An efficient method for searching and generating MOFs materials

Abstract

The structural space of MOFs is almost infinite. To perform molecular simulation calculations on all MOFs materials one by one, so as to select materials with outstanding performance, the computational cost is immeasurable. Therefore, the high-throughput computational screening of high-performance adsorption materials using only the Grand Canonical Ensemble Monte Carlo Simulation (GCMC) method can no longer meet the requirements. Traditional machine learning methods can only explore the target performance of materials in the original library, and do not give full play to the function of machine learning algorithms to predict and design materials. The present invention proposes a method for searching high-performance MOFs materials based on an improved adaptive genetic algorithm combined with a high-precision machine learning algorithm as a fitness function. This method can efficiently search and generate MOFs materials with high target performance outside the original database. This method provides a new idea for the search and generation of porous materials, and can also be widely used in other gas adsorption systems, which is of great significance for large-scale high-performance material screening and reverse design of materials.

Description

Technical Field

[0001] The invention relates to the field of porous material screening, and is a high-performance MOFs material searching and generating method based on an adaptive genetic algorithm. Background Art

[0002] Metal-organic frameworks (MOFs) are a new type of nanoporous material formed by self-assembly of organic ligands and metal ions or clusters through coordination bonds. MOFs materials are currently widely used in gas storage and separation, carbon dioxide capture, catalysis, biosensing, drug delivery and diagnosis. In the traditional field of computational materials, high-throughput screening of MOFs mainly uses Monte Carlo simulation and molecular dynamics simulation to achieve effective evaluation of MOFs material performance, so as to quickly discover the best target material and reveal its structure-activity relationship. However, the possible space of MOFs structures is almost infinite. To perform molecular simulation calculations on all MOFs materials one by one to select materials with outstanding performance, the computational cost is incalculable. Therefore, traditional simulation methods can no longer meet the current requirements for material screening.

[0003] In recent years, with the development of artificial intelligence technology, machine learning algorithms have been increasingly widely used in many fields. At present, in the field of materials, there are many methods that combine machine learning to analyze the structure-activity relationship between material properties and performance. By building a machine learning model, researchers can evaluate the performance of materials more quickly to accelerate the screening of target performance materials. At the same time, machine learning models can also better explore the basic laws behind materials, help workers discover more interesting problems, and thus promote the rapid development of the materials preparation industry.

[0004] Since the performance of materials in the MOFs database is skewed, most materials have poor performance. Traditional machine learning algorithms predict the performance of all materials, which also causes a certain amount of computing cost consumption; more importantly, the application of traditional machine learning research methods can only explore the target performance of materials in the original library. Therefore, breaking through the limitation of the limited number of original databases on the basis of accelerated computational screening of high-performance materials is an urgent problem to be solved in the field of material screening.

[0005] To solve the above problems, we proposed a method for searching and generating high-performance MOFs materials based on an adaptive genetic algorithm. This method can efficiently search and generate MOFs materials with high target performance outside the original library. Summary of the invention

[0006] The purpose of the present invention is to provide an efficient MOFs material search and generation method. Specifically, an improved genetic algorithm is combined with a high-precision machine learning model to search for MOFs individuals with high target performance, and then the materials are assembled and generated in combination with MGPNN software, and finally high-performance MOFs materials outside the original library are generated.

[0007] To achieve the above process, the present invention provides a method for generating MOFs materials based on an improved adaptive genetic algorithm. Figure 1 As shown, the following steps are included:

[0008] Step 1: Prediction model building

[0009] A variety of features related to the assembly of porous materials are obtained, and the original material data set is established by combining different types of physical features. The features in the data set include: potential interpenetration ability, actual interpenetration ability, inorganic nodes, first organic units, second organic units, and functional groups.

[0010] Based on the collected data set, a high-precision machine learning regression model is efficiently trained as the initial model using the Bayesian automatic parameter adjustment method. The mean absolute error (MAE) and determination coefficient (R 2 ) is used as the evaluation index between the model prediction value and the actual value. The calculation formula of the mean absolute error is as follows:

[0011]

[0012] The determination coefficient is as follows (2):

[0013]

[0014] Where m represents the number of samples, Represents the average of all predicted values, y i and They represent the true value of the sample and the predicted value of the model respectively.

