Atomic feature updating method based on graph convolutional layer general inter-atomic potential energy model

Abstract

The invention discloses an atomic feature updating method based on a graph convolutional layer general inter-atomic potential energy model, and belongs to the technical field of chemical and material science combined with a neural network. According to the method, the model can automatically distinguish contribution weights of different atom types to interaction information, the problem of feature confusion in a traditional splicing method is avoided, and therefore the multi-body action law in a complex material system and a distance attenuation weight mechanism designed based on the physical law are more accurately captured, and the attenuation characteristics of inter-atom influences are considered in the Message convergence stage. The injection of the priori knowledge guides the model to learn the physical law, the dependence of the model on the training data volume is reduced, and the generalization ability and stability of the model under the limited data condition are remarkably improved especially in the modeling of a rare material system.

Description

Technical Field

[0001] The present invention belongs to the technical field of chemistry and materials science combined with neural networks, and specifically relates to an atomic feature updating method based on a universal interatomic potential energy model of a graph convolutional layer. Background Art

[0002] Universal interatomic potentials (UIPs) are a class of potential energy models applicable to a wide range of material systems, used to describe the interactions between atoms. Compared to potential functions dedicated to specific materials or elements, UIPs have broader applicability, enabling accurate energy and force calculations across diverse elemental combinations, structural types, and environmental conditions. These potential energy models are typically parameterized by combining classical potential functions, first-principles calculation data, and machine learning methods to improve accuracy and generalizability. In recent years, UIP models based on graph neural networks (such as MACE, CHGNet, and Eqv2) have also been shown to generalize well across a variety of downstream tasks.

[0003] The existing general interatomic potential energy models are mainly divided into four categories: conservative-equivariant model (which satisfies both energy conservation and equivariance), non-conservative-equivariant model (which does not satisfy energy conservation but satisfies equivariance), conservative-invariant model (which satisfies both energy conservation and rotational invariance), and non-conservative-invariant model (which does not satisfy energy conservation but satisfies rotational invariance).

[0004] In conservative potential energy models, forces are calculated as negative derivatives 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 potential energy models. On the other hand, some networks improve efficiency by directly predicting forces using separate force heads. Although models that directly predict forces can achieve high accuracy, their non-conservative nature can lead to significant errors in certain property prediction tasks, such as molecular dynamics simulations in NVT ensembles.

[0005] Equivariant series models incorporate sufficient prior knowledge (rotational equivariance) to achieve superior performance even on small datasets. However, due to high training costs (GPU memory and training time), conservative equivariant models are difficult to scale up. Invariant models, on the other hand, have relatively low training costs, but their architecture lacks sufficient inductive bias, making them more prone to overfitting when training on smaller datasets.

[0006] For example, during the message construction phase, the existing invariance model concatenates neighbor atoms, central atoms, and side information, then transforms the dimensions through a nonlinear layer. During this process, the neighbor atom information, central atom information, and side information are each assigned different weights to ultimately form the message. This construction approach only considers assigning different weights to different features (neighbor atoms, central atoms, and edges). In reality, different atom types have different effects on the message. This inductive bias should be considered during the message construction phase.

[0007] In the message aggregation stage, the Message Passing mechanism of CHGNet adopts an isotropic information aggregation method, that is, when constructing the interaction information between atoms, equal weight is given to all neighboring atoms. Although this design ensures translational symmetry, it violates the basic physical law of the material system - the strength of the interaction between atoms decays exponentially with distance. Within a range of 5Å, about 78% of the interatomic distances are concentrated in the range of 2-3Å, and the contribution of atoms exceeding 4Å 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 having to passively learn the distance decay law through a large amount of data, which significantly increases the difficulty of training, especially in small sample scenarios (such as rare material systems), where the generalization performance drops by 32.7%. Summary of the Invention

[0008] In response to the above-mentioned deficiencies in the prior art, the present application provides an atomic feature updating method and system based on a universal interatomic potential energy model of a graph convolutional layer.

[0009] In a first aspect, the present application proposes an atomic feature updating method based on a universal interatomic potential energy model of a graph convolutional layer, comprising the following steps:

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

[0011] Calculating the attention weights of the center atom and the neighbor atoms in the atom type feature through the Message-Embed module, and generating a segment index list and a center-neighbor mapping table;

[0012] Performing a two-stage aggregation mechanism based on the segment index list and the center-neighbor mapping table to generate initial feature information;

[0013] The interatomic distances are processed by the Message-Agg module to obtain feature update weights, and the initial feature information is bitwise multiplied by the feature update weights to generate updated final atomic features.

