Fluoride laser crystal optical property prediction method based on deep learning

Abstract

The invention discloses an optical property prediction method of a fluoride laser crystal based on deep learning, which belongs to the technical field of deep learning and material science and comprises the following steps: constructing a source domain data set of the fluoride crystal with a crystal stability label, and training a source domain model by taking structural parameters and a chemical formula of the fluoride crystal as input to obtain a source domain model; obtaining a crystal stability prediction model; the constructed optical property prediction model comprises a feature extractor and a detection head which are connected in sequence, and the feature extractor is a structure branch GNN and a component branch CoTAN in the crystal stability prediction model; the detection head is used for mapping the features extracted by the feature extractor into a prediction result of the optical properties of the crystal; and inputting the obtained structure parameters and the chemical formula of the fluoride laser crystal to be predicted into the optical property prediction model to obtain an optical property prediction result. According to the method, the crystal stability prediction model is introduced into the optical property prediction process through transfer learning, the problem of scarcity of crystal optical data is solved, and the model prediction precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, and more particularly to a method for predicting optical properties of fluoride laser crystals based on deep learning. BACKGROUND

[0002] As an important class of functional materials, fluoride laser crystals have shown great application potential in many fields such as laser technology, optical communication, medical and cosmetic, military and defense, etc. In the field of laser technology, their excellent optical properties can provide support for achieving high-power, high-efficiency and high-stability laser output, and promote the development of laser towards smaller size and higher performance. In optical communication, they can be used to develop new optical signal processing and transmission devices, and improve the capacity and reliability of communication systems. In medical treatment, they can help to develop more accurate laser treatment equipment and improve the effectiveness and safety of disease treatment. In military applications, they can provide key material basis for advanced laser weapons and other equipment. Therefore, accurately predicting the optical properties of fluoride laser crystals, such as absorption spectrum, emission spectrum, refractive index, laser damage threshold, etc., is of great significance for optimizing crystal material design, improving laser system performance and expanding its application range, and is a key link to promote the technological progress in related fields.

[0003] The existing methods for predicting the optical properties of fluoride laser crystals mainly include empirical formula and traditional machine learning prediction methods. The prediction method based on empirical formula, such as calculating the thermal decomposition temperature of fluoride laser crystals through their chemical energy, only considers a single structural parameter of fluoride laser crystals, but ignores the complex factors of molecular structure, resulting in large prediction error. The prediction method based on traditional machine learning, such as using support vector regression and random forest models, needs to manually extract the structural features of the crystals, which usually relies on expert experience for feature extraction, and the feature selection is insufficient, resulting in poor generalization ability of the model.

[0004] In practical applications, the optical property data of fluoride laser crystals is often very scarce and high in acquisition cost, and the high-quality data available for training the optical property prediction model of fluoride laser crystals is relatively scarce. The quality and quantity of data directly affect the performance and generalization ability of the model, and further affect the prediction accuracy of the model. SUMMARY

[0005] To address the problems existing in the above-mentioned fields, this invention proposes a deep learning-based method for predicting the optical properties of fluoride laser crystals. The constructed source domain model learns the intrinsic relationship between structure, composition, and stability. The trained source domain model is then transferred to the optical property prediction task as a crystal stability prediction model, achieving effective knowledge transfer. By leveraging the crystal essence information contained in the stability data, additional prior knowledge is provided for optical property prediction, which to some extent compensates for the lack of optical data and thus improves the prediction accuracy of the optical property prediction model.

[0006] To address the aforementioned technical problems, this invention discloses a method for predicting the optical properties of fluoride laser crystals based on deep learning, comprising the following steps: A source domain model is constructed, comprising a structural branch GNN, a compositional branch CoTAN, and a fully connected layer. The structural branch GNN extracts the atomic diagram, line diagram, and dihedral diagram of the fluoride crystal based on structural parameters, and determines the structural features of the fluoride crystal based on these diagrams. The compositional branch CoTAN obtains and fuses the chemical embedding matrix, positional embedding matrix, and stoichiometric matrix of the fluoride crystal based on its chemical formula to generate the compositional features of the fluoride crystal. The fully connected layer concatenates and fuses the structural and compositional features, and outputs the crystal stability prediction results through mapping. Construct a source domain dataset of fluoride crystals with crystal stability labels; based on the source domain dataset, train a source domain model with the structural parameters and chemical formula of fluoride crystals as input, and obtain a crystal stability prediction model; An optical property prediction model is constructed, comprising a feature extractor and a detection head connected in sequence. The feature extractor is the structure branch GNN and the composition branch CoTAN in the crystal stability prediction model. The detection head is used to map the features extracted by the feature extractor to the prediction results of the crystal optical properties. Obtain the structural parameters and chemical formula of the fluoride laser crystal to be predicted, input them into the optical property prediction model, and obtain the optical property prediction results.

