A method, apparatus, and device for training quantum chemistry models based on machine learning

By obtaining the three-dimensional coordinates of molecules from a non-equilibrium structure database and training a quantum chemical model using a neural network model and a semi-empirical quantum chemical model, the problem of insufficient accuracy of existing models is solved, and higher analytical precision and transferability are achieved.

CN118212996BActive Publication Date: 2026-11-17SHENZHEN BAY LAB
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311419239.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-10-30
Publication Date
2026-11-17
Estimated Expiration
2043-10-30

AI Technical Summary

Technical Problem

Existing machine learning quantum chemistry models are insufficient in terms of accuracy, making it difficult to effectively balance computational efficiency and precision.

Method used

By obtaining the three-dimensional coordinates of molecules from a non-equilibrium structure database, calculating the descriptive parameters of the molecules using a neural network model, and training a quantum chemical model by combining it with a semi-empirical quantum chemical model, the model parameters are optimized through a molecular property loss function to improve the accuracy of the model.

Benefits of technology

It improves the accuracy and transferability of quantum chemical models, enabling them to better analyze the structural properties and interactions of molecules, and is applicable to the study of a variety of molecular and chemical properties.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118212996B_ABST
    Figure CN118212996B_ABST
Patent Text Reader

Abstract

A machine learning-based quantum chemical model training method, apparatus, and medium include: calculating a first descriptive parameter of a molecule based on its three-dimensional coordinates; inputting the first descriptive parameter into a neural network model to obtain a second descriptive parameter; calculating the enthalpy of formation of the molecule based on the second descriptive parameter and the single-center parameter in a semi-empirical quantum chemical model; adding the enthalpy of formation of the molecule to the forces acting on atoms in the quantum chemical model to be trained as the molecular property of the model; constructing the quantum chemical model based on the neural network model and the semi-empirical quantum chemical model; calculating the molecular property loss function between the model molecular property and the reference molecular property; and using the molecular property loss function to update the model parameters in the neural network model of the quantum chemical model to be trained, resulting in a trained quantum chemical model. The trained quantum chemical model can accurately analyze the molecular properties of molecules in various fields.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and specifically to a method, apparatus, and device for training quantum chemical models based on machine learning. Background Technology

[0002] Quantum chemical (QC) models are an effective method for studying the physicochemical properties of atomic systems and have been widely adopted in various research disciplines. For example, in physical chemistry, they are used to analyze the thermodynamics (such as enthalpy of formation and free energy) and structural properties (such as bond lengths, bond angles, and dipole moments) of molecules; in biochemistry, they are used to study the interactions of biomolecules. Balancing efficiency and accuracy in QC models has always been a hot topic and a challenge in this field.

[0003] Quantum chemistry methods at different scales, such as Hartree-Fock (HF), Density Functional Theory (DFT), Semiempirical QC (SQC), and Quantum Mechanics / Molecular Mechanics (QM / MM), have balanced computational efficiency and accuracy from multiple perspectives and have wide applications in fields such as biology and chemistry. In recent years, with the rapid development and widespread application of machine learning (ML) algorithms, new methods and applications based on the combination of ML and QC have emerged in large numbers. ML-QC methods not only rival MM methods in computational efficiency but also demonstrate great potential in the accuracy of predicting QC properties.

[0004] The combination of ML and QC methods is very flexible, mainly divided into two categories: specialized ML-QC models and general-purpose ML-QC models. However, most ML-QC models lack sufficient accuracy. Summary of the Invention

[0005] The main technical problem this invention addresses is the insufficient accuracy of the ML-QC model.

[0006] According to the first aspect, one embodiment provides a machine learning-based quantum chemical model training method for analyzing the structural properties and interactions of molecules, including:

[0007] Obtain molecules and their corresponding three-dimensional coordinates from a preset non-equilibrium structure database, and calculate the first description parameter of the molecule based on the three-dimensional coordinates of the molecule;

[0008] The first descriptive parameters of the molecule are input into a preset neural network model to obtain the second descriptive parameters of the molecule; wherein:

[0009] The first description parameter is the structural descriptor corresponding to each atom in the molecule, the second description parameter is the correction value of the single-center parameter corresponding to each atom in the molecule, and the neural network model is a model constructed based on the atom type corresponding to each atom in the molecule; or, the first description parameter is the thermometer code corresponding to the distance of each atom pair contained in the molecule, the second description parameter is the correction value of the two-center two-electron integral of each atom pair in the molecule, and the neural network model is a model constructed based on the atom type corresponding to each atom pair in the molecule;

[0010] The enthalpy of formation of the molecule is calculated based on the second descriptive parameter of the molecule and the single-center parameter in the preset semi-empirical quantum chemical model. The enthalpy of formation of the molecule and the atomic force corresponding to the quantum chemical model to be trained are added together to obtain the molecular property of the model. The quantum chemical model to be trained is constructed based on the neural network model and the semi-empirical quantum chemical model.

[0011] The molecular property loss function value between the model molecular properties and the preset reference molecular properties is calculated based on the preset molecular property loss function. The model parameters in the neural network model of the quantum chemical model to be trained are updated using the molecular property loss function value to obtain the trained quantum chemical model.

[0012] In one embodiment, calculating the first descriptive parameter of the molecule based on the three-dimensional coordinates corresponding to the molecule includes:

[0013] The atomic density field corresponding to each atom in the molecule is calculated based on the three-dimensional coordinates of the molecule and the preset atomic density field formula.

[0014] The atomic density field is expanded into an expansion formula containing a preset radial basis function, a preset spherical harmonic function, and expansion coefficients, and the values ​​corresponding to the expansion coefficients are calculated.

[0015] The corresponding spectral vector is calculated using the value of the expansion coefficient and the preset spectral vector formula. A local power spectral vector is constructed based on the spectral vector. The local power spectral vector is used as the structural descriptor corresponding to the atom, and the structural descriptor corresponding to each atom in the molecule is used as the first description parameter of the molecule.

[0016] In one embodiment, expanding the atomic density field into an expansion comprising a preset radial basis function, a preset spherical harmonic function, and expansion coefficients includes:

[0017]

[0018] Where, ρz (r) represents the atomic density field of the current atom in the molecule, z represents the relative position coordinates between the nuclei of the neighboring atoms and the current atom, r represents the three-dimensional coordinates of the molecule, n represents the order of the preset radial basis function, l represents the angular momentum quantum number, and m represents the magnetic quantum number. G represents the expansion coefficient. n (r) represents the radial basis function, Y lm (θ,φ) represents the preset spherical harmonic function, where θ represents the polar angle and φ represents the azimuth angle.

[0019] In one embodiment, calculating the first descriptive parameter of the molecule based on the three-dimensional coordinates corresponding to the molecule includes:

[0020] The thermometer code corresponding to the distance of each atom pair contained in the molecule is calculated according to a preset thermometer coding formula, and the thermometer code corresponding to the distance of each atom pair contained in the molecule is used as the first descriptive parameter of the molecule; wherein:

[0021] The preset thermometer coding formula includes:

[0022]

[0023] in, This is the thermometer code corresponding to the distance between the atomic pairs formed by atoms i and j, where k is a preset parameter, and r1 represents the atomic distance between the atomic pairs calculated based on the three-dimensional coordinates r of the molecule. k For a pre-defined uniform network, σ represents the reciprocal of the pre-defined grid spacing.

