Graph neural network-based method and system for predicting formation energy of charged defects in iii-v semiconductor, electronic device, and storage medium

Abstract

A III-V semiconductor charged defect formation energy prediction method and system based on a graph neural network, an electronic device, and a computer readable storage medium. The method comprises: obtaining crystal structure, atomic feature, adjacency relationship, bond length, bond angle, and charge state data of a defect sample; inputting the atomic feature and the adjacency relationship into a first encoder to extract atomic neighborhood structure features, and inputting the bond length and the bond angle into a second encoder to extract bond length and bond angle relationship features; after fusing the two features, injecting a global feature node composed of atomic statistics and supercell scale before charge interaction; after explicit coding of the charge state, interacting and fusing with the structure features to obtain joint features; and sequentially performing candidate formation energy point training and regression training to obtain a prediction model. The present application fuses local topology and geometric information through a double-branch structure encoding, combines global feature enhancement and explicit charge condition coding, and improves the accuracy and stability of charged defect formation energy prediction.

Description

Technical Field

[0001] This invention relates to the fields of semiconductor defect calculation, materials informatics, artificial intelligence and graph neural network technology, and specifically to a method, system, electronic device and computer-readable storage medium for predicting the formation energy of charged defects in III-V semiconductors based on graph neural networks. Background Technology

[0002] III-V group semiconductor materials are widely used in light-emitting devices, detectors, high-frequency communication devices, and high-reliability electronic devices. Intrinsic defects, impurity defects, and their charged states in the materials affect carrier concentration, doping limits, defect compensation behavior, nonradiative recombination processes, and device stability. Defect formation energy is an important parameter for evaluating defect thermodynamic stability, doping feasibility, and defect engineering control schemes.

[0003] In existing technologies, defect formation energies are typically obtained through first-principles supercell calculations. For charged defects, the calculation process involves not only defect structure relaxation but also factors such as charge correction, potential alignment, and finite size effects. Therefore, the computational cost of a single defect state is high, making it difficult to efficiently support the screening of large-scale candidate defects, doped elements, or host material systems.

[0004] With the application of machine learning methods such as graph neural networks in material property prediction, existing methods can predict some material properties based on crystal structure diagrams. However, existing methods are mostly geared towards modeling the properties of neutral crystals or overall materials, typically using the structure diagram formed by atomic nodes and their adjacency relationships as the main input. For the task of predicting the formation energy of charged defects, the defect charge state has a direct impact on the formation energy. If the defect charge state is not used as an independent input variable in feature interaction, the model's ability to express the differences in formation energy under different charge states is limited.

[0005] Furthermore, the formation energy of charged defects is related not only to the local atomic neighborhood of the defect but also to the overall composition of the host material, the supercell scale, and local geometric relationships. Existing graph neural network methods based on local message passing have certain advantages in extracting topological information of atomic neighborhoods, but the overall characteristics of the host material and the statistical information at the supercell scale may be weakened during multi-layer message passing. At the same time, if only a single structural encoding path is used, it is difficult to simultaneously take into account the topological information of atomic neighborhoods and local geometric relationships such as bond lengths and bond angles.

[0006] Finally, charged defect data typically exhibits characteristics such as limited sample size, wide distribution of formation energy labels, and significant differences between different host systems. Directly employing single-stage regression training may result in insufficient discriminative power of the learned features in the energy dimension, thus affecting prediction stability and regression accuracy. Furthermore, when switching between different encoder structures, existing methods often lack a unified structure that maintains the stability of the charge encoding path and the overall feature fusion path of the host material, limiting the model's scalability across different data scales and structural description methods.

[0007] Therefore, it is necessary to propose a method for predicting the formation energy of charged defects that can simultaneously integrate local structure information, bond length and bond angle relationship information, overall characteristics of the host material, and defect charge state information, and combine it with a training mechanism suitable for the characteristics of defect data to improve the accuracy, stability, and scalability of the prediction of the formation energy of charged defects in III-V semiconductors. Summary of the Invention

[0008] This invention addresses the problems of insufficient utilization of defect charge states, insufficient representation of overall host material characteristics, insufficient utilization of bond length and bond angle relationship information, and inconvenience in replacing structural encoders in existing methods for predicting the formation energy of charged defects in III-V semiconductors. It provides a method, system, electronic device, and computer-readable storage medium for predicting the formation energy of charged defects in III-V semiconductors based on graph neural networks. This method can comprehensively utilize local defect structure information, bond length and bond angle relationship information, overall host material characteristics, and defect charge state information within the same prediction framework, thereby improving the accuracy, stability, and applicability of charged defect formation energy prediction.

