Crystal structure generation method and device, electronic equipment and computer readable storage medium

Abstract

The application provides a crystal structure generation method and device, equipment, electronic equipment and a computer readable storage medium. The method comprises the following steps: obtaining a first crystal structure set of a predicted molecule; performing feature extraction on the first crystal structure set as a training set to obtain a feature vector of each crystal structure in the training set; training a neural network according to the feature vector of each crystal structure in the training set; generating a second crystal structure of the predicted molecule according to the neural network; the feature vector comprises: a cell parameter or a degree of freedom of a cell, a flexible angle of a molecular conformation or a molecular conformation, a position and a posture of the molecular conformation in the cell, or a central position and a direction of the molecular conformation in the cell; and the application can improve the generation efficiency of the crystal structure of the predicted molecule.

Description

Methods, apparatus, electronic devices and computer-readable storage media for generating crystal structures Technical Field

[0001] This application relates to the field of deep learning, and more particularly to a method and apparatus for generating crystal structures, electronic devices, and computer-readable storage media. Background Technology

[0002] Rapidly generating different types of crystal structures is one of the most crucial steps in crystal form prediction. Currently, the main methods for generating crystal structures in industry are generally based on traditional search algorithms such as global, local, and random search algorithms. The search space for crystal structures is enormous; traditional global search algorithms are very time-consuming, and local search algorithms are prone to getting trapped in local optima. Especially for crystals with multiple flexible dimensions, the search space increases exponentially with the increase in the degrees of freedom of the crystal structure. Summary of the Invention

[0003] This application can improve the efficiency of generating the crystal structure of the molecule to be predicted, and provides a crystal structure generation method, including:

[0004] Step A: Obtain the first set of crystal structures of the molecule to be predicted, the first set of crystal structures includes multiple crystal structures;

[0005] Step B: Using the first crystal structure set as the training set, feature extraction is performed to obtain the feature vectors of each crystal structure in the training set.

[0006] Step C: Train and generate a neural network based on the feature vectors of each crystal structure in the first crystal structure set;

[0007] Step D: Generate a second set of crystal structures for the molecule to be predicted based on the neural network, wherein the second set of crystal structures includes at least one crystal structure;

[0008] The feature vector includes:

[0009] Unit cell parameters or degrees of freedom of the unit cell,

[0010] The flexible angle of molecular conformation or molecular conformation, and

[0011] The position and orientation of a molecular conformation within a unit cell, or the central position and orientation of a molecular conformation within a unit cell.

[0012] Optionally, the method further includes:

[0013] Step E: Filter out the effective crystal structure of the molecule to be predicted from the second crystal structure set. When the difference between the effective crystal structure of each of the filtered molecules and all the crystal structures in the first crystal structure set is less than a first preset threshold, output the effective crystal structure of the filtered molecules as the crystal structure of the molecule to be predicted.

[0014] Optionally, the difference is determined based on the root mean square deviation between the filtered effective crystal structures and the crystal structures in the training set.

[0015] Optionally, step E further includes:

[0016] When the difference between the effective crystal structure of at least one filtered-out molecule to be predicted and all crystal structures in the first crystal structure set is not less than a first preset threshold, the effective crystal structure of the at least one filtered-out molecule to be predicted is added to the training set, and steps B to E are executed again.

[0017] Optionally, obtaining the first set of crystal structures for the molecule to be predicted includes:

[0018] Multiple effective crystal structures of the predicted molecule are generated using a predetermined number of random algorithms.

[0019] Optionally, the neural network is an adversarial neural network (WGAN) model; the adversarial neural network model includes a generator for generating the crystal structure of the proposed molecule and a discriminator for determining whether the crystal structure of the proposed molecule generated by the generator is close to the real one.

[0020] Optionally, training the neural network based on the feature vectors of each crystal structure in the training set includes:

[0021] In a uniformly distributed space, a z-vector is randomly sampled and input into the generator to generate the crystal structure of the molecule to be predicted. Then, the discriminator uses the Wasserstein distance between the crystal structure of the molecule to be predicted generated by the generator and multiple crystal structures in the first crystal structure set as the cost function. If the cost function meets the preset standard, the training stops. If the cost function does not meet the preset standard, the same step is repeated to further train the generator and the discriminator.