[0024] In one embodiment, calculating the enthalpy of formation of the molecule based on the second descriptive parameter of the molecule and the single-center parameter in a preset semi-empirical quantum chemical model includes:

[0025] The single-electron matrix, two-electron matrix, and nucleo-nucleo interaction matrix are calculated based on the second descriptive parameter of the molecule and the single-center parameter in the preset semi-empirical quantum chemical model, respectively; the single-electron matrix includes a single-center single-electron matrix and a two-center single-electron matrix; the two-electron matrix includes a single-center two-electron matrix and a two-center two-electron matrix;

[0026] A linear transformation is performed on the dual-center single-electron matrix in the single-electron matrix and the dual-center dual-electron matrix in the dual-electron matrix to obtain the single-electron matrix and dual-electron matrix in the molecular coordinate system.

[0027] The two-electron integral is calculated based on the two-electron matrix in the molecular coordinate system and the preset electron density matrix. The two-electron integral is then added to the single-electron matrix in the molecular coordinate system to obtain the interaction matrix.

[0028] Calculate the integral corresponding to the interaction matrix, convert the integral corresponding to the interaction matrix into the initial electron energy according to matrix diagonalization, calculate the difference between the initial electron energy and the electron energy after self-consistent field iteration, and obtain the energy difference. When the energy difference meets the preset energy threshold, the electron energy after self-consistent field iteration is used as the output electron energy.

[0029] The enthalpy of formation of the molecule is calculated based on the output electron energy and the nucleus-nucleus interaction matrix.

[0030] In one embodiment, the linear transformation of the two-center single-electron matrix in the single-electron matrix and the two-center two-electron matrix in the two-electron matrix to obtain the single-electron matrix and the two-electron matrix in the molecular coordinate system includes:

[0031] Construct a local biatomic coordinate system and a molecular coordinate system; wherein, there is a mapping relationship between the local biatomic coordinate system and the molecular coordinate system and the atoms in the molecule;

[0032] The double-center single-electron matrix and the double-center double-electron matrix in the single-electron matrix of the local diatomic coordinate system are transformed to the molecular coordinate system using a pre-calculated rotation matrix to obtain the single-electron matrix and double-electron matrix in the molecular coordinate system.

[0033] In one embodiment, the preset molecular property loss function includes:

[0034]

[0035] Where loss is the value of the molecular property loss function, loss penalty For the penalty loss function, N geo Represents the number of molecular conformations, E ML F represents the enthalpy of formation in the molecular properties of the model. ML This represents the forces acting on atoms in the model molecule. F represents the enthalpy of formation in the properties of the reference molecule. DFT The force on the atom in the reference molecule property is represented by w, which represents a preset training experience value, and λ is the value of the reference molecule. penalty V is the penalty coefficient, N is the number of single-center correction parameters in the molecule, and V is the number of correction parameters. i w represents the parameter value corresponding to the single-center correction parameter contained in the molecule. i This represents the empirical value corresponding to multiple training sessions for the molecule.

[0036] In one embodiment, before calculating the molecular property loss function value between the model molecular property and the preset reference molecular property based on the preset molecular property loss function, the method further includes:

[0037] The target enthalpy of formation and the forces acting on the target atom corresponding to the quantum chemical model to be trained were calculated using density functional theory.

[0038] When the preset training dataset contains molecules that meet the preset conformation conditions, the target generation enthalpy and the force on the target atom corresponding to the quantum chemical model to be trained are added together to obtain the properties of the reference molecule.

[0039] When the preset training dataset does not contain molecules that meet the preset conformation conditions, the energy deviation corresponding to the molecule is extracted, and the difference between the energy deviation corresponding to the molecule and the target enthalpy of formation corresponding to the quantum chemical model to be trained is calculated to obtain the calculated enthalpy of formation. The calculated enthalpy of formation and the force on the target atom are added together to obtain the reference molecule properties.

[0040] In one embodiment, the unbalanced structure database includes the ANI-1 database.

[0041] In one embodiment, the atomic force corresponding to the quantum chemical model to be trained is calculated according to a preset machine learning framework; wherein, the atomic force is the derivative of the formation enthalpy of the molecule with respect to the three-dimensional coordinate value of the molecule.

[0042] According to a second aspect, one embodiment provides a method for molecular property analysis, characterized in that it includes:

[0043] Obtain the three-dimensional coordinates of the molecule to be analyzed;

[0044] The three-dimensional coordinates of the molecule to be analyzed are input into a trained quantum chemical model to obtain the molecular properties of the molecule to be analyzed; the molecular properties include the enthalpy of formation and / or the forces acting on atoms; the quantum chemical model is trained using a quantum chemical model training method.

[0045] The machine learning-based quantum chemical model training method, apparatus, and medium according to the above embodiments include obtaining molecules and their corresponding three-dimensional coordinates from a preset non-equilibrium structure database, and calculating a first descriptive parameter of the molecule based on the corresponding three-dimensional coordinates. The preset non-equilibrium structure database includes chemical and conformational spaces; therefore, the first descriptive parameter of the molecule calculated from the molecule and its corresponding three-dimensional coordinates can capture all chemical reactions and properties of substances within the molecule, thus providing greater accuracy. The first descriptive parameter of the molecule is input into a preset neural network model to obtain a second descriptive parameter of the molecule. The enthalpy of formation corresponding to the molecule is calculated based on the second descriptive parameter of the molecule and the single-center parameter in the preset semi-empirical quantum chemical model. The enthalpy of formation corresponding to the molecule is added to the atomic forces corresponding to the quantum chemical model to be trained as the model molecule property. The model molecule property incorporates a description of atomic forces, and training the quantum chemical model to be trained based on the model molecule property can improve the accuracy of the trained model. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the training process of a machine learning-based quantum chemistry model according to an embodiment of the present invention.

[0047] Figure 2 This is a flowchart illustrating the training process of a machine learning-based quantum chemistry model, as shown in one embodiment.

[0048] Figure 3 A flowchart illustrating the training process of a machine learning-based quantum chemistry model, as shown in another embodiment.

[0049] Figure 4 A flowchart illustrating the training process of a machine learning-based quantum chemistry model, as shown in another embodiment.

[0050] Figure 5 A flowchart illustrating the training process of a machine learning-based quantum chemistry model, as shown in another embodiment.

[0051] Figure 6 A flowchart illustrating the training process of a machine learning-based quantum chemistry model, as shown in another embodiment.

[0052] Figure 7 This is a flowchart illustrating the forward propagation process in a machine learning-based quantum chemistry model training method in one embodiment.

[0053] Figure 8 This is a flowchart illustrating the combination of a neural network model and a preset semi-empirical quantum chemical model in a quantum chemical model to be trained, as shown in one embodiment.