[0007] Preferably, the structural branching GNN extracts the atomic diagram, line diagram, and dihedral diagram of the fluoride crystal based on structural parameters, and determines the structural features of the fluoride crystal based on the atomic diagram, line diagram, and dihedral diagram, specifically including: The structural branch GNN is based on the edge-gated attention graph neural network EGAT, which includes two levels of gated attention blocks; Based on structural parameters, with a preset atomic pair threshold as the cutoff radius, only the interaction between atomic pairs ≤ the threshold is considered. If the distance exceeds the threshold, it indicates no correlation. By capturing the many-body interaction between atoms, the atomic diagram, line diagram and dihedral diagram of the fluoride laser crystal are extracted. For an atomic graph, atoms are nodes, and the connections between atoms are edges. Edges only connect pairs of atoms whose distance is ≤ a threshold. The weight or feature of the edge is the bond length, and the bond length feature graph is encoded. For a line graph, nodes are bonds and edges are bond angles, encoding a three-body interaction graph; For a dihedral graph, nodes are bonds and edges are dihedral angles, encoding a four-body interaction graph; The bond length feature map, three-body interaction map, and four-body interaction map are input into the edge-gated attention map neural network EGAT. The edge information is updated through two-level gated attention blocks, adjacent edges, and node information, thereby updating the interactions between multiple bodies and outputting the structural feature vector of the multiple body interactions and periodic features of the fluoride crystal.

[0008] Preferably, the compositional branch CoTAN is based on chemical formulas, and obtains and fuses the chemical intercalation matrix, positional intercalation matrix, and stoichiometric matrix of the fluoride crystal to generate the compositional characteristics of the fluoride crystal, specifically including: The component branch CoTAN includes a Transformer, an embedding layer, a positional encoding layer, and a fully connected layer; By identifying the element symbols contained in the chemical formula, the corresponding element can be found in the periodic table and its atomic number can be obtained. Physical descriptors for fluoride crystals are obtained based on the properties of the elements in the chemical formula. The atomic fraction of each element is obtained by dividing the number of atoms of each element by the total number of atoms in the chemical formula. The atomic fractions are input into the Transformer, which encodes the atomic fractions through a self-attention mechanism to generate a stoichiometric matrix. In the embedding layer, atomic numbers are mapped to high-dimensional vectors, and residual connections and multi-layer nonlinear transformations are performed on the high-dimensional vectors to obtain the chemical embedding matrix; Obtain the physical descriptor vector and position index of the element. Use the sine / cosine function in the position encoding layer to generate the position encoding of the position index. Concatenate the physical descriptor vector and the position encoding through a fully connected layer to output the position embedding matrix of the physical descriptor. The chemical embedding matrix and the stoichiometric matrix are added element-wise and then concatenated with the position embedding matrix to generate a composition matrix. Composition features are then extracted from the composition matrix.

[0009] Preferably, the construction of the source domain dataset of fluoride crystals with crystal stability tags specifically includes: The structural parameters and chemical formulas of fluoride crystals are obtained, and by determining the crystal stability, stability labels are labeled on the structural parameters and chemical formulas of some fluoride crystals to construct a source domain dataset of fluoride crystals with crystal stability labels.

[0010] Preferably, the step of training a source domain model based on the source domain dataset, using the structural parameters and chemical formula of the fluoride crystal as input, to obtain a crystal stability prediction model specifically includes: The Adam optimizer, coupled with periodic learning rate scheduling, is used to train the source domain model with the MSE of multi-task learning as the loss function. The loss function uses mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 Three evaluation metrics are used. By determining the number of iterations and the initial learning rate, the evaluation metrics in the loss function are fine-tuned through iterative training until each evaluation metric meets the preset iteration conditions. The source domain model that satisfies the preset iteration conditions for each evaluation index is used as the crystal stability prediction model.

[0011] Preferably, the detection head is used to map the features extracted by the feature extractor to the predicted results of the crystal optical properties, specifically including: The detection head is a multi-output structure, and each type of optical property in the multi-output structure corresponds to a fully connected layer; Structural and compositional features were extracted using the structure branch GNN and composition branch CoTAN in the crystal stability prediction model, respectively. The extracted structural and compositional features are input into the detection head, which performs classification and detection through multiple output structures and outputs optical property prediction results.

[0012] Preferably, obtaining the structural parameters and chemical formula of the fluoride crystal specifically includes: Based on historical documents and materials databases, the chemical formula of fluoride crystals was determined; Structural parameters of fluoride crystals were extracted using the Materials Project database.

