阅读说明：本技术 机器学习方法、系统及设备 (Machine learning method, system and equipment ) 是由 李金金 汪志龙 张海阔 任嘉豪 刘金云 于 2020-05-08 设计创作，主要内容包括：本发明提供一种机器学习方法、系统及设备,所述机器学习方法包括：进行第一性原理计算,以获取吸附材料与多个重金属离子相结合的最优结构；在具有最优结构的吸附材料上随机放置重金属离子,获取带有重金属离子的复合结构,并计算不同重金属离子对应的复合结构的吸附能,以构建数据集；构建任意带有重金属离子的复合结构的主吸附能预测模型；训练剩余带有重金属离子的复合结构的吸附能预测模型；统计已训练带有重金属离子的复合结构的吸附能预测模型的统计学数据,以评估吸附材料对重金属离子的吸附能力。本发明准确评估吸附材料对重金属离子的吸附能力,以此来大幅降低吸附材料的设计成本,为设计新型吸附材料来吸附重金属离子提供有力指导。(The invention provides a machine learning method, a machine learning system and machine learning equipment, wherein the machine learning method comprises the following steps: performing first principle calculation to obtain an optimal structure of the adsorption material combined with a plurality of heavy metal ions; randomly placing heavy metal ions on an adsorption material with an optimal structure to obtain a composite structure with the heavy metal ions, and calculating adsorption energy of the composite structure corresponding to different heavy metal ions to construct a data set; constructing a main adsorption energy prediction model of any composite structure with heavy metal ions; training an adsorption energy prediction model of the residual composite structure with the heavy metal ions; and (4) counting statistical data of the adsorption energy prediction model of the trained composite structure with the heavy metal ions to evaluate the adsorption capacity of the adsorption material on the heavy metal ions. The method accurately evaluates the adsorption capacity of the adsorption material on the heavy metal ions, so that the design cost of the adsorption material is greatly reduced, and powerful guidance is provided for designing a novel adsorption material to adsorb the heavy metal ions.)

1. A machine learning method is characterized by being used for predicting the adsorption capacity of different heavy metal ions on an adsorption material; the machine learning method comprises the following steps:

performing first principle calculation to obtain an optimal structure of the adsorption material combined with a plurality of heavy metal ions;

randomly placing heavy metal ions on an adsorption material with an optimal structure to obtain a composite structure with the heavy metal ions, and calculating adsorption energy of the composite structure corresponding to different heavy metal ions to construct a data set for machine learning;

constructing a main adsorption energy prediction model of any composite structure with heavy metal ions; the main adsorption energy prediction model is used for predicting adsorption energy;

training an adsorption energy prediction model of a composite structure with residual heavy metal ions according to the constructed main adsorption energy prediction model;

and (4) counting statistical data of all the adsorption energy prediction models with trained composite structures with heavy metal ions to evaluate the adsorption capacity of the adsorption material on the heavy metal ions.

2. The machine learning method of claim 1, wherein prior to the step of performing first principles calculations, the machine learning method further comprises: and establishing a structure model of the adsorption material.

3. The machine learning method of claim 2, wherein the step of performing a first principle calculation to obtain an optimal structure of the adsorption material in combination with the plurality of heavy metal ions comprises:

continuously iterating by taking energy as an optimization target according to a conjugate gradient descent mode to optimize the atomic positions of the adsorption material and the heavy metal ions; and when the energy difference between two adjacent times is smaller than a preset energy difference threshold value, setting the structure with the minimum energy as an optimal structure for combining the adsorption material and a heavy metal ion.

4. The machine learning method of claim 2, wherein the step of calculating adsorption energies of composite structures corresponding to different heavy metal ions to construct a data set for machine learning comprises:

selecting any heavy metal ion from different heavy metal ions;

according to the first preset calculated number, calculating the adsorption energy of the optimal structure corresponding to the selected heavy metal ions;

according to the second preset calculated number, calculating the adsorption energy of the optimal structure corresponding to the residual heavy metal ions; and the first preset calculated number is greater than the second preset calculated number.

5. The machine learning method of claim 4,

the adsorption energy of the optimal structure corresponding to the selected heavy metal ions is the integral energy of the optimal structure formed by combining the selected heavy metal ions and the adsorption material, the energy of the adsorption material and the energy of the selected heavy metal ions;

the adsorption energy of the optimal structure corresponding to the residual heavy metal ions is the overall energy of the optimal structure, namely the energy of the adsorption material, of the residual heavy metal ions, and the energy of one heavy metal ion in the residual heavy metal ions.

6. The machine learning method of claim 4, wherein the constructing of the main predictive model of adsorption energy of any composite structure with heavy metal ions is constructing of a predictive model of adsorption energy of a selected composite structure with heavy metal ions:

calculating the local chemical environment of each atom according to the constructed data set, and forming the local chemical environments of all atoms into a matrix so as to construct the input of a main adsorption energy prediction model;

and performing matrix operation on the input of the main adsorption energy prediction model and the sub-networks corresponding to each atom to obtain the contribution values of the sub-networks, summarizing the contribution values of all the sub-networks into adsorption energy to serve as the output of the main adsorption energy prediction model, and obtaining the trained model parameters.

7. The machine learning method of claim 6, wherein the step of training the adsorption energy prediction model of the remaining complex structure with heavy metal ions based on the constructed main adsorption energy prediction model comprises introducing transfer learning to the constructed main adsorption energy prediction model; the method comprises the following steps:

initializing an adsorption energy prediction model of the residual composite structure with the heavy metal ions by using the trained model parameters;

and adjusting the model parameters of the adsorption energy prediction model of the residual composite structure with the heavy metal ions.

