the normalized proportions to enhance the distinguishability of proportion differences and improve the robustness of training results; the elemental composition of the matrix material is represented by bond pairs.

[0134] Vectorized encoding of dopant ion chemical formulas involves obtaining the chemical formula field of the dopant ion in each data entry and converting it into a fixed-length element vector using a predefined dopant ion chemical formula embedding function.

[0135] In step 2, the predefined dopant ion chemical formula embedding function constructs an ordered vector using all elements of the periodic table. Natural language processing is used to parse the chemical formula of the matrix material into a string, and regular expressions are employed to identify the element symbols and their corresponding molar ratios. A power function transformation is applied to the dopant ion ratio values ​​to increase the difference between different doping ratios and enhance their impact on the model output. The processed ratio values ​​are then filled into the embedding vector at the positions corresponding to the elements, with unrelated element positions filled with zeros, resulting in a fixed-dimensional sparse vector of dopant ions. The elemental composition of the dopant ions is represented by bond-value pairs.

[0136] Since the length of the periodic table is fixed, the generated matrix material and dopant ion embedding vectors have fixed dimensions and structural sparsity, which can ensure the consistency of the vector space during model training and solve the problem of inconsistent input dimensions caused by the variation of chemical formula lengths of different materials and dopant ions.

[0137] Dimensionality reduction is performed on the high-dimensional embedding vectors of the matrix material and dopant ions to compress the feature dimension and improve computational efficiency. Dimensionality reduction methods include applying algorithms such as principal component analysis and autoencoders to the embedding vectors, and mapping the original high-dimensional sparse vectors to a new set of feature vectors through linear or nonlinear transformations to form low-dimensional dense vectors that still retain the main information.

[0138] Principal component analysis (PCA) was used to reduce the dimensionality of the chemical formula of the matrix material, yielding its dense vector. Similarly, PCA was used to reduce the dimensionality of the chemical formula of the dopant ions, yielding their dense vectors. The dense vectors of the matrix material and the dopant ions both have dimensions between 15 and 40.

[0139] Step 3, the preprocessing of performance indicators, includes:

[0140] In preprocessing, a minimal constant term is added to the performance metrics to bias them and prevent the generation of zero values ​​after transformation. A logarithmic transformation is then performed on the adjusted performance metrics, using a base-10 logarithmic transformation. Based on the logarithmically transformed performance metrics, normalization or standardization methods are applied to distribute the performance metrics within a uniform numerical range, facilitating the convergence and generalization of the neural network model; the normalized or standardized performance metrics range between 0 and 1. Min-max normalization is used for normalization, but Z-score standardization or other scaling methods can also be employed.

[0141] Step 4 divides the preprocessed input features and performance metrics into training, validation, and test sets. The model is trained using the training set, validated using the validation set, and the test set is not involved in the training or validation process, but is only used to evaluate the model after the final training is complete. Specific steps include:

[0142] The data is divided into layers based on performance indicators, and the labels after layering are used as the basis for division to ensure that the performance indicator values ​​are evenly distributed in the divided data subsets. Specifically, the quantile binning method can be used for division.

[0143] Step 5 constructs a hybrid artificial neural network, including dynamic clustering of matrix materials, nonlinear transformation of dopant ions, interaction between matrix materials and dopant ions, feature fusion and attention mechanisms, and a fully connected network. Specific steps include:

[0144] A dynamic clustering submodule for matrix materials is constructed to structurally enhance the dense vectors of the matrix materials. Specifically, the dense vectors of the matrix materials are first input into this module, and then multiple initial cluster centers are introduced as reference points. By calculating the similarity between the current input vector and each cluster center, a cluster feature vector representing the material is generated. During the training of the hybrid artificial neural network, the initial cluster centers are used as weights for dynamic optimization through gradient backpropagation, realizing the learnable function of the cluster centers and dynamically adjusting the cluster centers to which different input vectors belong.

[0145] Several dynamic cluster centers are introduced, and their initialization parameters can be selected using methods such as random normal distribution or linear distribution, with random normal distribution being preferred. In this embodiment, the number of dynamic cluster centers is between 8 and 30.