[0013] Preferably, it further includes a deep learning-based optical property prediction system for fluoride laser crystals, comprising: A source domain model construction module is used to construct a source domain model, which includes a structural branch GNN, a compositional branch CoTAN, and a fully connected layer. The structural branch GNN extracts the atomic diagram, line diagram, and dihedral diagram of the fluoride crystal based on structural parameters, and determines the structural features of the fluoride crystal based on these diagrams. The compositional branch CoTAN obtains and fuses the chemical embedding matrix, positional embedding matrix, and stoichiometric matrix of the fluoride crystal based on the chemical formula to generate the compositional features of the fluoride crystal. The fully connected layer concatenates and fuses the structural and compositional features, and outputs the crystal stability prediction result through mapping. A source domain dataset of fluoride crystals with crystal stability labels is constructed. Based on the source domain dataset, the source domain model is trained using the structural parameters and chemical formula of the fluoride crystal as input to obtain the crystal stability prediction model. An optical property prediction module is used to construct an optical property prediction model. The optical property prediction model includes a feature extractor and a detection head connected in sequence. The feature extractor is the structure branch GNN and the composition branch CoTAN in the crystal stability prediction model. The detection head is used to map the features extracted by the feature extractor to the prediction results of the crystal's optical properties. The structural parameters and chemical formula of the fluoride laser crystal to be predicted are obtained, input into the optical property prediction model, and the optical property prediction results are obtained.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention proposes a deep learning-based method for predicting the optical properties of fluoride laser crystals. By constructing a source domain model and learning the intrinsic relationship between structure, composition, and stability, it simultaneously extracts features from both structural and compositional dimensions. This allows for a comprehensive and multi-layered characterization of the intrinsic properties of fluoride crystals. Even in situations where optical data is scarce, it provides rich and discriminative feature information for the optical property prediction model, laying a solid foundation for accurate optical property prediction. The trained source domain model is then transferred to the optical property prediction task as a crystal stability prediction model, achieving effective knowledge transfer. Through pre-training and transfer learning, the reliance on large amounts of labeled optical data is reduced, fully utilizing relatively easily obtainable crystal stability-related data. This transfer learning strategy leverages the essential crystal information contained in stability data to provide additional prior knowledge for optical property prediction, compensating to some extent for the lack of sufficient optical data. The feature extractor of the constructed optical property prediction model is the structure branch GNN and the composition branch CoTAN from the crystal stability prediction model. Because this feature extractor is pre-trained, it possesses a certain feature extraction capability. During the optical property prediction process, it can more quickly map the input data to a meaningful feature space, further alleviating the training difficulties caused by the scarcity of optical data, thereby improving the model's prediction accuracy and adapting to small-scale datasets. The transfer learning framework proposed in this invention can adapt to material systems lacking crystal structure information. Attached Figure Description

[0015] Figure 1 This is a flowchart of the optical property prediction method for fluoride laser crystals based on deep learning proposed in this invention. Figure 2 This is a diagram of the source domain model architecture proposed in this invention; Figure 3 A diagram illustrating the transfer learning framework provided in this embodiment of the invention; Figure 4 The training curves for the source domain model provided in the embodiments of the present invention. Detailed Implementation

[0016] The following will refer to the appendices in the embodiments of the present invention. Figures 1-4 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0017] Example like Figure 1 As shown, this invention proposes a method for predicting the optical properties of fluoride laser crystals based on deep learning, comprising the following steps: S1: Construct the source domain model, which includes the structural branch GNN, the component branch CoTAN, and the fully connected layer; The structure branch GNN extracts the atomic map, line map, and dihedral map of the fluoride crystal based on structural parameters, and determines the structural features of the fluoride crystal based on these maps. The composition branch CoTAN obtains the chemical embedding matrix, position embedding matrix, and stoichiometric matrix of the fluoride crystal based on the chemical formula and fuses them to generate the compositional features of the fluoride crystal. The fully connected layer splices and fuses the structural and compositional features, and outputs the crystal stability prediction results through mapping. S2: Construct a source domain dataset of fluoride crystals with crystal stability labels; based on the source domain dataset, train the source domain model with the structural parameters and chemical formula of the fluoride crystals as input, and obtain a crystal stability prediction model; S3: Construct an optical property prediction model, which includes a feature extractor and a detection head connected in sequence. The feature extractor is the structure branch GNN and the composition branch CoTAN in the crystal stability prediction model; the detection head is used to map the features extracted by the feature extractor to the prediction results of the crystal optical properties. S4: Obtain the structural parameters and chemical formula of the fluoride laser crystal to be predicted, input them into the optical property prediction model, and obtain the optical property prediction results.

[0018] In step S1, the constructed source domain model includes a structural branch GNN, a component branch CoTAN, and a fully connected layer, wherein: The structural branch GNN is based on the edge-gated attention graph neural network EGAT, which includes two levels of gated attention blocks.

