Atomic feature updating method based on graph convolutional layer general interatomic potential energy model

Abstract

The application discloses an atomic feature updating method based on a general interatomic potential model of a graph convolution layer, belongs to the field of chemistry and material science combined with neural networks, and discloses an improved invariance general interatomic potential model architecture, wherein a physically guided inductive bias mechanism is introduced, so that the model can automatically distinguish the contribution weight of different atomic types to interactive information, the feature confusion problem in a traditional splicing method is avoided, and therefore, the multi-body action law in a complex material system can be more accurately captured, and a distance decay weight mechanism designed based on a physical law considers the decay characteristics of the influence between atoms in a Message convergence stage. The injection of such prior knowledge guides the model to learn the physical law, reduces the dependence of the model on the amount of training data, and significantly improves the generalization ability and stability of the model under the condition of limited data, especially in the modeling of rare material systems.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of chemistry and materials science combined with neural networks, and particularly relates to an atomic feature updating method based on a general interatomic potential model of a graph convolution layer. BACKGROUND

[0002] Universal Interatomic Potentials are a class of potential energy models that can be applied to a variety of material systems to describe the interaction forces between atoms. Compared with specific material or element-specific potential functions, universal interatomic potentials have more extensive applicability and can provide accurate energy and force calculations under different element combinations, structure types and environmental conditions. Such potential energy models are usually parameterized by combining classical potential functions, first-principle calculation data and machine learning methods to improve accuracy and universality. In recent years, universal interatomic potential models based on graph neural networks (such as MACE, CHGNet, Eqv2, etc.) have also been proven to have good generalization in various downstream tasks.

[0003] Existing universal interatomic potential models can be mainly divided into four categories: conservative-isotropic models (satisfying both energy conservation and isotropy), non-conservative-isotropic models (not satisfying energy conservation but satisfying isotropy), conservative-invariant models (satisfying both energy conservation and rotational invariance), and non-conservative-invariant models (not satisfying energy conservation but satisfying rotational invariance).

[0004] Conservative potential models calculate forces by taking the negative derivative of the potential energy surface with respect to atomic positions. However, predicting forces as derivatives requires an additional backpropagation step, which increases the computational cost of the potential model. On the other hand, some networks directly predict forces by using separate force heads, thereby improving efficiency. Although models that directly predict forces can achieve high accuracy, they can cause significant errors in certain property prediction tasks (such as molecular dynamics simulations under the NVT ensemble) due to their non-conservative nature.

[0005] Isotropic series models introduce sufficient prior knowledge (rotational isotropy) to enable superior performance even on small data sets. However, the high training cost (memory, training time) makes it difficult to scale the parameter size of conservative isotropic models. Invariant models have relatively low training costs, but due to the lack of sufficient inductive bias in the model architecture, they are more likely to overfit when the training data is small.

[0006] As in the Message construction phase, the existing invariance model splices neighbor atoms, central atoms and edge information, and then changes the dimension through a nonlinear layer. In this process, neighbor atom information, central atom information and edge information are respectively assigned different weights to finally form the Message. This construction method only considers assigning different weights to different features (neighbor atoms, central atoms, edges), but in fact, different atomic types also have different effects on the Message. Therefore, this inductive bias should be considered in the Message construction phase.

[0007] In the Message aggregation phase, the isotropic information aggregation method is used in the Message Passing mechanism of CHGNet, that is, equal weights are assigned to all neighbor atoms when constructing the interaction information between atoms. Although this design ensures translational symmetry, it violates the basic physical law of material systems, that is, the strength of atomic interaction decreases exponentially with distance. Within a range of 5 Å, about 78% of the interatomic distances are concentrated in the 2-3 Å interval, and the contribution of atoms beyond 4 Å has decayed to less than 5% of the initial value. The existing model does not encode this prior knowledge into the neural network architecture, resulting in the model passively learning the distance decay law through a large amount of data, which significantly increases the training difficulty, especially in small sample scenarios (such as rare material systems), and the generalization performance decreases by 32.7%. SUMMARY

[0008] In view of the above problems of the prior art, the present application provides an atomic feature updating method and system based on a general interatomic potential model of a graph convolution layer.

