A method for predicting porphyrin MOF band gap based on deep transfer learning

Abstract

The application provides a method for predicting the band gap of porphyrin MOF based on deep transfer learning. First, a small data set of porphyrin MOF is constructed, and the Average SOAP Kernel is used to expand the data set of porphyrin MOF. Then, the GCN is pre-trained by using the QMOF database to obtain a pre-training model. The pre-training model is fine-tuned by using a transfer learning method to obtain a porphyrin MOF band gap prediction model. Finally, the band gap is predicted by using the porphyrin MOF band gap prediction model. The application can predict the band gap of porphyrin MOF with high accuracy. The application combines deep learning and transfer learning to predict the band gap of porphyrin MOF, and only depends on crystallographic information files instead of manually made geometric descriptors and chemical descriptors. The band gap of a given porphyrin MOF is efficiently and accurately predicted, the prediction accuracy and efficiency are improved, and the time and economic cost of measuring the band gap of porphyrin MOF are reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of functional materials intelligent prediction, and particularly relates to a method for predicting a band gap of a porphyrin MOF based on deep transfer learning. BACKGROUND

[0002] In recent years, the overconsumption of fossil fuels has led to high emissions of greenhouse gases, carbon dioxide entering the atmosphere and global energy crisis. By using semiconductor materials to simulate natural photosynthesis to achieve photocatalytic CO2 reduction into valuable solar fuels, it is considered to be one of the best solutions to solve the above problems. Porphyrin MOF has adjustable band structure, rich active sites, large specific surface area, uniform cavity and other advantages, so that porphyrin MOF becomes a very promising photocatalytic material.

[0003] In order to realize photocatalytic CO2 reduction, the photocatalyst should have a suitable band gap value. The traditional method to obtain the band gap value of the material is to calculate by density functional theory (DFT), but due to the large system and complex structure of porphyrin MOF, the time and economic cost of DFT calculation is very large.

[0004] Deep learning has been proven to be an excellent method to find the hidden relationship between input features and labels, and many deep learning methods have been applied to MOF and good results have been achieved. However, data dependence is one of the most serious problems in deep learning. Compared with traditional machine learning methods, deep learning is very dependent on massive training data, because it needs a large amount of data to understand the potential patterns of the data. When predicting the band gap value of porphyrin MOF, due to the large system and complex structure of porphyrin MOF, the labeling cost is very high, so the data collection of porphyrin MOF is complex and expensive, which makes it very difficult to build a large-scale, high-quality annotated porphyrin MOF dataset required by traditional deep learning.

[0005] Transfer learning relaxes the assumption that training data must be independent and identically distributed with test data, and encourages us to use transfer learning to solve the problem of insufficient training data. In transfer learning, training data and test data do not need to be independent and identically distributed. It does not need to start from zero training of the model in the target domain, which can significantly reduce the demand for training data and training time in the target domain. However, the fine-tuning set of transfer learning also needs a certain amount of data, which cannot be too small.

[0006] Crystallographic information file (CIF) contains detailed information of each crystal, such as cell parameters, atomic coordinates, etc. It is the most commonly used file format for crystal structure description, analysis, dissemination and expression. The CIF file describes in detail the type and three-dimensional coordinate information of each atom in the MOF unit cell. This three-dimensional coordinate information actually contains various physical and chemical properties of the MOF itself. SUMMARY

[0007] In order to solve the problem of requiring a large number of annotated data sets for traditional deep learning to predict material performance, the application provides a method for predicting the band gap of porphyrin MOF based on deep transfer learning. The application first designs and constructs a porphyrin MOF data set by changing the porphyrin central metal on the collected porphyrin MOF data set, and then realizes data expansion of the porphyrin MOF data set by using Average SOAP Kernel. The problem of insufficient data for fine-tuning of transfer learning is solved, and then deep transfer learning is used to predict the band gap of porphyrin MOF, so as to finally solve the problem of insufficient training data for traditional deep learning.

[0008] The application combines deep learning and transfer learning methods, and can efficiently and accurately predict the band gap value of a given porphyrin MOF without constructing additional descriptors and using only CIF files as input. The application can significantly reduce the prediction cost, lower the use threshold and reduce the prediction time.

[0009] The technical scheme of the application is as follows:

[0010] 1) Collect a plurality of porphyrin-based MOF materials from the Cambridge Crystal Structure Database, optimize the original crystallographic information files of the plurality of porphyrin-based MOF materials, and calculate the corresponding band gaps to obtain the corresponding optimized crystallographic information files and band gaps of the plurality of porphyrin-based MOF materials, wherein the optimized crystallographic information files and the corresponding band gaps of the plurality of porphyrin-based MOF materials constitute an initial porphyrin MOF data set;

[0011] 2) Select a plurality of MOF samples with high similarity to porphyrin MOF from the QMOF database and form an expanded MOF data set, fuse the expanded MOF data set with the initial porphyrin MOF data set to obtain a fused porphyrin MOF data set, and mark the QMOF database from which the expanded MOF data set is screened out as a screened QMOF database;

[0012] 3) Pre-train a deep learning model using the screened QMOF database to obtain a pre-trained model;

[0013] 4) Fine-tune the pre-trained model using a transfer learning method according to the fused porphyrin MOF data set to obtain a porphyrin MOF band gap prediction model;

[0014] 5) Input the corresponding optimized crystallographic information file of the porphyrin-based MOF material to be predicted into the porphyrin MOF band gap prediction model to obtain the band gap of the porphyrin-based MOF material to be predicted.