8. The machine learning method according to claim 6, further comprising using any one of the trained adsorption energy prediction models as a general evaluation framework to predict the adsorption capacity of the adsorption material on any heavy metal ions through first-nature principle calculation, adsorption energy prediction model training, transfer learning and energy average statistics.

9. A machine learning system is characterized by being used for predicting the adsorption capacity of different heavy metal ions on an adsorption material; the machine learning system includes:

the calculation module is used for performing first-principle calculation to obtain an optimal structure of the adsorption material combined with the heavy metal ions;

the first construction module is used for randomly placing heavy metal ions on the adsorption material with the optimal structure, acquiring a composite structure with the heavy metal ions, and calculating adsorption energy of the composite structure corresponding to different heavy metal ions to construct a data set for machine learning;

the second construction module is used for constructing a main adsorption energy prediction model of any composite structure with heavy metal ions; the main adsorption energy prediction model is used for predicting adsorption energy;

the training module is used for training the residual adsorption energy prediction model with the heavy metal ion composite structure according to the constructed main adsorption energy prediction model;

and the statistical module is used for counting the statistical data of the adsorption energy prediction models of all the trained composite structures with the heavy metal ions so as to evaluate the adsorption capacity of the adsorption material on the heavy metal ions.

10. A machine learning device, comprising: a processor and a memory;

the memory is configured to store a computer program, and the processor is configured to execute the computer program stored by the memory to cause the machine learning device to perform the machine learning method according to any one of claims 1 to 8.

Technical Field

The invention belongs to the technical field of environmental pollution assessment, and particularly relates to a machine learning method, a machine learning system and machine learning equipment.

Background

Heavy metal pollution is an extremely serious environmental problem, and waste water, sludge and the like containing heavy metal ions cause serious harm to human survival and physical and psychological health through soil and air, especially a food chain. In recent years, the adsorption technology has been developed significantly, and the nano material has important application prospect in heavy metal wastewater treatment due to the characteristics of high adsorption capacity, simple separation, high cyclic regeneration efficiency and the like. The adsorbing material has various active components such as pore diameter, groups and the like, and achieves the purpose of adsorbing heavy metal ions by forming ionic bonds or covalent bonds with the adsorbed metal ions. Therefore, after the organic modification is carried out on the adsorption material, the active sites of the adsorption material are further increased, and a better adsorbent can be designed. Finding or developing an ideal adsorbing material with a strong adsorbing effect on heavy metal ions becomes a current research hotspot and difficulty.

The existing research technology is mainly divided into experimental measurement and theoretical calculation. In order to develop a material with strong heavy metal ion adsorption capacity, researchers experimentally synthesize a novel candidate material under specific conditions, prepare a solution of heavy metal ions, and judge the adsorption capacity of the candidate material by measuring the concentrations of the heavy metal ions before and after adsorption. The technology can evaluate the adsorption capacity of an adsorption material and has the defects of long time consumption, high cost and high difficulty. In order to reduce trial and error costs, more and more researchers are beginning to assist in the design of new materials at a previous stage by means of theoretical calculations.

In particular, the prior art comprises the following drawbacks:

firstly, the experiment cost is high; the prior art requires expensive chemical material costs in preparing the adsorbing material, measuring the amount of adsorption, and the like, using experiments.

Secondly, the efficiency is low; the method has complicated steps of experimental preparation, synthesis, measurement and the like, and generally requires precision instruments such as XPS and the like for operation, so the experimental period is long.

Third, the recognition of the adsorption sites is inadequate; the specific adsorption sites of heavy metal ions on the adsorption material are difficult to measure experimentally, and theoretically, only the optimal adsorption sites are researched, so that the overall adsorption capacity of the adsorption material is lack of understanding.

Fourthly, theoretical research on the adsorption energy of any site requires a large amount of calculation based on the first principle, and is long in time consumption and high in cost.

Fifth, the generalization ability is poor; in the prior art, research can be usually carried out only on certain heavy metal ions, an ideal adsorption material is designed, and the applicability to other heavy metal ions is not strong.

Therefore, how to provide a machine learning method, system and device to solve the defects of high cost, low efficiency, large energy consumption, inconvenient use, poor generalization capability and the like in the prior art becomes a technical problem to be solved urgently by the technical staff in the field.

Disclosure of Invention

In view of the above disadvantages of the prior art, an object of the present invention is to provide a machine learning method, system and device, which are used to solve the problems of high cost, low efficiency, high energy consumption, inconvenient use and poor generalization capability in the prior art.

In order to achieve the above objects and other related objects, the present invention provides a machine learning method for predicting adsorption capacities of different heavy metal ions on an adsorption material; the machine learning method comprises the following steps: performing first principle calculation to obtain an optimal structure of the adsorption material combined with a plurality of heavy metal ions; randomly placing heavy metal ions on an adsorption material with an optimal structure to obtain a composite structure with the heavy metal ions, and calculating adsorption energy of the composite structure corresponding to different heavy metal ions to construct a data set for machine learning; constructing a main adsorption energy prediction model of any composite structure with heavy metal ions; the main adsorption energy prediction model is used for predicting adsorption energy; training an adsorption energy prediction model of a composite structure with residual heavy metal ions according to the constructed main adsorption energy prediction model; and (4) counting statistical data of all the adsorption energy prediction models with trained composite structures with heavy metal ions to evaluate the adsorption capacity of the adsorption material on the heavy metal ions.

In an embodiment of the invention, before the step of performing the first principle calculation, the machine learning method further includes: and establishing a structure model of the adsorption material.

