Crystal geometry and symmetry prediction method and device based on physical information fusion neural network

Abstract

The application discloses a crystal geometric shape and symmetry prediction method and device based on a physical information fusion neural network, and comprises the following steps: constructing training samples; each training sample comprises a calculation feature of a chemical formula and a real label corresponding to the chemical formula; the real label comprises a space group category real label and a lattice parameter real label; a training model is constructed, comprising a first splicing layer, an encoder, a reparameterization layer, a second splicing layer and a decoder group; the training model is trained by using a training sample set; and the crystal geometric shape and symmetry prediction is predicted by using a prediction model. In the training model, a physical information conditional variational autoencoder is used to encode multiple stable crystal structures into a Gaussian distribution, so that the problem that one chemical formula corresponds to multiple stable crystal structures in crystal structure prediction is solved. The method and the device improve the accuracy of the crystal geometric shape and symmetry prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of crystal materials, and particularly relates to a crystal geometric shape and symmetry prediction method and device based on physical information fusion neural network. BACKGROUND

[0002] The symmetry and geometry of a crystal fundamentally determine its various physical and chemical properties. However, crystal structure prediction, including the determination of symmetry and the optimization of crystal structure, remains a persistent challenge. Due to changes in environmental conditions, such as different pressures and temperatures, there can be multiple crystal structures corresponding to one chemical formula, which has been ignored or poorly handled in previous studies. The geometry and symmetry of a crystal can be described by a space group g and three cell edge lengths {a, b, c} and their angles {α, β, γ}.

[0003] With the development of machine learning, various machine learning methods have been applied to crystal structure prediction, including support vector machines, neural networks, random forests, etc. These methods all focus on a specific material series, which makes it difficult for these models to be generalized to other materials. The CRYSPNet disclosed in the document (Liang, Haotong, et al. "CRYSPNet: Crystal structure predictions via neural networks." Physical Review Materials 4.12 (2020): 123802.) classifies all crystal materials by lattice category and trains a neural network-based model for each category of material. The MLatticeABC disclosed in the document (Li, Yuxin, et al. "MLatticeABC: generic lattice constant prediction of crystal materials using machine learning." ACS omega 6.17 (2021): 11585-11594.) proposes a series of new features, also classifies all crystal materials by lattice category, and establishes a random forest model for each category of material. CRYSPNet and MLatticeABC both use lattice category as prior knowledge for material classification; and they both only predict a set of crystal geometric shapes and symmetries for one chemical formula, and cannot solve the problem of multiple crystal structures corresponding to one chemical formula. SUMMARY

[0004] In view of the above, the present application aims to provide a crystal geometry and symmetry prediction method and device based on physical information fusion neural network, which encodes multiple stable structures corresponding to a chemical formula as a Gaussian distribution to improve the accuracy of crystal geometry and symmetry prediction.

[0005] To achieve the above-mentioned application purposes, the embodiments of the present application provide a crystal geometry and symmetry prediction method based on physical information fusion neural network, comprising the following steps:

[0006] Constructing training samples: each training sample includes the calculated features of a chemical formula and the real labels corresponding to the chemical formula, and the real labels include space group category real labels and lattice parameter real labels;

[0007] Constructing a training model: the training model includes a first splicing layer, an encoder, a reparameterization layer, a second splicing layer, and a decoder group, wherein the decoder group includes a space group decoder, an a decoder, a b decoder, an alpha decoder, a beta decoder, and a gamma decoder, the calculated features of the chemical formula and the real labels are sequentially input into the first splicing layer for splicing, and then the average value and the variance are obtained by encoding through the encoder, the average value and the variance are input into the reparameterization layer to obtain a first feature vector, the first feature vector and the calculated features are input into the second splicing layer for splicing to obtain a second feature vector, and the second feature vector is input into the decoder group to be decoded to obtain the space group category prediction value and the prediction value of the lattice parameters a, b, c, alpha, beta, and gamma, thereby realizing the prediction of the crystal geometry and symmetry;

[0008] Training the training model using the training sample set: the average value and the variance are output to the KL divergence as a first loss, the space group category prediction value and the space group category real label are output to a second cross-entropy loss function to obtain a second loss, and the prediction value of the lattice parameters a, b, c, alpha, beta, and gamma and the lattice parameter real label are output to a physical information fusion regression loss function to obtain a third loss, the total loss constructed based on the first loss, the second loss, and the third loss is used for training, and the second splicing layer and the decoder group with determined parameters constitute a prediction model;

[0009] Using the prediction model to predict the crystal geometry and symmetry.

