中文   |   English   |   Email Alert
Fig/Tab
Advanced Search
image

Machine Learning Accelerated High-Throughput Computational Screening of Metal-Organic Frameworks

Wei Li , Tiangui Liang , Yuanchuang Lin , Weixiong Wu , Song Li

Progress in Chemistry ›› 2022, Vol. 34 ›› Issue (12) : 2619 -2637.

PDF
Progress in Chemistry ›› 2022, Vol. 34 ›› Issue (12) : 2619-2637 DOI: 10.7536/PC220524
research-article
Scan to read & Share
Machine Learning Accelerated High-Throughput Computational Screening of Metal-Organic Frameworks
Author information +
History +
PDF

Abstract

Metal-organic frameworks (MOFs) with ultrahigh surface area, large pore volume and tunable pore environment are regarded as promising adsorbents in gas adsorption and separation. With the exploding number of possible MOFs, it is an imperative challenge to discover high-performing MOFs for a specific application. High-throughput computational screening (HTCS) has been frequently adopted to identify the suitable adsorbents with remarkably reduced computational time and cost in the past decades. Recently, machine learning (ML) is used to accelerate the HTCS process and explore the structure-property relationship. In this review, we summarize the ML algorithms for gas storage, separation and catalysis by MOFs. Four categories of descriptors widely used in ML are reviewed, including geometrical descriptor (i.e., pore size, surface area), topological descriptor (i.e., pore connectivity, cavity size), chemical descriptor (i.e., atom type, degree of unsaturation), and energy-based descriptor (i.e., cohesive energies, Voronoi energies). Unsupervised learning and supervised learning, such as linear regression (LR), artificial neural network (ANN), support vector machine (SVM), and random forest (RF) are introduced and a complete overview of how these ML algorithms are effectively utilized to assist MOF discovery is provided. Recent research progress on ML algorithms in various applications, such as H2, CH4 storage, carbon capture, noble gas, Cx/Cy separation and catalysis are presented in this work. It is anticipated that ML will play a more vital role in identifying top candidates with increasing number of MOFs. Thus, this review aims to outline fundamental knowledge of ML algorithms for MOF discovery in the fields of gas storage, separation and catalysis.

Contents

1 Introduction

2 Methodology

2.1 High-throughput computational screening

2.2 Machine learning

3 Machine learning accelerated MOFs screening

3.1 Gas storage

3.2 Gas separation

3.3 MOFs physicochemical property

4 Conclusion and outlook

Graphical abstract

Keywords

metal-organic frameworks / machine learning / high-throughput computational screening / molecular simulation / adsorption

Cite this article

Download citation ▾
Wei Li, Tiangui Liang, Yuanchuang Lin, Weixiong Wu, Song Li. Machine Learning Accelerated High-Throughput Computational Screening of Metal-Organic Frameworks. Progress in Chemistry, 2022, 34(12): 2619-2637 DOI:10.7536/PC220524

登录浏览全文

4963

注册一个新账户 忘记密码

1 Introduction

Porous materials have a network structure composed of interconnected or closed pores. Traditional porous materials, including activated carbon, zeolite, molecular sieve, etc., are common adsorbents[1][2][3]. In recent years, Metal-Organic Frameworks (MOFs), Covalent organic Frameworks (COFs), carbon nanotubes and other new porous materials have attracted much attention. Among them, MOFs are periodic crystal structures composed of inorganic metal units (metal ions, clusters or chains) connected with organic ligands (carboxylate, phosphate, or nitrogen-containing ligands, etc.), which have the advantages of good thermal stability (~ 500 ℃), high specific surface area (>8000 m2/g), low density (~0.2 g/cm3) and abundant nanopores (3 3 Å ~ 100 100 Å)[4][5]. Through the combination of different structural units, connection modes and topological structures, researchers can regulate the physical and chemical properties to prepare MOFs materials with excellent properties, which can be used in gas adsorption and separation, heat pump, catalysis, battery and other fields[6⇓⇓~9][10,11][12⇓~14][15].
The development of MOFs is very rapid. As shown in Figure 1A, there are more than 90,000 experimental MOFs (eMOFs) synthesized so far, most of which are included in the Cambridge Structural Database (CSD)[16]. With the development of high-performance computing and simulation technology, researchers have also generated MOFs materials according to chemical laws. For example, in 2011, Wilmer et al. Designed and constructed a database of 137,953 hypothetical MOFs (hMOFs) based on the "top-down" method[17]. In 2017, Col Colón et al. Used Edge-Transitive Topological Nets algorithm and ToBaCCo code to construct a database consisting of 41 Topological structures and 13,512 hMOFs[18]. However, there are still nearly unlimited kinds of MOFs that can be designed and synthesized. On the one hand, such a large number of MOFs enables us to find MOFs suitable for various applications. On the other hand, only relying on experiments to search for and prepare high-performance MOFs requires a long cycle and high research and development costs, which is like looking for a needle in a haystack.
High-Throughput Computational Screening (HTCS) based on molecular modeling is the most efficient research method for evaluating and optimizing MOFs. The general process is shown in Fig. 1B: 1) constructing or selecting a screening database of MOFs; 2) pre-screening by using characteristics such as specific surface area or pore size; 3) Molecular simulation to predict the adsorption, separation, catalysis and other properties of MOFs; And 4) calculate that measurement index for comprehensive evaluation in combination with the actual application conditions to obtain the high-performance MOFs[4].
However, with the explosive growth of MOFs, even high-throughput screening technology is not enough to comprehensively evaluate their application performance in various scenarios, and the massive and multi-dimensional data generated by high-throughput screening also need to be analyzed in depth with the help of big data technology. Machine Learning (ML) is in line with this demand, and using it to assist high-throughput screening is a research hotspot in recent years. Machine learning is based on the statistical rules hidden in high-dimensional data. By training high-throughput screening data sets, it can predict the adsorption capacity, selectivity, isothermal adsorption curve and other material characteristics of MOFs adsorbents, and assist in accelerating the discovery of high-performance adsorbents. Commonly used Machine learning algorithms include Artificial Neural Networks (ANN), Gradient Boosting Machine (GBM), etc[19,20][21,22][23][24][25].
In this paper, the research progress of machine learning-assisted high-throughput screening to develop new adsorbents is summarized. Because most of the algorithms in this kind of research are based on open source software such as TensorFlow, Pytorch and Scikit-learn, this paper focuses on the structure-activity relationship discovered by these algorithms, or the predictive performance and generalization ability of their training models[26][27][28]. The development trend and challenges of big data technology in the field of MOFs research and development are discussed.

2 Method

On the basis of high-throughput screening, the general flow of machine learning-assisted research is shown in Figure 2 above. The high-throughput screening part includes a. Database selection, prescreening, and B. Molecular simulation and other processes, combined with the machine learning part including C. Descriptor selection, parameter optimization and d. Model training, performance prediction, and e. Data analysis, etc., can be carried out in the future. Experiments verify the performance of target MOFs or further propose new high-performance structure synthesis and design strategies.

2.1 High throughput screening technique

Selecting the appropriate database is the basis of high-throughput screening research, and it will also affect the performance of machine learning algorithms. MOFs databases are mainly divided into two categories: eMOFs and hMOFs[4]. Among them, eMOFs are compiled from crystal structure files collected in experiments, such as the CoRE MOFs (Computation-Ready Experimental MOFs) database established by Snurr et al., which have complete synthesis methods and are characterized by significant structural diversity[29]. The database of hMOFs is generated according to chemical laws, such as the MTV MOF (Multivariate MOF) database constructed by Li Song et al., which has clear material structure and obvious regularity[30].
When pre-screening the database, Zeo + +, RASPA, LAMMPS, etc. Are generally used to calculate or collect the geometric properties, topological structure, chemical information, adsorption potential distribution and other parameters of MOFs in the database, and the appropriate range is selected according to the application scenario to exclude the unqualified MOFs samples and reduce the amount of calculation[31][32][33]. Commonly used include maximum Pore Diameter (Largest Cavity Diameter, LCD), restricted pore Diameter (Pore Limited Diameter, PLD), Density (ρ), porosity (Void Fraction, VF), pore volume (Largest Cavity Diameter), specific Surface Area (Accessible Surface Area, ASA), Gravimetric specific Surface Area (Gravimetric Surface Area, GSA) and so on.
Molecular simulation technology is the main method to predict the performance of MOFs in high-throughput screening, generally including Grand Canonical Monte Carlo (GCMC) equilibrium adsorption properties, Molecular Dynamics (MD) dynamic diffusion properties[5,34][35]. In addition, there are also some studies based on classical DFT and lattice GCMC to predict the properties of MOFs and carry out high-throughput screening studies[36][37].

2.1.1 Monte Carlo simulation of MOFs.

GCMC simulations can be used to model the adsorption behavior of MOFs. In the simulation, the chemical potential of the adsorbate is specified, the volume (V), temperature (T) and chemical potential (μ) in the system are kept constant, and the kinetic properties of the gas molecules in the MOFs are not considered. The equilibrium distribution of the system is obtained by the microscopic States of the gas molecules in the MOFs, such as insertion, deletion, movement and rotation[38]. Through Monte Carlo simulation, the thermodynamic properties of the material can be obtained. From the equation of state or independent bulk simulation, it is easy to obtain the relationship between the adsorption capacity of MOFs and the partial pressure ratio, so as to obtain the isothermal adsorption curve of gases in the material[39]. In the canonical ensemble with fixed temperature, volume and number of molecules, data such as Henry's constant and binding energy can also be calculated by GCMC simulation[40].