[0146] For each dense vector of matrix material in the input model, the cosine similarity algorithm is used to calculate the similarity to obtain the matrix material similarity vector. Other similarity algorithms, such as Euclidean distance, can also be used.

[0147] A nonlinear transformation submodule for doped ions is constructed to extract the features of the chemical formula of doped ions. Specifically, the dense vector of doped ions is input into one or more fully connected networks, and the transformation vector of matrix ions is obtained by using activation functions, batch normalization, and discarding.

[0148] The dimension of the matrix ion nonlinear transformation vector is between 128 and 768.

[0149] A submodule for interaction between matrix material and dopant ions is constructed. The dense vector of the matrix material is transformed by dimension mapping to map it to the same dimension space as the dense vector of the dopant ions, and the two are then interacted.

[0150] A single fully connected neural network is used to map the matrix material to the same dimensionality as the dopant ion embedding vector, resulting in a dimension-aligned matrix material transformation vector. Element-wise multiplication is then performed to multiply the dimension-aligned matrix material dense vector with the dopant ion dense vector, forming a linear interaction vector. A single fully connected network is then used to extract nonlinear interaction information; alternatively, a multi-layer fully connected network can be used to transform the linear interaction vector into a nonlinear interaction vector.

[0151] A feature fusion and attention mechanism submodule is constructed to fuse the unique thermal vector of material category, the matrix material similarity enhancement vector, the dopant ion nonlinear transformation vector, and the nonlinear interaction vector between matrix material and dopant ion to obtain the initial fusion vector.

[0152] Based on the input format of the attention mechanism, the initial fusion vector is transformed to obtain the attention mechanism fusion input vector. The transformation from the initial fusion vector to the attention mechanism fusion input vector uses the Reshape algorithm. Then, using a multi-head attention mechanism, the attention feature vector of the attention mechanism fusion input vector is extracted.

[0153] The attention feature vector is concatenated with the initial fusion vector again to obtain the final fusion vector.

[0154] A fully connected network submodule is constructed, which consists of multiple layers of fully connected artificial neural networks. The first layer of the artificial neural network receives the final fusion vector from the attention module. Subsequent layers can use fully connected layers or residual connection structures to improve deep modeling capabilities, and are optimized by combining regularization, normalization and activation functions.

[0155] The number of fully connected layers in the fully connected network module ranges from 3 to 10. The number of neurons in each fully connected layer ranges from 128 to 1024. The activation function of the output layer in the fully connected network module is a linear function. The dropout rate in the fully connected network module is 10. -6 Up to 10 -3 between.

[0156] Step 6 optimizes the hyperparameters in the hybrid artificial neural network, selecting the hyperparameter set with the minimum error on the test set. Specific steps include:

[0157] A network search method can be used to select the hyperparameter set, or other methods, such as the hyperband algorithm, can be employed. Early stopping mechanisms and learning rate scheduling mechanisms can be used to avoid overfitting.

[0158] In step 7, based on the optimal hyperparameter set, a multi-fold cross-validation method is used to train the hybrid artificial neural network, and the trained model and preprocessor files are saved. Specific steps include:

[0159] Using the selected optimal hyperparameter set, a multi-fold cross-validation method is employed for final model training; 5-10 fold cross-validation is used for training; the trained model and preprocessor files are saved. The trained hybrid artificial neural network is saved as a reloadable keras file; the preprocessor file is saved as a reloadable pkl file.

[0160] In step 8, the model and preprocessor are loaded, the type of material to be predicted, the chemical formula of the matrix material and the chemical formula of the dopant ions are input, the scintillation performance of the material is predicted, and the raw data file is generated.

[0161] In step 9, based on the original data file, the performance curves of the scintillator material are plotted as the doping ion ratio changes. This system predicts the optical yield of the scintillator material, and the coefficient of determination (R²) between the predicted and actual values ​​is used. 2 The coefficient of determination (R²) between the predicted and actual values ​​of the decay time index of scintillator materials is greater than 0.9. 2 () greater than 0.95.