[0015] In the 1), the porphyrin central metal of each porphyrin-based MOF material is changed to obtain a plurality of new porphyrin-based MOF materials, and the optimized crystallographic information files and the corresponding band gaps of the plurality of new porphyrin-based MOF materials together with the optimized crystallographic information files and the corresponding band gaps of the plurality of collected porphyrin-based MOF materials constitute an initial porphyrin MOF dataset.

[0016] In the 2), the global similarity between each MOF sample in the QMOF database and each porphyrin-based MOF material in the initial porphyrin MOF dataset is calculated by using the atomic position average smoothing overlap kernel method, the global similarity between each MOF sample in the QMOF database and each porphyrin-based MOF material in the initial porphyrin MOF dataset is summed and averaged to obtain the average similarity of each MOF sample in the QMOF database, and the MOF samples in the plurality of QMOF databases are selected in descending order of the average similarity and constitute an expanded MOF dataset.

[0017] In the 3), the deep learning model is a graph convolutional neural network GCN.

[0018] The beneficial effects of the present application are:

[0019] The method for predicting the porphyrin MOF band gap based on deep transfer learning of the present application first changes the porphyrin central metal on the collected porphyrin MOF structure to design and construct an initial porphyrin MOF dataset, and then uses the Average SOAPKernel to realize data expansion of the porphyrin MOF dataset, which solves the problem of insufficient data for fine-tuning in transfer learning, and then uses deep transfer learning to predict the porphyrin MOF band gap, thereby solving the problem of insufficient training data in traditional deep learning. At the same time, the present application analyzes the CIF file instead of constructing a feature descriptor, which saves a lot of data preprocessing time. The universality of the CIF file also greatly reduces the use threshold of the present method, and only an arbitrarily optimized CIF can be used for prediction. The present method improves the prediction accuracy and efficiency and reduces the time and economic cost of measuring the porphyrin MOF. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a flowchart of the method of the present application.

[0021] Figure 2 is a prediction result graph of the porphyrin MOF band gap prediction model on the prediction set when the PMOF-plus368 of the present application is fine-tuned. DETAILED DESCRIPTION

[0022] The technical solutions of the present application will be further described in detail below in combination with the drawings:

[0023] As Figure 1As shown, the present application comprises the following steps:

[0024] 1) Collect existing multiple porphyrin-based MOF materials from the Cambridge Crystal Structure Database, use density functional theory (DFT) to optimize the original crystallographic information files of the multiple porphyrin-based MOF materials and calculate the corresponding band gaps, obtain the optimized crystallographic information files and the band gaps corresponding to the multiple porphyrin-based MOF materials, and form an initial porphyrin MOF dataset from the optimized crystallographic information files and the corresponding band gaps of the multiple porphyrin-based MOF materials.

[0025] 1) In the specific implementation, the structure of the porphyrin-based MOF material is correct, and there is no free solvent. In the specific implementation, the central metal of each porphyrin-based MOF material can be changed multiple times to obtain multiple new porphyrin-based MOF materials, and the optimized crystallographic information files and the corresponding band gaps of the multiple new porphyrin-based MOF materials are combined with the optimized crystallographic information files and the corresponding band gaps of the multiple porphyrin-based MOF materials collected to form the initial porphyrin MOF dataset.

[0026] 2) Select multiple MOF samples with high similarity to porphyrin MOF from the QMOF (Quantum MOF) database and form an expanded MOF dataset, fuse the expanded MOF dataset with the initial porphyrin MOF dataset to obtain a fused porphyrin MOF dataset, and record the QMOF database after screening out the expanded MOF dataset as a screened QMOF database.

[0027] 2) In the specific implementation, the global similarity between each MOF sample in the QMOF database and each porphyrin-based MOF material in the initial porphyrin MOF dataset is calculated using the Average smooth overlap of atomic positions Kernel (Average SOAP Kernel), the global similarity between each MOF sample in the QMOF database and each porphyrin-based MOF material in the initial porphyrin MOF dataset is summed and averaged, the average similarity of each MOF sample in the QMOF database is obtained, and multiple MOF samples in the QMOF database are selected in descending order of the average similarity and combined to form the expanded MOF dataset.

[0028] 3) The screened QMOF database is used to pre-train a deep learning model to obtain a pre-trained model; the deep learning model is a graph convolutional neural network (GCN).

[0029] 4) According to the fused porphyrin MOF dataset, a porphyrin MOF band gap prediction model is obtained by fine-tuning the pre-trained model using a transfer learning method; during the fine-tuning process, the learning rate, batch size, and loss rate of the pre-trained model are adjusted.

[0030] 5) input the optimized crystallographic information file corresponding to the porphyrin-based MOF material to be predicted into the porphyrin MOF band gap prediction model, and obtain the band gap of the porphyrin-based MOF material to be predicted.