[0009] According to a first aspect of the present invention, a method for predicting the formation energy of charged defects in group III-V semiconductors based on graph neural networks is provided. The method includes: acquiring crystal structure data, atomic feature data, atomic adjacency relationship data, bond length data, bond angle data, and defect charge state data of a defect sample; constructing an atomic neighborhood graph based on the atomic feature data and atomic adjacency relationship data, and extracting atomic neighborhood structural features using a first structural encoder; constructing a bond length and bond angle relationship graph based on the bond length data and bond angle data, and extracting bond length and bond angle relationship features using a second structural encoder; and fusing the atomic neighborhood structural features and the bond length and bond angle relationship features to obtain a fused structural feature.

[0010] According to the prediction method of the present invention, after obtaining the fused structural features, a global feature node is constructed based on the statistical information of all atomic features in the defect sample and the supercell-scale information. Before the defect charge state participates in feature fusion, the global feature node is fused with the fused structural features to obtain the structural features after fusion of global features. The global feature node can be composed of at least one of the mean, standard deviation, variance, number of supercell atoms, and their transformation values ​​of all atomic features. By introducing the global feature node, the overall composition information and supercell-scale information of the host material can be supplemented, reducing the problem of insufficient expression of the overall material features when relying solely on local atomic neighborhood information for prediction.

[0011] According to the prediction method of the present invention, the defect charge state is encoded as an independent input variable to obtain charge features. These charge features are then interactively fused with structural features obtained by fusing global features to obtain joint features containing defect charge state information. The charge features can be obtained through multilayer perceptrons, embedded coding, or continuous numerical mapping. The interactive fusion can employ attention mechanisms, gated fusion, feature modulation, conditional normalization, or combinations thereof. Thus, the same defect structure can generate mutually distinguishable feature results under different defect charge states, thereby improving the model's ability to express the differences in formation energy between different charge states.

[0012] According to the prediction method of the present invention, the model training process includes two stages. The first stage is candidate formation energy value point training, in which the range of the target defect formation energy is divided into multiple candidate formation energy value points, and basis function expansion is performed on the candidate formation energy value points. Then, a weighted predicted formation energy is constructed based on the similarity between the joint features of the samples and the features of the candidate formation energy value points, and the model is trained by combining classification loss and regression loss. The second stage is formation energy regression training. Based on the training results of the first stage, the defect formation energy label is used as a supervision signal to train the regression output part or the model as a whole to obtain a prediction model for outputting the predicted value of the defect formation energy.

[0013] In one embodiment, the first and second structure encoders are configured as pluggable, independent modules. While maintaining the charge encoding path, global feature node path, and regression output path unchanged, at least one of the first and second structure encoders can be replaced with other graph neural network encoders or crystal structure feature extractors suitable for crystal structure data. Therefore, this invention can adapt to different data scales, different crystal structure description methods, and different model implementation methods.

[0014] According to a second aspect of the present invention, a system for predicting the formation energy of charged defects in group III-V semiconductors based on graph neural networks is provided. The system includes a data acquisition module, a dual-branch structure encoding module, a structure fusion module, a global feature node construction module, a charge encoding module, an interactive fusion module, and a training and prediction module. The data acquisition module is used to acquire defect sample data; the dual-branch structure encoding module is used to extract atomic neighborhood structure features and bond length and bond angle relationship features respectively; the structure fusion module is used to obtain fused structure features; the global feature node construction module is used to construct and fuse global feature nodes; the charge encoding module is used to encode the defect charge state as charge features; the interactive fusion module is used to obtain joint features containing defect charge state information; and the training and prediction module is used to train the model and output predicted defect formation energy values.

[0015] According to a third aspect of the present invention, an electronic device is provided, the electronic device comprising a processor and a memory, the memory storing a computer program, which, when executed by the processor, implements the above-described method for predicting the formation energy of charged defects in III-V semiconductors based on graph neural networks.

[0016] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when executed by a processor, the computer program implements the above-described method for predicting the formation energy of charged defects in III-V semiconductors based on graph neural networks.