[0054] Figure 9 This is a structural diagram of the neural network model in the quantum chemical model to be trained in one embodiment;

[0055] Figure 10 This is a flowchart of the forward propagation process in a machine learning-based quantum chemistry model training method in another embodiment;

[0056] Figure 11 This is a flowchart illustrating the combination of a neural network model and a preset semi-empirical quantum chemical model in another embodiment of the quantum chemical model to be trained.

[0057] Figure 12 This is a structural diagram of the neural network model in the quantum chemical model to be trained in another embodiment;

[0058] Figure 13 A table showing the single-center parameters corresponding to atoms in the quantum chemical model to be trained;

[0059] Figure 14 This is a table showing the two-center, two-electron integrals corresponding to the atom pairs in the quantum chemical model to be trained. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of this application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to this application are not shown or described in the specification. This is to avoid obscuring the core parts of this application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0061] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0062] The serial numbers assigned to components in this document, such as "first" and "second," are used only to distinguish the described objects and have no sequential or technical meaning. The terms "connection" and "linkage" used in this application, unless otherwise specified, include both direct and indirect connections (linkages).

[0063] The machine learning-based quantum chemical model training method, apparatus, and medium according to the above embodiments include obtaining molecules and their corresponding three-dimensional coordinates from a preset non-equilibrium structure database, and calculating the first descriptive parameters of the molecules based on the corresponding three-dimensional coordinates. The preset non-equilibrium structure database includes chemical and conformational spaces; therefore, the first descriptive parameters of the molecules calculated from the molecules and their corresponding three-dimensional coordinates can capture all chemical reactions and the characteristics of substances in the molecules, thus providing greater accuracy. The first descriptive parameters of the molecules are input into a preset neural network model to obtain the second descriptive parameters of the molecules. The enthalpy of formation corresponding to the molecules is calculated based on the second descriptive parameters and the single-center parameters in the preset semi-empirical quantum chemical model. The enthalpy of formation corresponding to the molecules is added to the atomic forces corresponding to the quantum chemical model to be trained to obtain the model molecular properties. The preset neural network model is constructed based on the atomic type corresponding to each atom or the atomic type corresponding to each pair of atoms in the molecule, thus possessing strong transferability. The model molecular properties incorporate descriptions of atomic forces, and training the quantum chemical model to be trained based on these model molecular properties can improve the accuracy of the trained model. A well-trained quantum chemical model is not limited to a specific molecule, chemical reaction, or chemical property; it is also transferable and can be used for the research and analysis of a variety of molecules and chemical properties.

[0064] Please refer to Figure 1 Some embodiments of the present invention provide a quantum chemical model training method based on machine learning, including steps S10 to S40, which are described in detail below.

[0065] Step S10: Obtain the molecule and its corresponding three-dimensional coordinates from the preset non-equilibrium structure database, and calculate the first description parameter of the molecule based on the corresponding three-dimensional coordinates.

[0066] In some embodiments, the nonequilibrium structure database includes the ANI-1 database. The ANI-1 database is a nonequilibrium structure database suitable for training machine learning models. This database covers both chemical and conformational spaces and is obtained through normal mode sampling (NMS). In conformational space, a single molecule can contain a series of different nonequilibrium conformations.

[0067] Please refer to Figure 2 In some embodiments, step S10, which calculates the first description parameters of the molecule based on the three-dimensional coordinates of the molecule, includes steps S11 to S13, which will be described in detail below.

[0068] Step S11: Calculate the atomic density field corresponding to each atom in the molecule based on the three-dimensional coordinates of the molecule and the preset atomic density field formula.

[0069] In some embodiments, the preset atomic density field formula includes:

[0070]

[0071] Where, ρ z (r) represents the atomic density field of the current atom in the molecule, z represents the relative position coordinates between the nucleus of the neighboring atom and the nucleus of the current atom, |z| represents the relative distance between the current atom and the position of the neighboring atom, with the nucleus of the current atom as the reference point, and σ 2 R represents the variance of the Gaussian function. i The position of the neighboring atom is represented by r, and the three-dimensional coordinates of the molecule are represented by r.

[0072] In some embodiments, the atomic structure is transformed into an atomic density field ρ for each atom using an unnormalized Gaussian field centered on each atom. z .

[0073] Step S12: Expand the atomic density field into an expansion formula containing a preset radial basis function, a preset spherical harmonic function and expansion coefficients, and calculate the value corresponding to the expansion coefficients.

[0074] In some embodiments, the atomic density field can be represented by a set of orthonormal radial basis functions g. n Spherical harmonic function Y lm The expanded form includes:

[0075]

[0076] Where, ρ z (r) represents the atomic density field of the current atom in the molecule, z represents the relative position coordinates between the nuclei of the neighboring atoms and the current atom, r represents the three-dimensional coordinates of the molecule, n represents the order of the preset radial basis function, l represents the angular momentum quantum number, and m represents the magnetic quantum number. G represents the expansion coefficient. n (r) represents the radial basis function, Y lm (θ,φ) represents the preset spherical harmonic function, where θ represents the polar angle and φ represents the azimuth angle.

[0077] In some embodiments, the angular quantum number l is typically a non-negative integer. The range of the quantum number m is determined by the angular quantum number l, and is typically an integer between -l and l. The polar angle θ is typically a real value representing the angle from the positive semi-axis to a point. The azimuth angle φ is typically a real value representing the angle of clockwise rotation from the reference direction to a point.

[0078] In some embodiments, the values ​​corresponding to the expansion coefficients are calculated, including:

[0079]

[0080] in, G represents the expansion coefficient. n (r) represents the radial basis function, Y lm (θ,φ) represents the predefined spherical harmonic function, where θ represents the polar angle, φ represents the azimuth angle, and ρ z (r) represents the atomic density field of the current atom in the molecule, z represents the number of neighboring atoms, r represents the three-dimensional coordinate value of the molecule, n represents the order of the preset radial basis function, l represents the angular quantum number, m represents the magnetic quantum number, and R represents the neighboring atom.

[0081] Step S13: Calculate the corresponding spectral vector using the values ​​of the expansion coefficients and the preset spectral vector formula. Construct a local power spectral vector based on the spectral vector. Use the local power spectral vector as the structural descriptor corresponding to the atom, and use the structural descriptor corresponding to each atom in the molecule as the first description parameter of the molecule.

[0082] In some embodiments, the preset spectral vector formula includes:

[0083]

[0084] in, Let n' represent the spectral vectors for atomic number pairs Z1 and Z2, n' represent the order of the preset radial basis function, l represent the angular quantum number, m represent the magnetic quantum number, Z1 and Z2 represent different atomic numbers, and n' represent the order of the radial basis function for atomic number Z2.

[0085] In some embodiments, in the spectral vector It contains radial basis functions from nn′ to nn max and angular function values ​​from l to lmax. Here, nn max represents the maximum number of radial basis functions, and lmax represents the maximum value of the angular quantum number.

