A method for analyzing molecular potential energy field based on multi-element feature information transmission
By using multivariate feature information transfer and autoencoder technology, a molecular potential field analysis model was constructed, which solved the problem of the universality of molecular dynamics models across different atomic systems and achieved higher computational accuracy and stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- QIANWAN INST OF CNITECH
- Filing Date
- 2022-11-04
- Publication Date
- 2026-05-12
AI Technical Summary
In existing technologies, molecular dynamics models lack universality across different atomic systems, resulting in complex and difficult-to-integrate methods for describing local environments, which reduces the accuracy of calculation results.
A method based on multivariate feature information transmission is adopted. By obtaining element feature vectors and structural feature vectors, and using an undercomplete autoencoder and a continuous convolutional filter, combined with an artificial neural network, a molecular potential energy field analysis model is constructed to handle the local environmental characteristics of different material systems.
This improves the model's versatility and accuracy across different material systems, enhances its ability to describe local environments, and ensures the stability and accuracy of the calculation results.
Smart Images
Figure CN115662528B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of molecular dynamics technology, and more specifically to a molecular potential field analysis method based on the transmission of multivariate feature information. Background Technology
[0002] In the field of theoretical materials calculations, molecular dynamics, as a microscopic simulation method, can predict the thermodynamic properties of materials and the evolution of individual defects within materials. A potential energy field is a prerequisite for molecular dynamics calculations; an accurate potential energy field is essential to ensure the accuracy of the calculation results. Currently, various algorithms that utilize machine learning methods to construct potential energy fields typically develop separate potential energy fields for different atomic systems, leading to diverse and complex descriptions of various local environments that are difficult to integrate, thus reducing the overall model's universality across different systems. Summary of the Invention
[0003] To improve the model's universality across different systems and enable the integration of various local environment description methods, this invention proposes a molecular potential field analysis method based on multivariate feature information transfer, comprising the following steps:
[0004] S1: Obtain the element feature vector of the target molecular formula after processing the atomic number of the embedded matrix;
[0005] S2: Obtain the structural feature vector of the target molecular formula after processing the atomic coordinates using the atomic centrosymmetry function;
[0006] S3: An encoding network is constructed by an undercomplete autoencoder, and the multivariate feature vector composed of element feature vectors and structural feature vectors is processed to obtain atomic representations;
[0007] S4: Obtain the inter-atomic distance based on the atomic representation and aggregate neighboring atoms. Update the atom representation after neighbor aggregation using a continuous convolution filter.
[0008] S5: Based on the updated atomic representation, predict the atomic potential energy using an artificial neural network algorithm, and obtain the total potential energy of the system based on the atomic potential energy of each atom in the target molecular formula.
[0009] Furthermore, in step S1, the specific method for obtaining the element feature vector is as follows: obtain the corresponding one-hot code based on the atomic number, and use the mapping result obtained by multiplying it with the randomly generated embedding matrix as the element feature vector.
[0010] Furthermore, in step S2, the atomic centrosymmetry function includes a radial function and an angular function. The structural feature vector is obtained by integrating the calculation results of the radial function and the angular function, wherein:
[0011] The radial function is,
[0012]
[0013] In the formula, The radial function of the i-th atom in the target molecular formula is calculated, where e is the natural constant, η is the calculation parameter given based on the material system, and R... ij R is the Euclidean distance between the i-th atom and the j-th atom. s f is the preset cutoff radius. c () is the cosine cutoff function;
[0014] The angle function is,
[0015]
[0016] In the formula, ζ represents the calculated angular function of the i-th atom in the target molecular formula, where ζ is the calculation parameter given based on the material system, and θ is the angular function of the i-th atom. ijk Let λ be the angle formed by the i-th atom, the j-th atom, and the k-th atom, and let λ be a preset angle weighting coefficient. ik R is the Euclidean distance between the i-th atom and the k-th atom. jk Let be the Euclidean distance between the j-th atom and the k-th atom.