2.1.2 Molecular dynamics simulation

MD simulations were performed in conditions of constant volume, temperature, or pressure[40]. According to the principle of Newtonian mechanics, MD simulation establishes the motion equation of all gas molecules, solves the dynamic parameters through the interaction potential, and then obtains a series of microscopic States generated by the motion trajectory of gas molecules in MOFs with time[38]. The kinetic properties such as diffusion coefficient and conductivity were further determined from the molecular trajectories of MD simulation[41][42].

2.2 Machine learning method

Although high-throughput screening methods based on molecular modeling can evaluate the performance of thousands of adsorbents, they are still inadequate with the explosive growth of the number of MOFs and the continuous expansion of application scenarios. Machine learning (ML) can assist high-throughput screening technology, optimize descriptors, build data sets to train machine learning models, efficiently predict various properties of MOFs, and analyze structure-activity relationships.
The pre-step of ML model construction is the selection and transformation of descriptors, and the transformation should keep the target attributes unchanged[43,44]. Commonly used descriptors are shown in Figure 3 below, including geometric parameters, topology, chemical properties, and energy potential information. Geometrical parameters are commonly used descriptors, which are generally available in pre-screening, including pore volume, pore size, density, and porosity[4]. Topology includes pore connectivity, pore morphology and pore distribution.
For example, Krishnapriyan et al. Developed the Persistence Image Approach, which combines topological descriptors with regularized structural features and vectorizes them as the input of machine learning models, thus successfully predicting the adsorption performance of zeolites[45]. The above descriptors are close to the experimental measurement data and can intuitively reflect the basic properties of MOFs, but the machine learning model based on this training has poor generalization. In terms of chemical properties, such as atom type, unsaturated metal sites, metal cluster ratio, atomic electronegativity and Henry coefficient (Henry’s constant,KH) are also commonly used descriptors. Such data are widely distributed, and their characteristics should be transformed, such as logarithmic normalization of KH. Furthermore, scholars have also deeply analyzed the microscopic information of molecular internal potential Energy, such as Cohesive Energy, Voronoi Energy and energy distribution function, in order to improve the prediction accuracy and generalization of the model[46][47][48]. For example, Rosen et al. Used Density Functional Theory (DFT) to construct a QMOFs database with quantum chemical properties of materials, which contains the atomic type, energy, geometric parameters, charge Density, rotation density and other energy information of 15,713 eMOFs, so as to train the ML algorithm to accurately predict the energy band gap of MOFs[46].
In order to avoid overfitting, Cross-Validation methods such as k-Flod are used in machine learning training. The data set is divided into training data set, validation data set and test set, and the appropriate proportion is selected according to the training results[49]. Generally speaking, the ratio of the training set to the test set varies between 20% and 80%.
ML model selection is mainly divided into supervised learning and unsupervised learning, in which supervised learning mainly performs tasks such as regression and classification on descriptor data[49]. The regression task can build a model that relates the input variables of the training set to the output variables to predict the performance of the MOFs in the test set[5]. Such algorithms include: Least Absolute value convergence and Selection Operator algorithm (Least Absolute Shrinkage and Selection Operator, LASSO), Linear Regression (LR), Kernel Ridge Regression (Kernel Ridge Regression, KRR), artificial neural network, Deep Learning (DL) and so on. The classification task classifies the data set into subsets according to the output variables, and determines the corresponding value range of the descriptor and the classification criteria for measuring its performance[5]. Common classification algorithms include Decision Tree (DT), Random Forest (RF), Support Vector Machines (SVM), Bayesian Optimization (BO), etc. Unsupervised learning is mainly used to explore unclear relationships and performance parameters, such as structure-activity relationships, including K-means Cluster, Principal Component Analysis (PCA) and Singular Value Decomposition (SVD). In the research of machine learning-assisted high-throughput screening, there are also other algorithms derived. The basic principles of common algorithms are introduced below and briefly shown in Figure 4.

2.2.1 Multiple linear regression

Multivariate Linear Regression (MLR) is suitable for predicting large-scale systems with Linear relationships among multiple factors[50]. In the adsorption model with weak interaction, the properties of MOFs and their properties are also influenced by many factors, for example, the saturated adsorption capacity is linearly related to the pore volume, so it is widely used[4].

2.2.2 K-nearest neighbor

K-Nearest Neighbor (KNN) is an algorithm for classification and regression by measuring the distance between different feature values in the "training set"[51]. In the MOFs database, there are a large number of structures with similar properties, which can be classified and predicted by this algorithm.

2.2.3 Decision tree

Decision tree is a traditional supervised learning method, which is based on the attributes of the data in the test set, analogy and "branch structure" to establish a decision model to quantitatively give the relationship between the structural characteristics and performance of MOFs, but the decision tree algorithm is not suitable for dealing with large-scale data and continuous variables[52][4]. Boosted Decision Tree (BDT) improves the accuracy of DT by creating an integrated Tree model, building each regression tree in a step-by-step manner, and using a loss function to measure the prediction error[53].

2.2.4 Random forest

Random forest is an extension of decision tree algorithm[54]. Multiple decision trees are constructed from the training set and averaged to achieve more accurate predictions.

2.2.5 Support vector machine

Support vector machine (SVM) is a data classification-oriented method, which can distinguish the structure of the data set according to certain rules through mathematical transformation methods, and is suitable for complex linear or nonlinear structure-activity relationship. It is more flexible than DT and KNN algorithms, and can often obtain more accurate prediction results, so it is the most widely used[55]. In addition, Least Square Support Vector Machine (LSSVM) is one of the most commonly used methods for predicting parameters and modeling in complex systems[56].

2.2.6 Neural network

Artificial neural network algorithm is to find out the relationship between the structural characteristics and performance, build a neural network, and get a suitable model to accurately predict the performance of the test set by constantly adjusting the function parameters in the hidden layer. There are many algorithms derived from Neural Networks, including Deep Neural Networks (DNN), Recurrent Neural Network (RNN), Back Propagation Neural Network (BPNN) and so on, which have been used to screen high-performance MOFs[57][58][59].

2.2.7 Evolutionary algorithm

Evolutionary algorithms mainly include Genetic Algorithm (GA), Evolution Strategies (ES) and Evolution Programming (EP)[60]. GA is a random global search algorithm to find the optimal solution, which is similar to the law of gene inheritance in nature. The new structure is inherited from the gene of the previous generation of high-performance materials. After many generations, it can predict the structure of MOFs with higher performance or develop design strategies[61].

2.2.8 Model evaluation parameter

The prediction accuracy of regression machine learning models is evaluated mainly by the following parameters: (1) Absolute Error (Mean Absolute Error, MAE); (2) Mean Square Error (Mean Square Error, MSE); (3) Root-Mean-Square Error (RMSE); (4) Mean Absolute proportional Error (Mean Absolute Percentage Error, MAPE); and (5) coefficient of determination (Mean Absolute Error). Where n is the number of samples in the data set, xi is the descriptor data of the model input, yi is the predicted value of the model output, and y ¯ is the mean value of the predicted data[20]. Generally speaking, the MAE, MSE and RMSE of an excellent regression model are close to MAPE 0,R2 close to 1. The parameters for evaluating the Accuracy of the classification machine learning algorithm model include: (1) Accuracy (ACC), where TP is a true positive example, FP is a false positive example, TN is a true negative example, and FN is a false negative example; (2) Area Under the Curve (AUC) of Receiver Operating Characteristic (ROC). ACC and AUC are close to 1, indicating that the classification is accurate. It is worth mentioning that the regression evaluation index can also be used to predict and evaluate the classification algorithm.
M A E ( x , y ) = 1 n ∑ i = 1 n | x i - y i |
M S E ( x , y ) = 1 n ∑ i = 1 n | ( x i - y i ) 2
R M S E ( x , y ) = 1 n ∑ i = 1 n | ( x i - y i ) 2
M A P E ( x , y ) = 100 n ∑ i = 1 n x i - y i x i
R ( x , y ) 2 = 1 - ∑ i = 1 n ( x i - y i ) 2 ( y i - y - ) 2
A C C = T P + T N T P + T N + F P + F N
A U C ( x , y ) = 1 2 ∑ i = 1 n - 1 ( x i + 1 - x i ) ( y i + 1 + y i )

3 Research

3.1 Gas storage

MOFs are huge lattice structures composed of central metal ions connected with organic ligands, which can form a three-dimensional network with high specific surface area[62]. This porous structure helps to adsorb and store a large number of gases, including hydrogen (H2), methane (CH4), carbon dioxide (CO2), nitrogen (N2) and so on[63,64][65][66][67].