[0162] In this embodiment, the decay time of the scintillator material is predicted. The material type is selected as a single crystal, the matrix material chemical formula is LaBr3, the dopant ion is Ce ion, the dopant ion variation range is 0.02 to 0.22, and the step size is 0.0001. The specific steps are the same as the main steps of light yield prediction.

[0163] In this embodiment, the difference is that the material performance index input to the model is the decay time, and the predictive performance output by the model is also the decay time.

[0164] The material category has a dimension of 2, including ceramics and single crystals. The matrix material density vector has a dimension of 35, and the dopant ion density vector has a dimension of 36. Fourteen initial cluster centers are introduced. The dopant ion nonlinear transformation submodule has 512 fully connected neurons, and the nonlinear interaction layer has 38 neurons. The initial fusion vector has a dimension of 601, with 4 attention heads and a key dimension of 32. The fully connected network submodule has 5 neural network layers with 256 neurons per layer, and a dropout rate of 2 × 10⁻⁶. -5 .

[0165] like Figure 6 As shown, the artificial intelligence model used to predict the decay time of scintillator materials has a determination coefficient of 0.96 on the test set.

[0166] In this embodiment, as Figure 7As shown, the predicted decay time curves of Ce-doped LaBr3 single crystal materials are obtained. In the literature, the decay times of Ce-doped LaBr3 with Ce ion concentrations of 0.05, 0.1, and 0.2% were 17, 15, and 16 nanoseconds, respectively, as determined by experimental characterization. Comparing the results of this literature with those of this embodiment shows that the system accurately predicts the decay time, and the Ce ion doping regulation is consistent.

[0167] In this embodiment, the total number of parameters of the hybrid artificial neural network is 1.0M, and the time consumed to perform the prediction is 6.10 seconds.

[0168] In summary, this system comprehensively covers data processing, model building, optimization, and visualization. Applying the aforementioned prediction methods, it can collaboratively and efficiently predict the properties of scintillator materials. The system's comprehensiveness enables it to process data from various types of scintillator materials, making it applicable to a wide range of materials. The precise collaboration between modules ensures the accuracy of the prediction results. The automated processing flow accelerates the prediction speed. The entire process requires minimal experimental operations, significantly reducing costs.

[0169] Example 3

[0170] In this embodiment, an electronic device is used, including a memory and a processor, wherein a computer program is stored in the memory. The specific structure of this electronic device is as follows: Figure 8 As shown, the electronic device includes: an input device, a system bus, a memory, a central processing unit, a graphics processor, and a display device;

[0171] The computer program in the memory includes: a data acquisition module, a data input module, a data preprocessing module, a hybrid artificial neural network model, a network optimization module, and a visualization module;

[0172] Furthermore, the hybrid artificial neural network model includes a matrix material dynamic clustering submodule, a dopant ion nonlinear transformation submodule, a matrix material and dopant ion interaction submodule, a feature fusion and attention mechanism submodule, and a fully connected network submodule.

[0173] The physical entity of the computer device performs the same method as in Embodiment 1 to complete the performance prediction of the scintillator material.

[0174] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the properties of scintillator materials based on artificial intelligence, characterized in that, The method steps include: S1. Collect basic characterization parameters of multiple sets of scintillator materials, use them as input data after the first preprocessing process, collect the corresponding scintillator material performance data, use them as output data after the second preprocessing process, establish a one-to-one mapping relationship between the preprocessed input data and the output data, and construct a scintillator material mapping dataset. S2. Construct a hybrid artificial neural network model, which includes a matrix material dynamic clustering submodule, a dopant ion nonlinear conversion submodule, a matrix material and dopant ion interaction submodule, a feature fusion and attention mechanism submodule, and a fully connected network submodule. The submodules work together to capture the regulation law of dopant ions on the performance of scintillator materials. S3. Based on the scintillator material mapping dataset, train and validate the constructed hybrid artificial neural network model to obtain the trained hybrid artificial neural network model; S4. After the basic characterization parameters of the scintillator material to be predicted are processed through the first preprocessing step, they are input into the trained hybrid artificial neural network model, and the performance data of the scintillator material is output.