[0022] Optionally, the preset standard is: the fluctuation range of the Wasserstein distance for a preset number of consecutive times is within a preset range.

[0023] This application also provides a crystal structure generation apparatus, comprising:

[0024] First crystal structure set acquisition module: used to obtain the first crystal structure set of the molecule to be predicted, the first crystal structure set including multiple crystal structures;

[0025] Feature extraction module: used to extract features using the first crystal structure set as the training set, and obtain the feature vectors of each crystal structure in the training set respectively;

[0026] The feature vector includes:

[0027] The unit cell parameters or the degrees of freedom of the unit cell,

[0028] The flexible angle or molecular conformation of the molecular conformation, and

[0029] The position and orientation of the molecular conformation within the unit cell, or the central position and orientation of the molecular conformation within the unit cell.

[0030] Training module: used to train and generate a neural network based on the feature vectors of each crystal structure in the training set;

[0031] Crystal structure generation module: used to generate a second set of crystal structures of the molecule to be predicted based on the neural network, the second set of crystal structures including at least one crystal structure.

[0032] This application also provides an electronic device, including: a processor and a memory, wherein executable code is stored in the memory, and when the executable code is executed by the processor, the processor performs any of the methods described above.

[0033] This application also provides a computer-readable storage medium having executable code stored thereon, which, when executed by a processor in an electronic device, causes the electronic device to perform any of the methods described above.

[0034] In this embodiment, by using a simplified feature vector, a neural network is trained based on the feature vector using a deep learning method, which enables the rapid generation of crystal structures. Compared with the prior art that uses the coordinates of atoms in three-dimensional space as feature vectors, the number of parameters of the feature vector in this embodiment is significantly reduced, which can greatly reduce the search space, improve efficiency, and reduce computational costs. Attached Figure Description

[0035] Figure 1 is a schematic diagram of an embodiment of the crystal structure generation method of this application;

[0036] Figure 2 is a schematic diagram of an embodiment of the electronic device in this application;

[0037] Figure 3 is a schematic diagram of an embodiment of the crystal structure generation apparatus of this application. Detailed Implementation

[0038] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0039] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0040] It should be understood that although the terms "first," "second," "third," etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0041] As shown in Figure 1, Figure 1 is a schematic diagram of an embodiment of the crystal structure generation method of this application. The method includes:

[0042] Step A: Obtain the first set of crystal structures for the molecule to be predicted.

[0043] The first set of crystal structures includes multiple crystal structures. Optionally, a crystal structure can be a set composed of unit cells containing the same parallelepiped shape arranged adjacent to each other in the front, back, left, right, top, and bottom directions. There are various ways to obtain multiple crystal structures of the molecule to be predicted, such as the methods described in the paper "Crystal Structure Prediction via Particle Swarm Optimization Phys. Rev. B 82, 094116 (2010)" by Yanchao Wang, Jian Lv, Li Zhu, and Yanming Ma, or the paper "Particle-swarm structure prediction on clusters J. Chem. Phys. 137, 084104 (2012)" by Jian Lv, Yanchao Wang, Li Zhu, and Yanming Ma.

[0044] Optionally, a predetermined number of valid crystal structures for the predicted molecules can be generated using a random algorithm to obtain multiple crystal structures from the first crystal structure set of the predicted molecules described in step A. Valid crystal structures include those in which the internal atoms, ions, and molecules are arranged in a three-dimensional periodic pattern in space, the bonding relationships between atoms within the same molecule satisfy the bonding relationship characteristics displayed by the two-dimensional structure of the molecule, and a set of basis vectors corresponding to the three-dimensional periodicity and the corresponding unit cell crystal structure can be found. Crystal structures generated by random algorithms are not limited by the search space and can generate potentially existing valid crystal structures with less computational power and at a faster speed.