[0019] The structure-branched GNN, based on structural parameters and using a preset atomic pair threshold of 8 Å as the cutoff radius, only considers interactions when the distance between atomic pairs is ≤ the threshold; those exceeding the threshold are considered to have no direct correlation. By capturing many-body interactions between atoms, it extracts atomic diagrams, line diagrams, and dihedral diagrams of fluoride laser crystals, such as... Figure 2 As shown.

[0020] For the atomic graph, atoms are nodes, and the connections between atoms are edges. Edges only connect atomic pairs with a distance ≤ 8 Å. The weight or feature of the edge is the bond length. The bond length feature map is encoded and expanded into a multidimensional feature map through radial basis functions to capture subtle differences in bond length.

[0021] For line graphs, nodes are bonds and edges are bond angles. The three-body interaction graph is encoded, and the bond angles are used as features of the edges to describe the spatial arrangement of the three atoms, reflecting local symmetry.

[0022] For a dihedral diagram, nodes are bonds and edges are dihedral angles, encoding a four-body interaction diagram. The dihedral angles describe the torsional conformation of four atoms, reflecting a more complex spatial arrangement in the crystal.

[0023] The bond length feature map, three-body interaction map, and four-body interaction map are input into the 12-layer edge-gated attention map neural network EGAT. The edge information is updated through two-level gated attention blocks, adjacent edges, and node information, and then the multi-body interactions are updated. The feature dimension after network processing is reduced by average pooling. The fully connected layer maps the dimensionality-reduced features and outputs the structured feature vector of multi-body interactions and periodic features of the crystal structure, i.e., the structural data.

[0024] The 8Å cutoff radius is a core parameter in the many-body interaction diagram of crystal structures. By limiting the range of interactions, a balance between computational efficiency and accuracy is achieved.

[0025] In atomic diagrams, line diagrams, and dihedral diagrams, 8Å ensures a complete description of the local structure, while combining atomic type embedding and geometric feature encoding to provide an efficient data representation method for predicting the stability and optical properties of crystals.

[0026] Atomic diagrams encode bond lengths (two-body interactions), line diagrams encode bond angles (three-body interactions), and dihedral diagrams encode torsional conformations (four-body interactions), forming a multi-scale structural description from local to global. This multi-diagram fusion strategy avoids the limitations of single-structure descriptions and significantly improves the model's ability to resolve complex crystal structures.

[0027] Graph Neural Networks (GNNs) automatically learn high-order features of nodes (atoms / bonds) and edges (interactions) through message passing mechanisms, and capture long-range structural dependencies (such as inter-layer stacking patterns) through multi-layer aggregation, thereby more accurately associating structure and stability.

[0028] The component branch of CoTAN includes a Transformer, an embedding layer, a positional coding layer, and a fully connected layer.

[0029] The composition branch of CoTAN is based on chemical formulas in the stability prediction dataset. It identifies the element symbols contained in the chemical formulas, finds the corresponding elements in the periodic table, and obtains their atomic numbers. It obtains the physical descriptor of the fluoride laser crystal through experimental measurement. By counting the number of atoms of each element in the chemical formula and the total number of atoms in the chemical formula, it divides the number of atoms of each element by the total number of atoms to obtain the atomic fraction of each element. Atomic numbers are mapped to high-dimensional vectors using Mat2Vec, and the high-dimensional vectors are then subjected to residual connections (ResNet) and multi-layer nonlinear transformations to obtain the chemical embedding matrix. Obtain the physical descriptor vector and position index of the element, generate the position code of the position index using sine / cosine function, concatenate the physical descriptor vector and the position code through a fully connected layer, and output the position embedding matrix of the descriptor; The atomic fractions are input into the Transformer, which encodes the atomic fractions through a self-attention mechanism to generate a stoichiometric matrix.

[0030] The chemical embedding matrix and the stoichiometric matrix are added element-wise to generate the composition matrix; the composition matrix is ​​then concatenated with the positional embedding matrix to extract the composition features.

[0031] Extracting component features from the component matrix specifically includes: The generated component matrix is ​​processed through a 6-layer multi-head attention module to gradually extract more abstract and advanced component feature representations, resulting in an updated component matrix.

[0032] The updated component matrix is ​​input into DenseNet, and through multi-layer convolution and pooling operations, deep features in the component data are further extracted to uncover more subtle and essential component feature patterns, and finally the component data is output.

[0033] That is, after feature weighting and long-distance dependency capture through 6 layers of multi-head attention modules, an updated composition matrix is ​​obtained. Then, DenseNet performs deep feature extraction and reuse to finally generate composition data for material property prediction, forming a complete and efficient composition feature processing flow.

[0034] The fully connected layer concatenates and fuses the structural feature vectors of the fluoride crystal extracted by the structural branch GNN and the compositional features extracted by the compositional branch CoTAN. It then outputs the stability prediction results (material properties) through mapping and uses the trained source domain model as the crystal stability prediction model.