[0014] In some embodiments, the calculating the attention weights of the center atom and the neighbor atoms in the atom type feature by the Message-Embed module and generating a segment index list and a center-neighbor mapping table includes:

[0015] Input Layer atom type characteristics , the atomic type characteristics Central atom features and neighbor atom features ;

[0016] The central atom features and neighbor atom features Transform through linear layers to generate central atomic intermediate features and intermediate features of neighboring atoms , the transformation formula is:

[0017]

[0018] Among them, the weight matrix of the linear layer is , the bias term is ;

[0019] The central atom intermediate feature and intermediate features of neighboring atoms Execute Seg-Softmax mechanism to generate attention weights The Seg-Softmax mechanism is used to dynamically adjust the weight of information according to the different types of atoms. The calculation formula of the Seg-Softmax mechanism is:

[0020]

[0021] in, is the feature of the central atom / neighboring atom after the linear layer change, is the row index, is the column index, Indicates that the index is The set of all rows of Represents the i-th atomic number of the segment index list, the attention weight Including center attention weight and neighbor attention weights ;

[0022] The process of generating the segment index list is as follows: 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. , , Indicates the largest atomic number in the data set;

[0023] The center-neighbor mapping table generation process is as follows: define the center-neighbor mapping table , the first column represents the central atom index list , the second column represents the neighbor atom index list .

[0024] In some embodiments, a 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] Edge Features Transformed by the linear layer to generate edge intermediate features ;

[0026] The middle feature of the edge With the center attention weight and neighbor attention weights Perform bitwise multiplication respectively to generate weighted central atomic features and neighbor atom features ;

[0027] Aggregate the weighted central atom features and neighboring atom features to generate compressed central atom features and neighbor atom features .

[0028] In some embodiments, a 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 central atomic features and neighbor atom features Perform splicing to generate spliced features , the splicing formula is:

[0030]

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

[0032]

[0033] in, , and are the trainable weights of the first and second layers in the MLP,

[0034] , is the Sigmoid function;

[0035] Atomic features will be fused Index list by central atom and a list of neighbor atom indices Expand to and , the expanded formula is:

[0036]

[0037] The expanded central atom features , neighbor atom features and edge features Splicing to generate initial feature information , the splicing formula is:

[0038] .

[0039] In some embodiments, the interatomic distances are processed by the Message-Agg module to obtain feature update weights, and the initial feature information is bitwise multiplied by the feature update weights to generate updated final atomic features. The step of generating the feature update weights includes:

[0040] Given the interatomic distance , first apply the Bessel function to process it to generate a smoothly decaying weight , the formula is:

[0041]

[0042] in is the cutoff radius, set to 6, The base dimensions, represents the highest dimension expanded using Bessel functions, Represents the envelope function, which implements smooth attenuation with respect to distance. The calculation formula is as follows:

[0043]

[0044] in, ;

[0045] The processed distance features Transform through a trainable linear layer to generate feature update weights .

[0046] In some embodiments, the inter-atomic distance is processed by the Message-Agg module to obtain a feature update weight, and the initial feature information is bitwise multiplied by the feature update weight to generate an updated final atomic feature, including an edge feature update and an atomic feature update, wherein the edge feature update is:

[0047] Given the initial feature information , processed by the multi-layer perceptron to generate updated edge features , the formula for nonlinear transformation is:

[0048] .

[0049] In some embodiments, the atomic feature update is: before the aggregation operation, the initial feature information Update weights with the features Perform bitwise multiplication to obtain weighted initial feature information , the weighted initial feature information Perform aggregation operations to generate updated final atomic features , the aggregation operation represents the summation of all edges related to atom i.

[0050] In the second aspect, this application proposes an atomic feature update system based on a universal interatomic potential energy model of a graph convolutional 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 atom 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;

[0052] The index analysis module is used to calculate the attention weights of the central atom and the neighbor atoms in the atom type feature through the Message-Embed module, and generate a segment index list and a center-neighbor mapping table;

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

[0054] The atomic feature updating module is used to process the inter-atomic distances through the Message-Agg module to obtain feature update weights, and to multiply the initial feature information by the feature update weights to generate updated final atomic features.