[0009] The first aspect of the present application provides an atomic feature updating method based on a general interatomic potential model of a graph convolution layer, comprising the following steps:

[0010] Obtain atomic type features and edge features and input them into a Seg-Attention graph convolution layer, wherein the Seg-Attention graph convolution layer includes a Message-Embed module and a Message-Agg module;

[0011] Calculate the attention weights of the central atom and the neighbor atoms in the atomic type features through the Message-Embed module, and generate a segmented index list and a central-neighbor mapping table;

[0012] Based on the segmented index list and the central-neighbor mapping table, perform a two-stage aggregation mechanism to generate initial feature information;

[0013] The Message-Agg module processes the interatomic distance to obtain a feature update weight, and the initial feature information is multiplied by the feature update weight to generate updated final atomic features.

[0014] In some embodiments, the Message-Embed module calculates the attention weight of the center atom and the neighbor atom in the atomic type feature, and generates a segmented index list and a center-neighbor mapping table, including:

[0015] The first layer atomic type feature is inputted. The atomic type feature is divided into a center atomic feature and a neighbor atomic feature .

[0016] The center atomic feature and the neighbor atomic feature are transformed by a linear layer to generate a center atomic intermediate feature and a neighbor atomic intermediate feature , and the transformation formula is:

[0017]

[0018] wherein, the weight matrix of the linear layer is , and the bias term is .

[0019] The center atomic intermediate feature and the neighbor atomic intermediate feature are executed by a Seg-Softmax mechanism to generate an attention weight , and the Seg-Softmax mechanism is used to dynamically adjust the weight of information according to the difference of atomic types, and the Seg-Softmax mechanism calculation formula is:

[0020]

[0021] wherein, is the feature of the center atomic / neighbor atomic feature after linear layer change, is a row index, is a column index, represents a set of all rows with index , and represents the i-th atomic ordinal number of the segmented index list, and the attention weight includes a center attention weight and a neighbor attention weight .

[0022] The segment index list generation process is: according to the different types of atoms, the atoms are divided into segments, each segment corresponds to an atomic number, and a segment index list is obtained , , , wherein Z represents the largest atomic number in the data set;

[0023] The center-neighbor mapping table generation process is: defining a center-neighbor mapping table , the first list represents a center atom index list , the second list represents a neighbor atom index list .

[0024] In some embodiments, the two-stage aggregation mechanism is performed based on the segment index list and the center-neighbor mapping table to generate initial feature information, and the first stage includes:

[0025] The edge feature is transformed through the linear layer to generate an edge intermediate feature ;

[0026] The edge intermediate feature is multiplied by the center attention weight and the neighbor attention weight respectively to generate weighted center atom features and neighbor atom features ;

[0027] The weighted center atom features and neighbor atom features are aggregated respectively to generate compressed center atom features and neighbor atom features .

[0028] In some embodiments, the two-stage aggregation mechanism is performed based on the segment index list and the center-neighbor mapping table to generate initial feature information, and the second stage includes:

[0029] The compressed center atom features and neighbor atom features are spliced to generate spliced features , and the splicing formula is:

[0030]

[0031] The spliced features are nonlinearly transformed by a multi-layer perceptron to generate fused atom features , and the activation function of the MLP is SiLU, and the formula is:

[0032]

[0033] in, , and These are the trainable weights of the first and second layers in an MLP, respectively.

[0034] , For the Sigmoid function;

[0035] Fusion atomic features Based on the central atom index list and neighbor atomic index list Expand into and The extended formula is:

[0036]

[0037] Expanded central atom features Neighboring atomic characteristics Sum of edge features The data is then concatenated to generate initial feature information. The splicing formula is:

[0038] .

[0039] In some embodiments, the step of processing the inter-atomic distance through the Message-Agg module to obtain feature update weights, and then multiplying the initial feature information by the feature update weights bitwise to generate the updated final atomic features, includes the following steps:

[0040] Given interatomic distance First, the Bessel function is applied to process it, generating smoothly decaying weights. The formula is:

[0041]

[0042] in Set the cutoff radius to 6. The first base One dimension, This represents the highest dimension expanded using the Bessel function. The envelope function, which achieves smooth decay over distance, is calculated using the following formula:

[0043]

[0044] in, ;

[0045] The processed distance features The feature update weight is generated by transforming through a trainable linear layer .