[0017] Furthermore, in step S3,
[0018] When the target molecular formula is a simple system, the training function in the training process of the incomplete autoencoder is:
[0019] t′=D(X)=D(E(t))
[0020] In the formula, t′ is the reconstructed signal output by the undercomplete autoencoder, X is the atomic representation generated after the encoding network, D() is the decoder network, and E() is the encoding network;
[0021] When the target molecular formula is a multivariate system, the training function in the training process of the incomplete autoencoder is:
[0022]
[0023] In the formula, This is the atomic representation of the i-th class a element atom in the output of the encoding network. This is the atomic representation of the j-th class a element in the output of the encoding network. ε is the atomic representation of the k-th class b element atom output by the encoding network, and ε is the minimum distance value in the set spatial vector.
[0024] Furthermore, in step S4, when the material corresponding to the target molecular formula is a crystalline material, it is also necessary to calculate the periodic boundary condition distance, which is obtained by summing the squares of the interatomic distance vectors and then taking the square root.
[0025] Furthermore, the interatomic distance vector is obtained using the following formula:
[0026] D = j pos -i pos +S×Cell
[0027] In the formula, D is the interatomic distance vector, j pos Let i be the spatial coordinate of the j-th atom in the target molecular formula. pos Let be the spatial coordinates of the i-th atom in the target molecular formula, be the number of boundaries crossed between the i-th and j-th atoms in the target molecular formula, and Cell be the unit cell size.
[0028] Furthermore, in step S4, the update of the atom representation after neighbor aggregation by the successive convolutional filter is expressed by the following formula:
[0029]
[0030]
[0031] In the formula, C() represents the updated information obtained by processing the i-th atom in the target molecular formula after passing through the l-th interaction layer of a continuous convolution filter, where C() is the continuous convolution filter function, and D() is the updated information. ij X is the interatomic distance between the i-th atom and the j-th neighboring atom in the target molecular formula. j Let j be the atom representation of the j-th neighboring atom. This represents the atom of the i-th atom in the target molecular formula after being updated by the l-th interaction layer in a continuous convolution filter.
[0032] Furthermore, in step S5, the atomic potential energy prediction is expressed as the following formula:
[0033] y i =W out X i +b out
[0034] In the formula, y i The updated atomic potential energy W of the i-th atom in the target molecular formula. out To predict the weights, b out X is the predicted bias. i This represents the atom representation of the i-th atom in the target molecular formula after being completely updated by a continuous convolution filter.
[0035] Furthermore, in step S5, the total potential energy of the system is the sum of the atomic potential energies of each atom in the target molecular formula.
[0036] Compared with the prior art, the present invention has at least the following beneficial effects:
[0037] (1) The molecular potential field analysis method based on multi-feature information transmission described in this invention processes the two common local environmental features of elements and structures through coding technology, thereby ensuring the feature description of different material systems and improving the universality of the model.
[0038] (2) The problem of insufficient element feature discrimination in the single-element system is overcome by training with an undercomplete autoencoder, which makes the information contained in the atomic representation richer.
[0039] (3) By training the atomic representations of different elements in the multivariate system through the ternary loss function, the problem of the atomic representations of different elements being mixed in the vector space is avoided, so that the atoms in the vector space can be clustered according to the element type step by step.
[0040] (4) By calculating the periodic boundary conditions, the model can be flexibly applied in different systems and the neighboring atoms of different systems can be reasonably distinguished. Attached Figure Description
[0041] Figure 1 This is a step diagram of a molecular potential field analysis method based on the transmission of multivariate feature information. Detailed Implementation
[0042] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings to further illustrate the technical solutions of the present invention. However, the present invention is not limited to these embodiments.
[0043] Example 1
[0044] To address the shortcomings of existing technologies in molecular potential energy field research, avoid the difficulties in fusion caused by studying local features separately, and thus improve the model's versatility, for this reason, such as... Figure 1 As shown, this invention proposes a molecular potential field analysis method based on the transmission of multivariate feature information, comprising the following steps:
[0045] S1: Obtain the element feature vector of the target molecular formula after processing the atomic number of the embedded matrix;
[0046] S2: Obtain the structural feature vector of the target molecular formula after processing the atomic coordinates using the atomic centrosymmetry function;
[0047] S3: An encoding network is constructed by an undercomplete autoencoder, and the multivariate feature vector composed of element feature vectors and structural feature vectors is processed to obtain atomic representations;
[0048] S4: Obtain the inter-atomic distance based on the atomic representation and aggregate neighboring atoms. Update the atom representation after neighbor aggregation using a continuous convolution filter.