3.1.1 H2 Storage

H2 is an ideal new energy fuel, which is widely used in new energy vehicles and fuel cell technologies, but it is difficult to store and transport[68]. Compared with low-temperature compressed liquid hydrogen storage, hydrogen storage by physical adsorption has the advantages of energy saving and safety[68,69]. By adjusting the parameters such as van der Waals force between MOFs and H2 molecules, qualified MOFs can be designed to achieve efficient hydrogen storage[70].
Table 1 briefly summarizes the research progress of machine learning-assisted screening R & D of MOFs suitable for H2storage. The data set is mainly from the high-throughput screening results of GCMC, and the descriptors used are mainly energy potential information. On the whole, the training model has high prediction accuracy and obvious structure-activity relationship.
Cao et al. Trained three types of machine learning algorithms, including least squares support vector machine, Adaptive Neuro-Fuzzy Inference System (ANFIS), and artificial neural network[71]. Based on the Spearman and Pearson correlation coefficients, the effects of various features on the hydrogen storage capacity of MOFs were analyzed by using the "three-step" strategy (trial-and-error method, precision analysis method and ranking method), and it was found that the surface area and adsorption enthalpy were the most influential descriptors for the hydrogen storage capacity of MOFs[72]. The ANN model has the highest prediction accuracy, with a R2 of 0. 99, and the MAE and MSE are only 0. 05 wt% and 0. 059 wt%, respectively.
Bucior et al. Used the energy characteristics of the potential energy surface in the pores of MOFs as descriptors, and used the LASSO algorithm based on feature selection and regularization to predict the hydrogen storage capacity[73]. This descriptor uses the H2 as a probe to analyze the gridded potential energy of MOFs, and further simplifies the energy distribution into a histogram, each subset of which corresponds to a class of features in the regression model to predict the hydrogen adsorption performance of each space in the MOFs structure. A database of 2250 hMOFs was trained with this method for cross-validation, and the results in the training set of 1000 hMOFs were :MAE=2.3 g/L,RMSE=3.0 g/L,R2=0.96;1250, and the results in the test set of hMOFs were :MAE=2.4 g/L,RMSE=3.1 g/L,R2=0.96. The results show that the error of the model in predicting the adsorption capacity of H2 can be controlled within 3. 0 G/L, and the calculation speed is still more than 1000 times faster than that of GCMC simulation. Moreover, they used this method to find MFU-4l from 50,000 eMOFs, which has a hydrogen storage capacity of up to 47 G/L.
Giappa et al. First used ab-initio to calculate the bonding strength between hydrogen and 58 kinds of functional groups (C6H5-X(X are different types of functional groups), and found that the four functional groups, —OCONH2, —SO2NH2, —OPO3H2 and —OSO3H, have the strongest interaction[74]. Among them, the bonding strength of —OSO3H group modification is 80% higher than that of benzene. Then IRMOF-8, IRMOF-16 and IRMORF-14 were modified with these four functional groups, and the adsorption curves were obtained by GCMC simulation at 77 K and 300 K under pressure swing (0 bar ~ 100 bar) conditions, which were used as the training data set of the machine model. Using Coulomb Matrix, Bag of Bonds, and Smooth Overlap of Atomic Potentials as descriptors, kernel ridge regression, random forest, and support vector machine models were trained to predict the interaction strength between MOFs and Bag of Bonds. Among the results, the RF algorithm has the highest prediction accuracy, and its MAE is 1.99×10-4Ha, 1.16×10-4Ha and 2.10×10-4Ha under the three descriptors.
Ahmed et al. Combined 14 algorithms such as DT, LR and RF to predict the hydrogen storage capacity of MOFs[75]. The H2 adsorption isotherms of 98,695 MOFs were obtained by GCMC as a data set, and the adsorption mass and capacity of 918 734 MOFs were predicted by seven structural descriptors, including ρ, Va, GSA, SA, VF, LCD and PLD, under pressure swing (5 bar ~ 100 bar) and temperature swing (77 K, 298 K) conditions. Among them, the Extremely Randomized Trees (ERT) model stands out, taking the adsorption mass under pressure swing conditions as an example, its Extremely Randomized Trees is 0. 997, and the MAE and RMSE are only 0. 10 G/L and 0. 18 G/L. Further analysis shows that pore volume and porosity are the most important descriptors. A total of 8282 hMOFs with hydrogen storage properties exceeding the best record were identified using this algorithm, all of which are characterized by low density (<0.31 g/cm3), high specific surface area (>5300 m2/g), porosity (~ 0.90), and large pore volume (>3.3 cm3/g). This work is the most abundant research on machine learning-assisted screening of MOFs suitable for https://sorbent-ml.hymarc.org storage, and the model is displayed in the official website (https://sorbent-ml.hymarc.org), and users can quickly obtain the hydrogen storage performance after inputting the structural data of MOFs[76].

3.1.2 CH4 Storage

Natural gas is a transitional fuel from traditional fossil fuels to renewable energy, and its pollution is relatively low. Generally, natural gas contains 70% ~ 90% of CH4, which is the composition with the highest combustion energy among all hydrocarbons, and needs to be stored by adsorption and other means[77].
The research on CH4storage is shown in Table 2 below. Because CH4 is a non-polar molecule and its interaction with MOFs is relatively simple, it can be accurately predicted only by using geometric parameters and chemical information as descriptors, and there are many derivative algorithms used in this system.
Suyetin only used ASA, Va and ρ as descriptors to train the MLR model to predict the total and working adsorption capacity of CH4[20]. It is R2=0.979(298 K),R2=0.987(273 K),R2=0.990(240 K) in predicting Nbo topology MOFs; When predicting CoRE MOFs, the R2 can also reach 0.899, and the MAE and RMSE are 9.23 cm3/g and 12.60 cm3/g. Beauregard et al. combined the genetic algorithm with the random forest algorithm to construct the GARF algorithm. The main process is to use GA to construct 200 "chromosomes" carrying structural characteristics. Under the condition that the "population" size is 100, the MOFs with the best methane storage performance are predicted after 2000 times of iterative evolution, and the performance in each generation is predicted by combining RF[78]. The data of 130 398 hMOFs were used to train the model according to the ratio of 1 ∶ 9, and 15 MOFs with excellent performance were selected, and their methane adsorption capacity was up to 414 cm3/g,R2=0.95,MAPE=1.37 cm3/g. Taw et al. Coded the properties, metal clusters, organic ligands and functional groups of MOFs into Bayesian optimization algorithm, and divided them into structure-related parts and individual feature-related parts[84]. Then the LR + GP model is formed by linear regression with the structural correlation part and combining the individual characteristics with Gaussian Process (GP) at the same time. Both models use the same 20 MOFs as the starting point for Bayesian optimization, and add a training set material specified by the sampling function at each iteration. It can be found that the LR + GP model can search for structures with excellent performance faster at 35 bar. This method can be used to train 457,000 hMOFs structures, and only 0.1% of the data is needed as input, and the top 20 structures can be reliably found.
In another study, Kim et al. Constructed a Monte Carlo machine learning (MC-ML) model to predict the methane adsorption performance of 4951 CoRE MOFs[82]. Unlike most of the above studies, which are based on GCMC simulation to obtain descriptor data, this work collects data such as pore volume and isothermal adsorption curve in the experiment as input, and is dedicated to predicting the adsorption curve under other conditions. Among the four algorithms: Linear SVM, Nonlinear SVM, decision tree and random forest, RF has the highest prediction accuracy at 308 K and 318 K, and its R2=0.947,MSE=0.918 kJ/mol,MAPE=3.903 kJ/mol.
Inspired by the feature transformation in the machine learning modeling step, researchers do not directly use geometric parameters, topology, chemical properties and other descriptors as model inputs, but form new descriptors through data processing. For example, Lee et al. Used 648 Node Building Blocks (NBBs), 219 Edge Building Blocks (EBBs) and 1775 topologies to develop a pormake, generated a database of about 247 trillion hMOFs, and built a MOF-NET model based on it to explore the best methane storage structure using deep learning algorithms[83]. As shown in Figure 5, by mapping the topological unit to a random initial vector, the vector is iterated during the training process, and the result is put into the Multispecies Genetic Algorithm With Fitness Approximation (MSGA-FA) for screening. The results show that 964 hMOFs in the database have CH4storage performance greater than 200 cm3/cm3, and 96 of them exceed the methane storage record (208 cm3/cm3).
Gurnani et al. Trained Gravimetric Uptake Model (GUM), Geometric Model (GM) and Non-GM (NGM) to predict the methane adsorption performance of 137,953 hMOFs[19]. 30% of the data is set as the test set, and the top-down method is used to remove the chemical descriptors with lower weights, and the feedforward neural network maps the effective descriptors to the prediction results. R2=0.99,RMSE=9.05 cm3/g predicted by the GUM model at 298 K, 35 bar. Furthermore, k-Fold cross-validation was used to predict the methane adsorption curves of GM and NGM models. The difference is that the geometric descriptors are removed in the NGM model, as shown in Fig. 6, the R2=0.95,RMSE=18.64 cm3/g of the R2=0.99,RMSE=7.83 cm3/g,NGM model of the final GM model. The high accuracy of the NGM model saves the cost of a large number of computational geometric descriptors and provides a shortcut method for screening high-performance MOFs.