[0046] In some embodiments, the feature update weight is generated by processing the interatomic distance through the Message-Agg module, and the initial feature information is multiplied by the feature update weight to generate updated final atomic features, including edge feature updates and atomic feature updates, and the edge feature updates are:

[0047] The initial feature information is given The updated edge features are generated by processing through the multi-layer perceptron The formula of the nonlinear transformation is:

[0048] .

[0049] In some embodiments, the atomic feature updates are: before the aggregation operation, the initial feature information is multiplied by the feature update weight to obtain weighted initial feature information The weighted initial feature information is aggregated to generate updated final atomic features The aggregation operation represents summing all edges related to atom i.

[0050] The second aspect of the present application proposes an atomic feature update system based on a general interatomic potential model of a graph convolution layer, including a feature input module, an index analysis module, a two-stage aggregation module, and an atomic feature update module:

[0051] The feature input module is used to obtain atomic type features and edge features and input them into a Seg-Attention graph convolution layer, and the Seg-Attention graph convolution layer includes a Message-Embed module and a Message-Agg module.

[0052] The index analysis module is used to calculate the attention weight of the center atom and the neighbor atom in the atomic type feature through the Message-Embed module, and generate a segmented index list and a center-neighbor mapping table.

[0053] The two-stage aggregation module is used to perform a two-stage aggregation mechanism based on the segmented index list and the center-neighbor mapping table to generate initial feature information.

[0054] The atomic feature updating module is configured to obtain a feature updating weight by processing an interatomic distance through the Message-Agg module, multiply the initial feature information with the feature updating weight bit by bit, and generate an updated final atomic feature.

[0055] In a third aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0056] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program implements the steps of the above method when executed by a processor.

[0057] The present application has the following beneficial effects:

[0058] The improved invariance universal interatomic potential model architecture introduced in the present application improves the modeling accuracy of atomic interaction by introducing a physically guided inductive bias mechanism: in the Message construction stage, the atomic type-aware attention mechanism is used to realize the differentiated representation of the chemical characteristics between different elements. This design enables the model to automatically distinguish the contribution weight of different atomic types to the interaction information, avoids the feature confusion problem in the traditional splicing method, and thus more accurately captures the many-body interaction rules in complex material systems, and breaks through the small sample learning ability: the distance decay weight mechanism designed based on physical laws considers the decay characteristics of the interatomic influence in the Message aggregation stage. The injection of this prior knowledge guides the model to learn the physical laws, reduces the dependence of the model on the amount of training data, and significantly improves the generalization ability and stability of the model under limited data conditions, especially in the modeling of rare material systems. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The present application is a general flowchart.

[0060] Figure 2 The input and output relationship of the Seg-Attention graph convolution layer.

[0061] Figure 3 The system principle block diagram of the present application. DETAILED DESCRIPTION

[0062] Exemplary embodiments of the present application will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments set forth herein; on the contrary, these embodiments are provided so that the present application can be more thoroughly understood and the scope of the present application can be accurately conveyed to those skilled in the art.

[0063] The first aspect of the present application proposes an atomic feature updating method based on a general interatomic potential model of a graph convolution layer, as shown in the following formula (1): Figure 1 The method comprises the following steps:

[0064] S100: Obtain atomic type features and edge features and input them into a Seg-Attention graph convolution layer, wherein the Seg-Attention graph convolution layer comprises a Message-Embed module and a Message-Agg module;

[0065] As shown in the following formula (2), the Seg-Attention graph convolution layer is composed of two core modules: the Message-Embed module and the Message-Agg module. These two modules are respectively responsible for processing atomic type features and edge features, and finally generating updated atomic features and edge features. Figure 2

[0066] S200: Calculate attention weights of a center atom and neighbor atoms in the atomic type features by using the Message-Embed module, and generate a segmented index list and a center-neighbor mapping table;

[0067] In some embodiments, the step of calculating attention weights of a center atom and neighbor atoms in the atomic type features by using the Message-Embed module, and generating a segmented index list and a center-neighbor mapping table comprises:

[0068] input the first-layer atomic type features , the atomic type features are divided into center atom features and neighbor atom features ;