[0049] S5: Based on the updated atomic representation, predict the atomic potential energy using an artificial neural network algorithm, and obtain the total potential energy of the system based on the atomic potential energy of each atom in the target molecular formula.
[0050] To avoid the difficulties in integrating various methods for describing local environmental features when dealing with different atomic systems, this invention improves upon this by proposing a method for studying molecular potential fields based on universal local environmental features. The most common local environmental features among different molecules are elemental and structural features. To obtain these two types of features while considering the inherent defects of molecules, this invention employs the following approach.
[0051] Regarding elemental characteristics, this invention first sets the atomic number z = (z1, z2, ..., z) of the target molecular formula. n The process involves converting the atoms into a one-hot code (where n is the total number of atoms in the target molecular formula), and then multiplying the one-hot code by the embedding matrix to obtain the mapping result. The corresponding formula is as follows:
[0052] a i =H i M e
[0053] In the formula, a i H is the mapping result of the atomic number of the i-th atom in the target molecular formula. i M is the one-hot code of the i-th atom. e The embedding matrix is randomly generated.
[0054] Here, after the atomic number is processed by the embedding matrix, it is mapped from a single numerical value to a set of vectors. The advantage of this change is that:
[0055] 1. The expansion of vector form can increase the proportion of element features in subsequent processing, and in the subsequent joint processing with the central symmetric function, the two can occupy a similar proportion;
[0056] 2. The mapping process maps data to a high-dimensional space, enabling the algorithm to capture more potential features.
[0057] For example, with three elements, in the state of atomic number, the algorithm can only distinguish their size. However, in the form of vectors, the positions of two elements with similar properties are closer in space. Adding more elements can form a high-dimensional space with rich meaning, which is something that one-dimensional data cannot achieve.
[0058] Regarding structural features, this invention processes the atomic coordinates R = (r1, r2, ..., r) of the target molecular formula using atomic centrosymmetry functions. n This allows us to extract structural features of the atomic system, such as crystal system, symmetry, and dislocations. Specifically, the atomic centrosymmetry function includes radial and angular functions. By integrating the calculation results of the radial and angular functions, we obtain the structural feature vector, where:
[0059] The radial function is,
[0060]
[0061] In the formula, The radial function of the i-th atom in the target molecular formula is calculated, where e is the natural constant, η is the calculation parameter given based on the material system, and R... ij R is the Euclidean distance between the i-th atom and the j-th atom. s f is the preset cutoff radius. c () is the cosine cutoff function;
[0062] The angle function is,
[0063]
[0064] In the formula, ζ represents the calculated angular function of the i-th atom in the target molecular formula, where ζ is the calculation parameter given based on the material system, and θ is the angular function of the i-th atom. ijk Let λ be the angle formed by the i-th atom, the j-th atom, and the k-th atom, and let λ be a preset angle weighting coefficient. ik R is the Euclidean distance between the i-th atom and the k-th atom. jk Let be the Euclidean distance between the j-th atom and the k-th atom.
[0065] By acquiring and combining elemental and structural features—two types of general local environment features—in the two methods described above into corresponding multivariate feature vectors, we can ensure universality while enriching the expression of the atomic local environment in the final vectors. This is more conducive to the algorithm gaining more diverse physicochemical insights during training and helps the algorithm achieve a certain degree of extrapolation ability. In other words, even when some data not learned in training appears, it can still make stable and accurate predictions by relying on the physicochemical information contained in the previous multivariate feature vectors.
[0066] Considering that in a single-element system, since there is only one type of atom, the elemental features obtained after multivariate feature vectors are insufficient in their discriminative power, only expressing information about the element type. This makes the input and output of the encoding network easily converge. Therefore, this invention also proposes to use an undercomplete autoencoder to form a multivariate feature vector for the encoding network to process atoms. Specifically,
[0067] When the target molecular formula is a simple system, the training function in the training process of the incomplete autoencoder is:
[0068] t′=D(X)=D(E(t))
[0069] In the formula, t′ is the reconstructed signal output by the undercomplete autoencoder, X is the atomic representation generated after the encoding network, D() is the decoder network, and E() is the encoding network.