[0010] In one embodiment, the reparameterization layer reparameterizes the average value μ and the variance σ to obtain the first feature vector z using the following formula:

[0011] z=μ+σ∈

[0012] wherein ∈ is random noise taken from a Gaussian distribution with a mean of 0 and a variance of 1.

[0013] In one embodiment, the first loss is calculated by the following formula:

[0014]

[0015] wherein μ and σ denote the mean and variance of the encoder output, respectively.

[0016] In one embodiment, the second loss is calculated by the following formula:

[0017]

[0018] wherein N is the number of crystals, K is the number of space group categories, t i (g) is the true label of the i-th crystal for the space group category g, y g,i is the predicted value of the i-th crystal for the space group category g.

[0019] In one embodiment, the third loss is calculated by the following formula:

[0020]

[0021] wherein r = {a, b, c, a, β, γ} represents a set of 6 lattice parameters, D = 6 represents the number of lattice parameter data in r, N is the number of crystals, r ij denotes the true label of the j-th lattice parameter of the i-th crystal, denotes the predicted value of the j-th lattice parameter of the i-th crystal, denotes the predicted value of the space group category, p ij (·) denotes the indicator of the j-th lattice parameter of the i-th crystal, denotes the indication result of .

[0022] In one embodiment, the indication result of is obtained by looking up the following table:

[0023]

[0024] when is the 1st-2nd space group category, the value is [1, 1, 1, 1, 1, 1];

[0025] when is the 3rd-15th space group category, the value is [1, 1, 1, 0, 0, 1];

[0026] when is the 16th-74th space group category, [1, 1, 1, 0, 0, 0];

[0027] When is the 75th-142nd space group class, [1, 0, 1, 0, 0, 0];

[0028] When is the 146th, 148th, 155th, 160th-161st, 166th-167th space group class, [1, 0, 1, 0, 0, 0];

[0029] When is the 143rd-145th, 147th, 149th-154th, 156th-159th, 162nd-165th, 168th-194th space group class, [1, 0, 1, 0, 0, 0];

[0030] When is the 195th-230th space group class, [1, 0, 0, 0, 0, 0].

[0031] In an embodiment, the encoder adopts a fully connected neural network, and each decoder in the decoder group adopts a fully connected neural network.

[0032] In an embodiment, the prediction by the prediction model on the crystal geometry and symmetry prediction comprises:

[0033] Randomly sampling noise from a Gaussian distribution as a first feature vector;

[0034] Inputting the calculated features of the chemical formula to be predicted and the first feature vector into the prediction model, and performing splicing through a second splicing layer to obtain a second feature vector, the second feature vector being decoded through the decoding group to obtain a space group class prediction value and prediction values of lattice parameters a, b, c, α, β and γ, wherein the prediction of symmetry is realized based on the space group class prediction value, and the prediction of crystal geometry is realized based on the prediction values of the lattice parameters a, b, c, α, β and γ.

[0035] To achieve the above-mentioned purposes, the embodiment further provides a crystal geometry and symmetry prediction device based on a physical information fusion neural network, comprising a sample construction module, a model construction module, a training module and an application prediction module,

[0036] The sample construction module is used to construct training samples: each training sample comprises calculated features of a chemical formula and a real label corresponding to the chemical formula, and the real label comprises a space group class real label and lattice parameter real labels;