In an embodiment of the present invention, the step of performing the first principle calculation to obtain the optimal structure of the adsorbing material combined with the plurality of heavy metal ions includes: continuously iterating by taking energy as an optimization target according to a conjugate gradient descent mode to optimize the atomic positions of the adsorption material and the heavy metal ions; and when the energy difference between two adjacent times is smaller than a preset energy difference threshold value, setting the structure with the minimum energy as an optimal structure for combining the adsorption material and a heavy metal ion.

In an embodiment of the invention, the step of calculating the adsorption energy of the composite structure corresponding to different heavy metal ions to construct the data set for machine learning includes: selecting any heavy metal ion from different heavy metal ions; according to the first preset calculated number, calculating the adsorption energy of the optimal structure corresponding to the selected heavy metal ions; according to the second preset calculated number, calculating the adsorption energy of the optimal structure corresponding to the residual heavy metal ions; and the first preset calculated number is greater than the second preset calculated number.

In an embodiment of the present invention, the adsorption energy of the optimal structure corresponding to the selected heavy metal ion is the overall energy of the optimal structure formed by combining the selected heavy metal ion and the adsorption material-the energy of the selected heavy metal ion; the adsorption energy of the optimal structure corresponding to the residual heavy metal ions is the overall energy of the optimal structure, namely the energy of the adsorption material, of the residual heavy metal ions, and the energy of one heavy metal ion in the residual heavy metal ions.

In an embodiment of the present invention, the constructing of the main adsorption energy prediction model of any composite structure with heavy metal ions is constructing of an adsorption energy prediction model of a selected composite structure with heavy metal ions: calculating the local chemical environment of each atom according to the constructed data set, and forming the local chemical environments of all atoms into a matrix so as to construct the input of a main adsorption energy prediction model; and performing matrix operation on the input of the main adsorption energy prediction model and the sub-networks corresponding to each atom to obtain the contribution values of the sub-networks, summarizing the contribution values of all the sub-networks into adsorption energy to serve as the output of the main adsorption energy prediction model, and obtaining the trained model parameters.

In an embodiment of the invention, the step of training the residual adsorption energy prediction model with the composite structure of heavy metal ions according to the constructed main adsorption energy prediction model includes introducing transfer learning to the constructed main adsorption energy prediction model; the method comprises the following steps: initializing an adsorption energy prediction model of the residual composite structure with the heavy metal ions by using the trained model parameters; and adjusting the model parameters of the residual adsorption energy prediction model with the composite structure with the heavy metal ions until the model parameters of the two are the same.

In an embodiment of the invention, the machine learning method further includes using any one of the trained adsorption energy prediction models as a general evaluation framework, and predicting the adsorption capacity of the adsorption material on any heavy metal ion by calculating, training, transferring learning and counting an energy average value according to a first principle.

The invention provides a machine learning system for predicting the adsorption capacity of different heavy metal ions on an adsorption material; the machine learning system includes: the calculation module is used for performing first-principle calculation to obtain an optimal structure of the adsorption material combined with the heavy metal ions; the first construction module is used for randomly placing heavy metal ions on the adsorption material with the optimal structure, acquiring a composite structure with the heavy metal ions, and calculating adsorption energy of the composite structure corresponding to different heavy metal ions to construct a data set for machine learning; the second construction module is used for constructing a main adsorption energy prediction model of any composite structure with heavy metal ions; the main adsorption energy prediction model is used for predicting adsorption energy; the training module is used for training the residual adsorption energy prediction model with the heavy metal ion composite structure according to the constructed main adsorption energy prediction model; and the statistical module is used for counting the statistical data of the adsorption energy prediction models of all the trained composite structures with the heavy metal ions so as to evaluate the adsorption capacity of the adsorption material on the heavy metal ions.

A final aspect of the present invention provides a machine learning apparatus including: a processor and a memory; the memory is used for storing a computer program, and the processor is used for executing the computer program stored by the memory to enable the machine learning device to execute the machine learning method.

As described above, the machine learning method, system and device according to the present invention have the following advantages:

firstly, the method avoids complicated and expensive experimental links, is completely based on theoretical calculation, reduces the cost, reduces the dependence on the calculation of a first principle, and shortens the research period.

Secondly, a high-precision and high-efficiency adsorption energy prediction model is established through a machine learning method, accurate prediction is carried out on any adsorption site of a certain heavy metal ion on the adsorption material, and theoretical guidance is improved for designing a novel adsorption material.

Thirdly, the invention combines the transfer learning method, so that the prediction model is suitable for other heavy metal ions and has strong generalization capability.

Fourthly, the model constructed by the method is simple and can be further widely applied in the industry.

Drawings

Fig. 1 is a flowchart illustrating a machine learning method according to an embodiment of the invention.

FIG. 2A shows g-C of the present invention₃N₄A plan view of an optimal structure combined with three heavy metal ions of Pb (II), Hg (II) and Cd (II).

FIG. 2B shows g-C of the present invention₃N₄Oblique view of the optimal structure combined with three heavy metal ions of Pb (II), Hg (II) and Cd (II).

FIG. 3 shows Pb (II)/g-C constructed based on local chemical environment according to the present invention₃N₄Schematic diagram of the main adsorption energy prediction model.

FIG. 4 shows the present inventionClear utilization of a transfer learning algorithm to obtain Hg (II)/g-C₃N₄And Cd (II)/g-C₃N₄Schematic diagram of a model framework for predicting adsorption energy.

Fig. 5 is a schematic structural diagram of a machine learning system according to an embodiment of the invention.