2. The method for predicting the performance of scintillator materials based on artificial intelligence according to claim 1, characterized in that, The basic characterization parameters in S1 include the material type, the chemical formula of the matrix material, and the chemical formula of the dopant ions.

3. The method for predicting the performance of scintillator materials based on artificial intelligence according to claim 2, characterized in that, The first preprocessing process in S1 specifically includes: a vectorization encoding process and a dimensionality reduction process; The vectorization encoding process includes a vectorization encoding process for material categories, a vectorization encoding process for the chemical formula of the matrix material, and a vectorization encoding process for the chemical formula of the dopant ions. The specific process of vectorizing material categories is as follows: According to the preset mapping rules, the material category field in each data entry is converted into a unique thermal vector of the material category; The vectorization encoding process for the chemical formula of the base value material specifically includes: First, based on all elements in the periodic table, an ordered vector is constructed as the initial embedding vector. Then, natural language processing is used to parse the chemical formula of the base material, and regular expressions are used to identify the element symbols and their corresponding molar ratios in the chemical formula. The extracted molar ratios are then normalized and logarithmically or exponentially processed to obtain the first weighted ratio value. Based on this first weighted ratio value, values ​​are assigned to the corresponding positions of the initial embedding vector, and the remaining non-corresponding positions are assigned zero, thus obtaining the high-dimensional sparse vector of the matrix material. The vectorization encoding process for the chemical formula of doped ions specifically includes: First, based on all elements in the periodic table, an ordered vector is constructed as the initial embedding vector. Then, natural language processing is used to parse the chemical formula of the doped ions into strings. Regular expressions are used to identify the element symbols and their corresponding molar ratios in the chemical formula of the doped ions. The extracted molar ratios are then transformed by a power function to obtain a second weighted ratio value. Based on this first weighted ratio value, values ​​are assigned to the corresponding positions in the initial embedding vector, and the remaining non-corresponding positions are assigned zero, thus obtaining a high-dimensional sparse vector of the doped ions. The dimensionality reduction process specifically includes: Using a dimensionality reduction algorithm, the high-dimensional sparse vectors of the matrix material and the dopant ions are reduced in dimensionality. The high-dimensional sparse vectors are then mapped to a new set of feature vectors through linear or nonlinear transformations, thereby outputting the dense vectors of the matrix material and the dense vectors of the dopant ions. The second preprocessing process in S1 specifically includes: adding a minimal constant term for biasing, logarithmic transformation, and normalization or standardization.

4. The method for predicting the performance of scintillator materials based on artificial intelligence according to claim 3, characterized in that, In the hybrid artificial neural network model constructed in S2, the matrix material dynamic clustering submodule is used to structurally enhance the dense vectors of the matrix material. Its specific processing includes: The dense vector of the matrix material is input into the dynamic clustering submodule of the matrix material. Then, multiple initial cluster centers are introduced into this submodule as reference points. By calculating the similarity between the current input vector and each cluster center, the clustering feature vector representing the material is generated and output. In the training process of hybrid artificial neural networks, the initial cluster centers are used as weights and dynamically optimized through gradient backpropagation to realize the learnable function of the cluster centers and dynamically adjust the cluster centers to which different input vectors belong; the initial cluster centers are selected by random normal distribution or linear distribution method.

5. The method for predicting the performance of scintillator materials based on artificial intelligence according to claim 4, characterized in that, In the hybrid artificial neural network model constructed in S2, the dopant ion nonlinear conversion submodule consists of one or more fully connected networks, used to extract the features of the dopant ion chemical formula. Its specific processing includes: The dense vector of doped ions is input into the doped ion nonlinear transformation submodule, and after activation function, batch normalization and discarding, the doped ion nonlinear transformation vector is output.