[0037] The model construction module is configured to construct a training model, the training model comprising a first concatenation layer, an encoder, a reparameterization layer, a second concatenation layer, and a decoder set, wherein the decoder set comprises a space group decoder, an a decoder, a b decoder, an alpha decoder, a beta decoder, and a gamma decoder, and the calculated features of a chemical formula and a true label are sequentially input to the first concatenation layer for concatenation, and then the average value and the variance are obtained by encoding through the encoder, the average value and the variance are input to the reparameterization layer to obtain a first feature vector, the first feature vector and the calculated features are input to the second concatenation layer for concatenation to obtain a second feature vector, and the second feature vector is input to the decoder set to obtain a space group class prediction value and prediction values of lattice parameters a, b, c, alpha, beta, and gamma, so as to realize the prediction of the crystal geometry and symmetry;

[0038] The training module is configured to train the training model by using a training sample set: the average value and the variance are output to a KL divergence as a first loss, the space group class prediction value and the space group class true label are output to a second cross-entropy loss function to obtain a second loss, and the prediction values of the lattice parameters a, b, c, alpha, beta, and gamma and the lattice parameter true label are output to a regression loss function of physical information fusion to obtain a third loss, and the total loss constructed based on the first loss, the second loss, and the third loss is used for training, and the second concatenation layer and the decoder set determined by the parameters constitute a prediction model.

[0039] The application prediction module is configured to predict the crystal geometry and symmetry prediction by using the prediction model.

[0040] To achieve the above-mentioned purposes, the embodiment further provides a crystal geometry and symmetry prediction device based on physical information fusion neural network, which comprises a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, and the computer processor implements the above-mentioned crystal geometry and symmetry prediction method based on physical information fusion neural network when executing the computer program.

[0041] Compared with the prior art, the present application has at least the following beneficial effects:

[0042] In view of the technical problem that one chemical formula corresponds to multiple crystal structures in the prediction of crystal geometry and symmetry, the present application uses a physical information conditional variational autoencoder to encode multiple stable structures into a Gaussian distribution, thereby improving the accuracy of the prediction of crystal geometry and symmetry, including the prediction accuracy of lattice parameters and space groups. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative effort based on these drawings.

[0044] Figure 1 is a flowchart of the crystal geometry and symmetry prediction method based on the physical information fusion neural network provided by the embodiment;

[0045] Figure 2 is a training flowchart of the model provided by the embodiment;

[0046] Figure 3 is a network structure schematic diagram of the encoder provided by the embodiment;

[0047] Figure 4 is a network structure schematic diagram of the a decoder, the b decoder, the alpha decoder, the beta decoder and the gamma decoder provided by the embodiment;

[0048] Figure 5 is a network structure schematic diagram of the space group encoder provided by the embodiment;

[0049] Figure 6 is a flowchart of the crystal geometry and symmetry prediction based on the physical information fusion neural network provided by the embodiment;

[0050] Figure 7 is a structure schematic diagram of the crystal geometry and symmetry prediction device based on the physical information fusion neural network provided by the embodiment. DETAILED DESCRIPTION

[0051] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative effort based on these drawings.

[0052] Figure 1 is a flowchart of the crystal geometry and symmetry prediction method based on the physical information fusion neural network provided by the embodiment. As shown in Figure 1 , the crystal geometry and symmetry prediction method based on the physical information fusion neural network provided by the embodiment comprises the following steps:

[0053] Step 1, constructing a training sample.

[0054] In the embodiment, each training sample includes a calculated feature of a chemical formula and a true label corresponding to the chemical formula, and the true label includes a space group category true label and a lattice parameter true label.

[0055] The calculated feature of the chemical formula includes the following: atomic number, Mendeleev number, atomic weight, melting temperature, period table row and column, covalent radius, electronegativity, number of valence electrons in each orbital (s, p, d, f, total), number of unfilled electrons in each orbital (s, p, d, f, total), ground state band gap energy, ground state magnetic moment, stoichiometric p-norm (p = 0, 2, 3, 5, 7), element proportion fraction, electron proportion fraction in each orbital, energy band center, and ion characteristics (ability to form ionic compounds, ionic charge), maximum, minimum, average (all atoms divided by the number of element types) of atomic number, variance, and total number of atoms.