[0086] In some embodiments, a local power spectral vector is constructed based on the spectral vector. This local power spectral vector is used as the structural descriptor corresponding to the atom, and the structural descriptor corresponding to each atom in the molecule is used as the first descriptive parameter of the molecule. Using the local power spectral vector as the structural descriptor corresponding to the atom mainly involves constructing a matrix from the local power spectral vector according to a preset matrix size. The constructed matrix then serves as the structural descriptor corresponding to the atom. The preset matrix size is [n_positions, n_features], where n_positions represents the number of atoms contained in the molecule, and n_features refers to the value used by each atom to describe its surrounding chemical environment. n_features = ((n_elem_radial)*(n_elem_radial+1) / 2)*(lmax+1), where n_elem_radial = n_elem*nmax, lmax represents the maximum value of the angular quantum number, nmax represents the maximum number of radial functions and the angular quantum number from l to the maximum value lmax, n_elem represents the number of element types, and n_elem_radial represents the total number of radial functions for all element types.

[0087] In some embodiments, the structure descriptor is a SOAP structure descriptor, a molecular structure descriptor based on atomic density. It calculates the local power spectrum vector of a single atom in a molecule under the influence of surrounding atoms. Its advantage lies in the continuity of its derivatives, making it more suitable as input to neural networks compared to internal coordinate-based molecular structure descriptors, as it can accurately describe energy gradients. Because atomic density-based molecular structure descriptors are independent of the molecule's reference frame, they are not affected by coordinate selection or rotation, resulting in better stability and reliability, and a closer approximation of actual material properties. For example, by analyzing the degree of electron cloud overlap between atoms, information on the strength of intermolecular interactions can be obtained, which is highly valuable for studying chemical reactions or crystal structures.

[0088] In some embodiments, the first descriptive parameters of the molecule are calculated based on the three-dimensional coordinates corresponding to the molecule, including:

[0089] The thermometer code corresponding to the distance of each atom pair contained in the molecule is calculated according to a preset thermometer coding formula, and the thermometer code corresponding to the distance of each atom pair contained in the molecule is used as the first descriptive parameter of the molecule; where:

[0090] Preset thermometer coding formula, including:

[0091]

[0092] in, This is the thermometer code corresponding to the distance between the atomic pairs formed by atoms i and j, where k is a preset parameter, and r1 represents the atomic distance between the atomic pairs calculated based on the three-dimensional coordinates r of the molecule. k For a pre-defined uniform network, σ represents the reciprocal of the pre-defined grid spacing.

[0093] In some embodiments, the preset parameter k is 10, and the preset uniform network r k It is a uniform network consisting of 10 points. The 10 points refer to the points obtained directly by uniform sampling based on the defined minimum and maximum distances. The uniform network r k The range is from the minimum inter-atomic distance in the dataset to the cutoff distance, where the cutoff distance refers to the maximum distance defined for each type of atom pair based on the chemical environment. The preset grid spacing refers to the distance between atoms in a uniform network r. k The distance between adjacent points among 10 points.

[0094] Step S20: Input the first description parameters of the molecule into a preset neural network model to obtain the second description parameters of the molecule; wherein: the first description parameter is the structural descriptor corresponding to each atom in the molecule, the second description parameter is the correction value of the single-center parameter corresponding to each atom in the molecule, and the neural network model is a model constructed according to the atom type corresponding to each atom in the molecule; or, the first description parameter is the thermometer code corresponding to the distance of each atom pair contained in the molecule, the second description parameter is the correction value of the two-center two-electron integral of each atom pair in the molecule, and the neural network model is a model constructed according to the atom type corresponding to each atom pair in the molecule.

[0095] In some embodiments, when the first description parameter is the structural descriptor corresponding to each atom in the molecule, the second description parameter is the correction value of the single-center parameter corresponding to each atom in the molecule, and the neural network model is a model constructed based on the atom type corresponding to each atom in the molecule, the structural descriptor corresponding to each atom in the molecule is input into the model constructed based on the atom type corresponding to each atom in the molecule to obtain the correction value of the single-center parameter corresponding to each atom in the molecule.

[0096] In some embodiments, when the first description parameter is the thermometer code corresponding to the distance of each atom pair contained in the molecule, the second description parameter is the correction value of the two-center two-electron integral of each atom pair in the molecule, and the neural network model is a model constructed based on the atom type corresponding to each atom pair in the molecule, the thermometer code corresponding to the distance of each atom pair contained in the molecule is input into the model constructed based on the atom type corresponding to each atom pair in the molecule to obtain the correction value of the two-center two-electron integral of each atom pair in the molecule.

[0097] Please refer to Figure 13The table shown illustrates that light atoms and heavy atoms each contain different single-center parameters. Light atoms contain four single-center parameters, namely U... ss β s ξ s and G SS Heavy atoms contain 11 single-center parameters, namely U ss U pp β s β p ξ s ξ p G SS G SP H SP G PP and G P2 For the four atomic types C, N, O, and H, only H is a light atom, while C, N, and O are heavy atoms. Therefore, C, N, O, and H have 11, 11, 11, and 4 single-center parameters respectively. Adding the number of single-center parameters in each sub-neural network yields 37 single-center parameters, corresponding to 37 sub-neural networks. A neural network model is then constructed based on these 37 sub-neural networks. Each sub-neural network model contains one hidden layer with 8 nodes, and the Tanh activation function is used between the hidden layer and the output layer.

[0098] In some embodiments, the same sub-neural network can be used for atoms of the same atom type. For example, in the H2O molecule, H1 and H2 both belong to the H atom type, so H1 and H2 use the same sub-neural network. H has 4 single-center parameters, while O has 11 single-center parameters; therefore, the neural network model corresponding to the H2O molecule contains 15 sub-neural networks.

[0099] Please refer to Figure 14 The table shown, Figure 14 It includes the two-center, two-electron integral in the local coordinate system.

[0100] In some embodiments, a neural network model is constructed based on the atom type corresponding to each atom in the molecule or the atom type corresponding to each pair of atoms. Essentially, a separate sub-neural network is constructed for each parameter of each atom type, thus ensuring the transferability of the model.

[0101] Step S30: Calculate the enthalpy of formation of the molecule based on the second descriptive parameter of the molecule and the single-center parameter in the preset semi-empirical quantum chemical model. Add the enthalpy of formation of the molecule to the force on the atom corresponding to the quantum chemical model to be trained as the molecular property of the model. The quantum chemical model to be trained is constructed based on the neural network model and the semi-empirical quantum chemical model.

[0102] In some embodiments, a quantum chemical model to be trained is constructed based on a neural network model and a semi-empirical quantum chemical model. The semi-empirical quantum chemical model can be the AM1 (Austin Model 1) model, or other quantum chemical methods, such as PM3 (Parametric Method 3) and DFTB (Density Functional Tight Binding). In this scheme, the quantum chemical model to be trained is an ML-AM1 model constructed based on the neural network model corresponding to the machine learning algorithm and the semi-empirical quantum chemical model AM1.

[0103] In some embodiments, the AM1 model uses molecular orbital theory to describe the electronic structure of molecules. A key feature of the AM1 model is its reduction of complexity and computational cost through approximations of less empirically relevant molecular orbitals. By considering factors such as molecular symmetry and orbital overlap, the AM1 model has found wide application in chemistry, materials science, and biomedicine. Furthermore, by adding a Gaussian function to the nucleus-nuclear interactions while neglecting modified neglectodiatomic overlap (MNDO), the AM1 model achieved, for the first time, a description of hydrogen bonding interactions in a semi-empirical method.