3.1.3 Carbon capture

With the extensive use of fossil fuels, a large amount of CO2 is emitted into the atmosphere, causing problems such as the greenhouse effect. Therefore, the use of MOFs to adsorb and capture CO2 has become a frontier research hotspot[85].
In recent years, the progress of machine learning-assisted screening of MOFs suitable for CO2 adsorption is shown in Table 3, and a variety of descriptors of atomic distribution positions within the crystal structure have been developed.
It is often difficult to use MOFs to capture CO2 under the condition of water adsorption and pressure change. Deng et al. Used GCMC, MD simulation and back propagation neural network (BPNN), random forest (RF), decision tree (DT) and support vector machine (SVM) to study the adsorption and diffusion performance of CO2 in CoRE MOFs[22]. The RF model was found to predict the adsorption selectivity with the highest accuracy, which was R2=0.981. Combined with multivariate analysis, the importance of each descriptor was ranked as :Qst>ρ>LCD>SA≥VF>PLD, and finally 14 CoRE MOFs with excellent carbon capture performance were obtained, including REYCEF and JAHNEM.
Feature transformation has also been used to construct machine learning algorithm models for predicting the carbon capture performance of MOFs, and Burner et al. Constructed a deep learning model based on DNN to predict the CO2 adsorption performance and CO2/N2 selectivity of materials under low pressure environment (0.15 bar CO2,0.85 bar N2)[21]. They used 340,000 MOFs structures for training, which were divided into training, validation and test sets at a ratio of 7:1:2, and six geometric descriptors (such as SA, ρ, etc.), Atomic pockets (The-bag of Atoms, BOA), Chemical Motif Density (Chemical Motif Density, CMD), atomic Property weighted Radial Distribution Function (Atomic Property-weighted Radial Distribution Function, AP-RDF) were used as model inputs[89]. The AP-RDF descriptor is the most critical, representing the probability distribution of finding pairs of atoms within a given distance of the structure. The result of predicting the adsorption capacity is R2=0.965,RMSE=0.13 mmol/g, and the selectivity is R2=0.975 mmol/g,RMSE=10 mmol/g. Fanourgakis et al. Used the probability of a probe atom being adsorbed as a descriptor of the model, and constructed an RF model for predicting the adsorption of CO2, H2, and H2S molecules with dipole moments on nanomaterials[81]. The probability of a single particle being adsorbed by the porous material is found from the average Boltzmann factor, and the adsorption probability of probe atoms with different diameters (2.55Å, 3.3Å, 3.55Å, 4.4Å) to each MOFs is unique. The model divides 2932 CoRE MOFs into training set and test set according to the ratio of 1 ∶ 2, and the results show that the neutral particle with small dipole moment (0. 24 D) performs best in the model, and its R2=0.91~0.97.
Machine learning algorithms are also used to aid in the design of new materials. Zhang et al. Combined Monte Carlo Tree Search (MCTS) and recurrent neural network algorithm to study the effect of pore structure on the carbon capture performance of materials[87]. The information of 27 metal clusters and topological structures was extracted from eMOFs as Input, and then an organic linker was generated by MCTS. The nodes and Monte Carlo trees were formed by repeated selection, expansion, simulation and back propagation. One node corresponds to a Simplified Molecular Input Line Entry System (SMILES) symbol[90]. In the SMILES model, material molecules are represented as a sequence of strings, and the characters represent various chemical information, such as atoms, bonds, ring numbers, branches, or ions. RNN was used to train the SMILES model to obtain the organic ligands of the materials, and finally Zeo + + was used to design high-performance MOFs, which had good hydrophobicity in GCMC simulation and could work in humid environment. The structure-activity relationship shows that the pore size of MOFs with excellent carbon capture performance should be slightly smaller (such as 14.1818 Å) and the specific surface area should be slightly larger (such as 1750 m2/g).

3.2 Gas separation

Gas adsorption separation is another hot topic in the research of MOFs, especially in the capture and separation of Cx/Cy, noble gases and CO2[91]. Machine learning-assisted high-throughput screening technology is also an important means to explore the separation performance of a large number of MOFs for multi-component gas mixtures.

3.2.1 Noble gas separation

Radioactive rare gases have special applications in industrial production and daily life. At present, the separation of Xe/Kr by cryogenic distillation consumes a lot of energy. In the separation of rare gases by MOFs, the main feature is that the interaction between rare gases and MOFs is dominated by Van der Waals force, so most of the machine learning work only uses structural characteristics as descriptors, as shown in Table 4. Liang et al. First used GCMC to obtain the Xe/Kr gas separation performance of 303 991 GMOFs, and used LCD, PLD, Global cavity diameter (GCD), Global cavity diameter, ρ, SA, VF as descriptors to test the prediction performance of Ridge, LASSO, Elastic NET, SVM, BO, ANN, RF and XGB (Extreme Gradient Boosting) algorithms[92]. It was found that the R2 of adsorption capacity and selectivity predicted by XGB algorithm was 0. 951 and 0. 973 when the ratio of training set to testing set was 3 ∶ 7. The separation performance of 8 MOFs was better than that of Z11CBF-1000-2, and the selectivity of 38 MOFs was higher than that of SBMOF-1, with the highest value of 27. 68. The training results show that the weights of ρ, VF, Va and PLD are higher than those of other descriptors, and they are more important to predict the separation performance.
Ma et al. developed a deep neural network based on transfer learning, as shown in Figure 7. The idea is to fix the parameters of the input layer and the hidden layer after training, and change the relationship between the hidden layer and the output layer under different systems[93]. SA, GSA, VF, LCD and PLD were used as descriptors, and the DNN model was trained with 13,506 data at 100 bar and 243 K, and then transfer learning was used to predict the adsorption capacity of H2 and CH4 with a R2 of 0. 991 and 0. 980, while the R2 of predicting the separation selectivity of Xe/Kr was 0. 092. It is difficult to transfer the model trained by other gases to the Xe/Kr separation system based on the structural property descriptor. Anderson et al., based on the deep learning model of Multilayer Perceptron (MLP), used VF, GSA, LCD, PLD and PSSD (Pore Size Standard Deviation) as descriptors to predict the single-phase adsorption of Ar, Kr, Xe, Pore Size Standard Deviation and CH4 of 51,520 hMOFs, and the MDAPE was controlled at 9.0%, 9.8%, 11.1%, 7.8%, 9.4% and 10.4%, respectively[94]. Ideal Adsorbed Solution Theory (IAST) was then used to predict the separation performance of Ar/Kr, Kr/Xe, Ideal Adsorbed Solution Theory and N2/CH4, and it was found that 21 (Xe/Kr) of the Top 50 structures were identified by this method, indicating that machine learning combined with IAST can be used to evaluate the separation performance of mixed phases after predicting the single-phase adsorption performance, but it still needs further improvement[103].
Liu et al. introduced organic ligand atomic radius, electronegativity, ligand space, polarizability and other chemical information as descriptors, trained the BPNN model to predict the adsorption capacity, selectivity, separation potential and adsorption performance parameters of Xe/Kr, and divided the data obtained by GCMC into training, verification and test data sets according to 7 ∶ 1.5 ∶ 1.5[95]. The predicted R2 for different systems are all in the range of 0.94 ~ 0.98, and it is found that the length weight of organic ligand is the highest, about 58.2%. Therefore, it is still the development direction to incorporate more descriptors containing chemical information and improve the generalization ability of the model.

3.2.2 Cx/Cy separation

Similar to molecular sieves, the potential application of MOFs in chemical processes is the separation of Cx/Cy components such as homologues and azeotropes[104]. As shown in Table 4, recent studies have focused on predicting the separation performance of ethane/ethylene and other systems using supervised learning algorithms based on the combination of structural features and chemical information as descriptors. Halder et al. Used ρ, ASA, LCD, PLD four structural properties and 18 chemical information descriptors such as atom type and number to split the data set at a ratio of 8:92 to train RF to predict the ethane/ethylene selectivity of 115 302 hMOFs at 298 K and 1 bar, with a prediction accuracy R2 of about 0.89, and found that VF was the descriptor with the highest weight[96]. Wu et al. Used structure (ρ, GSA, SA, Va, PLD, LCD) and ligand defect ratio and chemical information (atomic number, metal cluster unsaturation and Zr element ratio) as descriptors, 10-Flod cross-validation, and 8:Linear regression, decision tree, random tree forest, support vector machine, K-nearest neighbor regression and gradient boosting tree (GBT) were compared to predict the properties of ethane/ethylene with various defects of UiO-66, such as Henry coefficient, adsorption heat, working adsorption capacity, elastic modulus and shear modulus[97]. The LR prediction accuracy is found to be high, with a R2 close to 1, and the most important descriptor is the Henry coefficient (KH), with a weight close to 40% at 1.0 bar, followed by GSA (23.5%), TDU (15.5%), DUC (8.9%) and Zr proportion. As shown in Fig. 8, the path leading to the target adsorbent for ethane/ethylene separation is thus proposed, that is, the structure with GSA≤1153.4 m2/g, Va≤0.54 cm3/g, PLD ≤ 5.55 Å, and LCD ≤ 8.77 Å has a probability of 84.2% to meet the performance requirement.
Among the other separation systems, Tang et al. Predicted the separation performance of up to 98 MOFs by using the GCMC simulation results of 66 pairs of near-azeotropes (C4, C7) of 6 MOFs as the training data set and 25 descriptors such as common structural properties as well as chemical information[98]. It is found that the error is larger when only MRGP is used at low pressure and low adsorption capacity, while the error can be effectively reduced by combining machine learning with IAST model. Qiao et al. Used BPNN and DT to predict the separation performance of p-xylene, m-xylene and o-xylene only based on the structural characteristics of VF, ASA, LCD, PLD and ρ[99]. The R2 predicted by BPNN is above 0. 9, and the MAE is about 0. 26. It is found that the weights of PLD and LCD descriptors are the highest, which are 0. 37 and 0. 28, while the other descriptors are between 0. 12 and 0. 16, which indicates that the pore has an important impact on the separation.