[0035] The source domain model constructed in this invention has a structure branch GNN output that can focus on the local-global structural features of fluoride crystals, and a composition branch CoTAN that can quantify the chemical composition and proportion of fluoride crystals. By splicing and fusing the two through a fully connected layer, features from different aspects can be integrated to achieve information complementarity and provide a more comprehensive basis for the final prediction.

[0036] The source domain model designed in this invention, which extracts features from both structural and compositional dimensions simultaneously, can comprehensively and multi-dimensionally characterize the intrinsic properties of fluoride crystals. Even when optical data is scarce, it can provide the model with rich and discriminative feature information, laying a solid foundation for accurately predicting optical properties.

[0037] The fusion of multi-source information enables the model to learn more essential feature representations of fluoride crystals, capturing the commonalities and differences between different crystals. When optical data is limited, this comprehensive feature extraction helps the model better adapt to unseen crystal samples, effectively reducing the risk of overfitting, enhancing the model's predictive ability on unknown data, and improving the model's generalization performance.

[0038] In step S2, relevant historical literature is retrieved through Web of Science and Google Scholar. Based on the historical literature and materials database, the chemical formula characterizing the stability of fluoride laser crystals is determined.

[0039] Based on the Materials Project database, structural parameters of fluoride crystals were extracted, and CIF files, thermodynamic parameters (formation energy, convex hull energy), and band information corresponding to the elements of fluoride crystals were extracted in a targeted manner. Fluoride crystals were screened, and those with bandgap values ​​in the range of (0, 10) and in which all crystals are non-metallic were selected. A stability prediction dataset for fluoride crystals was constructed, and a total of 9448 data samples were collected. The dataset was divided into three parts: training set (7560 data samples), validation set (944 data samples), and test set (944 data samples).

[0040] Based on the chemical formula and structural parameters of fluoride crystals, a stability prediction dataset for fluoride crystals is constructed.

[0041] By experimentally determining the stability of crystals, stability labels are labeled on the structural parameters and chemical formulas of some fluoride crystals in the stability prediction dataset of fluoride crystals, and a source domain dataset of fluoride crystals with crystal stability labels is constructed.

[0042] Input the stability prediction dataset into the pre-trained source domain model to train the source domain model and obtain the crystal stability prediction model.

[0043] In step S3, the present invention collects optical property data (LD pump wavelength range, emission wavelength range, upper level lifetime, stimulated emission cross section) of mid-wave infrared fluoride laser crystals, obtaining more than 100 valid literature data, searching for the corresponding CIF files of elements in relevant databases, constructing a dataset for predicting the optical properties of fluoride laser crystals, including the structural parameters and chemical formula of fluoride laser crystals, and using it as the target domain data.

[0044] The target domain data is input into the constructed optical property prediction model, and the optical property prediction results are obtained through prediction.

[0045] The constructed optical property prediction model comprises a feature extractor and a detector head connected sequentially. The feature extractor consists of the structural branch GNN and the compositional branch CoTAN from the crystal stability prediction model. The detector head is a multi-output structure, with each type of optical property corresponding to a fully connected layer. The detector head is used to map the features extracted by the feature extractor to the predicted results of the crystal's optical properties.

[0046] like Figure 3 As shown, the transfer learning framework provided in this embodiment of the invention includes three stages: pre-training, fine-tuning, and prediction. The specific process is as follows: During the pre-training phase Random initialization: Before pre-training begins, the parameters of the source domain model are randomly initialized as the starting point for training the source domain model.

[0047] Sufficient Formation Energy Data: Utilizing readily available and abundant formation energy data of fluoride crystals, combined with their structural parameters and chemical formulas, this data is input into a randomly initialized source domain model for pre-training. The structure branch (GNN) extracts structural features such as atomic diagrams, line diagrams, and dihedral diagrams based on structural parameters; the composition branch (CoTAN) obtains chemical embedding matrices, positional embedding matrices, and stoichiometric matrices based on chemical formulas and fuses them to generate compositional features. Together, these two branches enable the source domain model to learn the relevant knowledge of fluoride crystal structure, composition, and formation energy, resulting in a well-trained source domain model. This step corresponds to the "pre-training" stage in the diagram, and its purpose is to allow the source domain model to initially learn the characteristic representation of the crystal using sufficient formation energy data.

[0048] During the fine-tuning phase Scarce Optical Property Data: Due to the scarcity of optical property data for fluoride laser crystals, directly training the model is quite difficult. Therefore, the structural parameters and chemical formula of the fluoride laser crystal (i.e., the scarce optical property data) are used as the target domain data and directly input into the optical property prediction model for optical property prediction.

[0049] Fine-tuning: The structural branch GNN and the compositional branch CoTAN in the trained source domain model are used as feature extractors for the optical property prediction model and connected to the detection head in sequence to construct the optical property prediction model.