[0045] Step B: Using the first crystal structure set as the training set, feature extraction is performed to obtain the feature vectors of each crystal structure in the training set.

[0046] Each crystal structure's feature vector in the training set contains feature 1, feature 2, and feature 3. Feature 1 is the cell parameter or the degrees of freedom of the cell; feature 2 is the flexibility angle or molecular conformation; and feature 3 is the position and orientation of the molecular conformation within the cell, or the center position and orientation of the molecular conformation within the cell. Feature 1 in the feature vectors of different crystal structures of the molecule to be predicted in the training set can be all the same. For example, they can all be the cell parameters of the crystal structure of the molecule to be predicted, or they can all be the degrees of freedom of the crystal structure of the molecule to be predicted, or feature 1 in the feature vectors of some crystal structures of the molecule to be predicted can be the cell parameters of the crystal structure, while feature 1 in the feature vectors of other crystal structures can be the degrees of freedom of the crystal structure of the crystal structure. Similarly, features 2 and 3 in the feature vectors of different crystal structures of the molecule to be predicted in the training set are described similarly to feature 1 above.

[0047] Traditional crystal structure generation methods, such as stochastic or particle swarm optimization (PSO) algorithms, are time-consuming and prone to getting trapped in local optima. In this embodiment, a simplified feature vector is used. The neural network is trained by extracting feature vectors of multiple crystal structures of the molecule to be predicted. Since the feature vector includes the aforementioned features 1, 2, and 3 of multiple crystal structures of the molecule to be predicted, the neural network trained with this feature vector can generate effective crystal structures efficiently, quickly, and at low cost. In addition, compared with the prior art that uses the coordinates of each atom in three-dimensional space as the feature vector, the feature vector extracted in the three-dimensional space of the crystal structure in this embodiment can greatly reduce the spatial dimension. Moreover, the number of parameters is significantly reduced, which can greatly reduce the search space, improve efficiency, and reduce computational cost.

[0048] Step C: Train and generate a neural network based on the feature vectors of each crystal structure in the training set.

[0049] Optionally, the neural network trained based on the feature vectors of each crystal structure in the training set is a WGAN (Work-Adversarial Network) model. The WGAN model can quickly generate the crystal structure of the molecule to be predicted, and it can also improve the space for searching crystal structures and increase the diversity of generated crystal structures.

[0050] There are several ways to train a generative adversarial neural network (GAN) model. In one example, the GAN model includes two networks: one network generates data (i.e., the crystal structure of the molecule to be predicted), called the "generator"; the other network judges the authenticity of the data generated by the generator (i.e., whether it is close to reality), called the "discriminator." Training the GAN model based on feature vectors includes training both the generator and the discriminator based on the feature vectors. In one example, when training the GAN, the z-vector is first randomly sampled in a uniformly distributed space and input into the generator to generate the crystal structure of the molecule to be predicted. Then, the discriminator uses the Wasserstein distance between the crystal structure of the molecule to be predicted generated by the generator and multiple crystal structures in the first crystal structure set as the cost function. The process depends on whether the cost function meets a preset criterion to determine whether this step needs to be repeated to further train the generator and discriminator; that is, if the cost function meets the preset criterion, training stops; if the cost function does not meet the preset criterion, this step is repeated to further train the generator and discriminator.

[0051] The Wasserstein distance is used to characterize the probability distribution function of the feature vectors of the crystal structure of the molecule to be predicted. Also known as the Earth-Mover distance (EM distance), it measures the distance between two distributions. In this application example, the Wasserstein distance is used to measure the distance between the feature vector distributions of multiple newly generated crystal structures during training and the feature vector distributions of multiple crystal structures in the training set. Below is a formula for calculating the Wasserstein distance.

[0052]

[0053] Here, ∏(P1,P2) is the set of all possible joint distributions combining distributions P1 and P2. For each possible joint distribution γ, we can sample (x,y) ~ γ to a sample x and y, and calculate the distance ||xy|| between these samples. Therefore, we can calculate the expected value E(x,y)[||xy||] of the sample pair distance under this joint distribution γ. The lower bound that can be obtained for this expected value among all possible joint distributions is... This refers to the Wasserstein distance. E(x,y)[||xy||] can be solved as γ, representing the cost required to move pile P1 to pile P2 under this path planning. The Wasserstein distance represents the minimum cost under optimal path planning.