3.2.2 Other gas separation

In terms of other gas separation, Cho et al. Also predicted the CH4/H2S separation performance of CoRE-MOFs based on RF algorithm, and found that the combination of structural characteristics with chemical information such as bond angle, charge and dipole moment of metal elements could reduce the predicted RMSE from 0.3 to 0.1[100]. Yan et al. Used LCD, VF, SA, PLD, ρ, KH and Qst as descriptors to train RF, GBT and XGB models to predict the separation performance of 6013 CoRE MOFs at 1 bar and 298 K for O2/N2, including selectivity, working adsorption capacity, regeneration performance and adsorption performance, according to the ratio of training set, validation set and test set of 7 ∶ 1.5 ∶ 1.5[101]. The average R2 predicted by XGB is 0.93, which is slightly higher than that of GBT (0.92) and RF (0.91), and it is found that the weight of KH is the highest, and the separation performance is the best when it is located in the interval of 1.2~3.2×10-6mol/kg/Pa. Zhou et al. Used SVM, RF, GBT, and DNN algorithms to predict the D2/H2 separation performance of 12 723 CoRE MOFs under 17-Fold cross-validation, and found that the DNN model improved the most after adding metal element species to the PLD, LCD, and VF descriptors[102]. Overall, SVM is the best, and the analysis shows that LCD should be slightly larger than the size of D2, exerting the kinetic quantum sieving KQS separation effect. However, when PLD is close to H, repulsion occurs, and various structural features need to be optimized to achieve optimal performance.
In this kind of machine learning research, the separation of mixed phase gases is more complex, so structural properties and chemical information descriptors have been widely used. However, it is still necessary to further develop new descriptors and algorithms to improve the accuracy of prediction and model transfer.

3.3 Physicochemical properties of MOFs

3.3.1 Adsorption heat pump

Adsorption-driven Heat Pumps (AHPs) is a new device that realizes energy conversion through Adsorption and desorption between adsorbent and adsorbate (working fluid), which has attracted wide attention because of its energy-saving and environmental protection characteristics[105].
Li et al. Predicted the energy efficiency ratio (Coefficient of Performance, COP) of ethanol/CoRE-MOFs working pair in single-stage and cascade systems, and first used PCA analysis to find that pore size is the factor affecting its Performance[106]. The decision tree shows that the working adsorption capacity of the low temperature stage > 0.18 G/G is the necessary condition for the system COP to be greater than 1.6. Four ML algorithms (MLR, DT, GBM, and RF) were then used to predict and evaluate the COPs of more than 3 million cascaded AHPs (shown in Figure 9)[107]. In the case of introducing Henry coefficient as a chemical descriptor, 20% of the data is used as a training data set to compare the four algorithms. It is found that the prediction accuracy of RF is the highest and the (R2=0.95),DT is close to that of GBM algorithm, while the prediction effect of MLR algorithm is the worst. The analysis shows that the Va is the main factor affecting the COP in the low temperature level descriptor, while the Henry coefficient is the key factor determining the COP in the high temperature level descriptor.
Shi et al. Used four ML models (BPNN, DT, RF, and SVM) to predict the performance of 6013 methanol/CoRE-MOFs working pairs in heat pump applications[34]. Six classes of descriptors (ρ, Qst, LCD, VF, SA, and PLD) were taken as input. The algorithm with the highest accuracy was also RF (R = 0.86), followed by BPNN (R = 0.84). Further analysis obtained the weight of descriptors, which was consistent with the results of univariate analysis. Qst and ρ were the more important descriptors. DT elucidated the path leading to high-performance MOFs (ρ<885.10 kg/m3,Qst>42.16,ϕ<0.74). Then, Shi et al. Studied the COP of 6013 CoRE MOFs working pairs under three working conditions (heat pump, refrigeration and ice making) at heat source temperatures of 355 K and 365 K, respectively[108]. The descriptors of the structures with excellent performance were found to be clustered in the range of ρ=434~1582 kg/m3, Qst=34~35 kJ/mol, LCD = 7.00 Å ~ 22.00 Å, VF = 0.65 ~ 0.90, and SA=1550~3000 m2/cm3 by one-way analysis of variables. It was found that nearly 80% of the selected structures from 137,953 hMOFs met the above conditions. Meanwhile, they analyzed the economy of the application of MOFs working pairs in AHPs by using Equation 8, where Ctotal represents the total cost, Cequipment represents the equipment cost, Ccycle represents the cycle consumption cost, and CMOF represents the material cost. With the development of new high-performance MOFs materials, the proportion of Cequipment decreases while the Ccycle increases gradually, and further research shows that the Ctotal can be reduced by improving the Ccycle and other means. In addition, they trained six machine learning algorithms: KRR, BPNN, SVM, DT, RF and Gradient Boosting Regression Trees (GBRT) to predict Gradient Boosting Regression Trees, among which RF has the highest prediction accuracy, and its test set results are R = 0.876, MAE = 0.166, RMSE = 0.262. It is found that the metal type descriptors of MOFs have a low impact on the economy of AHPs, while the Qst and ρ have a significant impact. Finally, they experimentally synthesized and verified the economic cost of a class of MOF materials, DOTSOV03, which is consistent with the simulation results, as low as 1 $/kJ.
C t o t a l = C e q u i p m e n t + C c y c l e + C M O F

3.2 Catalysis field

MOFs have been widely used in the field of catalysis because of their good stability, porosity and periodicity[109]. The design of new catalysts needs to consider all kinds of factors, including temperature, pressure, time, pH value and so on. The relationship between these multiple variables and their effects on the catalytic performance of MOFs are extremely complex. However, the structural properties vary with these factors, so it is necessary to further study the catalytic performance of their structural properties (such as surface area, pore size, porosity, charge distribution, etc.) Under different conditions[110].
Liu et al. Summarized the data of about 100 literatures using MOFs to catalyze the fixation of carbon dioxide to cyclic carbonate[95]. The data sets of metal clusters and organic ligands of MOFs, and the reactants and products in the fixed reaction of Stochastic Gradient Descent were selected to construct the data set, and several ML algorithms such as SVM, KNN, DT, Stochastic Gradient Descent (SGD) and ANN were trained to predict the catalytic performance of MOFs for carbon dioxide at a ratio of 8 ∶ 2, and the advantages of several ML algorithms were compared from the perspectives of precision, recall and F1 score (harmonic mean of precision and recall). The results show that the accuracy of SVM, SGD and ANN models is higher, and their F1 scores are 0. 92, 0. 93 and 0. 92, respectively, while the accuracy of KNN and DT models is only 0. 82 and 0. 81. In order to further improve the accuracy of the training model, they predicted the catalytic performance of 1311 hMOFs by combining SVM, SGD and ANN, and found that the best metal ions in the catalytic fixation of CO2 were Mn, V, Cu, Ni, Zr and Y, and the corresponding ligands were TACTMB, TDCBPP, TCPP and H3L. It was speculated that the 24 MOFs combined by these 6 metals and 4 ligands might have excellent performance in the catalytic fixation of carbon dioxide.
Chen et al. Extracted the phase diagram and morphological relationship from the experimental images of powder X-ray diffraction and Scanning Electron Microscopy (SEM), and trained the Multivariate Decision Tree (MDT) model[111]. The phase boundary of the high-dimensional Synthetic space is predicted with a ratio of 9 ∶ 1, and the data are balanced by SMOTE (Synthetic Minority Oversampling Technique).From numerous influencing factors (temperature, HfCl4, H2BPDC, DMF,Concentration of H2O and HCO2H), the effects of temperature and concentration of H2O and HCO2H were found to be the most significant[112]. And 1000 different MDTs were constructed as a random forest to predict the phase distribution of NMOFs (mainly UiO-67, hcp-UiO-67, hxl-UiO-67) with the concentration of H2O and HCO2H as descriptors at a fixed temperature, and the average accuracy of the model reached 80%. Subsequently, they used transmission electron microscopy (TEM) and mass spectrometry (MS) to observe the HF6 versus HF12SBUs(Secondary Building Units,SBUs) components in the MOFs synthesis solution. The crystal growth mechanism of first forming SBUs, then exchanging groups on SBUs, and then crosslinking into a solid network was confirmed. At the same time, by studying the relationship between the phase diagram and the crystal morphology and its growth conditions, it is found that the formation of Hf12SBUs in solution is very important to determine the phase of the crystal. For this purpose, the CNN model was trained on the crystal thickness/length of 228 SEM scanning images, and 31 labeled images were predicted with an R2 of 0.81, and the MAE of length and thickness were 0.1 and 0.14, respectively. They also found that the thickness and lateral size of hxl-UiO increased with the increase of SBUs concentration in solution, and the growth of NMOFs crystals in specific forms could be quantitatively predicted by ligand solubility and regulator concentration. Based on this principle, synthesis steps could be designed to obtain NMOFs with enhanced matrix transport and improve their catalytic activity for olefin hydrogenation.