[0104] In some embodiments, the atomic forces corresponding to the quantum chemical model to be trained are calculated using a pre-defined machine learning framework. Here, the atomic force is the derivative of the enthalpy of formation of the molecule with respect to its three-dimensional coordinates. The pre-defined machine learning framework refers to the PyTorch framework. Since the entire process from the 3D coordinates of the molecule to the enthalpy of formation—that is, from coordinates to energy—is implemented within the PyTorch framework, the automatic differentiation function in PyTorch can be used to calculate the atomic forces. Adding the calculated atomic forces to the enthalpy of formation of the molecule to obtain the molecular property of the model, combined with the description of the atomic forces, allows the trained quantum chemical model to be better applied in fields such as molecular dynamics and reactivity prediction.

[0105] Please refer to Figure 3 In some embodiments, step S30, which calculates the enthalpy of formation of the molecule based on the second descriptive parameter of the molecule and the single-center parameter in the preset semi-empirical quantum chemical model, includes steps S31 to S35, which are described in detail below.

[0106] Step S31: Calculate the single-electron matrix, two-electron matrix, and nucleo-nucleus interaction matrix based on the second descriptive parameter of the molecule and the single-center parameter in the preset semi-empirical quantum chemical model, respectively; the single-electron matrix includes the single-center single-electron matrix and the two-center single-electron matrix; the two-electron matrix includes the single-center two-electron matrix and the two-center two-electron matrix.

[0107] In some embodiments, based on the second descriptive parameter of the molecule and U in a preset semi-empirical quantum chemical model ss U pp β s β p ξ s ξ p The single-electron matrix h is obtained by calculating the single-center parameters. This single-electron matrix includes both single-center and double-center single-electron matrices; adding the single-center and double-center single-electron matrices yields the single-electron matrix h. Based on the second descriptive parameter of the molecule and the ξ in the pre-defined semi-empirical quantum chemical model... s ξ p G SS G SP H SP G PP and The two-electron matrix G is obtained by calculating the single-center parameter. This two-electron matrix includes both the single-center and two-center two-electron matrices; adding these two matrices yields the two-electron matrix G. This is based on the second descriptive parameter of the molecule and the single-center parameter G in the pre-defined semi-empirical quantum chemical model. SS The nucleus-nuclear interaction matrix was calculated. Nuclear-nuclear interaction matrix It can be obtained by a two-center two-electron integral (S A S A |S B S B Approximately: Among them, (S) A S A |S B S B Z represents a two-center, two-electron integral. A and Z B These represent the atomic numbers corresponding to atoms A and B, respectively.

[0108] In some embodiments, G SS G PP H SP G P2 The integral representing a single-center two-electron configuration is also denoted as (SS|SS), (SS|PP), (SP|SP), (PP|PP), and (PP|P,P,). Each integral represents the energy of the interaction between the electron 1 density distribution and the electron 2 density distribution, where electron 1 is the product of the first two atomic orbitals, and electron 2 is the product of the last two atomic orbitals. ξ s and ξ p Represents the orbital index, β s and βp This represents the single-electron nuclear resonance integral.

[0109] In some embodiments, the single-electron matrix h describes the motion and energy of a single electron, which can be understood as describing the state of each electron in a ground-state orbit of an atomic clock, and can also be called the single-electron wavefunction. The two-electron matrix G describes the interaction between two electrons, and can be regarded as a matrix representation of the product of the two electron wavefunctions. Therefore, both the single-electron matrix h and the two-electron matrix G are used to describe the motion of electrons.

[0110] Step S32: Perform a linear transformation on the single-electron matrix with two centers and the double-electron matrix with two centers in the single-electron matrix to obtain the single-electron matrix and the double-electron matrix in the molecular coordinate system.

[0111] Please refer to Figure 4 In some embodiments, step S32 performs a linear transformation on the dual-center single-electron matrix in the single-electron matrix and the dual-center dual-electron matrix in the dual-electron matrix to obtain the single-electron matrix and dual-electron matrix in the molecular coordinate system, including steps S321 to S322, which are described in detail below.

[0112] Step S321: Construct a local biatomic coordinate system and a molecular coordinate system; wherein, there is a mapping relationship between the local biatomic coordinate system and the molecular coordinate system and the atoms in the molecule.

[0113] In some embodiments, a local diatomic coordinate system is established, and the coordinate axes corresponding to the two atoms are defined as the x-axis, y-axis, and z-axis. In this case, the dual-center portions of the single-electron matrix h and the two-electron matrix G are described relative to the local diatomic coordinate system. A molecular coordinate system is constructed, and the coordinate axes corresponding to the atoms coinciding with the local diatomic coordinate system are defined as the x′-axis, y′-axis, and z′-axis.

[0114] Step S322: Using the pre-calculated rotation matrix, transform the double-center single-electron matrix in the single-electron matrix and the double-center double-electron matrix in the double-electron matrix in the local diatomic coordinate system to the molecular coordinate system to obtain the single-electron matrix and double-electron matrix in the molecular coordinate system.

[0115] In some embodiments, a rotation matrix is ​​calculated and used to transform the coordinates in the local diatomic coordinate system to the molecular coordinate system. This can be achieved by choosing appropriate Euler angles. The bicentric portions of the single-electron matrix h and the bielectron matrix G are linearly transformed using the rotation matrix, i.e., rotated from the local diatomic coordinate system to the molecular coordinate system. Here, the bicentric portion refers to the bicentric single-electron matrix in the single-electron matrix and the bicentric bielectron matrix in the bielectron matrix.

[0116] In some embodiments, linear transformations can be used to perform further quantum chemical calculations in molecular coordinates, such as the Hartree-Fock method or density functional theory, to obtain the electronic structure and properties of the molecule.

[0117] Step S33: Calculate the two-electron integral based on the two-electron matrix in the molecular coordinate system and the preset electron density matrix. Add the two-electron integral to the single-electron matrix in the molecular coordinate system to obtain the interaction matrix.

[0118] In some embodiments, the two-electron integral is calculated based on the two-electron matrix in molecular coordinates and a preset electron density matrix, including:

[0119]

[0120] Where (μν,λσ) represents the two-center two-electron repulsion integral in the two-electron matrix, D represents the preset electron density matrix, and G μν This represents a two-electron integral.

[0121] In some embodiments, the dual-electron integral G μν It can also be expressed as G(D), that is, the two-electron integral G(D) is calculated by the two-electron matrix G and the preset electron density matrix D in the molecular coordinate system. The electron density matrix D is a diagonal matrix composed of molecular orbital coefficients, which is used to represent the electron density on each molecular orbital.

[0122] In some embodiments, the two-electron integral G(D) and the single-electron matrix h in molecular coordinates are added together to obtain the interaction matrix F.

[0123] In some embodiments, the interaction matrix F is the Fock matrix in quantum chemistry, which describes the interactions of electrons in a given quantum state. Specifically, the Fock matrix includes the repulsive and attractive forces between electrons, as well as their self-interactions. The elements in the Fock matrix are determined by the Hamiltonian of the Schrödinger equation. By solving the Fock matrix, the wavefunction of the electron can be obtained, thus revealing the electron distribution and energy.