[0056] The crystal geometry and symmetry corresponding to each chemical formula are represented by the lattice parameters and the space group category, respectively. When the space group category is determined, the relationship between the corresponding lattice parameters is determined, and then the crystal system can be determined, and then the symmetry can be determined. The space group category true label is represented as g, and the lattice parameter true label is represented as a, b, α, β, and γ.

[0057] Step 2, constructing a training model, the training model includes a first concatenation layer, an encoder, a reparameterization layer, a second concatenation layer, and a decoder group.

[0058] As shown in Figure 2 , the training model constructed in the embodiment includes a first concatenation layer, an encoder, a reparameterization layer, a second concatenation layer, and a decoder group. The decoder group includes a space group decoder, an a decoder, a b decoder, an α decoder, a β decoder, and a γ decoder. The calculated feature of the chemical formula and the true label [g, a, b, α, β, γ] are input into the first concatenation layer in sequence for concatenation, and then the average μ and the variance σ are obtained by encoding through the encoder. The average μ and the variance σ are input into the reparameterization layer to obtain a first representation vector. The first representation vector and the calculated feature are input into the second concatenation layer for concatenation to obtain a second representation vector. The second representation vector is input into the decoder group to obtain the space group category prediction value and the prediction value of the lattice parameters a, b, c, α, β, and γ . The prediction of the crystal geometry and symmetry is realized.

[0059] In the embodiment, the encoder is composed of a fully connected neural network, as shown in Figure 3 , the number of neurons in the first layer of the fully connected neural network is 512, and the second layer is composed of two parts, each part containing 64 neurons.

[0060] In the embodiment, the reparameterization layer reparameterizes the mean value μ and the variance σ to obtain the first representation vector z by using the following formula:

[0061] z = μ + σ ∈

[0062] where ∈ is random noise taken from a Gaussian distribution with a mean of 0 and a variance of 1.

[0063] In the embodiment, the a decoder, the b decoder, the c decoder, the a decoder, the β decoder, and the γ decoder in the decoder group are all composed of a fully connected neural network, as shown in FIG. 3, and the number of neurons in each layer of the fully connected neural network is 256, 512, 256, 128, 64, and 1, respectively. Figure 4

[0064] In the embodiment, the spatial decoder is composed of a fully connected neural network, as shown in FIG. 4, and the number of neurons in each layer of the fully connected neural network is 256, 512, 256, 128, 64, and 230, respectively. Figure 5

[0065] Step 3: training the training model by using the training samples, and determining the second splicing layer and the decoder group as the prediction model.

[0066] In the embodiment, when the training model is trained, the total loss function used includes a first loss, a second loss, and a third loss, wherein the mean value μ and the variance σ are output to the KL divergence as the first loss

[0067]

[0068] The spatial group category prediction value and the spatial group category true label are output to the second cross-entropy loss function to obtain the second loss

[0069]

[0070] where N is the number of crystals, K is the number of spatial group categories, t i (g) is the true label of the spatial group category of the i th crystal, y g,i is the prediction value of the spatial group category of the i th crystal.

[0071] The prediction value of the lattice parameters a, b, c, a, β, and γ and the lattice parameter true label are output to the regression loss function of the physical information fusion to obtain the third loss

[0072]

[0073] ​​where r={a, b, c, a, b, g} represents a set of 6 lattice parameters, D=6 represents the number of lattice parameter data in r, N is the number of crystals, r ij represents the true label of the jth lattice parameter of the ith crystal, represents the predicted value of the jth lattice parameter of the ith crystal, represents the predicted value of the space group category, p ij (·) represents the indicator of the jth lattice parameter of the ith crystal, represents the indication result of . Table 1 is the indicator, represents the indication result of is obtained by looking up the following table:

[0074] Table 1

[0075]

[0076] When is the 1st-2nd space group category, takes the value [1, 1, 1, 1, 1, 1];

[0077] When is the 3rd-15th space group category, takes the value [1, 1, 1, 0, 0, 1];

[0078] When is the 16th-74th space group category, takes the value [1, 1, 1, 0, 0, 0];

[0079] When is the 75th-142nd space group category, takes the value [1, 0, 1, 0, 0, 0];

[0080] When is the 146th, 148th, 155th, 160th-161st, 166th-167th space group category, takes the value [1, 0, 1, 0, 0, 0];

[0081] When is the 143rd-145th, 147th, 149th-154th, 156th-159th, 162nd-165th, 168th-194th space group category, takes the value [1, 0, 1, 0, 0, 0];

[0082] When is the 195th-230th space group category, takes the value [1, 0, 0, 0, 0, 0].