[0055] In a third aspect, the present application proposes an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable 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 proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0057] Beneficial effects of the present invention:

[0058] The proposed improved invariant universal interatomic potential model architecture improves the accuracy of atomic interaction modeling by introducing a physics-guided inductive bias mechanism: in the message construction stage, the atom-type-aware attention mechanism is used to achieve differentiated characterization of the chemical properties between different elements. This design enables the model to automatically distinguish the contribution weights of different atom types to the interaction information, avoiding the problem of feature confusion in traditional splicing methods, thereby more accurately capturing the laws of multi-body interactions in complex material systems and breaking through the small sample learning ability: the distance decay weight mechanism designed based on physical laws takes into account the decay characteristics of interatomic influences in the message aggregation stage. This injection of prior knowledge guides the model to learn physical laws, reduces the model's dependence on the amount of training data, and significantly improves the model's generalization ability and stability under limited data conditions, especially in the modeling of rare material systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is the overall flow chart of the present invention.

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

[0061] Figure 3 This is a system principle block diagram of the present invention. DETAILED DESCRIPTION

[0062] The following will describe exemplary embodiments of the present invention in more detail with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein; rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0063] In the first aspect, this application proposes an atomic feature updating method based on a universal interatomic potential energy model of a graph convolutional layer, such as Figure 1 As shown, the following steps are included:

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

[0065] like Figure 2 As shown in Figure 1, the Seg-Attention graph convolutional layer consists of two core modules: the Message-Embed module and the Message-Agg module. These two modules are responsible for processing atomic type features and edge features, respectively, and ultimately generate updated atomic features and edge features.

[0066] S200: Calculating the attention weights of the center atom and the neighbor atoms in the atom type feature through the Message-Embed module, and generating a segment index list and a center-neighbor mapping table;

[0067] In some embodiments, the calculating the attention weights of the center atom and the neighbor atoms in the atom type feature by the Message-Embed module and generating a segment index list and a center-neighbor mapping table includes:

[0068] Input Layer atom type characteristics , the atomic type characteristics Central atom features and neighbor atom features ;

[0069] The central atom features and neighbor atom features Transform through linear layers to generate central atomic intermediate features and intermediate features of neighboring atoms , the transformation formula is:

[0070]

[0071] Among them, the weight matrix of the linear layer is , the bias term is ;

[0072] The central atom intermediate feature and intermediate features of neighboring atoms Execute Seg-Softmax mechanism to generate attention weights The Seg-Softmax mechanism is used to dynamically adjust the weight of information according to the different types of atoms. The calculation formula of the Seg-Softmax mechanism is:

[0073]

[0074] in, is the feature of the central atom / neighboring atom after the linear layer change, is the row index, is the column index, Indicates that the index is The set of all rows of Represents the i-th atomic number of the segment index list, the attention weight Including center attention weight and neighbor attention weights ;

[0075] The process of generating the segment index list is as follows: 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. , , Indicates the largest atomic number in the data set;

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

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

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

[0079] Furthermore, for a certain system with 4 atoms (numbered 0, 1, 2, and 3), and a cutoff radius of 6, a center-neighbor correspondence table is generated:

[0080] . Then at this time , .

[0081] To calculate the central atom characteristics ( ) as an example, at this time , obtained through linear layer transformation For the first List , the features corresponding to rows 1 and 2 are both atom 0, so the softmax calculation is performed on rows 1 and 2 of each column to obtain the softmax score of atom 0, while rows 3 and 4 correspond to atom 1, rows 5, 6, and 7 correspond to atom 2, and row 8 corresponds to atom 3. Similarly, the softmax calculation is performed on each segment to obtain the softmax score. Finally, the Seg-softmax output is returned (Similarly, if the neighbor atomic feature is input, then the return ).

[0082] The inductive bias of this mechanism is reflected in forcing the model to focus only on the local feature differences of elements of the same type: by strictly limiting the Softmax calculation to groups of atoms of the same type, it implicitly assumes that there are comparable interactions between elements of the same type (such as the similarity of chemical properties), and the feature differences of heterogeneous 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 type.

[0083] S300: Executing a two-stage aggregation mechanism based on the segment index list and the center-neighbor mapping table to generate initial feature information;

[0084] In some embodiments, a 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:

[0085] Edge Features Transformed by the linear layer to generate edge intermediate features ;

[0086] The middle feature of the edge With the center attention weight and neighbor attention weights Perform bitwise multiplication respectively to generate weighted central atomic features and neighbor atom features ;

[0087] Aggregate the weighted central atom features and neighboring atom features to generate compressed central atom features and neighbor atom features ;

[0088] In some embodiments, a 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:

[0089] The compressed central atomic features and neighbor atom features Perform splicing to generate spliced features , the splicing formula is:

[0090]

[0091] The spliced features are transformed nonlinearly through a multi-layer perceptron to generate fused atomic features. , the activation function of MLP is SiLU, and the formula is:

[0092]

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

[0094] , is the Sigmoid function;

[0095] Atomic features will be fused Index list by central atom and a list of neighbor atom indices Expand to and , the expanded formula is:

[0096]

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

[0098] .