Description of the element reference numerals

5	Machine learning system
		51	First building Block
52	Computing module
		53	Second building Block
54	Training module
		55	Statistical module
56	Prediction module
		S11～S18	Step (ii) of

Detailed Description

The embodiments of the present invention are described below with reference to specific embodiments, and other advantages and effects of the present invention will be easily understood by those skilled in the art from the disclosure of the present specification. The invention is capable of other and different embodiments and of being practiced or of being carried out in various ways, and its several details are capable of modification in various respects, all without departing from the spirit and scope of the present invention. It is to be noted that the features in the following embodiments and examples may be combined with each other without conflict.

It should be noted that the drawings provided in the following embodiments are only for illustrating the basic idea of the present invention, and the components related to the present invention are only shown in the drawings rather than drawn according to the number, shape and size of the components in actual implementation, and the type, quantity and proportion of the components in actual implementation may be changed freely, and the layout of the components may be more complicated.

The technical principle of the machine learning method, the system and the equipment is as follows:

the present invention uses an adsorbent material (e.g., g-C)₃N₄Adsorbing material) as a research object, on the basis of first-principle calculation, investigating the adsorption strength relationship of different sites of heavy metal ions, and establishing the adsorbing material (for example, g-C) by combining with a deep neural network₃N₄Adsorbent material) for a certain heavy metal ion, and introducing transfer learning into the prediction of adsorption energy to obtain adsorbent material (e.g., g-C)₃N₄Adsorption material) to other heavy metal ions. Fusion of big data statistics to accurately evaluate adsorbent materials (e.g., g-C)₃N₄Adsorbent material) adsorption capacity for heavy metal ions. Therefore, the design cost of the adsorption material is greatly reduced, and powerful guidance is provided for designing a novel adsorption material to adsorb heavy metal ions. The invention is not only applicable to adsorbent materials (e.g., g-C)₃N₄Adsorption material), also be applicable to the adsorption capacity prediction of other adsorption materials to different heavy metal ions.

Example one

The embodiment provides a machine learning method for predicting the adsorption capacity of different heavy metal ions on an adsorption material; the machine learning method comprises the following steps:

performing first principle calculation to obtain an optimal structure of the adsorption material combined with any heavy metal ion;

randomly placing heavy metal ions on an adsorption material with an optimal structure to obtain a composite structure with the heavy metal ions, and calculating adsorption energy of the composite structure corresponding to different heavy metal ions to construct a data set for machine learning;

constructing a main adsorption energy prediction model of any composite structure with heavy metal ions; the main adsorption energy prediction model is used for predicting adsorption energy;

training an adsorption energy prediction model of a composite structure with residual heavy metal ions according to the constructed main adsorption energy prediction model;

and (4) counting statistical data of all the adsorption energy prediction models with trained composite structures with heavy metal ions to evaluate the adsorption capacity of the adsorption material on the heavy metal ions.

The machine learning method provided by the present embodiment will be described in detail below with reference to the drawings. The machine learning method is used for predicting the adsorption capacity of different heavy metal ions on the adsorption material. In practical application, the adsorbing material can adopt g-C_xN_yE.g. Si₃N₄、Ge₃N₄、C₃P₄、Si₃P₄、MoS₂、C₂N₂、g-C₃N₄(C₃N₄Having a 5-phase structure in which the g-phase is g-C₃N₄A typical polymer semiconductor) or the like.

Please refer to fig. 1, which is a flowchart illustrating a machine learning method according to an embodiment. As shown in fig. 1, the machine learning method specifically includes the following steps:

s11, establishing the g-C₃N₄The structural model of (1).

In particular, g-C of a parallelogram is established₃N₄And (5) a structural model.

S12, performing first principle calculation to obtain g-C₃N₄The optimal structure combined with three heavy metal ions. In this embodiment, the three heavy metal ions are pb (ii), hg (ii), cd (ii), respectively. Please refer to fig. 2A and 2B, which are g-C₃N₄A plan view and an oblique view of the optimal structure combined with three heavy metal ions of Pb (II), Hg (II) and Cd (II).

In this embodiment, the first sexual principle refers to an algorithm for directly solving the schrodinger equation after approximation processing according to the principle of interaction between atomic nuclei and electrons and the basic motion law thereof and by applying the quantum mechanics principle and starting from specific requirements. Often used to calculate ground state properties of material systems, etc.

Specifically, step S11 is to continuously iterate according to a conjugate gradient descent manner with energy as an optimization target to optimize atomic positions of the adsorption material and the multiple heavy metal ions; when the difference between the energies of two adjacent times is less than a preset energy difference threshold (for example, less than 1e-6eV), the structure with the smallest energy is set as the optimal structure for the compound semiconductor to bind with a heavy metal ion.

In this embodiment, the conjugated gradient descent method means that in the optimization process, the structures of the adsorption material and the heavy metal ions, i.e. their atomic xyz coordinates, are used as input, so that their energies can be calculated. The derivative of the energy with respect to the xyz coordinate yields the force of each atom in the xyz direction, referred to as the atomic force. According to the stress, the atoms move, and new coordinates, namely new structures, are generated after the atoms move. The optimization stops until the energy difference between the two structures is less than a threshold.