3.3.3 Stability

With the continuous expansion of the application fields of MOFs, researchers are increasingly concerned about whether their structures can be stable in practical application environments.
Nandy et al. Collected thousands of literature data on the stability of MOFs, as shown in Figure 10, and used Natural Language Processing (NLP) to identify the stability of 2179 MOFs in the CoRE-MOFs database after solvent activation, and the thermal stability of 3132 structures under thermogravimetric analysis[113]. The crystal files of MOFs were collected simultaneously, SBUs and organic ligands were sorted out, and Zeo + + was used to calculate information such as pore size, specific surface area, pore volume and density. Based on the cif file, the modified autocorrelation function was established by the connection of the central atoms of the metal cluster or the N atoms in the porphyrin ring, which was used as a descriptor to predict the solvent activation stability and thermal stability of the collection structure by ANN, and its R2=0.76,AUC=0.79 showed that the MAE of thermal stability was 47 ℃. The thermal stability of MOFs was further studied, and the AUC was increased to 0. 81 and the MAE was reduced to 44 ℃ after adding the Gaussian process based on the clustering principle[114].
Water stability is another property of particular interest. Batra et al. Divided MOFs into four categories according to their water Stability: Stable, High Kinetic Stability, Low Kinetic Stability, and Unstable[115]. Recursive Feature Elimination (RFE) was then used to reduce the dimensionality of the optimized features according to the three chemical properties of MOFs, namely, metal clusters, organic ligands, and metal-ligand molar ratios, and the RFE test accuracy of the two and three models was improved from 80% and 64% to 83% and 71%, respectively. Based on these two features, SVM, RF and GB machine learning models were trained to predict the water stability of 207 MOFs under 5-Fold cross-validation, and the R2 of the two models were 88% and 71%, respectively. The water stability of other 10 kinds of MOFs was predicted, and the R2 of the second and third kinds of MOFs were 0. 9 and 0. 6, respectively. The structure-activity relationship analysis using the two-class RF model shows that MOFs with large atomic radius (> 1.44 Å) and low ionization potential (< 7.5 eV) metal atoms, as well as low number of six-membered rings (< 0.04) and high cyclic divalent node (> 0.01) ligands, have high water stability.

3.3.4 Electrochemical property

The electrochemical properties of MOFs significantly affect their application prospects in the fields of electrocatalysis, resistance sensing and electrode materials, while most MOFs structures are nearly insulating due to their high porosity and covalent/coordination bonding.Therefore, researchers have begun to explore the electrochemical properties of MOFs, especially the point charge, in order to redesign or modify them to obtain suitable structures[116].
Raza et al. constructed a Message Passing Neural Network (MPNN) model[117]. The ratio of training, validation and test sets is 7:1:2. The local bonding environment of atoms in MOFs is used as a descriptor to predict the point charge of each atom. The stochastic gradient descent method is used to optimize the model, and it is found that the model has the best prediction effect under the hyperparameters of R = 4 (element embedding dimension), K = 10 (hidden node feature dimension), and t = 30 (message passing time steps). Based on this, the charge of 2266 MOFs calculated by DFT was predicted, and the MAE between the predicted value and DDEC (Density Derived Electrostatic and Chemical) charge was 0. 025, while the prediction time of this method was about 3 s, which significantly reduced the computational cost.
Rosen et al. Predicted the conduction bands of 15,713 MOFs using descriptors such as geometry, energy, band gap, charge density, density of States, partial charge, spin density, and bond order[46]. Kernel ridge regression, crystal map convolutional neural network, and smooth overlap model of atomic positions were trained after data set allocation in 8 ∶ 2 ratio, and the results showed that the prediction effect of crystal map convolutional neural network model was the best, and its R2=0.88,MAE=0.27 eV. This model can predict MOFs materials with desired electronic structure characteristics, and screen out samples with low band gap in order to obtain MOFs with high conductivity. Finally, it is found that the band gap of Ag(DCl)2(DCl=2,5-Cl,Cl-N,N0-dicyanoquinone diamine) is 0.71 eV, which has high conductivity. The high band gap Sr[C2H4(SO3]2] (8.36 eV) is nearly insulating. Korolev et al. Designed a Gradient Boosting Decision Trees (GBDT) model by combining DDEC and Qeq charge calculation methods[118]. The intrinsic properties of the corresponding site element (such as atomic number, covalent radius and solution temperature) and the positional information of the adjacent covalent/coordination environment (such as the number of atomic neighbors and topological distribution) are used as descriptors.The GBDT model was trained to predict the charge of some structures in CoRE MOFs by using a 9 ∶ 1 ratio of the training set to the test set and 10-Fold cross-validation, and the predicted MAD was found to be as low as 0. 01 e compared with the DDEC charge. The difference between the atomic local environment descriptors of MOFs and COFs is shown in Figure 11. Under this condition, the model is applied to CURATED COFs, and the prediction accuracy is still high, with a MAD of only 0.05 e. In addition, the simulation results of carbon dioxide adsorption characteristics calculated by this method are consistent with the results of DDEC charge simulation, which shows the effectiveness and high accuracy of this method.
Kancharlapalli et al. Used chemical information such as atomic properties (electronegativity, first ionization potential) and its coordination environment (number of adjacent atoms, average bonding distance, average electronegativity of bonding atoms, average first ionization potential of bonding atoms) as descriptors.Linear regression, neural network, gradient boosted decision tree and random forest were trained to predict 10 under 10-Fold cross-validation with a ratio of 8 ∶ 2.143 CoRE MOFs possess the DDEC Charge of the ASR (All Solvent Removed) structure with an RF training accuracy of All Solvent Removed, and the CM5 (Charge Model 5) charge is predicted with a higher accuracy of R2=0.9969,MAE=0.010[119]. The DDEC and CM5 charges on another 404 MOFs were predicted by these two models, and the R2 were 0. 9899 and 0. 9932, and the MAE were 0. 026 and 0. 013, respectively, indicating that the RF model can be applied to the charge prediction of other porous crystals and zeolites.

3.3.5 Other

The experimental specific surface area of MOFs is usually calculated by gas adsorption method (BET). However, for structures with large specific surface area, this method may overestimate the specific surface area by considering non-monolayer adsorbed molecules as monolayer adsorption (i.e., pore filling pollution problem). Datar et al. Proposed a data-driven method to accurately predict the surface area of MOFs[120]. They trained the LASSO model based on the isothermal adsorption curves of 328 MOFs at a ratio of 4:6 to predict their actual monolayer area. It is found that the model can predict the true monolayer area of most MOFs, and the deviation is less than 20%. Only 18 MOFs (5. 5%) were predicted by the model with a deviation of more than 20%, but the deviation was controlled below 30%, which was more accurate than the BET method. They further simulated the N2(77 K) isothermal adsorption curves of 68 MOFs. As shown in Fig. 12, the true monolayer area deviation predicted by ML was found to be all below 20%, while the BET method overestimated the area of 20 of these materials. The larger the area, the higher the deviation, up to 77%.
Syah et al. Studied the composite Layered Double Hydroxide composed of metal-organic framework and Layered Double Hydroxide (LDH)[121]. Temperature and initial thallium (Tl) concentration were used as descriptors to train the ANN model to predict the equilibrium concentration of heavy metal ions in the solution, and the R2 predicted by the ANN model were all greater than 0. 999. Further analysis showed that the initial Tl concentration in the solution had a greater impact on the equilibrium concentration than temperature.

4 Conclusion and prospect

To sum up, machine learning-assisted high-throughput screening of metal-organic frameworks has developed rapidly in recent years. The reason is that the number of MOFs has shown an explosive growth trend, and the scope and conditions of their applications have been expanding. High-throughput screening is the mainstream method for goal-oriented discovery of high-performance target MOFs and exploration of structure-activity relationship. However, the screening process generates a large amount of data, has many feature dimensions, and consumes significant computing resources, which is complementary to machine learning. The combination of the two is like adding wings to a tiger, which can efficiently and accurately study the properties of MOFs and their performance in various applications, deeply clarify their structure-activity relationship, and guide the design strategy.
In terms of descriptors, the current research mainly uses supervised learning to screen MOFs materials, and the core of the model data set is generally all kinds of descriptors with labels. Among them, the descriptors of geometric parameters and chemical properties are easy to obtain, while the descriptors of topological structure and energy potential information are difficult to obtain due to factors such as technology, time and computational cost, and the process of labeling is complex, which brings many difficulties to the training of supervised learning models. Therefore, it is necessary to continue to develop more accessible descriptors and increase the input features and information of the model, so as to improve the performance of the machine learning model and find the target MOFs in various applications more efficiently and accurately. On the other hand, other machine learning algorithms, such as semi-supervised learning and unsupervised learning models, can be developed to train ML models to predict and discover excellent MOFs in unlabeled data sets.However, this may require more massive data to provide implicit relationships, more accurate machine learning algorithms to accurately discover materials, and matching experimental verification strategies. In terms of the MOFs structure database used by machine learning, although the prediction accuracy of the current machine learning model is relatively high, almost all of them are the original database. Especially when training and predicting the performance of CoRE MOFs, the algorithm based on tree classification usually has better performance, such as RF, which may lead to poor generalization performance of the model. The migration ability of the algorithm needs to be further enhanced, for example, the machine learning model trained by eMOFs is used to predict the performance of hMOFs, even COFs, Zeolite and other porous structures. And because of the "black box" characteristics of most machine learning models, the machine learning models trained by research usually only stay at the level of predicting the optimal MOFs in the existing database[122]. Therefore, it is the proper meaning of the future development of machine learning to develop optional, optimizable and interpretable models, such as LR, DT, KNN, etc., so as to directly introduce the design strategies of new high-performance MOFs, or to clearly reveal the implicit relationship characteristics of materials.
In addition, there is often a gap between the simulation results and the experimental synthesis, and researchers have screened out a large number of excellent hMOFs, but the actual experimental results may not be satisfactory. When developing machine learning algorithms, it is possible to reduce such errors by introducing the differences between experiments and simulations, such as solvent activation and crystal defects, as descriptors. Moreover, although machine learning algorithms and molecular simulation are combined in high-throughput screening, most of the current studies are still relatively independent processes, and machine learning analysis and prediction are generally used after molecular simulation. It is envisaged that machine learning will be deeply integrated into the molecular simulation or experimental process, for example, some parameters can be embedded into the molecular simulation calculation or experimental design process after being predicted only by machine learning. In a word, machine learning-assisted high-throughput screening of MOFs still faces some problems, but it has developed rapidly in descriptor acquisition, algorithm development and model application, and will be an indispensable key technology for efficient design and development of MOFs in the future.