[0050] By using scarce optical property data as the target domain data and directly inputting it into the optical property prediction model, the model can be adapted to the optical property prediction task, thus obtaining the prediction results. This corresponds to the "fine-tuning" step in the diagram, where a transfer learning strategy is used to leverage the knowledge of the trained source domain model to assist the optical property prediction model in its predictions.

[0051] Material properties: The predicted output includes the optical properties and other material properties.

[0052] Traditional deep learning models typically require a large amount of labeled data for training. However, the method proposed in this invention reduces the dependence of optical property prediction models on a large amount of labeled optical data through pre-training and transfer learning, effectively solving the problem of model training difficulties caused by the scarcity of optical data.

[0053] The structural branch GNN and the compositional branch CoTAN in the trained source domain model are used as feature extractors for the optical property prediction model to extract structural and compositional features for the target domain data, respectively. The extracted structural and component features are input into the multi-output structure, and the output is the optical property prediction result.

[0054] The optical property prediction model reuses the structure branch GNN and composition branch CoTAN from the crystal stability prediction model as feature extractors, avoiding the need to retrain the feature extraction part from scratch in the optical property prediction task. This sharing mechanism fully utilizes the feature extraction capabilities already learned by the crystal stability prediction model, prevents redundant learning, significantly improves data utilization efficiency, and allows limited optical data to play a greater role.

[0055] Since the feature extractor is pre-trained and has a certain feature extraction capability, it can map the input data to a meaningful feature space more quickly during the optical property prediction process, further alleviating the training difficulties caused by the scarcity of optical data.

[0056] This invention employs the Adam optimizer with a 1-cycle learning rate schedule, using MSE as the loss function to ensure regression gradient stability, and repeatedly trains the source domain model multiple times to ensure statistical robustness. This embodiment uses mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 Three evaluation metrics; please refer to the training process of the source domain model. Figure 4 It presents three types of core indicator curves during the source domain model training process. Figure 4 In the diagram, subplot (a) shows the Loss curve during the training of the source domain model, illustrating the changes in loss values ​​between the training and test sets. The final Loss for the training set is 0.1805, and the Loss for the test set is 0.0572. Subplot (b) shows the MAE curve during the training of the source domain model, presenting the fluctuations in the mean absolute error between the two classes of data. The final MAE for the training set is 0.2060, and the MAE for the test set is 0.1160. Subplot (c) shows the R... 2 The curve reflects the change in the model's ability to capture patterns in the two types of data, and the final training set R 2 0.8220, test set R 2The curves clearly show the performance trend of the source domain model on different data as the training rounds progress. After hyperparameter tuning, the optimal hyperparameter combination is 148 iterations, an initial learning rate of 0.00005, and a batch size of 64. The performance is evaluated on the trained source domain model using the test set, and the MAE is 0.1160.

[0057] As can be seen from this embodiment, the method proposed in this invention can achieve high-precision prediction of stability and optical properties without the need to synthesize fluoride laser crystals.

[0058] The proposed optical property prediction method for fluoride laser crystals reduces reliance on large amounts of labeled optical data through pre-training and transfer learning. This method fully utilizes readily available crystal stability data to pre-train the source domain model, enabling it to learn the intrinsic relationship between structure, composition, and stability. The trained source domain model is then transferred as a crystal stability prediction model to the optical property prediction task, achieving effective knowledge transfer. This transfer learning strategy leverages the essential crystal information contained in the stability data to provide additional prior knowledge for optical property prediction, compensating to some extent for the lack of sufficient optical data.

[0059] This invention also proposes a deep learning-based optical property prediction system for fluoride laser crystals, comprising: The source domain model construction module is used to build the source domain model, which includes a structural branch GNN, a compositional branch CoTAN, and a fully connected layer. The structural branch GNN extracts the atomic diagram, line diagram, and dihedral diagram of the fluoride crystal based on structural parameters, and determines the structural features of the fluoride crystal based on these diagrams. The compositional branch CoTAN obtains and fuses the chemical embedding matrix, positional embedding matrix, and stoichiometric matrix of the fluoride crystal based on the chemical formula to generate the compositional features of the fluoride crystal. The fully connected layer concatenates and fuses the structural and compositional features, and outputs the crystal stability prediction result through mapping. A source domain dataset of fluoride crystals with crystal stability labels is constructed. Based on the source domain dataset, the source domain model is trained using the structural parameters and chemical formula of the fluoride crystal as input to obtain the crystal stability prediction model. The optical property prediction module is used to construct an optical property prediction model. The optical property prediction model includes a feature extractor and a detection head connected in sequence. The feature extractor is the structure branch GNN and the composition branch CoTAN in the crystal stability prediction model. The detection head is used to map the features extracted by the feature extractor to the prediction results of the crystal's optical properties. The module obtains the structural parameters and chemical formula of the fluoride laser crystal to be predicted, inputs them into the optical property prediction model, and obtains the optical property prediction results.