In this embodiment, in the process of optimizing the structure by using the conjugate Gradient descent algorithm, the Generalized Gradient Approximation (GGA) which is mainstream at present (which belongs to a class of Density Functional Theory, Density Functional Theory (DFT) is a quantum mechanical method for researching the electronic structure of a multi-electron system), the Density Functional Theory has wide application in physics and chemistry, and is particularly used for researching the properties of molecules and condensed states, and is one of the most common methods in the fields of condensed state physical computing materials science and computational chemistry), the gradient of the density function is added on the basis of the original functional, the gradient of the density function can be understood as the change of electron kinetic energy or electron density along with space, and the Perew-Berke-Ernzerhof (PBE) exchange correlation functional in GGA can predict more accurate system energy, binding energy and activation energy. When the electronic state of the atom in the calculation system is processed, namely ion-electron interaction energy is calculated, a virtual potential, namely a pseudo potential, is introduced when the numerical value calculation is carried out on the energy band structure, and the pseudo potential is selected to project the extended wave projected extended-wave (PAW).

S13, in g-C with optimal structure₃N₄The heavy metal ions are randomly arranged (in this embodiment, pb (ii), hg (ii), cd (ii)) to obtain a composite structure with the heavy metal ions, and the adsorption energy of the composite structure corresponding to different heavy metal ions is calculated to construct a data set for machine learning. In this embodiment, the data set includes g-C of optimal structure₃N₄And corresponding adsorption energy.

Specifically, the S13 includes the following steps:

selecting any heavy metal ion from different heavy metal ions.

In this embodiment, Pb (II) is selected

And calculating the adsorption energy of the optimal structure corresponding to the selected heavy metal ions according to the first preset calculated number.

In this embodiment, the optimal structure corresponding to the selected heavy metal ion is Pb (II)/g-C₃N₄。

The number of the first preset calculation is set to 7000-10000, and the 7000-10000 structures can ensure a huge machine learning data set, so that the final prediction precision is higher.

The adsorption energy of the optimal structure corresponding to the selected heavy metal ions is the overall energy of the optimal structure formed by combining the selected heavy metal ions and the adsorption material-the energy of the selected heavy metal ions,

the calculation formula is specifically as follows:

ΔE＝E_sub+met-E_sub-E_met

wherein, Delta E is the adsorption energy of the optimal structure corresponding to the selected heavy metal ions, E_sub+metFor the overall energy of the optimum structure of the combination of the selected heavy metal ions and the adsorption material, E_subEnergy of adsorption material, E_metIs the energy of the selected heavy metal ions.

According to the second preset calculated number, calculating the adsorption energy of the optimal structure corresponding to the residual heavy metal ions; and the first preset calculated number is greater than the second preset calculated number.

In this example, the remaining heavy metal ions are Hg (II) and Cd (II).

The adsorption energy of the optimal structure corresponding to the residual heavy metal ions is the overall energy of the optimal structure, namely the energy of the adsorption material, of the residual heavy metal ions, and the energy of one heavy metal ion in the residual heavy metal ions.

S14, constructing a main adsorption energy prediction model of any composite structure with heavy metal ions; the main adsorption energy prediction model is used for predicting adsorption energy.

In this example, Pb (II)/g-C was constructed₃N₄The main adsorption energy prediction model.

Specifically, the S14 includes:

and calculating the local chemical environment of each atom according to the constructed data set, and forming the local chemical environments of all atoms into a matrix so as to construct the input of a main adsorption energy prediction model. In this embodiment, the local chemical environment of the atom is selected as the structural feature, which can well satisfy the translational invariance, rotational invariance, and displacement invariance in the periodic material.

Wherein, the calculation process of calculating the local chemical environment of each atom is as follows:

local chemical environment G of ith atom and its neighbor atom j_ijExpressed as:

e(x)＝x/||x||

the subscript in the above formula represents the difference, x_ijDenotes the difference between the X coordinates of the ith atom and the jth atom, a_(i)And b_(i)Denotes the two atoms closest to the ith atom, R_iaDenotes the difference between the coordinates of the i-th atom and the coordinate vector of the a-atom, R_ibRepresenting the difference between the coordinates of the ith atom and the coordinate vector of the b atom. The calculated cutoff distance between neighboring atoms is set toTwo atoms are at a distance of less thanIs considered a neighbor atom; r_ijIs the difference between the coordinate vectors of the ith atom and the jth atom, i.e., (x)_ij,y_ij,z_ij) Can also be written as (x)_i-x_j,y_i-y_j,z_i-z_j). Where j is a neighbor atom of i, i.e., an atom that is less than a threshold distance from i; r_ia(i)Where a (i) represents the atom nearest to the i-th atom; r_ib(i)Where b (i) represents the atom second closest to the ith atom; e (X) represents a direction vector of the vector X. I X I is (X)²+y²+z²) And 0.5 is the length of the vector. e (X) ═ X | |, y/| | X |, z/| | X | | |) is referred to as"rotation matrix".

And performing matrix operation on the input of the main adsorption energy prediction model and the sub-networks corresponding to each atom to obtain the contribution values of the sub-networks, summarizing the contribution values of all the sub-networks into adsorption energy to serve as the output of the main adsorption energy prediction model, and obtaining the trained model parameters. FIG. 3 shows Pb (II)/g-C constructed based on local chemical environment₃N₄Schematic diagram of the main adsorption energy prediction model. As shown in FIG. 3, for Pb (II)/g-C₃N₄There are 3 sub-networks, Pb, C, N respectively. The neural network has a structure of 3-5 hidden layers, each layer comprises 10-50 neurons, iteration is carried out by an Adam optimizer (an optimization algorithm used for solving an optimization problem) in the training process, 32-128 are selected as batch training units, the initial learning rate is 0.002-0.003 in the training process, and namely the step length of each parameter updating is realized. Before training, the parameter values represented by the neurons in the neural network are initialized randomly, and then the parameters of the neurons are adjusted according to the output-input correspondence, namely the training process. After a neural network is trained, a model parameter is obtained, for a new structure, after a matrix of a local chemical environment is calculated, the matrix and the model parameter are used for matrix operation, and then the adsorption energy can be predicted, wherein the process is very fast and is approximately 10000 times faster than that of DFT calculation.