[0015] Based on the above evaluation indicators, the performance of the model is evaluated on the test set, and the effects of multiple machine learning models are compared to obtain the final prediction model. This prediction model is used as the fitness function of the adaptive genetic algorithm to evaluate the searched MOFs materials.

[0016] Step 2: Apply an adaptive genetic algorithm (Tangent Adaptive Genetic Algorithm, referred to as TAGA) proposed in the present invention to search for high-performance MOFs individuals. Figure 2 This is the algorithm flow chart of TAGA. The algorithm flow of TAGA algorithm is as follows:

[0017] (1) The population individuals of the adaptive genetic algorithm are encoded with real numbers. A chromosome represents a feature subset, so a chromosome can be represented as G i ={g1,g2,…gi,…gn}, where n is the characteristic number.

[0018] (2) Initialize a population of size N. The initial population is randomly drawn from the original data set, and the current evolutionary generation is set to 0;

[0019] (3) Use the machine learning model as the fitness function to evaluate all chromosomes in the initial population, calculate the performance value of each individual, and based on the elite strategy, the individuals with the top two material performance in the initial population are directly passed on to the next generation;

[0020] (4) The tournament selection operator is used to perform selection operations on the population, and the tournament size Tsize = 2.

[0021] (5) Adjust the individual crossover rate P according to the adaptive adjustment strategy proposed by TAGA c and mutation rate P m , and use single-point crossover and single-point mutation strategies to perform crossover and mutation operations. c and P m The adaptive adjustment strategy is as follows:

[0022]

[0023]

[0024] Among them, P cmin is the minimum crossover rate, P cmax is the maximum crossover rate, P mmax is the maximum mutation rate, P mmin is the minimum mutation rate, f avg is the average fitness of the population, f′ is the fitness of the better individual when crossover occurs, f is the fitness value of the individual to be mutated, A=0.002.

[0025] (6) Determine whether the individual generated in (5) is a new individual. If the condition is met, it is directly passed into the new generation group; otherwise, the newly generated individual is discarded;

[0026] (7) If a new generation population consisting of N new individuals has been generated, go to (8), otherwise return to (4);

[0027] (8) If the maximum number of evolution generations is reached, the algorithm ends, otherwise the number of iterations is increased by 1 and returns to (4);

[0028] Step 3: Based on the high-performance MOF individuals found by TAGA, the building blocks after genetic encoding are decoded into actual building block files, and the three building blocks of inorganic units, organic units, and functional groups are used as inputs of the MGPNN software to assemble the materials to obtain the final MOFs materials.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) The present invention proposes an effective method for screening MOF materials, namely, a high-performance MOFs material optimization algorithm based on an improved adaptive genetic algorithm; this method breaks through the problem that traditional molecular simulation methods can only screen materials in the library, and has the advantages of a larger search space and faster speed;

[0031] (2) The present invention proposes an improved adaptive genetic algorithm, which can dynamically adjust the mutation rate and crossover rate of the genetic algorithm during the evolution process, thereby improving the "premature" situation of the genetic algorithm to a certain extent, and shows better search performance and search stability than some current adaptive genetic algorithms;

[0032] (3) The prediction model in the present invention can effectively reveal the structure-activity relationship between the structure of MOFs materials and the methane adsorption performance, and can effectively predict the methane adsorption value of MOFs materials;

[0033] (4) The method of the present invention is not limited to finding MOFs materials with high methane adsorption, but can also be extended to other gas adsorption systems or other porous materials; BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Method flow chart

[0035] Figure 2 Flowchart of TAGA’s search for high-performance MOFs

[0036] Figure 3 Prediction effect and accuracy of the XGBoost model on the test set

[0037] Figure 4 LightGBM prediction effect and accuracy on the test set

[0038] Figure 5 Schematic diagram of the structure of the top four MOFs materials in terms of assembled performance DETAILED DESCRIPTION

[0039] Table 1 10 standard test functions for TAGA implementation testing

[0040] Table 2 Comparison of the optimization performance of TAGA algorithm and other five adaptive genetic algorithms in solving the minimum value of the function

[0041] Table 3 Comparison of the performance standard deviation of the TAGA algorithm and the other five adaptive genetic algorithms in solving the minimum value of the function

[0042] Table 4 Comparison of predicted and actual values ​​of the top 4 materials in terms of performance

[0043] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific examples described herein are only used to explain the present invention and are not used to limit the present invention.