[0017] Compared with existing technologies, the present invention has at least the following beneficial effects: The present invention uses the defect charge state as an independent input variable to participate in feature fusion, which can distinguish the formation energy differences of the same defect structure under different defect charge states; by introducing the overall composition information and supercell-scale information of the host material through global feature nodes, it can improve the problem of insufficient overall feature expression when relying solely on local message passing; by extracting topological information and local geometric information through atomic neighborhood graphs and bond length and bond angle relationship graphs respectively, it can improve the ability to express local structural changes of defects; by combining candidate formation energy value point training and formation energy regression training in a two-stage training method, it can improve training stability and prediction accuracy under conditions of limited sample size and large distribution span of formation energy labels; simultaneously, by retaining the charge encoding path, global feature node path, and regression output path through the structure encoder interface, it is convenient to replace and extend between different structure encoders. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a charge-aware graph learning model architecture that introduces virtual global environment nodes, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the benchmark performance comparison and robustness evaluation of different model structures under a fixed dataset partitioning in an embodiment of the present invention. Figure 3 This is a bar chart comparing the effect of introducing a global token on improving model prediction error in embodiments of the present invention; Figure 4 This is a schematic diagram of the prediction accuracy and error distribution of the preferred model provided in the embodiments of the present invention; Figure 5 This is a schematic diagram of the verification analysis of global context node improvement of graph-level sample separability in an embodiment of the present invention; wherein (a) is a heatmap of graph embedding cosine similarity under different node strategies (no nodes, static nodes, attention nodes), and (b) and (c) are quantitative comparison bar charts of average off-diagonal cosine similarity and effective rank under the above strategies, respectively. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Where there is no conflict, the following embodiments and their technical features can be combined with each other.

[0020] In this embodiment of the invention, the global feature node refers to a feature node generated from the statistics of all atomic features in the defect sample and the supercell-scale information, used to provide the model with overall material information; the bond length and bond angle relationship graph refers to a graph structure established based on the bond lengths between atoms and the bond angle relationships formed by adjacent atoms, used to provide the model with local geometric relationship information. The above names are only used to clearly describe each processing object and should not be construed as limiting the specific network structure.

[0021] Example 1: A Method for Predicting the Formation Energy of Charged Defects in Group III-V Semiconductors Based on Graph Neural Networks S1. Obtain crystal structure data, atomic feature data, atomic adjacency relationship data, bond length data, bond angle data, and defect charge state data of the defect sample.

[0022] In this embodiment, the defect sample data comes from a first-principles calculation dataset based on density functional theory (DFT). Each defect sample includes at least the defect supercell structure (containing atomic coordinates and lattice parameters), atomic type information, atomic initial characteristics, atomic adjacency relationships, interatomic bond length characteristics, atomic three-body angle characteristics, defect charge state (q), and a defect formation energy label corrected by first-principles calculations.

[0023] Initial atomic features can be expressed as element embedding vectors, physicochemical property vectors, or combinations thereof; bond length features can be expressed using Gaussian function expansion; angle features can be expressed as angle values, angle cosine values, or their basis function expansion forms. The defect formation energy label can be a corrected defect formation energy obtained under preset chemical potential conditions, Fermi level reference conditions, charge correction methods, and potential alignment methods; in other embodiments, at least one of the above-mentioned chemical potential conditions, Fermi level reference conditions, charge correction methods, and potential alignment methods can also be used as model input features. After unified preprocessing, the above data is provided as input to subsequent steps.

[0024] S2. Construct an atomic neighborhood graph based on the atomic feature data and atomic adjacency relationship data, and extract the atomic neighborhood structural features through a first structure encoder; construct a bond length and bond angle relationship graph based on the bond length data and bond angle data, and extract the bond length and bond angle relationship features through a second structure encoder.

[0025] Specifically, the first structural encoder extracts the structural features of the atomic neighborhood by performing message passing, aggregation, and pooling operations on the atomic node features and adjacent edge information. The second structural encoder constructs a bond length and bond angle relationship graph based on bond length and bond angle information. The bond length and bond angle relationship graph can use chemical bonds, adjacent atom pairs, or edges as nodes or interaction units, and takes the angular relationship formed by three atoms as the coupling relationship, thereby simultaneously expressing the local geometric configuration corresponding to the bond length and the three-body angle.

[0026] Preferably, the first structural encoder can be an atomic neighborhood graph convolutional encoder, such as a CGCNN-type encoder; the second structural encoder can be a bond length and bond angle relationship encoder, such as a CEGANN-type encoder.

[0027] Furthermore, the first and second structure encoders are connected via a structure encoder interface. While maintaining the explicit charge encoding path, the global feature node path, and the regression output path unchanged, at least one of the first and second structure encoders can be replaced with other graph neural network encoders or crystal structure feature extractors suitable for crystal structure data.