[0060] This invention enables reliable prediction of the formation energy of fluoride laser crystals. Compared to traditional experimental measurement methods, the method proposed in this invention does not require the preparation of actual fluoride laser crystal samples; the formation energy can be effectively predicted based solely on their crystal structure parameters, thus saving experimental costs. It provides researchers with a simple and effective tool for predicting the formation energy of fluoride laser crystals to be synthesized.

[0061] Compared to traditional machine learning methods, this invention offers a simpler input method. Only the CIF file of the material needs to be input to effectively predict the formation energy, saving significant time spent searching for data. Furthermore, it automatically extracts the most suitable feature representation for the formation energy, eliminating tedious feature engineering. This enables faster and more accurate prediction of the formation energy, promoting innovative development in the field of fluoride laser crystal applications.

[0062] The optical property prediction method and system for fluoride laser crystals based on deep learning proposed in this embodiment can realize an efficient process of inputting the chemical formula and structural parameters of fluoride laser crystals and outputting accurate stability prediction values. Compared with traditional methods, it can reduce R&D costs and accelerate the development of new crystals.

[0063] In summary, the method proposed in this invention constructs a complete chain from crystal stability prediction to optical property prediction. It first pre-trains the model using stability data and then gradually transitions to the optical property prediction task. This step-by-step approach allows the model to fully utilize different types of data at different stages, progressively improving its understanding and predictive ability regarding fluoride crystal properties, thus maximizing the value of limited data.

[0064] In practical research and application of fluoride laser crystals, optical data is often scarce and expensive to obtain. The method proposed in this invention, through a series of innovative designs, can construct an effective predictive model even with scarce data, providing crucial technical support for the research and development of fluoride laser crystals. This reduces the research and development risks and costs caused by insufficient data, and has significant practical application value.

[0065] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0066] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A method for predicting the optical properties of fluoride laser crystals based on deep learning, characterized in that, Includes the following steps: A source domain model is constructed, which includes a structural branch GNN, a compositional branch CoTAN, and a fully connected layer. The structural branch GNN extracts the atomic diagram, line diagram, and dihedral diagram of the fluoride crystal based on structural parameters, and determines the structural features of the fluoride crystal based on the atomic diagram, line diagram, and dihedral diagram. The compositional branch CoTAN is based on chemical formulas, and obtains and fuses the chemical embedding matrix, positional embedding matrix and stoichiometric matrix of fluoride crystals to generate compositional features of fluoride crystals; the fully connected layer splices and fuses the structural features and compositional features, and outputs the crystal stability prediction results through mapping. Construct a source domain dataset of fluoride crystals with crystal stability labels; based on the source domain dataset, train a source domain model with the structural parameters and chemical formula of fluoride crystals as input, and obtain a crystal stability prediction model; An optical property prediction model is constructed, comprising a feature extractor and a detection head connected in sequence. The feature extractor is the structure branch GNN and the composition branch CoTAN in the crystal stability prediction model. The detection head is used to map the features extracted by the feature extractor to the prediction results of the crystal optical properties. Obtain the structural parameters and chemical formula of the fluoride laser crystal to be predicted, input them into the optical property prediction model, and obtain the optical property prediction results.

2. The method for predicting the optical properties of fluoride laser crystals based on deep learning according to claim 1, characterized in that, The structural branch GNN extracts atomic diagrams, line diagrams, and dihedral diagrams of fluoride crystals based on structural parameters, and determines the structural features of the fluoride crystals based on these diagrams, specifically including: The structural branch GNN is based on the edge-gated attention graph neural network EGAT, which includes two levels of gated attention blocks; Based on structural parameters, with a preset atomic pair threshold as the cutoff radius, only the interaction between atomic pairs ≤ the threshold is considered. If the distance exceeds the threshold, it indicates no correlation. By capturing the many-body interaction between atoms, the atomic diagram, line diagram and dihedral diagram of the fluoride laser crystal are extracted. For an atomic graph, atoms are nodes, and the connections between atoms are edges. Edges only connect pairs of atoms whose distance is ≤ a threshold. The weight or feature of the edge is the bond length, and the bond length feature graph is encoded. For a line graph, nodes are bonds and edges are bond angles, encoding a three-body interaction graph; For a dihedral graph, nodes are bonds and edges are dihedral angles, encoding a four-body interaction graph; The bond length feature map, three-body interaction map, and four-body interaction map are input into the edge-gated attention map neural network EGAT. The edge information is updated through two-level gated attention blocks, adjacent edges, and node information, thereby updating the interactions between multiple bodies and outputting the structural feature vector of the multiple body interactions and periodic features of the fluoride crystal.