[0044] In order to test the search performance and search stability of an improved adaptive genetic algorithm (TAGA) proposed in the present invention, we use the 10 minimum standard test functions in Table 1 as examples to compare the performance of TAGA and the existing five adaptive algorithms. Each test calculation is run 20 times, and the average results of the 20 calculations are taken for analysis. Table 2 gives the average optimal value of the 20 optimization calculations, and Table 3 gives the standard deviation of the 20 optimization calculations of each algorithm. The average optimal value can measure the search performance of each algorithm, and the standard deviation can measure the robustness and computational stability of each algorithm. From the comparison results, it can be seen that among the 10 standard test functions, the 20 average optimal values ​​of the 10 functions obtained by the TAGA optimization calculation are the smallest, and the error value with the theory is smaller. Moreover, TAGA also performs well in search stability among the 10 functions. The results show that both the algorithm's optimization performance and optimization stability show better results.

[0045] Table 1. Standard test functions

[0046]

[0047] Table 2. Average results of 20 optimization calculations of different algorithms (GOV is the global optimal value)

[0048]

[0049] Table 3. Standard deviation of 20 optimization calculations for different algorithms

[0050]

[0051] Note: TAGA is an adaptive genetic algorithm proposed in this invention.

[0052] In order to test the effectiveness of the method of the present invention for generating high-performance MOFs materials, we conducted tests in a methane adsorption system to find MOFs individuals with excellent methane adsorption performance outside the original MOFs database.

[0053] Step 1: The data set used in the embodiment is a dataset of 51,163 reduced hMOFs samples developed and coded by the Snner research group. Each MOF sample data in the dataset contains 6 structural features: potential interpenetration ability, actual interpenetration ability, inorganic unit, first organic unit, second organic unit and functional group. Based on these samples, the data is divided into training set and test set according to a 4:1 ratio, and the Bayesian automatic parameter adjustment method is used to adjust the parameters of the XGBoost and LightGBM models. The final accuracy of the two models on the test set is as follows: Figure 3 and Figure 4 The results show that the XGBoost model has better accuracy. The R 2 It reached 0.91, which can better evaluate and predict the performance of new MOFs individuals searched by genetic algorithms.

[0054] Step 2: Based on the TAGA genetic algorithm proposed in the present invention, the high-precision XGBoost model trained in step 1 is used as the fitness function of TAGA. The maximum population of TAGA is set to 200, and the maximum iterative evolution generation is set to 40. After searching, the model finally found 979 MOFs individuals that did not exist in the original library, of which 27 materials had a predicted performance exceeding 528 cm 3 / g.

[0055] Step 3: assemble the four materials predicted by TAGA to be the best using MGPNM software. Decode the features after gene encoding into actual building blocks, and use the inorganic units, first organic units, second organic units, and functional groups required for software assembly as input to assemble the materials, and finally obtain four new MOFs materials. Figure 5 Finally, in order to verify the real performance of the materials, we used GCMC simulation software to evaluate the performance of the assembled MOFs materials. The methane adsorption amount simulated by GCMC calculation of the four MOF materials is shown in Table 4. The results show that the highest performance of the four materials reached 580.73 cm 3 / g, both exceeded the performance of the best material in the original database (528cm 3 / g).

[0056] Table 4 XGBoost predicted adsorption values ​​and GCMC simulated adsorption values ​​of the top 4 materials

[0057]

[0058] The new method proposed by the present invention provides a new idea for searching and generating new porous materials, and can also be widely used in material screening of other gas adsorption systems. This method is of great significance for large-scale high-performance material screening and reverse design of materials.