[0028] S3. The atomic neighborhood structure features and the bond length and bond angle relationship features are fused to obtain the fused structure features.

[0029] After mapping the first and second structural features to a unified feature dimension, a fused structural feature is constructed using gated fusion, residual fusion, or weighted fusion. Preferably, the unified feature dimension can be set to 384 dimensions; this value is merely an example and does not constitute a limitation on the scope of protection of this invention.

[0030] S4. Construct global feature nodes based on the statistical information of all atomic features in the defect sample and the supercell scale information, and fuse the global feature nodes with the fused structural features before the defect charge state participates in feature fusion to obtain the structural features after fusing global features.

[0031] Specifically, the initial features of all atoms in the defect sample are statistically analyzed to obtain the mean vector and discreteness statistical vector of atomic features, and the total number of supercell atoms or their transformation values ​​are counted. The mean vector, discreteness statistical vector and supercell scale information are concatenated to generate the overall material feature vector. The overall material feature vector is mapped to a unified feature dimension through a multilayer perceptron to generate global feature nodes.

[0032] Preferably, the global feature node is composed of at least one of the mean, standard deviation, variance, supercell atomic number, and their transformed values ​​of all atomic features. When the initial atomic feature dimension is 92 dimensions, its mean vector and standard deviation vector can be calculated separately and concatenated with the atomic number feature in the form of log(1+N) to obtain a 185-dimensional overall material feature vector, which is then mapped to a unified 384-dimensional feature dimension. The above dimension configuration is only an example and can be adjusted according to the atomic feature dimension and model capacity in actual implementation.

[0033] S5. The defect charge state is used as an independent input variable for explicit charge encoding to obtain charge features. The charge features are then interactively fused with the structural features after fusing global features to obtain joint features containing defect charge state information.

[0034] Specifically, the defect charge state is taken as a scalar input and mapped to charge features using a multilayer perceptron, embedding coding, or continuous numerical mapping. Subsequently, through attention interaction, gating interaction, conditional normalization, or feature modulation, the charge features are coupled with structural features enhanced by global feature nodes to obtain joint features containing defect charge state information. Thus, the model can explicitly distinguish similar structural configurations under different defect charge states and express the charge-related differences in the formation energy of charged defects in the joint features.

[0035] Preferably, the explicit charge encoder can be a multilayer perceptron with a 1-dimensional input layer, 64-dimensional and 128-dimensional hidden layers, and a 384-dimensional output layer; the charge interaction module can adopt a combination of multi-head attention-based interaction and gating fusion to obtain the final joint features. The above number of layers and dimensions are only examples and do not constitute necessary limitations.

[0036] S6. Perform two-stage training based on the joint features containing defect charge state information to obtain a formation energy prediction model.

[0037] Specifically, the two-stage training includes: The first stage is candidate formation energy point training. Multiple candidate formation energy points are preset within the target formation energy range, and continuous-valued basis function expansion is performed on these candidate formation energy points. Similarity calculation is performed between the joint features of the samples and the features of the candidate formation energy points, and a weighted predicted formation energy is generated based on the similarity. Simultaneously, a classification loss (e.g., Cross-Entropy Loss) is constructed using the nearest neighbor candidate points of the true formation energy as supervision, and a regression loss (e.g., Smooth L1 Loss) is constructed using the difference between the true formation energy and the weighted predicted formation energy. By jointly optimizing the classification loss and regression loss, candidate formation energy point training is achieved, enabling the joint features to form a discriminative distribution along the energy dimension.

[0038] Preferably, 33 candidate formation energy points can be set within the target formation energy range, and 64 Gaussian functions can be used to perform continuous value expansion on the candidate formation energy points to enhance the expressive power of the energy axis neighborhood structure. The number of candidate formation energy points and basis functions mentioned above are only examples and can be adjusted according to the distribution of training data.

[0039] The second stage is regression fine-tuning. The joint features obtained in the first stage are input into the regression output module to form scalar labels as supervision signals for regression training. Training methods may include freezing the common network layers and training only the regression output module, partially fine-tuning the common network layers and interaction layers, or fine-tuning the entire model.

[0040] S7. Input the defect sample to be predicted into the trained formation energy prediction model and output the defect formation energy prediction value.