[0070] By using symmetric decoding technology, a decoder with a network structure symmetrical to the encoding network is constructed in a single-element system. This allows the most prominent and essential features to be obtained through continuous reconstruction during training, thereby enhancing the features of the central symmetric function.
[0071] When the target molecular formula is a multivariate system, the training function in the training process of the incomplete autoencoder is:
[0072]
[0073] In the formula, This is the atomic representation of the i-th class a element atom in the output of the encoding network. This is the atomic representation of the j-th class a element in the output of the encoding network. ε is the atomic representation of the k-th class b element atom output by the encoding network, and ε is the minimum distance value in the set spatial vector.
[0074] In a multivariate system, a ternary loss function is used to classify multivariate feature vectors according to element type. During training, a certain atomic input is first used as an anchor point, then another atomic input of the same type is taken as a positive sample, and another atomic input of a different type is taken as a negative sample. Through training, the relative positions of the three samples in the vector space are changed, thereby achieving element-based classification in the space and avoiding the problem of different element atoms being mixed in the vector space in a multivariate system.
[0075] Atomic representations can be obtained by processing multivariate feature vectors using a trained undercomplete autoencoder. Since materials are generally composed of unit cells, the obtained atomic representations also need to be aggregated based on inter-atomic distances. It's important to note that crystalline materials (metals, ceramics, etc.) have large-volume and continuous similar atomic distributions in their microstructure. Due to their large volume, they are often divided into small unit cells for computational simplification. This leads to a crucial point: atoms at the boundaries of these small unit cells lose many neighboring atoms. To compensate for this loss of neighboring atomic information, periodic boundary condition calculations are introduced, utilizing the similarity of atomic distributions in crystalline materials. The operation is similar to a snake game: a judgment distance, such as 0.5 nanometers, is initially set. For example, a sphere with a radius of 0.5 nanometers is drawn around the atom on the left boundary. This sphere encloses all its neighboring atoms, but due to the division, only half a sphere is visible; the other half is found on the right side. This is equivalent to the atom on the right moving out of the boundary and filling the gap on the left, thus obtaining information about the neighboring atoms of all boundary atoms. The mechanism is similar to a snake game. In specific calculations, eight identical unit cells are stacked on a cubic unit cell to form a 3×3 unit cell. The spherical surfaces of each atom are then drawn sequentially to obtain the neighboring atomic information. The periodic boundary condition distance can be obtained by summing the squares of the inter-atomic distance vectors and then taking the square root. The inter-atomic distance vector is obtained using the following formula:
[0076] D = j pos -i pos +S×Cell
[0077] In the formula, D is the interatomic distance vector, j pos Let i be the spatial coordinate of the j-th atom in the target molecular formula. pos Let be the spatial coordinates of the i-th atom in the target molecular formula, be the number of boundaries crossed between the i-th and j-th atoms in the target molecular formula, and Cell be the unit cell size.
[0078] After calculating the distances to all neighboring atoms (obtained using the Pythagorean theorem for inter-atom distance vectors), the neighboring atom representations are aggregated using a continuous convolution filter to generate updated information. The processing steps are as follows:
[0079]
[0080]
[0081] In the formula, C() represents the updated information obtained by processing the i-th atom in the target molecular formula after passing through the l-th interaction layer of a continuous convolution filter, where C() is the continuous convolution filter function, and D() is the updated information.ij X is the interatomic distance between the i-th atom and the j-th neighboring atom in the target molecular formula. j Let j be the atom representation of the j-th neighboring atom. This represents the atom of the i-th atom in the target molecular formula after being updated by the l-th interaction layer in a continuous convolution filter.
[0082] Finally, an artificial neural network algorithm was used to train the mapping between atomic representations, potential energy, and forces acting on each atom. The prediction formulas for the potential energy of the atomic system and the forces acting on each atom are as follows:
[0083] y i =W out X i +b out
[0084] In the formula, y i The updated atomic potential energy W of the i-th atom in the target molecular formula. out To predict the weights, b out X is the predicted bias. i Let be the atomic representation of the i-th atom in the target molecular formula after being completely updated by a continuous convolution filter. The total potential energy of the system can be obtained by summing the energies of each atom, and the interatomic forces can be obtained by taking the partial derivative of the total potential energy of the system with respect to the coordinates of each atom.