[0099] Among them, in the two stages, first and Perform aggregation operation, that is, compress the information of both to the atomic level ( ), and then splicing, MLP nonlinear changes and other operations. The advantage of this is to avoid using The dimension of the tensor is changed, resulting in additional memory overhead ( ). Actual measured data also shows that this method has little impact on model accuracy (accuracy is reduced by 1-3%, and video memory requirements are reduced by about 15%).

[0100] To avoid excessive loss of precision, the atomic features are then according to and Expand to and , and spliced with the edge features to generate the final message .

[0101]

[0102] S400: Processing the inter-atomic distances through the Message-Agg module to obtain feature update weights, bitwise multiplying the initial feature information by the feature update weights to generate updated final atomic features.

[0103] In some embodiments, the interatomic distances are processed by the Message-Agg module to obtain feature update weights, and the initial feature information is bitwise multiplied by the feature update weights to generate updated final atomic features. The step of generating the feature update weights includes:

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

[0105]

[0106] in is the cutoff radius, set to 6, The base dimensions, represents the highest dimension expanded using Bessel functions, Represents the envelope function, which implements smooth attenuation with respect to distance. The calculation formula is as follows:

[0107]

[0108] in, ;

[0109] The processed distance features Transform through a trainable linear layer to generate feature update weights .

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

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

[0112] .

[0113] In some embodiments, the atomic feature update is: before the aggregation operation, the initial feature information Update weights with the features Perform bitwise multiplication to obtain weighted initial feature information , the weighted initial feature information Perform aggregation operations to generate updated final atomic features , the aggregation operation represents the aggregation of all The relevant edges are summed, and the edge features and atomic features are updated through the Message Agg module. The aggregation module implemented in this way can guide the model to learn physical laws (neighborhood information decays with distance).

[0114] In the second aspect, this application proposes an atomic feature update system based on a universal interatomic potential energy model of a graph convolutional layer, such as Figure 2 As shown, it includes feature input module, index analysis module, two-stage aggregation module and atomic feature update module:

[0115] The feature input module is used to obtain atom 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 weights of the central atom and the neighbor atoms in the atom type feature through the Message-Embed module, and generate a segment index list and a center-neighbor mapping table;

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

[0118] The atomic feature updating module is used to process the inter-atomic distances through the Message-Agg module to obtain feature update weights, and to multiply the initial feature information by the feature update weights to generate updated final atomic features.

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

[0120] In a fourth aspect, the present application proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0121] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0122] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

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

[0124] In the embodiments provided in the present disclosure, it should be understood that the disclosed apparatus / computer equipment and methods can be implemented in other ways. For example, the apparatus / computer equipment embodiments described above are merely schematic. For example, the division of modules or units is merely a logical function division. In actual implementation, there may be other division methods. Multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection of the apparatus or unit, which may be electrical, mechanical or other forms.

[0125] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0126] In addition, the functional units in the various embodiments of the present disclosure may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0127] If the integrated module / unit is implemented as a software functional unit and sold or used as a standalone product, it can be stored in a computer-readable storage medium. Based on this understanding, the present disclosure can implement all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program can include computer program code, which can be in source code form, object code form, executable file, or some intermediate form. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, removable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunications signals, and software distribution media. It should be noted that the content included in computer-readable media can be appropriately increased or decreased based on the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, based on legislation and patent practice, computer-readable media does not include electrical carrier signals and telecommunications signals.

[0128] The above are only preferred embodiments of the present invention. It should be pointed out that various modifications and improvements made by those skilled in the art without departing from the present technical solution should also be deemed to fall within the scope of protection required by this solution.

Claims

1. An atomic feature updating method based on a universal interatomic potential energy model in a graph convolutional layer, characterized by: The following steps are involved: Obtain atom type features and edge features and input them into the Seg-Attention graph convolution layer, which includes a Message-Embed module and a Message-Agg module; Calculating the attention weights of the center atom and the neighbor atoms in the atom type feature through the Message-Embed module, and generating a segment index list and a center-neighbor mapping table; Performing a two-stage aggregation mechanism based on the segment index list and the center-neighbor mapping table to generate initial feature information; The interatomic distances are processed by the Message-Agg module to obtain feature update weights, and the initial feature information is bitwise multiplied by the feature update weights to generate updated final atomic features.