[0054] There can be multiple preset standards, which can be set as needed.

[0055] Optionally, the preset standard is that the fluctuation range of the Wasserstein distance for a preset number of consecutive training iterations is within a preset range. When the fluctuation range of the Wasserstein distance calculated after training the adversarial neural network model for a preset number of consecutive iterations is within the preset range, training of the generator and the discriminator is stopped. The preset number of iterations can be tens of times, such as 20, 25, 30, 35, etc. Optionally, the preset range can be a range smaller than a preset value, which can be set based on empirical values, such as 0.007, 0.008, 0.009, 0.0007, 0.0008, or 0.0009. When calculating the fluctuation range of the Wasserstein distance, it can be determined by using the Wasserstein distance value calculated after a preset number of consecutive training iterations, or by using the value of the Wasserstein distance before and after a preset number of decimal places, ignoring the value after the preset number of decimal places. For example, when determining whether the fluctuation range of the Wasserstein distance calculated after a preset number of consecutive training sessions is within the preset range, the fourth decimal place in the Wasserstein distance is ignored, and only the values ​​of the first and last three decimal places in each Wasserstein distance calculation are used to calculate the fluctuation range of the Wasserstein distance.

[0056] Step D: Generate a second set of crystal structures for the molecule to be predicted based on the neural network.

[0057] The second set of crystal structures includes at least one crystal structure. There are various ways to generate crystal structures, such as using existing techniques based on the neural network to generate the crystal structure of the predicted molecule. In one example, assuming the feature vector dimension is m, firstly, uniform sampling is performed in the m-dimensional space to obtain an m-dimensional vector following a uniform distribution. This vector is then input into a trained neural network model to obtain a new m-dimensional output vector. This output vector corresponds to the degree-of-freedom vector of a crystal structure, containing the aforementioned features 1, 2, and 3 of the crystal structure. Then, using the coordinate template of the crystal structure and this degree-of-freedom vector, a corresponding new crystal structure is generated.

[0058] In some examples, the crystal structure generated by the neural network cannot guarantee its validity, and optionally, the generated crystal structure may be filtered for validity. In one example, as shown in Figure 1, the crystal structure generation method further includes:

[0059] Step E: Filter out the effective crystal structure of the molecule to be predicted from the second crystal structure set. When the difference between the effective crystal structure of each of the filtered molecules and all the crystal structures in the first crystal structure set is less than a first preset threshold, output the effective crystal structure of the filtered molecules as the crystal structure of the molecule to be predicted.

[0060] Optionally, the difference is determined based on the root mean square deviation (RMSD) between the filtered effective crystal structures and the crystal structures in the training set.

[0061] For example, the degree of difference between two crystal structures can be calculated using the following formula.

[0062]

[0063] Where N represents the number of atoms, mi is the mass of the i-th atom, Xi is the coordinate vector of the target atom i, Yi is the coordinate vector of the reference atom i, and M is the total mass of the N atoms. When RMSD does not use mass weighting, all mi = 1 and M = N.

[0064] The first preset threshold can be pre-set based on experience, for example, a value within the range of 0 to 5, such as 0.5, 0.7, 0.8, 1.0, 1.2, 1.5, 1.8, etc. When the difference is less than the first preset threshold, it means that the similarity between the two crystal structures is relatively good.

[0065] Optionally, as shown in Figure 2, step E further includes:

[0066] When the difference between the effective crystal structure of at least one filtered-out molecule to be predicted and all crystal structures in the first crystal structure set is not less than a first preset threshold, the effective crystal structure of the at least one filtered-out molecule to be predicted is added to the training set, and steps B to E are executed again. For each filtered-out molecule to be predicted, the difference between the effective crystal structure and all crystal structures in the training set is calculated. Valid crystal structures that differ significantly from all crystal structures in the training set (e.g., not less than the first preset threshold) are added to the training set. This results in a more diverse first crystal structure set, allowing the trained neural network to explore a wider space and making it more likely to obtain the true crystal structure of the molecule to be predicted.