[0041] Specifically, during the inference phase, the defect sample to be predicted is input into the trained model, and sequentially undergoes atomic neighborhood structure encoding, bond length and bond angle relationship encoding, global feature node fusion, explicit charge encoding, and joint regression to output a predicted defect formation energy value. The predicted value and its ranking results can be used for rapid screening of candidate defect configurations, doping feasibility assessment, defect thermodynamic stability ranking, and priority scheduling of first-principles calculation tasks.

[0042] In one implementation, the model is trained and tested using a fixed data partition, and the prediction stability is evaluated through multiple repeated experiments. Figure 2 As shown, the method can achieve lower prediction errors and smaller fluctuations in repeatability in forming energy prediction tasks, and can be combined with encoders of different structures. Figure 3 As shown, the applicability of explicit charge encoding and global feature enhancement paths to charged defect samples is further verified by using the statistical errors of defect charge states and different charge subsets. Figure 4 As shown, there is a predetermined consistency between the model predictions and the first-principles energy labels, and the residual distribution can be used to characterize the prediction bias. Figure 5 As shown, the introduction of global feature nodes improves the discriminative power of sample features, making joint features easier to distinguish by the regression output module, thereby improving the predictive stability of the formation energy. The above verification results are used to illustrate the technical effects of the present invention and do not limit the implementation of the present invention on other defective datasets, structural encoders, or training strategies.

[0043] Based on the same inventive concept, this invention also provides a system for predicting the formation energy of charged defects in III-V semiconductors based on graph neural networks. The system includes a data acquisition module, a dual-branch structure encoding module, a structure fusion module, a global feature node construction module, an explicit charge encoding module, an interactive fusion module, and a training and prediction module. The data acquisition module acquires crystal structure data, atomic feature data, adjacency relationship data, bond length data, bond angle data, and defect charge state data of defect samples. The dual-branch structure encoding module extracts atomic neighborhood structure features and bond length and bond angle relationship features respectively through the structure encoder interface. The structure fusion module fuses the atomic neighborhood structure features and the bond length and bond angle relationship features to obtain fused structure features. The global feature node construction module constructs global feature nodes based on atomic feature statistics and supercell scale information, and generates structure features after fusing global features before defect charge state injection. The explicit charge encoding module encodes defect charge states as charge features. The interactive fusion module interacts and fuses the charge features with the structure features after fusing global features to obtain joint features containing defect charge state information. The training and prediction module performs candidate formation energy point training and formation energy regression training based on the joint features containing defect charge state information, and outputs the predicted value of defect formation energy.

[0044] Based on the same inventive concept, the present invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable by the processor. When the processor executes the computer program, it implements the above-described method for predicting the formation energy of charged defects based on graph neural networks. The electronic device can be a server, workstation, personal computer, graphics processing unit server, cloud computing node, or other computing device capable of performing model training and inference tasks.

[0045] Based on the same inventive concept, the present invention also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-described method for predicting the formation energy of charged defects based on graph neural networks. The computer-readable storage medium may include a read-only memory, random access memory, a magnetic disk, an optical disk, a solid-state drive, flash memory, or other non-transitory storage media capable of storing a computer program.

[0046] Any matters not covered in the above embodiments of the present invention are well-known in the art.

[0047] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A method for predicting the formation energy of charged defects in III-V semiconductors based on graph neural networks, characterized in that, include: Acquire crystal structure data, atomic feature data, adjacency relationship data, bond length data, bond angle data, and defect charge state data of defect samples; The atomic feature data and the adjacency relationship data are input into the first structure encoder to construct an atomic neighborhood graph, and the atomic neighborhood structure features are extracted through the first structure encoder. The bond length data and the bond angle data are input into the second structure encoder to construct a bond length and bond angle relationship diagram, and the bond length and bond angle relationship features are extracted through the second structure encoder. The atomic neighborhood structure features and the bond length and bond angle relationship features are fused to obtain the fused structure features; Global feature nodes are constructed based on the statistical information and supercell scale information of the atomic feature data. Before the defect charge state features participate in feature fusion, the global feature nodes are fused with the fused structural features to obtain the structural features after fusion of global features. The defect charge state data is used as a conditional variable independent of the structure input for explicit charge encoding to obtain charge characteristics; The charge features are interactively fused with the structural features obtained after fusing global features to obtain joint features that include defect charge state information; Based on the joint features containing defect charge state information, candidate formation energy point training and formation energy regression training are performed sequentially to obtain a formation energy prediction model. as well as The crystal structure data, atomic feature data, adjacency relationship data, bond length data, bond angle data, and defect charge state data of the defect sample to be predicted are input into the formation energy prediction model, and the predicted value of the defect formation energy of the defect sample to be predicted is output.