References

Publishing order | Descend order by publishing year | Descend order by cited within
[1]

Heidarinejad Z, Dehghani M H, Heidari M, Javedan G, Ali I, Sillanpää M. Environ Chem Lett, 2020, 18 (2): 393.
[2]

Van Speybroeck V, Hemelsoet K, Joos L, Waroquier M, Bell R G, Catlow C R A. Chem. Soc. Rev., 2015, 44 (20): 7044.
[3]

Wang H J, Wang M D, Liang X, Yuan J Q, Yang H, Wang S Y, Ren Y X, Wu H, Pan F S, Jiang Z Y. Chem. Soc. Rev., 2021, 50 (9): 5468.
[4]

Liu Z L, Li W liu H, Zuang X D, Li S. Acta Chim Sinica, 2019, 77 (4): 323.
[5]

(刘治鲁, 李炜, 刘昊, 庄旭东, 李松. 化学学报. 2019, 77(4): 323.)
[6]

Altintas C, Altundal O F, Keskin S, Yildirim R. J Chem Inf Model, 2021, 61 (5): 2131.
[7]

Ahmed A, Seth S, Purewal J, Wong-Foy A G, Veenstra M, Matzger A J, Siegel D J. Nat. Commun., 2019, 10 (1): 1.
[8]

Herm Z R, Bloch E D, Long J R. Chem. Mater., 2014, 26 (1): 323.
[9]

Tan K, Zuluaga S, Gong Q H, Gao Y Z, Nijem N, Li J, Thonhauser T, Chabal Y J. Chem. Mater., 2015, 27 (6): 2203.
[10]

Zhao X, Wang Y X, Li D S, Bu X H, Feng P Y. Adv. Mater., 2018, 30 (37): 1705189.
[11]

Janiak C. J Eur. J. Inorg. Chem, 2012, 2625: 2634.
[12]

Li W, Xia X X, Li S. ACS Appl. Mater. Interfaces, 2019, 12 (2): 3265.
[13]

Chen L Y, Xu Q. Matter, 2019, 1 (1): 57.
[14]

Jiao L, Wang Y, Jiang H L, Xu Q. Adv. Mater., 2018, 30 (37): 1703663.
[15]

Lee J, Farha O K, Roberts J, Scheidt K A, Nguyen S T, Hupp J T. Chem. Soc. Rev., 2009, 38 (5): 1450.
[16]

Chu F L, Hu J L, Wu C L, Yao Z G, Tian J, Li Z, Li C L. ACS Appl. Mater. Interfaces, 2018, 11 (4): 3869.
[17]

Moghadam P Z, Li A, Wiggin S B, Tao A, Maloney A G, Wood P A, Ward S C, Fairen-Jimenez D. Chem. Mater., 2017, 29 (7): 2618.
[18]

Wilmer C E, Leaf M, Lee C Y, Farha O K, Hauser B G, Hupp J T, Snurr R Q. Nat. Chem., 2012, 4 (2): 83.
[19]

Colón Y J, Gomez-Gualdron D A, Snurr R Q. Cryst. Growth Des., 2017, 17 (11): 5801.
[20]

Gurnani R, Yu Z Z, Kim C, Sholl D S, Ramprasad R. Chem. Mater., 2021, 33 (10): 3543.
[21]

Suyetin M. Faraday Discuss., 2021, 231: 224.
[22]

Burner J, Schwiedrzik L, Krykunov M, Luo J, Boyd P G, Woo T K. J. Phys. Chem. C, 2020, 124 (51): 27996.
[23]

Deng X M, Yang W Y, Li S H, Liang H, Shi Z N, Qiao Z W. Appl. Sci., 2020, 10 (2): 569.
[24]

Islamoglu T, Idrees K B, Son F A, Chen Z, Lee S-J, Li P, Farha O K. J. Mater. Chem. A, 2022, 10 (1): 157.
[25]

Jain A K, Mao J, Mohiuddin K M. Comput., 1996, 29 (3): 31.
[26]

Friedman J H. Ann Appl Stat, 2001, 29 (5): 1189.
[27]

Pang B, Nijkamp E, Wu Y N. J Educ Behav Stat, 2019, 45 (2): 227.
[28]

Chen K M, Cofer E M, Zhou J, Troyanskaya O G. Nat. Methods, 2019, 16 (4): 315.
[29]

Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V. J Mach Learn Res, 2011, 12: 2825.
[30]

Chung Y G, Camp J, Haranczyk M, Sikora B J, Bury W, Krungleviciute V, Yildirim T, Farha O K, Sholl D S, Snurr R Q. Chem. Mater., 2014, 26 (21): 6185.
[31]

Li S, Chung Y G, Simon C M, Snurr R Q. J. Phys. Chem. Lett, 2017, 8 (24): 6135.
[32]

Sarkisov L, Bueno-Perez R, Sutharson M, Fairen-Jimenez D. Chem. Mater., 2020, 32 (23): 9849.
[33]

Dubbeldam D, Calero S, Ellis D E, Snurr R Q. Mol Simul, 2016, 42 (2): 81.
[34]

Kadioglu O, Keskin S. Sep. Purif. Technol., 2018, 191: 192.
[35]

Shi Z N, Liang H, Yang W Y, Liu J, Liu Z L, Qiao Z W. Chem. Eng. Sci., 2020, 214: 115430.
[36]

Adatoz E, Avci A K, Keskin S. Sep. Purif. Technol., 2015, 152: 207.
[37]

Altundal O F, Haslak Z P, Keskin S. Ind. Eng. Chem. Res., 2021, 60 (35): 12999.
[38]

McDaniel J G, Li S, Tylianakis E, Snurr R Q, Schmidt J. J. Phys. Chem. C, 2015, 119 (6): 3143.
[39]

Zhong C L, Liu D H, Yang Q Y. Structure-Property Relationship and Design of Metal-Organic Frameworks. Beijing: Science Press, 2013. 35.
[40]

仲崇立, 刘大欢, 阳庆元. 金属-有机骨架材料的构效关系及设计. 北京: 科学出版社, 2013. 35.).
[41]

Keskin S, Liu J C, Rankin R B, Johnson J K, Sholl D S. Ind. Eng. Chem. Res., 2009, 48 (5): 2355.
[42]

Jiang J W, Babarao R, Hu Z Q. Chem. Soc. Rev., 2011, 40 (7): 3599.
[43]

Smit B, Maesen T L. Chem. Rev., 2008, 108 (10): 4125.
[44]

Düren T, Bae Y-S, Snurr R Q. Chem. Soc. Rev., 2009, 38 (5): 1237.
[45]

Wang M, Wang T, Cai P Q, Chen X D. Small Methods, 2019, 3 (5): 1900025.
[46]

Oliynyk A O, Mar A. Acc. Chem. Res., 2018, 51 (1): 59.
[47]

Krishnapriyan A S, Haranczyk M, Morozov D. J. Phys. Chem. C, 2020, 124 (17): 9360.
[48]

Rosen A S, Iyer S M, Ray D, Yao Z, Aspuru-Guzik A, Gagliardi L, Notestein J M, Snurr R Q. Matter, 2021, 4 (5): 1578.
[49]

Ward L, Liu R Q, Krishna A, Hegde V I, Agrawal A, Choudhary A, Wolverton C. Phys. Rev. B, 2017, 96 (2): 024104.
[50]

Chong S, Lee S, Kim B, Kim J. Coord Chem Rev, 2020, 423: 213487.
[51]

Moreno-Torres J G, Sáez J A, Herrera F. IEEE Trans Neural Netw Learn Syst, 2012, 23 (8): 1304.
[52]

Chung M K, Worsley K J, Nacewicz B M, Dalton K M, Davidson R J. NeuroImage, 2010, 53 (2): 491.
[53]

Zhang M-L, Zhou Z-H. Pattern Recognit, 2007, 40 (7): 2038.
[54]

Raccuglia P, Elbert K C, Adler P D, Falk C, Wenny M B, Mollo A, Zeller M, Friedler S A, Schrier J, Norquist A J. Nature, 2016, 533 (7601): 73.
[55]

Gligorov V V, Williams M. J. Instrum., 2013, 8 (02): P02013.
[56]

Belgiu M Dr?gu瘙塅 L. ISPRS J. Photogramm, 2016, 114: 24.
[57]

Cherkassky V, Ma Y. Neural Netw, 2004, 17 (1): 113.
[58]

Suykens J A, Vandewalle J. Neural Process Lett, 1999, 9 (3): 293.
[59]

Montavon G, Samek W, Müller K-R. Digit Signal Process, 2018, 73: 1.
[60]

Medsker L R, Jain L. Comput Aided Des Appl, 2001, 5: 64.
[61]

Erb R J. Pharm. Res., 1993, 10 (2): 165.
[62]

Angeline P J, Saunders G M, Pollack J B. IEEE Trans Neural Netw Learn Syst, 1994, 5 (1): 54.
[63]

Whitley D. Stat Comput, 1994, 4 (2): 65.
[64]

Sillar K, Hofmann A, Sauer J. J. Am. Chem. Soc., 2009, 131 (11): 4143.
[65]