And S15, training the residual adsorption energy prediction model with the heavy metal ion composite structure according to the constructed main adsorption energy prediction model.

In this embodiment, the S15 includes:

and initializing an adsorption energy prediction model of the residual composite structure with the heavy metal ions by using the trained model parameters.

And adjusting the model parameters of the model according to the adsorption energy prediction of the residual composite structure with the heavy metal ions until the model converges or the error RMSE of the model is less than 0.1.

Due to Hg (II)/g-C₃N₄And Cd (II)/g-C₃N₄Has small data quantity, adopts a transfer learning algorithm and utilizes the trained dataPb(II)/g-C₃N₄The model parameters of the main adsorption energy prediction model of (1) initialize Hg (II)/g-C₃N₄And Cd (II)/g-C₃N₄The model parameters of (1). To Hg (II)/g-C₃N₄And Cd (II)/g-C₃N₄The model parameters are finely adjusted, in order to ensure that the parameter oscillation amplitude is very small and achieve the purpose of fine adjustment, the learning rate is adjusted to be 0.001-0.0015, Hg (II)/g-C₃N₄And Cd (II)/g-C₃N₄The model parameters of (1) are the same as those of Pb (II)/g-C3N4, such as the number of neurons, and then the two models are trained. In this embodiment, migration learning is used in the adsorption energy prediction model, so that the problems of prediction inaccuracy, overfitting and the like caused by small data volume in the machine learning prediction of material properties can be solved.

FIG. 4 shows Hg (II)/g-C obtained by the transfer learning algorithm₃N₄And Cd (II)/g-C₃N₄Schematic diagram of a model framework for predicting adsorption energy.

S16, performing S14 and S15 in a loop such that the predicted root mean square error RMSE of the adsorption energy by the adsorption energy prediction model of the complex structure with the heavy metal ion is <0.1eV, i.e., the root mean square error of the predicted adsorption energy and the actual adsorption energy:

wherein M is the number of energies, also equals to the number of structures, corresponding to 7000-10000 in Pb; e_NN,iIs the energy of the ith structure predicted by the neural network, i.e., the ith predicted energy. E_DFT,iIs the energy of the ith structure, i.e., the ith calculated energy, calculated using the DFT (i.e., the first principle). E^RMSEThe calculation is the root mean square error of the two.

And S17, counting the statistical data of the adsorption energy prediction models of all the trained composite structures with the heavy metal ions to evaluate the adsorption capacity of the adsorption material for the heavy metal ions.

Specifically, the well-trained Hg (II)/g-C is utilized₃N₄And Cd (II)/g-C₃N₄The adsorption energy prediction model further predicts 7000-10000 corresponding structures, so that the adsorption energy numbers of the three systems are the same, and the minimum value, the maximum value, the standard deviation and the average value of the energy of the three systems are calculated and analyzed by utilizing statistical knowledge at the moment. The average value is stable and reliable, all energy data can be comprehensively considered, and the contained information is the most. Therefore, the average value of the energy is selected to measure the adsorption capacity.

S18, taking any one of the trained adsorption energy prediction models as an evaluation universal framework, and adopting the steps S11-S16 to predict the adsorption capacity of the adsorption material to any heavy metal ions.

For example, for a new heavy metal ion such as Cu, only a small amount of DFT calculation may be performed, and then the adsorption energy prediction model parameter of Pb is used to perform migration learning, so as to obtain an adsorption energy prediction model of Cu, and then 7000-10000 adsorption energies of Cu are rapidly predicted, and the average value is taken to obtain the overall adsorption capacity of the adsorption material to Cu.

The machine learning method has the following beneficial effects:

first, the machine learning method of this embodiment avoids tedious and expensive experimental links, is completely based on theoretical calculation, reduces cost, reduces the dependence on first-nature principle calculation, and shortens the research period.

Secondly, a high-precision and high-efficiency adsorption energy prediction model is established through a machine learning method, accurate prediction is carried out on any adsorption site of a certain heavy metal ion on the adsorption material, and theoretical guidance is improved for designing a novel adsorption material.

Thirdly, the machine learning method described in this embodiment is combined with a transfer learning method, so that the prediction model is suitable for other heavy metal ions, and the model has a strong generalization capability.

Fourth, the model constructed by the machine learning method of the present embodiment is simple and can be further widely applied in the industry.

Example two

The embodiment provides a machine learning system, which is used for predicting the adsorption capacity of different heavy metal ions on an adsorption material; the machine learning system includes:

the calculation module is used for performing first-principle calculation to obtain an optimal structure of the adsorption material combined with the heavy metal ions;

the first construction module is used for randomly arranging heavy metal ions on the adsorption material with the optimal structure, acquiring a composite structure with the heavy metal ions, and calculating adsorption energy of the composite structure corresponding to different heavy metal ions to construct a data set for machine learning;

the second construction module is used for constructing a main adsorption energy prediction model of any composite structure with heavy metal ions; the main adsorption energy prediction model is used for predicting adsorption energy;

the training module is used for training the residual adsorption energy prediction model with the heavy metal ion composite structure according to the constructed main adsorption energy prediction model;

and the statistical module is used for counting the statistical data of the adsorption energy prediction models of all the trained composite structures with the heavy metal ions so as to evaluate the adsorption capacity of the adsorption material on the heavy metal ions.