3. The method for predicting the optical properties of fluoride laser crystals based on deep learning according to claim 1, characterized in that, The compositional branch CoTAN, based on chemical formulas, obtains and fuses the chemical embedding matrix, positional embedding matrix, and stoichiometric matrix of fluoride crystals to generate compositional characteristics of fluoride crystals, specifically including: The component branch CoTAN includes a Transformer, an embedding layer, a positional encoding layer, and a fully connected layer; By identifying the element symbols contained in the chemical formula, the corresponding element can be found in the periodic table and its atomic number can be obtained. Physical descriptors for fluoride crystals are obtained based on the properties of the elements in the chemical formula. The atomic fraction of each element is obtained by dividing the number of atoms of each element by the total number of atoms in the chemical formula. The atomic fractions are input into the Transformer, which encodes the atomic fractions through a self-attention mechanism to generate a stoichiometric matrix. In the embedding layer, atomic numbers are mapped to high-dimensional vectors, and residual connections and multi-layer nonlinear transformations are performed on the high-dimensional vectors to obtain the chemical embedding matrix; Obtain the physical descriptor vector and position index of the element. Use the sine / cosine function in the position encoding layer to generate the position encoding of the position index. Concatenate the physical descriptor vector and the position encoding through a fully connected layer to output the position embedding matrix of the physical descriptor. The chemical embedding matrix and the stoichiometric matrix are added element-wise and then concatenated with the position embedding matrix to generate a composition matrix. Composition features are then extracted from the composition matrix.

4. The method for predicting the optical properties of fluoride laser crystals based on deep learning according to claim 1, characterized in that, The source domain dataset for constructing fluoride crystals with crystal stability labels specifically includes: The structural parameters and chemical formulas of fluoride crystals are obtained, and by determining the crystal stability, stability labels are labeled on the structural parameters and chemical formulas of some fluoride crystals to construct a source domain dataset of fluoride crystals with crystal stability labels.

5. The method for predicting the optical properties of fluoride laser crystals based on deep learning according to claim 4, characterized in that, The step of training a source domain model based on the source domain dataset, using the structural parameters and chemical formula of fluoride crystals as input, to obtain a crystal stability prediction model, specifically includes: The Adam optimizer, coupled with periodic learning rate scheduling, is used to train the source domain model with the MSE of multi-task learning as the loss function. The loss function uses mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 Three evaluation metrics are used. By determining the number of iterations and the initial learning rate, the evaluation metrics in the loss function are fine-tuned through iterative training until each evaluation metric meets the preset iteration conditions. The source domain model that satisfies the preset iteration conditions for each evaluation index is used as the crystal stability prediction model.

6. The method for predicting the optical properties of fluoride laser crystals based on deep learning according to claim 5, characterized in that, The detection head is used to map the features extracted by the feature extractor to the prediction results of the crystal optical properties, specifically including: The detection head is a multi-output structure, and each type of optical property in the multi-output structure corresponds to a fully connected layer; Structural and compositional features were extracted using the structure branch GNN and composition branch CoTAN in the crystal stability prediction model, respectively. The extracted structural and compositional features are input into the detection head, which performs classification and detection through multiple output structures and outputs optical property prediction results.

7. The method for predicting the optical properties of fluoride laser crystals based on deep learning according to claim 4, characterized in that, The acquisition of the structural parameters and chemical formula of fluoride crystals specifically includes: Based on historical documents and materials databases, the chemical formula of fluoride crystals was determined; Structural parameters of fluoride crystals were extracted using the Materials Project database.

8. A deep learning-based optical property prediction system for fluoride laser crystals, characterized in that, include: A source domain model construction module is used to construct a source domain model, which includes a structural branch GNN, a compositional branch CoTAN, and a fully connected layer. The structural branch GNN extracts the atomic diagram, line diagram, and dihedral diagram of the fluoride crystal based on structural parameters, and determines the structural features of the fluoride crystal based on these diagrams. The compositional branch CoTAN obtains and fuses the chemical embedding matrix, positional embedding matrix, and stoichiometric matrix of the fluoride crystal based on the chemical formula to generate the compositional features of the fluoride crystal. The fully connected layer concatenates and fuses the structural and compositional features, and outputs the crystal stability prediction result through mapping. A source domain dataset of fluoride crystals with crystal stability labels is constructed. Based on the source domain dataset, the source domain model is trained using the structural parameters and chemical formula of the fluoride crystal as input to obtain the crystal stability prediction model. An optical property prediction module is used to construct an optical property prediction model. The optical property prediction model includes a feature extractor and a detection head connected in sequence. The feature extractor is the structure branch GNN and the composition branch CoTAN in the crystal stability prediction model. The detection head is used to map the features extracted by the feature extractor to the prediction results of the crystal's optical properties. The structural parameters and chemical formula of the fluoride laser crystal to be predicted are obtained, input into the optical property prediction model, and the optical property prediction results are obtained.

Images