Claims

1. An efficient MOFs material search and generation method, characterized in that: include: Step 1: Build a high-precision machine learning model to predict the performance of MOFs materials, that is, based on the original material data set containing the characteristics of the building block information of MOFs materials, a machine learning model that can well reveal the structure-activity relationship of MOFs materials is trained by the Bayesian automatic parameter adjustment method; The dataset is a dataset of 51,163 reduced hMOFs samples developed and coded by Snner's research group. Each MOF sample in the dataset contains six structural features: potential interpenetration ability, actual interpenetration ability, inorganic unit, first organic unit, second organic unit and functional group. The method of training a machine learning model that can well reveal the structure-activity relationship of MOFs materials by using the Bayesian automatic parameter adjustment method is to divide the data into a training set and a test set according to a ratio of 4:1, and to adjust the parameters of the XGBoost and LightGBM models by using the Bayesian automatic parameter adjustment method; Step 2: Using the obtained high-precision machine learning model as the fitness function of the genetic algorithm, searching for MOFs materials with high methane adsorption based on the improved adaptive genetic algorithm; The adaptive genetic algorithm is specifically: (1) The population individuals of the adaptive genetic algorithm are encoded with real numbers. One chromosome represents a feature subset, so the chromosome can be represented as Gi = {g1, g2, …gi, …gn}, where n is the number of features; (2) Initialize a population of size N. The initial population is randomly drawn from the original data set, and the current evolutionary generation is set to 0; (3) Use the machine learning model as the fitness function to evaluate all chromosomes in the initial population, calculate the performance value of each individual, and based on the elite strategy, the individuals with the top two performance in the initial population are directly passed to the next generation; (4) Use the tournament selection operator to perform selection operations on the population, and the tournament size Tsize = 2; (5) The crossover rate and mutation rate of individuals are adjusted according to the adaptive adjustment strategy proposed by TAGA, and the crossover and mutation operations are performed using single-point crossover and single-point mutation strategies; (6) Determine whether the individual generated in (5) is a new individual. If the conditions are met, directly pass it into the new generation group; otherwise, discard the current newly generated individual; (7) If a new generation population consisting of N new individuals has been generated, go to (8), otherwise return to (4); (8) If the maximum number of evolution generations is reached, the algorithm ends, otherwise the number of iterations is increased by 1 and returns to (4); Step 3: Decode the searched high methane adsorption MOFs materials into corresponding basic building blocks of MOFs, and input them into MGPNN software for material assembly generation.

2. An efficient MOFs material search and generation method according to claim 1, characterized in that: The high-precision machine learning model is used as the fitness function of the genetic algorithm. Specifically, the trained high-precision machine learning model can well characterize the structure-activity relationship between MOFs materials. The genetic algorithm uses the same input features as machine learning as iterative chromosomes, so the machine learning model can be used as the fitness function of the genetic algorithm to evaluate newly generated individuals.

3. The efficient MOFs material search and generation method according to claim 1, characterized in that: The searched high methane adsorption MOFs materials are decoded into the corresponding basic components of MOFs. Specifically, in the genetic algorithm, the combination of 6 structural features that characterize the MOFs material information is used as the chromosome of the MOFs individual; therefore, after searching for high-performance individuals, they need to be decoded and restored to the corresponding structural information; four of the 6 genes are the input information required for material assembly and generation.

4. According to claim 1, an efficient MOFs material search and generation method is characterized in that the material is input into the MGPNN software for material assembly generation. Specifically, when the MGPNN software performs material assembly, it is necessary to input the .car file representing the MOFs structure of inorganic units, organic units, and functional groups for assembly.

5. An efficient MOFs material search and generation method according to claim 1, characterized in that: The specific adjustment strategy is: to dynamically adjust the mutation rate and crossover rate of the genetic algorithm through an improved adaptive adjustment strategy to alleviate the "premature" phenomenon of the genetic algorithm; c and P m for: Among them, P cmin is the minimum crossover rate, P cmax is the maximum crossover rate, P mmax is the maximum mutation rate, P mmin is the minimum mutation rate, f avg is the average fitness of the population, f′ is the fitness of the better individual when crossover occurs, f is the fitness value of the individual that mutates, A=0.002.

Images