[0067] This application also provides an electronic device, as shown in FIG2. The electronic device 20 includes a processor 21 and a memory 22. The memory 22 stores executable code. When the executable code is executed by the processor 21, the processor 21 performs any of the above methods.

[0068] This application also provides a crystal structure generation apparatus. Specifically, as shown in FIG3, FIG3 is a schematic diagram of an embodiment of the crystal structure generation apparatus 300 of this application. This crystal structure generation apparatus includes obtaining a first crystal structure set.

[0069] This application also provides a crystal structure generation apparatus 300, comprising:

[0070] First crystal structure set acquisition module 301: used to obtain the first crystal structure set of the molecule to be predicted, the first crystal structure set including multiple crystal structures;

[0071] Feature extraction module 302: used to extract features using the first crystal structure set as the training set, and obtain the feature vectors of each crystal structure in the training set respectively;

[0072] The feature vector includes:

[0073] The unit cell parameters or the degrees of freedom of the unit cell,

[0074] The flexible angle or molecular conformation of the molecular conformation, and

[0075] The position and orientation of the molecular conformation within the unit cell, or the central position and orientation of the molecular conformation within the unit cell.

[0076] Training module 303: used to train and generate a neural network based on the feature vectors of each crystal structure in the training set;

[0077] Crystal structure generation module 304: used to generate a second set of crystal structures of the molecule to be predicted based on the neural network, the second set of crystal structures including at least one crystal structure.

[0078] Optionally, the device 300 further includes:

[0079] Filtering module 305: Used to filter out the effective crystal structure of the molecule to be predicted from the second crystal structure set. When the difference between the effective crystal structure of each molecule to be predicted and all crystal structures in the training set is less than a first preset threshold, the effective crystal structure of the molecule to be predicted is output as the crystal structure of the molecule to be predicted.

[0080] Optionally, the difference is determined based on the root mean square deviation between the filtered effective crystal structures and the crystal structures in the training set.

[0081] Optionally, the filtering module 305 is further configured to:

[0082] When the difference between at least one filtered effective crystal structure of the proposed molecule and all crystal structures in the training set is not less than a first preset threshold, the effective crystal structure of the at least one filtered effective crystal structure of the proposed molecule is added to the training set. The feature extraction module 302 is then used to extract features from the training set to obtain feature vectors for each crystal structure in the training set. The training module 303 is then used to train a neural network based on the feature vectors of each crystal structure in the training set. The crystal structure generation module 304 is then used to generate a second crystal structure set of the proposed molecule based on the neural network. The second crystal structure set includes at least one crystal structure.

[0083] Optionally, the first crystal structure set 301 obtaining module is used to generate a preset number of valid crystal structures of the molecule to be predicted through a random algorithm.

[0084] Optionally, the neural network is an adversarial neural network (WGAN) model; the adversarial neural network model includes a generator for generating the crystal structure of the proposed molecule and a discriminator for determining whether the crystal structure of the proposed molecule generated by the generator is close to the real one.

[0085] Optionally, the training module 303 is used to randomly sample the z vector in a uniformly distributed space, input it into the generator to generate the crystal structure of the molecule to be predicted, and then the discriminator uses the Wasserstein distance between the crystal structure of the molecule to be predicted generated by the generator and multiple crystal structures in the first crystal structure set as the cost function. If the cost function meets the preset standard, the training stops; if the cost function does not meet the preset standard, the same step is repeated to further train the generator and the discriminator.

[0086] Optionally, the preset standard is: the fluctuation range of the Wasserstein distance for a preset number of consecutive times is within a preset range. This application also provides a computer-readable storage medium (or a non-transitory machine-readable storage medium or a machine-readable storage medium) storing executable code (or a computer program or computer instruction code), which, when executed by a processor of an electronic device (or a server, etc.), causes the processor to perform some or all of the steps of the method described above according to this application.