[0031] In a specific implementation, 17 porphyrin MOF materials (various organic ligands and various metal ions or clusters, various structures) are selected from the Cambridge Crystal Structure Database, and the porphyrin central metal of the 17 porphyrin-based MOF materials is changed multiple times to obtain 202 porphyrin-based MOF materials.

[0032] 34 porphyrin-based MOF materials are selected from the 202 porphyrin-based MOF materials and constitute PMOF-34 (with various porphyrin central metals, various structures, and uniform band gap distribution), which is used as a prediction set and is a unified standard for evaluating model performance. The remaining 168 porphyrin-based MOF materials constitute an initial porphyrin MOF data set, denoted as PMOF-168.

[0033] The first 200, 500, 800, 1100, 1400, 1700, and 2000 MOF samples with high average similarity in the QMOF database are sequentially taken, respectively merged with the PMOF-168 data set, and then the PMOF-plus368, PMOF-plus668, PMOF-plus968, PMOF-plus1268, PMOF-plus1568, PMOF-plus1868, and PMOF-plus2168 data sets are obtained, which are used as fine-tuning sets. The screened QMOF database is the QMOF data set after removing the first 2000 samples with high average similarity, and the pre-training model is obtained by pre-training the graph convolutional neural network GCN.

[0034] The merged porphyrin MOF data set is divided into a training set, a validation set, and a test set in a ratio of 8:0.5:1.5, and PMOF-34 is used as a prediction set.

[0035] After the model training is completed, the final effect of the model is evaluated on the test set and the prediction set, and R 2 , RMSE, and MAE are selected to quantitatively describe the accuracy of the model prediction results.

[0036] The model comprehensively tests the test set and the prediction set, and the performance is better when PMOF-plus368 and PMOF-plus968 are used as fine-tuning sets, and PMOF-plus368 needs to expand the data set, which achieves good performance. As Figure 2 shown, the horizontal coordinate of the coordinate system in the figure is the band gap value calculated by DFT, and the vertical coordinate is the band gap value predicted by the model. When PMOF-plus368 is used as a fine-tuning set, the R 2The point in the figure is mostly close to the straight line of y=x, and thus the method for predicting a porphyrin MOF band gap based on deep transfer learning can effectively predict the porphyrin MOF band gap.

[0037] Compared with a traditional data-driven method, the method can save the time for feature construction, and when the obtained porphyrin MOF data set is not much, the data set can be expanded by using the Average SOAP Kernel and deep transfer learning, so that the porphyrin MOF band gap can be accurately and efficiently predicted, and the time and economic cost is reduced.

[0038] The above only describes some embodiments of the present application, and it should be noted that those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments, or equivalently replace some technical features, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting the band gap of porphyrin MOF based on deep transfer learning, characterized in that: The steps include: 1) Collecting multiple porphyrin-based MOF materials from the Cambridge Crystal Structure Database, performing structural optimization on the original crystallographic information files of the multiple porphyrin-based MOF materials and calculating the corresponding band gaps, obtaining the optimized crystallographic information files and band gaps corresponding to the multiple porphyrin-based MOF materials, and forming an initial porphyrin MOF dataset from the optimized crystallographic information files and corresponding band gaps of the multiple porphyrin-based MOF materials; 2) screening multiple MOF samples with high similarity to porphyrin MOF from the QMOF database and forming an extended MOF dataset, fusing the extended MOF dataset with the initial porphyrin MOF dataset to obtain a fused porphyrin MOF dataset, and recording the QMOF database after filtering out the extended MOF dataset as the filtered QMOF database; 3) Pre-training the deep learning model using the screened QMOF database to obtain a pre-trained model; 4) Based on the fused porphyrin-MOF dataset, the pre-trained model was fine-tuned using the transfer learning method to obtain a porphyrin-MOF band gap prediction model; 5) inputting the optimized crystallographic information file corresponding to the porphyrin-based MOF material to be predicted into the porphyrin-based MOF band gap prediction model to obtain the band gap of the porphyrin-based MOF material to be predicted; In the step 1), the porphyrin central metal of each porphyrin-based MOF material is changed multiple times to obtain a corresponding plurality of new porphyrin-based MOF materials, and the optimized crystallographic information files and corresponding band gaps of the multiple new porphyrin-based MOF materials are combined with the collected optimized crystallographic information files and corresponding band gaps of the multiple porphyrin-based MOF materials to form an initial porphyrin MOF data set; In the above 2), the global similarity between each MOF sample in the QMOF database and each porphyrin-based MOF material in the initial porphyrin MOF dataset is calculated using the atomic position average smoothed overlapping kernel method, and the global similarities between each MOF sample in the QMOF database and each porphyrin-based MOF material in the initial porphyrin MOF dataset are summed and averaged to obtain the average similarity of each MOF sample in the QMOF database. Multiple MOF samples in the QMOF database are selected in descending order of the average similarity to form an expanded MOF dataset.

2. A method for predicting porphyrin MOF band gap based on deep transfer learning according to claim 1, characterized in that, In the above 3), the deep learning model is a graph convolutional neural network GCN.

Images