[0083] Specifically, the first loss The second loss And the third loss The total loss is constructed by the weighted sum of the first loss, the second loss and the third loss, the model is parameter optimized by using the total loss, and after the parameter optimization is completed, the second splicing layer and the decoder group determined by the parameters constitute a prediction model as shown in Figure 6 .

[0084] Step 4, predicting the crystal geometry and symmetry prediction by using the prediction model.

[0085] As shown in Figure 6 , when predicting the crystal geometry and symmetry prediction by using the prediction model, noise is randomly sampled from a Gaussian distribution with a mean of 0 and a variance of 1 as a first feature vector; the calculated features of the chemical formula to be predicted and the first feature vector are input into the prediction model, and the second feature vector is obtained by splicing through the second splicing layer, and the space group category prediction value And the predicted values of the lattice parameters a, b, c, alpha, beta and gamma And Among them, the symmetry prediction is realized based on the space group category prediction value , and the crystal geometry prediction is realized based on the predicted value And .

[0086] The crystal geometry and symmetry prediction method based on the physical information fusion neural network provided by the embodiment solves the problem that one chemical formula corresponds to multiple stable crystal structures by using the physical information conditional variational autoencoder to encode multiple stable crystal structures into a Gaussian distribution.

[0087] Based on the same inventive concept, as shown in Figure 7 , the crystal geometry and symmetry prediction device based on the physical information fusion neural network provided by the embodiment includes a sample construction module, a model construction module, a training module and an application prediction module.

[0088] The sample construction module is configured to construct training samples, each of which includes a calculated feature of a chemical formula and a true label corresponding to the chemical formula, and the true label includes a space group category true label and a lattice parameter true label. The model construction module is configured to construct a training model, which includes a first concatenation layer, an encoder, a reparameterization layer, a second concatenation layer, and a decoder group, wherein the decoder group includes a space group decoder, an a decoder, a b decoder, a c decoder, an alpha decoder, a beta decoder, and a gamma decoder. The training module is configured to train the training model using the training samples, wherein the average value and the variance are output to a KL divergence as a first loss, the space group category predicted value and the space group category true label are output to a second cross-entropy loss function to obtain a second loss, and the predicted values of the lattice parameters a, b, c, alpha, beta, and gamma and the lattice parameter true label are output to a regression loss function of physical information fusion to obtain a third loss, and the total loss constructed based on the first loss, the second loss, and the third loss is used for training. The application prediction module is configured to predict the crystal geometry and symmetry prediction using the prediction model.

[0089] It should be noted that the crystal geometry and symmetry prediction device based on the physical information fusion neural network provided in the above embodiments should be divided into the above functional modules for illustration, and the above functions can be completed by different functional modules according to needs, that is, the internal structure of the terminal or server is divided into different functional modules to complete all or part of the above described functions. In addition, the crystal geometry and symmetry prediction device based on the physical information fusion neural network provided in the above embodiments and the crystal geometry and symmetry prediction method based on the physical information fusion neural network belong to the same concept, and the specific implementation process is described in detail in the crystal geometry and symmetry prediction method based on the physical information fusion neural network, which will not be repeated here.

[0090] Based on the same inventive company, the embodiments further provide a crystal geometry and symmetry prediction device based on a physical information fusion neural network, which includes a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, and the computer processor implements the following steps when executing the computer program:

[0091] Step 1, constructing a training sample;

[0092] Step 2, constructing a training model, the training model comprising a first concatenation layer, an encoder, a reparameterization layer, a second concatenation layer, and a decoder group;

[0093] Step 3, training the training model using the training sample, the second concatenation layer and the decoder group determining parameters to form a prediction model;

[0094] Step 4, predicting the crystal geometry and symmetry using the prediction model.