[0069] transform the center atom features and the neighbor atom features through a linear layer to generate center atom intermediate features and neighbor atom intermediate features , and the transformation formula is as follows:

[0070]

[0071] wherein, the weight matrix of the linear layer is , and the bias term is ;

[0072] perform a Seg-Softmax mechanism on the center atom intermediate features and the neighbor atom intermediate features to generate attention weights​​ , the Seg-Softmax mechanism is used to dynamically adjust the weight of information according to the difference of atomic types, and the calculation formula of the Seg-Softmax mechanism is:

[0073]

[0074] wherein, is the feature of the central atom / neighbor atom feature after linear layer change, is the row index, is the column index, represents the set of all rows with index , represents the i-th atomic number of the segmentation index list, and the attention weight includes the central attention weight and the neighbor attention weight ;

[0075] The segmentation index list generation process is: according to the difference of atomic types, the atoms are divided into segments, each segment corresponds to an atomic number, and the segmentation index list is obtained, , represents the largest atomic number in the data set;

[0076] For example, if the system contains three atoms of hydrogen (H), carbon (C) and oxygen (O), then , the elements in are 0, 1, and 2, respectively, corresponding to H, C, and O.

[0077] The center-neighbor mapping table generation process is: define the center-neighbor mapping table , the first list represents the center atom index list , and the second list represents the neighbor atom index list .

[0078] For example, if there is an edge connecting atom 1 and atom 2 in the system, then one record in Neighbor_List is [1, 2].

[0079] Further, for a system with a total of 4 atoms (numbered 0, 1, 2, and 3), under the condition that the cutoff radius is 6, the center-neighbor correspondence table

[0080] is generated. , .

[0081] Taking the calculation of the central atom feature ( ) as an example, at this time After linear layer transformation, the result is obtained For the first in X List The features corresponding to rows 1 and 2 are both atoms 0. Therefore, softmax calculation is performed on rows 1 and 2 of each column to obtain the softmax score for atom 0. Rows 3 and 4 correspond to atoms 1, rows 5, 6, and 7 to atoms 2, and row 8 to atoms 3. Similarly, softmax calculation is performed on each segment to obtain the softmax score. Finally, the Seg-softmax output is returned. (Similarly, if the neighboring atomic features are the input, then return...) ).

[0082] The inductive bias of this mechanism is manifested in forcing the model to focus only on the local feature differences of elements of the same class: by strictly limiting the Softmax calculation to the same group of atoms, it implicitly assumes that there are comparable interactions between elements of the same class (such as similarity of chemical properties), while the feature differences of dissimilar elements should be decoupled, thereby guiding the model to ignore global irrelevant noise and focus on the relative importance analysis of features within the group, while naturally satisfying the substitution invariance of atoms of the same class.

[0083] S300: Based on the segmented index list and the center-neighbor mapping table, a two-stage aggregation mechanism is executed to generate initial feature information;

[0084] In some embodiments, the two-stage aggregation mechanism based on the segmented index list and the center-neighbor mapping table is used to generate initial feature information. The first stage includes:

[0085] edge features The linear layer is used to transform and generate edge intermediate features. ;

[0086] The middle feature of the edge With central attention weight Neighbor attention weight Perform positional multiplications to generate weighted central atom features. and neighboring atomic features ;

[0087] The weighted central atom features and neighboring atom features are aggregated separately to generate compressed central atom features. and neighboring atomic features ;

[0088] In some embodiments, the two-stage aggregation mechanism based on the segmented index list and the center-neighbor mapping table is used to generate initial feature information. The second stage includes:

[0089] compressed center atom features and neighbor atom features are spliced to generate spliced features The splicing formula is:

[0090]

[0091] The spliced features are subjected to nonlinear transformation by a multilayer perceptron to generate fused atom features The activation function of the MLP is SiLU, and the formula is:

[0092]

[0093] wherein, , and are trainable weights of the first and second layers of the MLP,

[0094] , is a Sigmoid function;

[0095] The fused atom features are expanded according to the center atom index list and the neighbor atom index list to and , and the expansion formula is:

[0096]

[0097] The expanded center atom features , the neighbor atom features and the edge features are spliced to generate initial feature information , and the splicing formula is:

[0098] .