The machine learning system provided by the present embodiment will be described in detail below with reference to the drawings. Please refer to fig. 5, which is a schematic structural diagram of a machine learning system in an embodiment. As shown in fig. 5, the machine learning system 5 includes a first constructing module 51, a calculating module 52, a second constructing module 53, a training module 54, a statistical module 55, and a predicting module 56.

The first building block 51 is used for building the g-C₃N₄The structural model of (1).

In particular, said first building block 51 establishes the g-C of a parallelogram₃N₄And (5) a structural model.

The calculation module 52 is configured to perform a first principle calculation to obtain g-C₃N₄The optimal structure combined with three heavy metal ions. In this embodiment, the three heavy metal ions are Pb (II) and Hg, respectively(II), Cd (II) and three ions. Please refer to fig. 2A and 2B, which are g-C₃N₄Plan view and oblique view of optimal structure combined with Pb (II), Hg (II) and Cd (II) heavy metal ions respectively

In this embodiment, the first sexual principle refers to an algorithm for directly solving the schrodinger equation after approximation processing according to the principle of interaction between atomic nuclei and electrons and the basic motion law thereof and by applying the quantum mechanics principle and starting from specific requirements. Often used to calculate ground state properties of material systems, etc.

Specifically, the calculation module 52 continuously iterates in a conjugate gradient descent manner with energy as an optimization target to optimize atomic positions of the adsorption material and the multiple heavy metal ions; when the difference between the energies of two adjacent times is less than a preset energy difference threshold (for example, less than 1e-6eV), the structure with the smallest energy is set as the optimal structure for the compound semiconductor to bind with a heavy metal ion.

The first building block 51 is also used for g-C with optimal structure₃N₄The heavy metal ions are randomly arranged (in this embodiment, pb (ii), hg (ii), cd (ii)) to obtain a composite structure with the heavy metal ions, and the adsorption energy of the composite structure corresponding to different heavy metal ions is calculated to construct a data set for machine learning. In this embodiment, the data set includes g-C of optimal structure₃N₄And corresponding adsorption energy.

Specifically, the first building module 51 selects any heavy metal ion from different heavy metal ions, and calculates the adsorption energy of the optimal structure corresponding to the selected heavy metal ion according to a first preset calculated number; according to the second preset calculated number, calculating the adsorption energy of the optimal structure corresponding to the residual heavy metal ions; and the first preset calculated number is greater than the second preset calculated number.

The adsorption energy of the optimal structure corresponding to the selected heavy metal ions is the overall energy of the optimal structure formed by combining the selected heavy metal ions and the adsorption material-the energy of the selected heavy metal ions,

the calculation formula is specifically as follows:

ΔE＝E_sub+met-E_sub-E_met

wherein, Delta E is the adsorption energy of the optimal structure corresponding to the selected heavy metal ions, E_sub+metFor the overall energy of the optimum structure of the combination of the selected heavy metal ions and the adsorption material, E_subEnergy of adsorption material, E_metIs the energy of the selected heavy metal ions.

The adsorption energy of the optimal structure corresponding to the residual heavy metal ions is the overall energy of the optimal structure, namely the energy of the adsorption material, of the residual heavy metal ions, and the energy of one heavy metal ion in the residual heavy metal ions.

The second construction module 53 is configured to construct a main adsorption energy prediction model of any composite structure with heavy metal ions; the main adsorption energy prediction model is used for predicting adsorption energy.

Specifically, the second building module 53 calculates the local chemical environment of each atom according to the built data set, and forms the local chemical environments of all atoms into a matrix to build the input of the main adsorption energy prediction model. In this embodiment, the local chemical environment of the atom is selected as the structural feature, which can well satisfy the translational invariance, rotational invariance, and displacement invariance in the periodic material.

Wherein, the calculation process of calculating the local chemical environment of each atom is as follows:

local chemical environment G of ith atom and its neighbor atom j_ijExpressed as:

e(x)＝x/||x||

the subscript in the above formula represents the difference, x_ijDenotes the difference between the X coordinates of the ith atom and the jth atom, a_(i)And b_(i)Denotes the two atoms closest to the ith atom, R_iaDenotes the difference between the coordinates of the i-th atom and the coordinate vector of the a-atom, R_ibRepresenting the difference between the coordinates of the ith atom and the coordinate vector of the b atom. The calculated cutoff distance between neighboring atoms is set toTwo atoms are at a distance of less thanIs considered a neighbor atom; (ii) a R_ijIs the difference between the coordinate vectors of the ith atom and the jth atom, i.e., (x)_ij,y_ij,z_ij) Can also be written as (x)_i-x_j,y_i-y_j,z_i-z_j). Where j is a neighbor atom of i, i.e., an atom that is less than a threshold distance from i; r_ia(i)Where a (i) represents the atom nearest to the i-th atom; r_ib(i)Where b (i) represents the atom second closest to the ith atom; e (X) represents a direction vector of the vector X. I X I is (X)²+y²+z²) And 0.5 is the length of the vector. The term "rotation matrix" refers to (X/| X |, y/| X |, z/| X |).

And performing matrix operation on the input of the main adsorption energy prediction model and the sub-networks corresponding to each atom to obtain the contribution values of the sub-networks, summarizing the contribution values of all the sub-networks into adsorption energy to serve as the output of the main adsorption energy prediction model, and obtaining the trained model parameters.

The training module 54 coupled to the first building module 51 and the second building module 53 is configured to train the adsorption energy prediction model of the composite structure with the remaining heavy metal ions according to the constructed main adsorption energy prediction model.