2. The method according to claim 1, wherein: The Message-Embed module is used to calculate the attention weights of the center atom and the neighbor atoms in the atom type feature, and to generate a segment index list and a center-neighbor mapping table, including: Input Layer atom type characteristics , the atomic type characteristics Central atom features and neighbor atom features ; The central atom features and neighbor atom features Transform through linear layers to generate central atomic intermediate features and intermediate features of neighboring atoms , the transformation formula is: Among them, the weight matrix of the linear layer is , the bias term is ; The central atom intermediate feature and intermediate features of neighboring atoms Execute Seg-Softmax mechanism to generate attention weights The Seg-Softmax mechanism is used to dynamically adjust the weight of information according to the different types of atoms. The calculation formula of the Seg-Softmax mechanism is: in, is the feature of the central atom / neighboring atom after the linear layer change, is the row index, is the column index, Indicates that the index is The set of all rows of Represents the i-th atomic number of the segment index list, the attention weight Including center attention weight and neighbor attention weights ; The process of generating the segment index list is as follows: 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. , , Indicates the largest atomic number in the data set; The center-neighbor mapping table generation process is as follows: define the center-neighbor mapping table , the first column represents the central atom index list , the second column represents the neighbor atom index list .

3. The method according to claim 2, wherein: 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: Edge Features Transformed by the linear layer to generate edge intermediate features ; The middle feature of the edge With the center attention weight and neighbor attention weights Perform bitwise multiplication respectively to generate weighted central atomic features and neighbor atom features ; Aggregate the weighted central atom features and neighboring atom features to generate compressed central atom features and neighbor atom features .

4. The method according to claim 3, wherein: 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: The compressed central atomic features and neighbor atom features Perform splicing to generate spliced features , the splicing formula is: The spliced features are transformed nonlinearly through a multi-layer perceptron to generate fused atomic features. , the activation function of MLP is SiLU, and the formula is: in, , and are the trainable weights of the first and second layers in the MLP, , is the Sigmoid function; Atomic features will be fused Index list by central atom and a list of neighbor atom indices Expand to and , the expanded formula is: The expanded central atom features , neighbor atom features and edge features Splicing to generate initial feature information , the splicing formula is: 。 5. The method according to claim 4, characterized in that: The interatomic distances are processed by the Message-Agg module to obtain feature update weights, and the initial feature information is bitwise multiplied by the feature update weights to generate updated final atomic features. The step of generating the feature update weights includes: Given the interatomic distance , first apply the Bessel function to process it to generate a smoothly decaying weight , the formula is: in is the cutoff radius, set to 6, The base dimensions, represents the highest dimension expanded using Bessel functions, Represents the envelope function, which implements smooth attenuation with respect to distance. The calculation formula is as follows: in, ; The processed distance features Transform through a trainable linear layer to generate feature update weights .

6. The method according to claim 5, characterized in that: The inter-atomic distance is processed by the Message-Agg module to obtain a feature update weight, and the initial feature information is bitwise multiplied by the feature update weight to generate an updated final atomic feature, including an edge feature update and an atomic feature update. The edge feature update is: Given the initial feature information , processed by the multi-layer perceptron to generate updated edge features , the formula for nonlinear transformation is: 。 7. The method according to claim 6, characterized in that: The atomic feature is updated as follows: before the aggregation operation, the initial feature information Update weights with the features Perform bitwise multiplication to obtain weighted initial feature information , the weighted initial feature information Perform aggregation operations to generate updated final atomic features , the aggregation operation represents the summation of all edges related to atom i.

8. An atomic feature update system based on a universal interatomic potential energy model in graph convolutional layers, characterized by: It includes feature input module, index analysis module, two-stage aggregation module and atomic feature update module: The feature input module is used to obtain atom 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; The index analysis module is used to calculate the attention weights of the central atom and the neighbor atoms in the atom type feature through the Message-Embed module, and generate a segment index list and a center-neighbor mapping table; The two-stage aggregation module is configured to perform a two-stage aggregation mechanism based on the segment index list and the center-neighbor mapping table to generate initial feature information; The atomic feature updating module is used to process the inter-atomic distances through the Message-Agg module to obtain feature update weights, and to multiply the initial feature information by the feature update weights to generate updated final atomic features.

9. 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, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images