[0124] In some embodiments, the Hamiltonian of the Fock matrix elements, i.e., the entire Fock matrix, is specifically expanded as follows:

[0125]

[0126]

[0127]

[0128] Among them, Fμμ F μν and F μλ U represents the Hamiltonian of the Fock matrix. μμ This is a single-center, single-electron term, representing the kinetic energy of the electron in the atomic orbital μ of atom A and the attractive potential energy of the electron due to the nucleus of atom A. It is a parameterized, fixed value, V. μμ,B This is a two-center single-electron attraction term, representing the attraction between the electrons in the μν atomic orbitals of atom A and the nucleus of atom B. Its value is equal to the corresponding two-electron integral. (μμ,νν) represents the Coulomb integral, (μν,μν) represents the repulsion integral, and β μλ The integral of the two-center single-electron nuclear resonance is represented by a parameterized fixed value, and (μν,λσ) represents the two-center two-electron repulsion integral. νν D λσ D μν and D λσ These represent the electron density matrices under different integrals.

[0129] In some embodiments, the coulomb integral (μμ,νν) = g μν The repulsion integral (μν,μν)=h μν .

[0130] Step S34: Calculate the integral corresponding to the interaction matrix. Convert the integral corresponding to the interaction matrix into the initial electron energy according to matrix diagonalization. Calculate the difference between the initial electron energy and the electron energy after self-consistent field iteration to obtain the energy difference. When the energy difference meets the preset energy threshold, the electron energy after self-consistent field iteration is used as the output electron energy.

[0131] In some embodiments, the steps for calculating the integral F(D) corresponding to the interaction matrix are consistent with those for calculating the two-electron integral G(D), and will not be repeated here. The integral F(D) corresponding to the interaction matrix is ​​converted into the electron energy E by matrix diagonalization. el Calculate the electron energy E el The energy difference is calculated by comparing the energy of the electron after the self-consistent field iteration (SCF) with the energy of the electron after the self-consistent field iteration. When the energy difference meets a preset energy threshold, the energy E of the electron after the self-consistent field iteration is set. el As output electron energy.

[0132] In some embodiments, electron energy Where D is the electron density matrix, F(D) is the integral corresponding to the interaction matrix, h represents the single-electron matrix, R represents the neighboring atoms, and Tr represents the trace of the matrix. The trace of the matrix is ​​equal to the sum of the eigenvalues ​​of the matrix, that is, the sum of the elements on the main diagonal of the matrix.

[0133] In some embodiments, in quantum chemistry, a preset energy threshold, namely the energy convergence threshold for self-consistent field iteration, is used to determine whether the self-consistent field iteration has reached convergence. When the energy change after each iteration is lower than this threshold, the computation can be considered to have achieved sufficient accuracy, and the self-consistent field iteration process can be stopped. The specific energy convergence threshold is affected by factors such as the quantum chemical method used, the basis set, and the computational objective. Generally, common energy convergence thresholds are between 10^-5 and 10^-8 Hartree. 1 Hartree = 627.5094 kcal / mol. In practical applications, a relatively lenient threshold (such as 10^-5) is usually set first, and then gradually reduced until the required accuracy is achieved.

[0134] Step S35: Calculate the enthalpy of formation of the molecule based on the output electron energy and the nucleus-nucleus interaction matrix.

[0135] In some embodiments, the output electron energy E el nuclear-nuclear interaction matrix By adding the results together, we obtain the total energy E. tot ,Right now enthalpy of formation corresponding to the molecule From total energy E tot The energy E required for the ionization of the valence electrons of atoms el A and atomization enthalpy experimental values Obtaining, can be represented as Among them, the energy E required to ionize the valence electrons of the atoms involved is... el A Based on the single-center parameter U SS U PP G SS G PP H SP G PP and G P2 Calculated.

[0136] Step S40: Calculate the molecular property loss function value between the model molecular properties and the preset reference molecular properties according to the preset molecular property loss function. Use the molecular property loss function value to update the model parameters in the neural network model of the quantum chemical model to be trained, and obtain the trained quantum chemical model.

[0137] Please refer to Figure 5 In some embodiments, before step S40 calculates the molecular property loss function value between the model molecular property and the preset reference molecular property according to the preset molecular property loss function, steps S41 to S43 are also included, which will be described in detail below.

[0138] Step S41: Calculate the target generation enthalpy and the force on the target atom corresponding to the quantum chemical model to be trained using density functional theory.

[0139] In some embodiments, the target enthalpy of formation corresponding to the quantum chemical model to be trained is calculated using density functional theory (DFT). and the target atom subjected to force F DFT .

[0140] Step S42: When the preset training dataset contains molecules that meet the preset conformation conditions, the target generation enthalpy and the force on the target atom corresponding to the quantum chemical model to be trained are added together to obtain the properties of the reference molecule.

[0141] In some embodiments, when the preset training dataset contains molecules that meet the preset conformation conditions, that is, when the preset training dataset contains molecules with experimentally generated enthalpy values, the target enthalpy of generation is determined. Among them, H exp This is the experimental value of the enthalpy of formation in actual experiments. Here, conformation refers to the spatial structure of the molecule.

[0142] Step S43: When the preset training dataset does not contain molecules that meet the preset conformation conditions, extract the energy deviation corresponding to the molecule, calculate the difference between the energy deviation corresponding to the molecule and the target enthalpy of formation corresponding to the quantum chemical model to be trained, and obtain the calculated enthalpy of formation. Add the calculated enthalpy of formation and the force on the target atom to obtain the reference molecule properties.

[0143] In some embodiments, linear regression is used to extract the energy deviation of each atom from the molecule, where the energy deviation E is... bais =∑ z=1,6,7,8 p z N z Where z represents the number of neighboring atoms, p z N represents the parameter indicating the contribution of an atom to the bias. z This indicates the preset number. The difference between the energy deviation corresponding to the molecule and the target enthalpy of formation based on the quantum chemical model to be trained is calculated to obtain the calculated enthalpy of formation. Among them, E bais For energy deviation, E DFT The target enthalpy of formation is the quantum chemical model to be trained, which is calculated using density functional theory (DFT).

[0144] In some embodiments, since the generation enthalpy of the semi-empirical method and the total energy of the density functional theory method are not on the same order of magnitude, it is necessary to perform steps S41 to S43 to preprocess the total energy according to the density functional theory method.

[0145] In some embodiments, the molecular property loss function value between the model molecular properties and the preset reference molecular properties is calculated based on a preset molecular property loss function, wherein the preset molecular property loss function includes:

[0146]

[0147]

[0148] Where loss is the value of the molecular property loss function, loss penalty For the penalty loss function, N geo Represents the number of molecular conformations, E ML F represents the enthalpy of formation in the molecular properties of the model. ML This represents the forces acting on atoms in the model molecule. F represents the enthalpy of formation in the properties of the reference molecule. DFT The force on the atom in the reference molecule property is represented by w, which represents a preset training experience value, and λ is the value of the reference molecule. penalty V is the penalty coefficient, N is the number of single-center correction parameters in the molecule, and V is the number of correction parameters. i w represents the parameter value corresponding to the single-center correction parameter contained in the molecule. i This represents the empirical value corresponding to multiple training sessions for the molecule.