[0095] In practical applications, the memory can be a proximal volatile memory such as RAM, and can also be a non-volatile memory such as ROM, FLASH, floppy disk, mechanical hard disk, etc., and can also be a remote storage cloud. The processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), i.e., the crystal geometry and symmetry prediction method based on physical information fusion neural network can be implemented through these processors.

[0096] The above detailed description of the specific embodiments of the present application has described the technical solutions and beneficial effects of the present application. It should be understood that the above description is only the most preferred embodiment of the present application and is not intended to limit the present application. Any modifications, supplements, and equivalent replacements made within the principle range of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting crystal geometry and symmetry based on physical information fusion neural network, characterized in that, The method comprises the following steps: constructing a training sample: each training sample comprises calculated features of a chemical formula and a true label corresponding to the chemical formula, and the true label comprises a space group category true label and a lattice parameter true label, wherein the calculated features of the chemical formula are as follows: atomic number, Mendeleev number, atomic weight, melting temperature, period table row and column, covalent radius, electronegativity, number of valence electrons in each orbital, number of unfilled electrons in each orbital, ground state band gap energy, ground state magnetic moment, stoichiometric p-norm, element proportion fraction, electron proportion fraction in each orbital, energy band center and ion characteristics, maximum, minimum, average and variance of atomic number, and total number of atoms; constructing a training model: the training model comprises a first splicing layer, an encoder, a reparameterization layer, a second splicing layer and a decoder group, wherein the decoder group comprises a space group decoder, an a decoder, a b decoder, a c decoder, an alpha decoder, a beta decoder and a gamma decoder, the calculated features of the chemical formula and the true label are sequentially input into the first splicing layer for splicing, and then the average and the variance are obtained by encoding through the encoder, the average and the variance are input into the reparameterization layer to obtain a first feature vector, the first feature vector and the calculated features are input into the second splicing layer for splicing to obtain a second feature vector, and the second feature vector is input into the decoder group to obtain a space group category prediction value and prediction values of lattice parameters a, b, c, alpha, beta and gamma through decoding, so as to realize prediction of crystal geometry and symmetry; training the training model by using the training sample: the average and the variance are output to a KL divergence as a first loss, the space group category prediction value and the space group category true label are output to a second cross-entropy loss function to obtain a second loss, and the prediction values of the lattice parameters a, b, c, alpha, beta and gamma and the lattice parameter true label are output to a regression loss function of physical information fusion to obtain a third loss, and the total loss constructed based on the first loss, the second loss and the third loss is used for training, and the second splicing layer and the decoder group determined by parameters constitute a prediction model; predicting the crystal geometry and symmetry by using the prediction model. 2.The method of claim 1, wherein, The reparameterization layer reparameterizes the average μ and the variance σ to obtain the first feature vector z by using the following formula: z = μ + σ ∈ where ∈ is random noise taken from a Gaussian distribution with a mean of 0 and a variance of 1. 3.The method of claim 1, wherein, first loss is calculated using the following equation: Wherein, μ and σ represent the mean and variance output by the encoder respectively. 4.The method of claim 1, wherein, second loss is calculated using the following equation: where N is the number of crystals, K is the number of space group categories, t i (g) is the true label of the i-th crystal space group category, y g,i is the predicted value of the i-th crystal space group category. 5.The method of claim 1, wherein, Third loss is calculated using the following equation: where r = {a, b, c, a, b, g} represents a set of 6 lattice parameters, D = 6 represents the number of lattice parameter data in r, N is the number of crystals, r ij represents the true label of the jth lattice parameter of the ith crystal, represents the predicted value of the jth lattice parameter of the ith crystal, represents the predicted value of the space group class, p ij (·) represents the indicator of the jth lattice parameter of the ith crystal, represents the indication result of the jth lattice parameter of the ith crystal. represents the indication result of the jth lattice parameter of the ith crystal. 6.The method of claim 5, wherein, indicates the result of the indication of is found by looking up the following table: When For the 1st-2nd space group class, takes the value [1, 1, 1, 1, 1, 1]; When for the 3rd-15th space group class, takes the value [1, 1, 1, 0, 0, 1]; When for the 16th-74th space group class, takes the value [1, 1, 1, 0, 0, 0]; When for the 75th-142nd space group class, takes the value [1, 0, 1, 0, 0, 0]; When For the 146, 148, 155, 160-161, 166-167 space group classes, Has the value [1, 0, 1, 0, 0, 0]; When for the 143-145, 147, 149-154, 156-159, 162-165, 168-194 space group classes, takes the value [1, 0, 1, 0, 0, 0]; When is the 195th-230th space group class, takes the value [1, 0, 0, 0, 0, 0].