Bobbitt N S, Snurr R Q. Mol Simul, 2019, 45: 1069.
[66]

Ren J, Musyoka N M, Langmi H W, Swartbooi A, North B C, Mathe M. Int. J. Hydrog. Energy, 2015, 40 (13): 4617.
[67]

Maryam P, Ehsan M, David W, Suib S L, Ranjan S. ACS Comb Sci, 2017, 19 (10): 640.
[68]

Fernandez M, Boyd P G, Daff T D, Aghaji M Z, Woo T K. J. Phys. Chem. Lett, 2014, 5 (17): 3056.
[69]

Fernandez M, Barnard A S. ACS Comb Sci, 2016, 18 (5): 243.
[70]

Lin X, Jia J H, Hubberstey P, Schrder M, Champness N R. CrystEngComm, 2007, 9 (6): 438.
[71]

Getman R B, Bae Y S, Wilmer C E, Snurr R Q. Chem. Rev., 2012, 112 (2): 703.
[72]

Gangu K K, Maddila S, Mukkamala S B, Jonnalagadda S B. J. Energy Chem., 2018, 30 (03): 140.
[73]

Cao Y, Dhahad H A, Zare S G, Farouk N, Anqi A E, Issakhov A, Raise A. Int. J. Hydrog. Energy, 2021, 46 (73): 36336.
[74]

Jiang Y C, Zhang G F, Wang J J, Vaferi B. Int. J. Hydrog. Energy, 2021, 46 (46): 23591.
[75]

Bucior B J, Bobbitt N S, Islamoglu T, Goswami S, Gopalan A, Yildirim T, Farha O K, Bagheri N, Snurr R Q. Mol. Syst. Des. Eng., 2019, 4 (1): 162.
[76]

Giappa R M, Tylianakis E, Di Gennaro M, Gkagkas K, Froudakis G E. Int. J. Hydrog. Energy, 2021, 46 (54): 27612.
[77]

Ahmed A, Siegel D J. Patterns, 2021, 2 (7): 100291.
[78]

Glasby L T, Moghadam P Z. Patterns, 2021, 2 (7): 100305.
[79]

Martin R L, Simon C M, Smit B, Haranczyk M. J. Am. Chem. Soc., 2014, 136 (13): 5006.
[80]

Beauregard N, Pardakhti M, Srivastava R, Modeling. J Chem Inf Model, 2021, 61 (7): 3232.
[81]

Pardakhti M, Nanda P, Srivastava R. J. Phys. Chem. C, 2020, 124 (8): 4534.
[82]

Wang R H, Zhong Y S, Bi L M, Yang M L, Xu D G. ACS Appl. Mater. Interfaces, 2020, 12 (47): 52797.
[83]

Fanourgakis G S, Gkagkas K, Tylianakis E, Froudakis G. J. Phys. Chem. C, 2020, 124 (13): 7117.
[84]

Kim S Y, Kim S I, Bae Y S. J. Phys. Chem. C, 2020, 124 (36): 19538.
[85]

Lee S, Kim B, Cho H, Lee H, Lee S Y, Cho E S, Kim J, Interfaces. ACS Appl. Mater. Interfaces, 2021, 13 (20): 23647.
[86]

Taw E, Neaton J B. Adv. Theory Simul., 2022, 5 (3): 2100515.
[87]

Zendehboudi S, Bahadori A, Lohi A, Elkamel A, Chatzis I. Energy Fuels, 2013, 27 (1): 401.
[88]

Dashti A, Bahrololoomi A, Amirkhani F, Mohammadi A H. J. CO2 Util., 2020, 41: 101256.
[89]

Zhang X Y, Zhang K X, Yoo H, Lee Y J. ACS Sustain. Chem. Eng., 2021, 9 (7): 2872.
[90]

Li S Y, Zhang Y J, Hu Y X, Wang B J, Sun S R, Yang X W, He H. J. Materiomics, 2021, 7: 1029.
[91]

Fernandez M, Trefiak N R, Woo T K. Journal of Physical Chemistry C, 2013, 117 (27): 14095.
[92]

Weininger D. J Chem Inf Model, 1988, 28 (1): 31.
[93]

Li J-R, Sculley J, Zhou H-C. Chem. Rev., 2012, 112 (2): 869.
[94]

Liang H, Jiang K, Yan T-A, Chen G H. ACS Omega, 2021, 6 (13): 9066.
[95]

Ma R, Colón Y J, Luo T. ACS Appl. Mater. Interfaces, 2020, 12 (30): 34041.
[96]

Anderson R, Biong A, Gomez-Gualdron D A. J. Chem. Theory Comput., 2020, 16 (2): 1271.
[97]

Liu Z W, Zhang K, Xia Q B, Wang X J, Huang B C, Xi H X. Chem. Eng. Sci., 2021, 243: 116772.
[98]

Halder P, Singh J K. Energy Fuels, 2020, 34 (11): 14591.
[99]

Wu Y, Duan H P, Xi H X. Chem. Mater., 2020, 32 (7): 2986.
[100]

Tang D, Gharagheizi F, Sholl D S. J. Phys. Chem. B, 2021, 125 (3): 926.
[101]

Qiao Z W, Yan Y L, Tang Y X, Liang H, Jiang J W. J. Phys. Chem. C, 2021, 125 (14): 7839.
[102]

Cho E H, Deng X P, Zou C L, Lin L-C. J. Phys. Chem. C, 2020, 124 (50): 27580.
[103]

Yan Y L, Shi Z N, Li H L, Li L F, Yang X, Li S H, Liang H, Qiao Z W. Chem. Eng. J., 2022, 427: 131604.
[104]

Zhou M S, Vassallo A, Wu J Z. J. Membr. Sci., 2020, 598: 117675.
[105]

Simon C M, Smit B, Haranczyk M. Comput Phys Commun, 2016, 200: 364.
[106]

Bao Z B, Chang G G, Xing H B, Krishna R, Ren Q L, Chen B L. Energy Environ. Sci., 2016, 9 (12): 3612.
[107]

De Lange M F, Verouden K J, Vlugt T J, Gascon J, Kapteijn F. Chem. Rev., 2015, 115 (22): 12205.
[108]

Li W, Xia X X, Cao M, Li S. J. Mater. Chem. A, 2019, 7 (13): 7470.
[109]

Li W, Xia X X, Li S. J. Mater. Chem. A, 2019, 7 (43): 25010.
[110]

Shi Z N, Yuan X Y, Yan Y L, Tang Y L, Li J J, Liang H, Tong L P, Qiao Z W. J. Mater. Chem. A, 2021, 9 (12): 7656.
[111]

McCarver G A, Rajeshkumar T, Vogiatzis K D. Coord Chem Rev, 2021, 436: 213777.
[112]

Erdem Günay M, Yıldırım R. Catal Rev Sci Eng, 2021, 63 (1): 120.
[113]

Chen P, Tang Z Y, Zeng Z M, Hu X F, Xiao L P, Liu Y, Qian X D, Deng C Y, Huang R Y, Zhang J Z, Bi Y L, Lin R K, Zhou Y, Liao H g, Zhou D, Wang C, Lin W B. Matter, 2020, 2 (6): 1651.
[114]

Chawla N V, Bowyer K W, Hall L O, Kegelmeyer W P. J Artif Intell Res, 2002, 16: 321.
[115]

Nandy A, Terrones G, Arunachalam N, Duan C, Kastner D W, Kulik H J. Sci. Data, 2022, 9 (1): 1.
[116]

Nandy A, Duan C, Kulik H J. J. Am. Chem. Soc., 2021, 143 (42): 17535.
[117]

Batra R, Chen C, Evans T G, Walton K S, Ramprasad R. Nat. Mach. Intell., 2020, 2 (11): 704.
[118]

Zanca F, Glasby L T, Chong S, Chen S, Kim J, Fairen-Jimenez D, Monserrat B, Moghadam P Z. J. Mater. Chem. C, 2021, 9: 13584.
[119]

Raza A, Sturluson A, Simon C M, Fern X. J. Phys. Chem. C, 2020, 124 (35): 19070.
[120]

Korolev V V, Mitrofanov A, Marchenko E I, Eremin N N, Tkachenko V, Kalmykov S N. Chem. Mater., 2020, 32 (18): 7822.
[121]

Kancharlapalli S, Gopalan A, Haranczyk M, Snurr R Q. J. Chem. Theory Comput., 2021, 17 (5): 3052.
[122]

Datar A, Chung Y G, Lin L-C. J. Phys. Chem. Lett, 2020, 11 (14): 5412.
[123]

Syah R, Al-Khowarizmi A, Elveny M, Khan A. Environ. Technol. Innov., 2021, 23: 101805.
[124]

Jablonka K M, Ongari D, Moosavi S M, Smit B. J. Am. Chem. Soc., 2020, 120 (16): 8066.

Funding

Fundamental Research Funds for the Central Universities(21621039)

GuangZhou Basic and Applied Basic Research Foundation(202201010433)

China Postdoctoral Science Foundation(2021M701413)

PDF (9950KB)

4008

Accesses

0

Citation

Detail
Sections
Recommended

/

〈 〉

Address of Editorial Office: No. 33 Beisihuan Xilu, Zhongguancun , Beijing 100190, China
Tel: +86 10 8262 7757
Fax: +86 10 8262 7757
E-mail: progchem@mail.las.ac.cn
http://www.progchem.ac.cn
    • 　
    Copyright © Editorial Office of Progress in Chemistry