[0099] wherein, in the two stages, first, the and the are subjected to aggregation operation, i.e., the information of the two is compressed to the atom level , and then splicing, MLP nonlinear transformation and other operations are performed. The advantage of this is to avoid using dimensional tensor for transformation, resulting in additional memory overhead . At the same time, the actual measurement data show that this way has little effect on the accuracy of the model (the accuracy is reduced by 1-3%, and the memory requirement is reduced by about 15%);

[0100] To avoid too much loss of accuracy, after that, the atomic features According to With Extended to With , and spliced with edge features, generate the final message .

[0101]

[0102] S400: The Message-Agg module processes the atomic distance to obtain a feature update weight, and the initial feature information is multiplied by the feature update weight to generate an updated final atomic feature.

[0103] In some embodiments, the Message-Agg module processes the atomic distance to obtain a feature update weight, and the initial feature information is multiplied by the feature update weight to generate an updated final atomic feature, and the generation of the feature update weight comprises:

[0104] Given the atomic distance , first apply the Bessel function to process it to generate a smooth and decaying weight , the formula is:

[0105]

[0106] Where is the cutoff radius, set to 6, represents the first dimension of the base, represents the highest dimension extended using the Bessel function, represents the envelope function, which realizes smooth decay on the distance, and the calculation formula is as follows:

[0107]

[0108] Where, ;

[0109] The processed distance feature is transformed through a trainable linear layer to generate a feature update weight .

[0110] In some embodiments, the Message-Agg module processes the atomic distance to obtain a feature update weight, and the initial feature information is multiplied by the feature update weight to generate an updated final atomic feature, including edge feature update and atomic feature update, and the edge feature update is:

[0111] Given the initial feature information , the initial feature information is processed by the multi-layer perception to generate updated edge features The formula of the nonlinear transformation is:

[0112] .

[0113] In some embodiments, the atomic feature update is: before the aggregation operation, the initial feature information is multiplied by the feature update weight to obtain weighted initial feature information The weighted initial feature information is aggregated to generate updated final atomic features The aggregation operation represents the summation of all edges related to the atom , thus the edge features and atomic features are updated by the Message Agg module, and the convergence module implemented in this way can guide the model to learn the physical law (neighbor information decays with distance).

[0114] The second aspect of the application proposes an atomic feature update system based on a general interatomic potential model of a graph convolution layer, as shown in Figure 2 , including a feature input module, an index analysis module, a two-stage aggregation module, and an atomic feature update module:

[0115] The feature input module is used to obtain atomic type features and edge features and input them into the Seg-Attention graph convolution layer, and the Seg-Attention graph convolution layer includes a Message-Embed module and a Message-Agg module.

[0116] The index analysis module is used to calculate the attention weight of the center atom and the neighbor atom in the atomic type feature through the Message-Embed module, and generate a segmented index list and a center-neighbor mapping table.

[0117] The two-stage aggregation module is used to perform a two-stage aggregation mechanism based on the segmented index list and the center-neighbor mapping table to generate initial feature information.

[0118] The atomic feature update module is used to process the interatomic distance by the Message-Agg module to obtain a feature update weight, multiply the initial feature information by the feature update weight, and generate updated final atomic features.

[0119] In a third aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor implements the steps of the above method when running the computer program.

[0120] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, wherein the computer program, when executed by a processor, implements the steps of the above method.

[0121] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is exemplified, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the above described functions. Each functional unit and module in the embodiment can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of software functional unit. In addition, the specific names of each functional unit and module are only for easy distinction, and do not limit the protection scope of the present application. The specific working process of the units and modules in the above system can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0122] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0123] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present disclosure.

[0124] In the embodiments of the present disclosure, it should be understood that the disclosed apparatus / computer device and method can be implemented in other manners. For example, the described apparatus / computer device embodiments are merely schematic. For example, the division of the modules or units is merely logical function division. There can be another division manner for the actual implementation, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different units, can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0125] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments.

[0126] In addition, each functional unit in the various embodiments of the present disclosure can be integrated in one processing unit, or each unit can exist physically as separate units, or two or more units can be integrated in one unit. The integrated unit can be implemented in the form of hardware, or in the form of software function units.