In this embodiment, the training module 54 is configured to initialize the adsorption energy prediction model of the remaining composite structure with the heavy metal ions by using the trained model parameters. And adjusting the model parameters of the residual adsorption energy prediction model with the composite structure with the heavy metal ions until the model parameters of the two are the same.

The statistical module 55 coupled to the training module 54 is configured to invoke the first building module 51 and the second building module 52 in a loop, so that the root mean square error RMSE predicted by the adsorption energy of the model for predicting the adsorption energy of the complex structure with heavy metal ions is less than 0.1eV, i.e. the root mean square error between the predicted adsorption energy and the actual adsorption energy:

the statistical module 55 is further configured to count statistical data of all the trained prediction models of adsorption energy of the composite structure with heavy metal ions, so as to evaluate the adsorption capacity of the adsorption material for heavy metal ions.

Specifically, the statistics module 55 utilizes trained Hg (II)/g-C₃N₄And Cd (II)/g-C₃N₄The adsorption energy prediction model further predicts 7000-10000 corresponding structures, so that the adsorption energy numbers of the three systems are the same, and the minimum value, the maximum value, the standard deviation and the average value of the energy of the three systems are calculated and analyzed by utilizing statistical knowledge at the moment. The average value is stable and reliable, all energy data can be comprehensively considered, and the contained information is the most. Therefore, the average value of the energy is selected to measure the adsorption capacity.

The prediction module 56 coupled to the first construction module 51, the calculation module 52, the second construction module 53, the training module 54, and the statistical module 55 is configured to use any one of the trained adsorption energy prediction models as a general evaluation framework, and call the first construction module 51, the calculation module 52, the second construction module 53, the training module 54, and the statistical module 55 to predict the adsorption capacity of the adsorption material to any heavy metal ion.

It should be noted that the division of the modules of the above system is only a logical division, and the actual implementation may be wholly or partially integrated into one physical entity, or may be physically separated. And the modules can be realized in a form that all software is called by the processing element, or in a form that all the modules are realized in a form that all the modules are called by the processing element, or in a form that part of the modules are called by the hardware. For example: the x module can be a separately established processing element, and can also be integrated in a certain chip of the system. In addition, the x-module may be stored in the memory of the system in the form of program codes, and may be called by one of the processing elements of the system to execute the functions of the x-module. Other modules are implemented similarly. All or part of the modules can be integrated together or can be independently realized. The processing element described herein may be an integrated circuit having signal processing capabilities. In implementation, each step of the above method or each module above may be implemented by an integrated logic circuit of hardware in a processor element or an instruction in the form of software. These above modules may be one or more integrated circuits configured to implement the above methods, such as: one or more Application Specific Integrated Circuits (ASICs), one or more microprocessors (DSPs), one or more Field Programmable Gate Arrays (FPGAs), and the like. When a module is implemented in the form of a Processing element scheduler code, the Processing element may be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. These modules may be integrated together and implemented in the form of a System-on-a-chip (SOC).

EXAMPLE III

The embodiment provides a machine learning apparatus, including: a processor, memory, transceiver, communication interface, or/and system bus; the memory is used for storing the computer program, the communication interface is used for communicating with other devices, and the processor and the transceiver are used for running the computer program to enable the machine learning device to execute the steps of the machine learning method.

The above-mentioned system bus may be a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Architecture (EISA) bus, or the like. The system bus may be divided into an address bus, a data bus, a control bus, and the like. For ease of illustration, only one thick line is shown, but this does not mean that there is only one bus or one type of bus. The communication interface is used for realizing communication between the database access device and other equipment (such as a client, a read-write library and a read-only library). The Memory may include a Random Access Memory (RAM), and may further include a non-volatile Memory (non-volatile Memory), such as at least one disk Memory.

The Processor may be a general-purpose Processor, and includes a Central Processing Unit (CPU), a Network Processor (NP), and the like; the Integrated Circuit may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other Programmable logic device, discrete Gate or transistor logic device, or discrete hardware components.

The protection scope of the machine learning method according to the present invention is not limited to the execution sequence of the steps illustrated in the embodiment, and all the solutions implemented by adding, subtracting, and replacing steps according to the principles of the present invention are included in the protection scope of the present invention.

The present invention also provides a machine learning system, which can implement the machine learning method of the present invention, but the implementation apparatus of the machine learning method of the present invention includes, but is not limited to, the structure of the machine learning system described in this embodiment, and all the structural modifications and substitutions of the prior art made according to the principles of the present invention are included in the scope of the present invention.

The machine learning method, the system and the equipment have the following beneficial effects:

firstly, the method avoids complicated and expensive experimental links, is completely based on theoretical calculation, reduces the cost, reduces the dependence on the calculation of a first principle, and shortens the research period.

Secondly, a high-precision and high-efficiency adsorption energy prediction model is established through a machine learning method, accurate prediction is carried out on any adsorption site of a certain heavy metal ion on the adsorption material, and theoretical guidance is improved for designing a novel adsorption material.

Thirdly, the invention combines the transfer learning method, so that the prediction model is suitable for other heavy metal ions and has strong generalization capability.

Fourthly, the model constructed by the method is simple and can be further widely applied in the industry. In conclusion, the present invention effectively overcomes various disadvantages of the prior art and has high industrial utilization value.

The foregoing embodiments are merely illustrative of the principles and utilities of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or change the above-mentioned embodiments without departing from the spirit and scope of the present invention. Accordingly, it is intended that all equivalent modifications or changes which can be made by those skilled in the art without departing from the spirit and technical spirit of the present invention be covered by the claims of the present invention.