7. The method of claim 1, wherein the method is based on a physical information fusion neural network. The encoder adopts a fully connected neural network, and each decoder in the decoder group adopts a fully connected neural network. 8.The method of claim 1, wherein, The prediction of the crystal geometry and symmetry by using the prediction model comprises: randomly sampling noise from a Gaussian distribution as a first feature vector; inputting the calculated features of a chemical formula to be predicted and the first feature vector into the prediction model, splicing through the second splicing layer to obtain a second feature vector, and decoding the second feature vector through the decoding group to obtain a space group category prediction value and prediction values of lattice parameters a, b, c, alpha, beta and gamma, wherein the symmetry is predicted based on the space group category prediction value, and the crystal geometry is predicted based on the prediction values of the lattice parameters a, b, c, alpha, beta and gamma. 9.A device for predicting crystal geometry and symmetry based on physical information fusion neural network, characterized in that, The method comprises a sample construction module, a model construction module, a training module, and an application prediction module. The sample construction module is configured to construct training samples, each of which comprises calculated features of a chemical formula and a true label corresponding to the chemical formula, the true label comprising a space group category true label and a lattice parameter true label, wherein the calculated features of the chemical formula are as follows: atomic number, Mendeleev number, atomic weight, melting temperature, period table row and column, covalent radius, electronegativity, number of valence electrons in each orbital, number of unfilled electrons in each orbital, ground state band gap energy, ground state magnetic moment, stoichiometric p-norm, element proportion fraction, electron proportion fraction in each orbital, energy band center and ion characteristics, maximum, minimum, average, and variance of atomic number, and total number of atoms. The model construction module is configured to construct a training model, the training model comprising a first concatenation layer, an encoder, a reparameterization layer, a second concatenation layer, and a decoder group, wherein the decoder group comprises a space group decoder, an a decoder, a b decoder, a c decoder, an alpha decoder, a beta decoder, and a gamma decoder, the calculated features of the chemical formula and the true label being sequentially input into the first concatenation layer for concatenation, then being encoded by the encoder to obtain an average value and a variance, the average value and the variance being input into the reparameterization layer to obtain a first representation vector, the first representation vector and the calculated features being input into the second concatenation layer for concatenation to obtain a second representation vector, the second representation vector being input into the decoder group to obtain a space group category prediction value and prediction values of lattice parameters a, b, c, alpha, beta, and gamma through decoding, thereby realizing prediction of crystal geometry and symmetry. The training module is configured to train the training model using the training samples: the average value and the variance are output to a KL divergence as a first loss, the space group category prediction value and the space group category true label are output to a second cross-entropy loss function to obtain a second loss, the prediction values of the lattice parameters a, b, c, alpha, beta, and gamma and the lattice parameter true label are output to a regression loss function of physical information fusion to obtain a third loss, and the training is performed based on a total loss constructed based on the first loss, the second loss, and the third loss, the second concatenation layer and the decoder group with determined parameters forming a prediction model. The application prediction module is configured to predict the crystal geometry and symmetry prediction using the prediction model.

10. A device for predicting crystal geometry and symmetry based on physical information fusion neural network, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, characterized in that, The computer processor executes the computer program to implement the crystal geometry and symmetry prediction method based on the physical information fusion neural network according to any one of claims 1-8.

Images