2. The method according to claim 1, characterized in that, The global feature node is composed of at least one of the mean, standard deviation, variance, number of atoms in the supercell, and transformed value of the number of atoms in the supercell, and is used to characterize the overall chemical and scale information of the host material where the defect sample is located.

3. The method according to claim 1, characterized in that, The explicit charge coding encodes the defect charge state as a conditional variable independent of the structural input. The explicit charge coding employs embedded coding, multilayer perceptron coding, continuous numerical mapping coding, or a combination thereof.

4. The method according to claim 1, characterized in that, The interaction and fusion of the charge features and the structural features after the fusion of global features adopt attention interaction, gating fusion, feature modulation, conditional normalization or a combination thereof to form a joint feature sensitive to different defect charge states.

5. The method according to claim 1, characterized in that, The fusion of the atomic neighborhood structure features and the bond length and bond angle relationship features includes at least one of the following: normalization, linear or nonlinear projection, splicing, weighted summation, residual fusion, and gated fusion.

6. The method according to claim 1, characterized in that, The candidate formation energy point training includes: Discretize the target defect formation energy range into multiple candidate formation energy points; The continuous values ​​of each candidate formation energy point are expanded using basis functions to obtain the candidate point features corresponding to each candidate formation energy point. The joint features containing defect charge state information are compared with the features of each candidate point to calculate similarity. The values ​​of each candidate formation energy point are then weighted and summed based on the calculated similarity to obtain the weighted predicted formation energy. A classification loss is constructed using the nearest neighbor candidate point of the true formation energy as the supervision target, and a regression loss is constructed using the difference between the true formation energy and the weighted predicted formation energy. The model parameters are trained by jointly optimizing the classification loss and the regression loss.

7. The method according to claim 1, characterized in that, The formation energy regression training is performed after the candidate formation energy points are trained. The defect formation energy scalar label is used as the supervision signal, and the formation energy regression model is obtained by freezing the common network layer training, fine-tuning part of the network or fine-tuning the whole model.

8. The method according to claim 1, characterized in that, The first structural encoder and the second structural encoder exist as independent graph representation learning branches; while keeping the charge feature interaction fusion path, the global feature node fusion path, and the formation energy regression training path unchanged, at least one of the first structural encoder and the second structural encoder can be replaced by other types of graph neural network encoders.

9. The method according to claim 1, characterized in that, The bond length and bond angle relationship diagram uses chemical bonds, adjacent atom pairs or edges as nodes or interaction units, and the bond angle between any two chemical bonds sharing the same central atom as the edge relationship or interaction relationship. The bond length information and bond angle information are used together to characterize the three-body geometry of the defect locality.

10. The method according to claim 1, characterized in that, The defect formation energy is a modified defect formation energy obtained under at least one of the following conditions: preset chemical potential conditions, Fermi level reference conditions, charge correction method, and potential alignment method. Alternatively, at least one of the preset chemical potential conditions, Fermi level reference conditions, charge correction methods, and potential alignment methods can be used as model input features.

11. The method according to claim 1, characterized in that, The defect samples include vacancies, interstitial atoms, substitutional atoms, recombination defects, or charged defects in III-V group semiconductor materials.

12. A system for predicting the formation energy of charged defects in III-V semiconductors based on graph neural networks, characterized in that, include: The data acquisition module is used to acquire crystal structure data, atomic feature data, adjacency relationship data, bond length data, bond angle data, and defect charge state data of defect samples. The dual-branch structure encoding module is used to input the atomic feature data and the adjacency relationship data into the first structure encoder to extract atomic neighborhood structure features, and to input the bond length data and the bond angle data into the second structure encoder to extract bond length and bond angle relationship features. The structure fusion module is used to fuse the atomic neighborhood structure features and the bond length and bond angle relationship features to obtain fused structure features; The global feature node construction module is used to construct global feature nodes based on atomic feature statistics and supercell scale information, and to generate structural features after fusing global features before defect charge state injection. An explicit charge encoding module is used to encode the defect charge state data as a charge feature as a conditional variable independent of the structural input; An interactive fusion module is used to interactively fuse the charge features with the structural features after fusing the global features to obtain joint features containing defect charge state information; The training and prediction module is used to sequentially perform candidate formation energy point training and formation energy regression training based on the joint features containing defect charge state information to obtain a formation energy prediction model, and input the defect sample to be predicted into the formation energy prediction model to output the defect formation energy prediction value.

13. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the method according to any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 11.

Images