[0149] In some embodiments, F DFT The weight is greater than The weight.

[0150] In some embodiments, the deviation between the molecular property loss function value calculated by the preset molecular property loss function and the high-level density functional theory method is used to train the quantum chemical model to be trained, achieving chemical accuracy, thereby making properties such as enthalpy of formation, dipole moment and optimized structural parameters comparable to experimental values.

[0151] In some embodiments, to limit the deviation of parameters between the final trained quantum chemical model and the initial AM1 model, a penalty term is set in the molecular property loss function to reduce the correction of parameters in the AM1 model. Specifically, a penalty loss function (loss) is set in the molecular property loss function. penalty .

[0152] In some embodiments, the parameters in the AM1 model refer to U. ss U pp β s βp ξ s ξ p G SS G SP H SP G PP and

[0153] In some embodiments, the trained quantum chemical model combines neural networks and semi-empirical quantum chemical methods. Simultaneously, the semi-empirical AM1 quantum chemical model is implemented within the PyTorch framework, ensuring the model's interpretability and chemical essence. Back-differentiation is used to calculate atomic forces within the PyTorch framework, and the structure descriptor is calculated from scratch. Therefore, it not only boasts high computational speed but also maintains quantum chemical accuracy. Furthermore, the neural network model in this method can be adjusted by increasing or decreasing the number of hidden layers, or replaced with other pre-trained neural networks, depending on the research system.

[0154] Please refer to Figure 6 Some embodiments of the present invention provide a method for molecular property analysis, including steps S50 to S60, which are described in detail below.

[0155] Step S50: Obtain the three-dimensional coordinates of the molecule to be analyzed.

[0156] In some embodiments, the three-dimensional coordinates of the molecule to be analyzed are those of molecules in different disciplines that require molecular property analysis.

[0157] Step S60: Input the three-dimensional coordinates of the molecule to be analyzed into the trained quantum chemical model to obtain the molecular properties of the molecule to be analyzed; the molecular properties include the enthalpy of formation and / or the forces on the atoms; the quantum chemical model is trained by the quantum chemical model training method.

[0158] In some embodiments, the molecular properties of the molecule may also be charge, electronegativity, and polarity.

[0159] In some embodiments, the quantum chemical model can be obtained by the training method described in any of the embodiments herein. For example, the molecular property loss function value between the molecular properties of the model and the molecular properties of a preset reference is calculated based on the preset molecular property loss function. The molecular property loss function value is then used to update the model parameters in the neural network model of the quantum chemical model to be trained, thereby obtaining the trained quantum chemical model.

[0160] Please refer to Figures 7-9 In some embodiments, Figure 7A flowchart of the forward propagation in a machine learning-based quantum chemistry model training method is proposed. Molecular coordinates refer to the three-dimensional coordinates of the molecule, and the single-atom SOAP structure descriptor is the first descriptive parameter of the molecule. After passing through a pre-defined neural network, the second descriptive parameter, i.e., the single-center parameter, is obtained. The single-center parameter is input into the AM1 model to obtain the DFT reference molecular property, which is the loss function value of the molecular property corresponding to the model to be trained, constructed from the neural network model and the semi-empirical quantum chemistry model. The DFT reference molecular property, i.e., the pre-defined reference molecular property, is used when calculating the molecular property loss function value. Figure 8 This corresponds to the flowchart of the combination of the neural network model and the preset semi-empirical quantum chemical model in the quantum chemical model to be trained. Figure 8 The ML model is a neural network model, and the AM1 layer is a semi-empirical quantum chemical model. Figure 9 The corresponding diagram is the structure diagram of the neural network model in the quantum chemistry model to be trained. The left side shows different types of atoms and their corresponding single-center parameters, while the right side shows the specific structure diagram of the neural network model.

[0161] Please refer to Figures 10-12 In some embodiments, Figure 10 A flowchart of the forward propagation in a machine learning-based quantum chemistry model training method is proposed. Molecular coordinates represent the three-dimensional coordinates of the molecule, and the atom-pair distance thermometer is encoded as the first descriptive parameter of the molecule. After passing through a pre-defined neural network, the second descriptive parameter, namely the two-center two-electron integral, is obtained. The two-center two-electron integral is input into the AM1 model to obtain the DFT reference molecular property, which is the loss function value of the molecular property corresponding to the model to be trained, constructed from the neural network model and the semi-empirical quantum chemistry model. The DFT reference molecular property, i.e., the pre-defined reference molecular property, is used when calculating the molecular property loss function value. Figure 11 This corresponds to the flowchart of the combination of the neural network model and the preset semi-empirical quantum chemical model in the quantum chemical model to be trained. Figure 8 The ML model is a neural network model, and the AM1 layer is a semi-empirical quantum chemical model. Figure 12 The corresponding diagram is the structure diagram of the neural network model in the quantum chemical model to be trained. The left side shows different types of atom pairs and the corresponding two-center two-electron integrals of the atom pairs, while the right side shows the specific structure diagram of the neural network model.

[0162] In some embodiments, the atoms include O, H1, and H2, and the atom pair thermometer code to be calculated includes OH1, OH2, and H1H2, which are pairwise combinations of all atoms. OH1 and OH2 correspond to... Figure 12 The left side shows the HO atom pair type, with H1 and H2 corresponding to the HH atom pair type. The right side shows the number of double-center double-electron integrals for different atom pair types.

[0163] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0164] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A method of training a quantum chemistry model based on machine learning, applied to analyzing structural properties and interactions of molecules, characterized in that, include: Obtain molecules and their corresponding three-dimensional coordinates from a preset non-equilibrium structure database, and calculate the first description parameter of the molecule based on the three-dimensional coordinates of the molecule; The first descriptive parameters of the molecule are input into a preset neural network model to obtain the second descriptive parameters of the molecule; wherein: The first description parameter is the structural descriptor corresponding to each atom in the molecule, the second description parameter is the correction value of the single-center parameter corresponding to each atom in the molecule, and the neural network model is a model constructed based on the atom type corresponding to each atom in the molecule; or, the first description parameter is the thermometer code corresponding to the distance of each atom pair contained in the molecule, the second description parameter is the correction value of the two-center two-electron integral of each atom pair in the molecule, and the neural network model is a model constructed based on the atom type corresponding to each atom pair in the molecule; The enthalpy of formation of the molecule is calculated based on the second descriptive parameter of the molecule and the single-center parameter in the preset semi-empirical quantum chemical model. The enthalpy of formation of the molecule and the atomic force corresponding to the quantum chemical model to be trained are added together to obtain the molecular property of the model. The quantum chemical model to be trained is constructed based on the neural network model and the semi-empirical quantum chemical model. The molecular property loss function value between the model molecular properties and the preset reference molecular properties is calculated based on the preset molecular property loss function. The model parameters in the neural network model of the quantum chemical model to be trained are updated using the molecular property loss function value to obtain the trained quantum chemical model.