[0085] In summary, the molecular potential field analysis method based on multi-feature information transmission described in this invention uses encoding technology to process two common local environmental features: elements and structures. This ensures the characteristic description of different material systems and improves the model's versatility.
[0086] The problem of insufficient feature discrimination in elemental systems is overcome by using an undercomplete autoencoder training method, resulting in richer information in the atomic representations. By training the atomic representations of dissimilar elements in a multivariate system using a ternary loss function, the problem of dissimilar element atomic representations mixing in the vector space is avoided, and the atoms in the vector space exhibit a step-by-step clustering effect based on element type.
[0087] By calculating periodic boundary conditions, the model can be flexibly applied in different systems and can reasonably distinguish neighboring atoms in different systems.
[0088] Example 2
[0089] In this embodiment, potential energy fields are constructed for eight organic molecules: benzene, toluene, malondialdehyde, salicylic acid, aspirin, ethanol, uracil, and naphthalene. In the parameter settings, the vector lengths calculated from the embedding matrix and the atomic centrosymmetric function are both 64, resulting in a combined multivariate feature vector with a length of 128. After compression and integration by the encoding module, the output atomic representation length is 64, and the number of interaction layers for message passing updates is 6. The encoding module iterates 50 times during training with a learning rate of 0.1. The overall learning rate is set to 0.01 during the entire training process.
[0090] During the learning process, 10,000 atomic structures were used for training, and 2,000 atomic structures were used for testing and verification. The final force field prediction errors are shown in Table 1.
[0091] Table 1: Prediction Error Table for Organic Molecules
[0092]
[0093] As shown in the error results above, the overall error is very small, meeting the chemical precision requirement of 1 kcal / mol. The system energy and atomic forces predicted using this embodiment are very close to the results calculated using first-principles calculations, demonstrating the reliability of this method in organic molecular systems.
[0094] Example 3
[0095] In this embodiment, potential energy fields are constructed for ceramics and metals. The dataset includes titanium dioxide, hafnium dioxide, copper, and nickel. In the parameter settings, the encoding module undergoes 50 training rounds, the encoding network is a two-layer design, the input vector length is 128, and the output atom representation length is 64. The interaction layers for copper and nickel are 2, while the interaction layers for titanium dioxide and hafnium dioxide are 4. Furthermore, during the encoding module training, a decoder technique is used for copper and nickel, while a ternary loss function technique is used for hafnium dioxide and titanium dioxide. The prediction errors obtained from the training are shown in Table 2.
[0096] Table 2: Prediction Error Table for Crystal Materials
[0097]
[0098] As can be seen from the error results shown above, the overall error is very small, indicating that this example has good applicability in crystalline materials such as metals and ceramics.
[0099] Example 4
[0100] In this embodiment, a potential energy field is constructed for a complex ternary alloy system, using iron-chromium-aluminum and aluminum-copper-magnesium material systems. In the parameter settings, the encoding module undergoes 50 training rounds, the encoding network is a two-layer design, the input vector length is 128, and the output atom representation length is 64. The number of interaction layers during training is 6, and the encoding module is trained using a ternary loss function technique.
[0101] The prediction results are shown in Table 3:
[0102] Table 3: Prediction Error Table for Ternary Alloys
[0103]
[0104] As shown in the error results above, the overall error not only meets the requirements of chemical accuracy, but also demonstrates better adaptability to the iron-chromium-aluminum dataset with incomplete component coverage and limited data volume. The overall error is significantly lower than that of the aluminum-copper-magnesium dataset, indicating that this implementation example can achieve better performance even with low data quality.
[0105] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.
[0106] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0107] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0108] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.