[0127] If the integrated module / unit is implemented in the form of software function units and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on such understanding, the present disclosure implements all or part of the processes in the above-described embodiment methods, which can also be completed by computer programs instructing related hardware. The computer program can be stored in a computer readable storage medium, and when the processor executes the computer program, the steps of the above-described various method embodiments can be implemented. The computer program can include computer program code, which can be in the form of source code, object code, executable file or some intermediate form. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), electric carrier wave signal, telecommunication signal and software distribution medium, etc. It should be noted that the computer readable medium can include appropriate contents according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to the legislation and patent practice, the computer readable medium does not include electric carrier wave signal and telecommunication signal.

[0128] The above merely is the preferred embodiment of the present application, it is pointed out that, for the person skilled in the art without departing from the prior art, several deformation and improvement technical solutions should also be considered to fall into the scope of the present application.

Claims

1. An atomic feature update method based on a general inter-atomic potential energy model of graph convolutional layers, characterized in that: Includes the following steps: Atom type features and edge features are obtained and input into the Seg-Attention graph convolutional layer, which includes a Message-Embed module and a Message-Agg module; The Message-Embed module calculates the attention weights of the center atom and its neighboring atoms in the atom type features, and generates a segmented index list and a center-neighbor mapping table, including: The input number Layer Atom Type Characteristics The atomic type features Divided into central atom characteristics and neighboring atomic features ; The central atom feature and neighboring atomic features Transformation is performed through a linear layer to generate intermediate features of the central atom. Intermediate features of neighboring atoms The transformation formula is: The weight matrix of the linear layer is: The bias term is ; The intermediate features of the central atom Intermediate features of neighboring atoms Execute the Seg-Softmax mechanism to generate attention weights The Seg-Softmax mechanism is used to dynamically adjust the weight of information based on different atom types. The calculation formula for the Seg-Softmax mechanism is as follows: in, The features are those of the central atom / neighboring atoms after being transformed by a linear layer. For row index, For column indexes, Indicates that the index is The set of all rows, The attention weight represents the i-th atomic number of the segmented index list. Including central attention weights Neighbor attention weight ; The process of generating the segmented index list is as follows: Based on the different atom types, the atoms are divided into... Each segment corresponds to an atomic number, resulting in a segmented index list. , , This represents the largest atomic number in the dataset; The process of generating the center-neighbor mapping table is as follows: Define the center-neighbor mapping table. The first column represents the list of central atom indices. The second column represents the list of neighbor atomic indices. ; Based on the segmented index list and the center-neighbor mapping table, a two-stage aggregation mechanism is executed to generate initial feature information. The first stage includes: edge features The linear layer is used to transform and generate edge intermediate features. ; The middle feature of the edge With central attention weight Neighbor attention weight Perform positional multiplications to generate weighted central atom features. and neighboring atomic features ; The weighted central atom features and neighboring atom features are aggregated separately to generate compressed central atom features. and neighboring atomic features ; The second phase includes: The compressed central atom features and neighboring atomic features Perform splicing to generate spliced ​​features. The splicing formula is: The concatenated features are then subjected to a nonlinear transformation using a multilayer perceptron to generate fused atomic features. The activation function of the MLP is SiLU, and the formula is: in, , and These are the trainable weights of the first and second layers in an MLP, respectively. , For the Sigmoid function; Fusion atomic features Based on the central atom index list and neighbor atomic index list Expand into and The extended formula is: Expanded central atom features Neighboring atomic characteristics Sum of edge features The data is then concatenated to generate initial feature information. The splicing formula is: ； The feature update weights are obtained by processing the inter-atomic distances through the Message-Agg module. The initial feature information is then multiplied bitwise with the feature update weights to generate the updated final atomic features. The steps for generating the feature update weights include: Given interatomic distance First, the Bessel function is applied to process it, generating smoothly decaying weights. The formula is: in Set the cutoff radius to 6. The first base One dimension, This represents the highest dimension expanded using the Bessel function. The envelope function, which achieves smooth decay over distance, is calculated using the following formula: in, ; Processed distance features The transformation is performed through a trainable linear layer to generate feature update weights. .