6. The method for predicting the performance of scintillator materials based on artificial intelligence according to claim 5, characterized in that, In the hybrid artificial neural network model constructed in S2, the matrix material and dopant ion interaction submodule is used to perform dimension alignment operations on the matrix material dense vector and to perform interactive processing on the matrix material dense vector and the dopant ion dense vector. The specific processing includes: Input the matrix material density vector and the dopant ion density vector into the matrix material and dopant ion interaction submodule; By using a fully connected neural network, the matrix material dense vector is mapped to the same dimensional space as the dopant ion dense vector, thus obtaining the dimension-aligned matrix material dense vector; Element-wise multiplication is used to multiply the dimension-aligned matrix material dense vector and the dopant ion dense vector element-wise to form a linear interaction vector; One fully connected network is used to extract the nonlinear interaction information between the dimension-aligned matrix material dense vector and the doped ion dense vector to form a nonlinear interaction vector; or a multi-layer fully connected network is used to convert the linear interaction vector into a nonlinear interaction vector and output the nonlinear interaction vector.

7. The method for predicting the performance of scintillator materials based on artificial intelligence according to claim 6, characterized in that, In the hybrid artificial neural network model constructed in S2, the feature fusion and attention mechanism submodule is used to fuse vectors and extract attention feature vectors. Its specific processing includes: The material category unique thermal vector, cluster feature vector, doped ion nonlinear transformation vector, and nonlinear interaction vector are input into the feature fusion and attention mechanism submodule and fused to obtain the initial fusion vector. The Reshape algorithm is used to convert the initial fusion vector into the input format of the attention mechanism, thus obtaining the fusion input vector of the attention mechanism. By utilizing a multi-head attention mechanism, we extract the attention feature vector from the fused input vector. The attention feature vector is concatenated with the initial fusion vector again to output the final fusion vector.

8. The method for predicting the performance of scintillator materials based on artificial intelligence according to claim 7, characterized in that, In the hybrid artificial neural network model constructed in S2, the fully connected network submodule consists of multiple layers of fully connected artificial neural networks. The first layer of the artificial neural network receives the final fusion vector from the attention module, and subsequent layers adopt fully connected layers or residual connection structures, and are optimized by combining regularization, normalization and activation functions.

9. The method for predicting the properties of scintillator materials based on artificial intelligence according to claim 1, characterized in that, The specific process of training, validating, and testing the constructed hybrid artificial neural network model in S3 includes: S31. Divide the scintillator material mapping dataset into a training set, a validation set, and a test set; S32. Define the set of hyperparameter values ​​to be searched, and use the network search method, with mean squared error as the loss function, to select the hyperparameter combination with the smallest error based on the training set. S33. Using the selected optimal hyperparameter combination, the model is validated and trained based on the validation set using a multi-fold cross-validation method. The test set is not used in the training and validation process, but only for evaluating the model after training is completed.

10. A scintillator material performance prediction system based on artificial intelligence, characterized in that, The system operates using an artificial intelligence-based method for predicting the properties of scintillator materials as described in any one of claims 1-9. The system includes a data acquisition module, a data input module, a data preprocessing module, a hybrid artificial neural network model, a network optimization module, and a visualization module. The data acquisition module is used to collect basic characterization parameters and corresponding scintillator material performance data of multiple sets of scintillator materials. In conjunction with the data preprocessing module, the basic characterization parameters are used as input data after a first preprocessing process, and the material performance data are used as output data after a second preprocessing process. A one-to-one mapping relationship is established between the preprocessed input data and the output data to construct a scintillator material mapping dataset. The data input module is used to input the basic characterization parameters of the scintillator material to be predicted into the trained hybrid artificial neural network model; The hybrid artificial neural network model is used to capture the regulatory mechanism of dopant ions on the performance of scintillator materials. It includes a matrix material dynamic clustering submodule, a dopant ion nonlinear conversion submodule, a matrix material and dopant ion interaction submodule, a feature fusion and attention mechanism submodule, and a fully connected network submodule. The network optimization module trains and validates the constructed hybrid artificial neural network model based on the scintillator material mapping dataset, and obtains the trained hybrid artificial neural network model. The visualization module is connected to the trained hybrid artificial neural network model and is used to visualize the output of the trained hybrid artificial neural network model.

Images