Claims
1. A molecular potential field analysis method based on multivariate feature information transmission, characterized in that, Including the following steps: S1: Obtain the element feature vector of the target molecular formula after processing the atomic number of the embedded matrix; S2: Obtain the structural feature vector of the target molecular formula after processing the atomic coordinates using the atomic centrosymmetry function; S3: An encoding network is constructed by an undercomplete autoencoder, and the multivariate feature vector composed of element feature vectors and structural feature vectors is processed to obtain atomic representations; S4: Obtain the inter-atomic distance based on the atomic representation and aggregate neighboring atoms. Update the atom representation after neighbor aggregation using a continuous convolution filter. S5: Based on the updated atomic representation, predict the atomic potential energy using an artificial neural network algorithm, and obtain the total potential energy of the system based on the atomic potential energy of each atom in the target molecular formula; In step S3, When the target molecular formula is a simple system, the training function in the training process of the incomplete autoencoder is: In the formula, X is the reconstructed signal output by the undercomplete autoencoder, D() is the decoder network, and E() is the encoding network. When the target molecular formula is a multivariate system, the training function in the training process of the undercomplete autoencoder is: In the formula, This is the atomic representation of the i-th class a element atom in the output of the encoding network. This is the atomic representation of the j-th class a element in the output of the encoding network. This is the atomic representation of the k-th class b element in the output of the encoding network. This is the minimum distance value in the defined spatial vector.
2. The molecular potential field analysis method based on multivariate feature information transmission as described in claim 1, characterized in that, In step S1, the specific method for obtaining the element feature vector is as follows: obtain the corresponding one-hot code based on the atomic number, and use the mapping result obtained by multiplying it with the randomly generated embedding matrix as the element feature vector.
3. The molecular potential field analysis method based on multivariate feature information transmission as described in claim 1, characterized in that, In step S2, the atomic centrosymmetry function includes a radial function and an angular function. The structural feature vector is obtained by integrating the calculation results of the radial function and the angular function, wherein: The radial function is, In the formula, The result of the radial function calculation for the i-th atom in the target molecular formula. It is a natural constant. Based on the calculation parameters given by the material system, Let be the Euclidean distance between the i-th atom and the j-th atom. For the preset cutoff radius, It is a cosine cutoff function; The angle function is, In the formula, The result of the angle function calculation for the i-th atom in the target molecular formula. Based on the calculation parameters given by the material system, Let be the angle formed by the i-th atom, the j-th atom, and the k-th atom. For preset angle weighting coefficients, Let be the Euclidean distance between the i-th atom and the k-th atom. Let be the Euclidean distance between the j-th atom and the k-th atom.
4. The molecular potential field analysis method based on multivariate feature information transmission as described in claim 1, characterized in that, In step S4, when the material corresponding to the target molecular formula is a crystalline material, it is also necessary to calculate the periodic boundary condition distance, which is obtained by summing the squares of the interatomic distance vectors and then taking the square root.
5. The molecular potential field analysis method based on multivariate feature information transmission as described in claim 4, characterized in that, The interatomic distance vector is obtained using the following formula: In the formula, D is the interatomic distance vector. Let J be the spatial coordinates of the j-th atom in the target molecular formula. Let S be the spatial coordinates of the i-th atom in the target molecular formula, S be the number of boundaries crossed between the i-th and j-th atoms in the target molecular formula, and Cell be the unit cell size.
6. The molecular potential field analysis method based on multivariate feature information transmission as described in claim 1, characterized in that, In step S4, the update of the atom representation after neighbor aggregation by the successive convolutional filter is expressed by the following formula: In the formula, This refers to the updated information obtained by processing the i-th atom in the target molecular formula after passing through the l-th interaction layer of a continuous convolution filter. For continuous convolution filter functions, The distance between the i-th atom and the j-th neighboring atom in the target molecular formula is given by the given atom. Let j be the atom representation of the j-th neighboring atom. This represents the atom of the i-th atom in the target molecular formula after being updated by the l-th interaction layer in a continuous convolution filter.
7. The molecular potential field analysis method based on multivariate feature information transmission as described in claim 1, characterized in that, In step S5, the atomic potential energy prediction is expressed by the following formula: In the formula, The updated atomic potential energy of the i-th atom in the target molecular formula. To predict weights, The predicted bias. This represents the atom representation of the i-th atom in the target molecular formula after being completely updated by a continuous convolution filter.
8. The molecular potential field analysis method based on multivariate feature information transmission as described in claim 7, characterized in that, In step S5, the total potential energy of the system is the sum of the atomic potential energies of each atom in the target molecular formula.