[0087] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for generating a crystal structure, characterized in that, include: Step A: Obtain a first set of crystal structures for the molecule to be predicted, the first set of crystal structures including multiple crystal structures; Step B: Use the first set of crystal structures as a training set to extract features, and obtain feature vectors for each crystal structure in the training set; Step C: Train a neural network based on the feature vectors of each crystal structure in the training set; wherein the neural network is an adversarial neural network model; Step D: Generate a second set of crystal structures for the molecule to be predicted based on the neural network, the second set of crystal structures including at least one crystal structure; wherein the feature vectors include: unit cell parameters or degrees of freedom of the unit cell, flexibility angle of the molecular conformation, and position and orientation of the molecular conformation in the unit cell, or center position and orientation of the molecular conformation in the unit cell.

2. The method according to claim 1, characterized in that, The method further includes: step E: filtering out the effective crystal structure of the molecule to be predicted from the second crystal structure set; when the difference between the effective crystal structure of each of the filtered molecules and all crystal structures in the training set is less than a first preset threshold, the effective crystal structure of the filtered molecules to be predicted is output as the crystal structure of the molecule to be predicted.

3. The method according to claim 2, characterized in that, The difference is determined based on the root mean square deviation between the filtered effective crystal structures and the crystal structures in the training set.

4. The method according to claim 2, characterized in that, Step E further includes: when the difference between the effective crystal structure of at least one filtered-out molecule to be predicted and all crystal structures in the training set is not less than a first preset threshold, adding the effective crystal structure of the at least one filtered-out molecule to the training set, and executing steps B to E again.

5. The method according to claim 1, characterized in that, The process of obtaining the first set of crystal structures for the molecule to be predicted includes: generating a predetermined number of valid crystal structures for the molecule to be predicted using a random algorithm.

6. The method according to claim 1, characterized in that, The adversarial neural network model is a WGAN model; the adversarial neural network model includes a generator for generating the crystal structure of the proposed molecule and a discriminator for judging whether the crystal structure of the proposed molecule generated by the generator is close to the real one.

7. The method according to claim 6, characterized in that, The step of training the neural network based on the feature vectors of each crystal structure in the training set includes: randomly sampling the z vector in a uniformly distributed space, inputting it into the generator to generate the crystal structure of the molecule to be predicted, and then using the Wasserstein distance between the crystal structure of the molecule to be predicted generated by the generator and multiple crystal structures in the first crystal structure set as the cost function. If the cost function meets the preset standard, the training stops; if the cost function does not meet the preset standard, the step is repeated to further train the generator and the discriminator.

8. The method according to claim 7, characterized in that, The preset standard is: the fluctuation range of the Wasserstein distance for a preset number of consecutive times is within a preset range.

9. A crystal structure generation apparatus, characterized in that, include: First crystal structure set acquisition module: used to obtain a first crystal structure set of the molecule to be predicted, the first crystal structure set including multiple crystal structures; Feature extraction module: used to perform feature extraction using the first crystal structure set as a training set, respectively obtaining feature vectors of each crystal structure in the training set; wherein, the feature vectors include: unit cell parameters or degrees of freedom of the unit cell, flexible angle of the molecular conformation, and position and orientation of the molecular conformation in the unit cell, or center position and orientation of the molecular conformation in the unit cell; Training module: used to train a neural network based on the feature vectors of each crystal structure in the training set; wherein, the neural network is an adversarial neural network model; Crystal structure generation module: used to generate a second crystal structure set of the molecule to be predicted based on the neural network, the second crystal structure set including at least one crystal structure.

10. An electronic device, characterized in that, include: processor; And a memory having executable code stored thereon, which, when executed by the processor, causes the processor to perform the method as described in any one of claims 1 to 8.

11. A computer-readable storage medium having executable code stored thereon, which, when executed by a processor in an electronic device, causes the electronic device to perform the method as claimed in any one of claims 1 to 8.

Images