2. The method according to claim 1, characterized in that: The feature update weights are obtained by processing the inter-atomic distances through the Message-Agg module. The initial feature information is then multiplied bitwise with the feature update weights to generate the updated final atomic features, including edge feature updates and atomic feature updates. The edge feature updates are as follows: Given the initial feature information The updated edge features are generated by processing the data through the multilayer perceptron. The formula for nonlinear transformation is: 。 3. The method according to claim 2, characterized in that: The atomic features are updated as follows: before the aggregation operation, the initial feature information is updated. Update weights with the features Perform a bitwise multiplication to obtain the weighted initial feature information. The weighted initial feature information Perform aggregation operations to generate updated final atomic features. The aggregation operation represents summing all edges associated with atom i.

4. An atomic feature update system based on a general inter-atomic potential energy model of graph convolutional layers, characterized in that: It includes a feature input module, an index analysis module, a two-stage aggregation module, and an atomic feature update module: The feature input module is used to obtain atomic type features and edge features and input them into the Seg-Attention graph convolutional layer. The Seg-Attention graph convolutional layer includes a Message-Embed module and a Message-Agg module. The index analysis module is used to calculate the attention weights of the center atom and its neighboring atoms in the atom type features through the Message-Embed module, and to generate a segmented index list and a center-neighbor mapping table, including: The input number Layer Atom Type Characteristics The atomic type features Divided into central atom characteristics and neighboring atomic features ; The central atom feature and neighboring atomic features Transformation is performed through a linear layer to generate intermediate features of the central atom. Intermediate features of neighboring atoms The transformation formula is: The weight matrix of the linear layer is: The bias term is ; The intermediate features of the central atom Intermediate features of neighboring atoms Execute the Seg-Softmax mechanism to generate attention weights The Seg-Softmax mechanism is used to dynamically adjust the weight of information based on different atom types. The calculation formula for the Seg-Softmax mechanism is as follows: in, The features are those of the central atom / neighboring atoms after being transformed by a linear layer. For row index, For column indexes, Indicates that the index is The set of all rows, The attention weight represents the i-th atomic number of the segmented index list. Including central attention weights Neighbor attention weight ; The process of generating the segmented index list is as follows: Based on the different atom types, the atoms are divided into... Each segment corresponds to an atomic number, resulting in a segmented index list. , , This represents the largest atomic number in the dataset; The process of generating the center-neighbor mapping table is as follows: Define the center-neighbor mapping table. The first column represents the list of central atom indices. The second column represents the list of neighbor atomic indices. ; The two-stage aggregation module is used to execute a two-stage aggregation mechanism based on the segmented index list and the center-neighbor mapping table to generate initial feature information. The first stage includes: edge features The linear layer is used to transform and generate edge intermediate features. ; The middle feature of the edge With central attention weight Neighbor attention weight Perform positional multiplications to generate weighted central atom features. and neighboring atomic features ; The weighted central atom features and neighboring atom features are aggregated separately to generate compressed central atom features. and neighboring atomic features ; The second phase includes: The compressed central atom features and neighboring atomic features Perform splicing to generate spliced ​​features. The splicing formula is: The concatenated features are then subjected to a nonlinear transformation using a multilayer perceptron to generate fused atomic features. The activation function of the MLP is SiLU, and the formula is: in, , and These are the trainable weights of the first and second layers in an MLP, respectively. , For the Sigmoid function; Fusion atomic features Based on the central atom index list and neighbor atomic index list Expand into and The extended formula is: Expanded central atom features Neighboring atomic characteristics Sum of edge features The data is then concatenated to generate initial feature information. The splicing formula is: ； The atomic feature update module is used to process the inter-atomic distance through the Message-Agg module to obtain feature update weights, and then multiply the initial feature information by the feature update weights bitwise to generate the updated final atomic features. The steps for generating the feature update weights include: Given interatomic distance First, the Bessel function is applied to process it, generating smoothly decaying weights. The formula is: in Set the cutoff radius to 6. The first base One dimension, This represents the highest dimension expanded using the Bessel function. The envelope function, which achieves smooth decay over distance, is calculated using the following formula: in, ; Processed distance features The transformation is performed through a trainable linear layer to generate feature update weights. .

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-3.

6. 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 steps of the method as described in any one of claims 1-3.

Images