2. The method of claim 1, wherein, The step of calculating the first descriptive parameter of the molecule based on the three-dimensional coordinates corresponding to the molecule includes: The atomic density field corresponding to each atom in the molecule is calculated based on the three-dimensional coordinates of the molecule and the preset atomic density field formula. The atomic density field is expanded into an expansion formula containing a preset radial basis function, a preset spherical harmonic function, and expansion coefficients, and the values ​​corresponding to the expansion coefficients are calculated. The corresponding spectral vector is calculated using the value of the expansion coefficient and the preset spectral vector formula. A local power spectral vector is constructed based on the spectral vector. The local power spectral vector is used as the structural descriptor corresponding to the atom, and the structural descriptor corresponding to each atom in the molecule is used as the first description parameter of the molecule.

3. The method as described in claim 2, characterized in that, Expanding the atomic density field into an expansion containing a preset radial basis function, a preset spherical harmonic function, and expansion coefficients includes: Where, ρ z (r) represents the atomic density field of the current atom in the molecule, z represents the relative position coordinates between the nuclei of the neighboring atoms and the current atom, r represents the three-dimensional coordinates of the molecule, n represents the order of the preset radial basis function, l represents the angular momentum quantum number, and m represents the magnetic quantum number. G represents the expansion coefficient. n (r) represents the radial basis function, Y lm (θ,φ) represents the preset spherical harmonic function, where θ represents the polar angle and φ represents the azimuth angle.

4. The method as described in claim 1, characterized in that, The step of calculating the first descriptive parameter of the molecule based on the three-dimensional coordinates corresponding to the molecule includes: The thermometer code corresponding to the distance of each atom pair contained in the molecule is calculated according to a preset thermometer coding formula, and the thermometer code corresponding to the distance of each atom pair contained in the molecule is used as the first descriptive parameter of the molecule; wherein: The preset thermometer coding formula includes: in, This is the thermometer code corresponding to the distance between the atomic pairs formed by atoms i and j, where k is a preset parameter, and r1 represents the atomic distance between the atomic pairs calculated based on the three-dimensional coordinates r of the molecule. k For a pre-defined uniform network, σ represents the reciprocal of the pre-defined grid spacing.

5. The method as described in claim 1, characterized in that, The step of calculating the enthalpy of formation of the molecule based on the second descriptive parameter of the molecule and the single-center parameter in the preset semi-empirical quantum chemical model includes: The single-electron matrix, two-electron matrix, and nucleo-nucleo interaction matrix are calculated based on the second descriptive parameter of the molecule and the single-center parameter in the preset semi-empirical quantum chemical model, respectively; the single-electron matrix includes a single-center single-electron matrix and a two-center single-electron matrix; the two-electron matrix includes a single-center two-electron matrix and a two-center two-electron matrix; A linear transformation is performed on the dual-center single-electron matrix in the single-electron matrix and the dual-center dual-electron matrix in the dual-electron matrix to obtain the single-electron matrix and dual-electron matrix in the molecular coordinate system. The two-electron integral is calculated based on the two-electron matrix in the molecular coordinate system and the preset electron density matrix. The two-electron integral is then added to the single-electron matrix in the molecular coordinate system to obtain the interaction matrix. Calculate the integral corresponding to the interaction matrix, convert the integral corresponding to the interaction matrix into the initial electron energy according to matrix diagonalization, calculate the difference between the initial electron energy and the electron energy after self-consistent field iteration, and obtain the energy difference. When the energy difference meets the preset energy threshold, the electron energy after self-consistent field iteration is used as the output electron energy. The enthalpy of formation of the molecule is calculated based on the output electron energy and the nucleus-nucleus interaction matrix.

6. The method as described in claim 5, characterized in that, The linear transformation of the single-electron matrix and the two-electron matrix in the single-electron matrix to obtain the single-electron matrix and the two-electron matrix in the molecular coordinate system includes: Construct a local biatomic coordinate system and a molecular coordinate system; wherein, there is a mapping relationship between the local biatomic coordinate system and the molecular coordinate system and the atoms in the molecule; The double-center single-electron matrix and the double-center double-electron matrix in the single-electron matrix of the local diatomic coordinate system are transformed to the molecular coordinate system using a pre-calculated rotation matrix to obtain the single-electron matrix and double-electron matrix in the molecular coordinate system.

7. The method as described in claim 1, characterized in that, The preset molecular property loss function includes: Where loss is the value of the molecular property loss function, loss penalty For the penalty loss function, N geo Represents the number of molecular conformations, E ML F represents the enthalpy of formation in the molecular properties of the model. ML This represents the forces acting on atoms in the model molecule. F represents the enthalpy of formation in the properties of the reference molecule. DFT The force on the atom in the reference molecule property is represented by w, which represents a preset training experience value, and λ is the value of the reference molecule. penalty V is the penalty coefficient, N is the number of single-center correction parameters in the molecule, and V is the number of correction parameters. i w represents the parameter value corresponding to the single-center correction parameter contained in the molecule. i This represents the empirical value corresponding to multiple training sessions for the molecule.

8. The method as described in claim 1, characterized in that, Before calculating the molecular property loss function value between the model molecular property and the preset reference molecular property based on the preset molecular property loss function, the method further includes: The target enthalpy of formation and the forces acting on the target atom corresponding to the quantum chemical model to be trained were calculated using density functional theory. When the preset training dataset contains molecules that meet the preset conformation conditions, the target generation enthalpy and the force on the target atom corresponding to the quantum chemical model to be trained are added together to obtain the properties of the reference molecule. When the preset training dataset does not contain molecules that meet the preset conformation conditions, the energy deviation corresponding to the molecule is extracted, and the difference between the energy deviation corresponding to the molecule and the target enthalpy of formation corresponding to the quantum chemical model to be trained is calculated to obtain the calculated enthalpy of formation. The calculated enthalpy of formation and the force on the target atom are added together to obtain the reference molecule properties.

9. The method as described in claim 1, characterized in that, The unbalanced structure database includes the ANI-1 database.

10. The method as described in claim 1, characterized in that, The atomic force corresponding to the quantum chemical model to be trained is calculated according to a preset machine learning framework; wherein, the atomic force is the derivative of the formation enthalpy of the molecule with respect to the three-dimensional coordinate value of the molecule.

11. A method for molecular property analysis, characterized in that, include: Obtain the three-dimensional coordinates of the molecule to be analyzed; The three-dimensional coordinates of the molecule to be analyzed are input into a trained quantum chemical model to obtain the molecular properties of the molecule to be analyzed; the molecular properties include the enthalpy of formation and / or the forces on the atoms; the quantum chemical model is trained by the quantum chemical model training method as described in any one of claims 1 to 10.

12. A computer-readable storage medium, characterized in that, The medium stores a program that can be executed by a processor to implement the method as described in any one of claims 1-10.

Citation Information

Patent Citations

  • Method and apparatus for analysis of molecular combination based on computational estimation of electrostatic affinity using basis expansions

    US20050119835A1

  • Method of training prediction model for determining molecular binding force

    US20220130496A1