Metal organic framework performance prediction method, system, equipment, medium and product

Abstract

The invention discloses a metal organic framework performance prediction method, system and device, a medium and a product, and relates to the field of metal organic framework performance prediction, and the method comprises the steps: obtaining the crystal structure data of a to-be-predicted metal organic framework; and generating a 3D energy grid image of the to-be-predicted metal organic framework by using the methane molecular probe. According to the crystal structure data and the 3D energy grid image, determining the performance of the to-be-predicted metal organic framework by using a metal organic framework performance prediction model; according to the model, an initial performance prediction model is pre-trained by using a pre-training data set in combination with a pre-training task so as to learn universal feature representation. Performing fine tuning on the pre-training model by using a fine tuning data set of a specific task so as to adapt to a target performance prediction task; the initial performance prediction model comprises a crystal diagram convolutional neural network with a pooling layer removed and a multi-modal Transform model. According to the invention, high-precision metal organic framework performance prediction is realized.

Description

Technical Field

[0001] The present application relates to the field of metal organic framework performance prediction, and in particular to a metal organic framework performance prediction method, system, equipment, medium and product. Background Art

[0002] Metal-Organic Frameworks (MOFs) are a class of crystalline porous materials. Their highly porous structure makes them potentially useful in gas storage, separation, and purification. Their porous structure can be used to encapsulate drug molecules and control drug release, making MOFs potentially useful in drug delivery systems. In addition, they can also be used as catalysts and to make electronic devices. In recent years, the value of MOFs has been continuously explored, and the research on single-metal MOFs has received widespread recognition and attention from the international scientific community.

[0003] MOFs have a broad chemical space, and in principle an unlimited number of MOFs can be synthesized. Therefore, constructing the structure-property relationship of single-metal MOFs through artificial intelligence models can effectively explore this broad chemical space and identify the best candidate materials with the desired properties.

[0004] For bimetallic MOFs, the performance of bimetallic MOFs cannot be accurately predicted due to the small number of data sets. Summary of the invention

[0005] The purpose of this application is to provide a metal organic framework performance prediction method, system, equipment, medium and product to improve the accuracy of predicting the performance of bimetallic MOFs.

[0006] To achieve the above objectives, this application provides the following solutions: In a first aspect, the present application provides a method for predicting the performance of a metal organic framework, comprising: Acquiring crystal structure data of a metal-organic framework to be predicted; the metal-organic framework to be predicted is a bimetallic metal-organic framework or a monometallic metal-organic framework; Using a methane molecular probe to generate a 3D energy grid image of the metal-organic framework to be predicted; According to the crystal structure data of the metal organic framework to be predicted and the 3D energy grid image, the performance of the metal organic framework to be predicted is determined using a metal organic framework performance prediction model; the performance includes gas adsorption capacity, gas diffusivity, electronic properties, stability and thermal stability; wherein the metal organic framework performance prediction model is obtained by fine-tuning the pre-trained performance prediction model using a fine-tuning dataset; the fine-tuning dataset includes the crystal structure data of the fine-tuned metal organic framework, the 3D energy grid image of the fine-tuned metal organic framework and the corresponding performance labels; the crystal structure data of the fine-tuned metal organic framework is the crystal structure data of a hypothetical metal organic framework of a single metal generated by the PORMAKE tool based on the ToBaCCo database, the CoREMOF database and the RCSR database, or is generated by the PORMAKE tool based on experimental data. The invention discloses a method for preparing a pre-trained performance prediction model of a hypothetical metal-organic framework of a bimetallic structure; the experimental data include bimetallic building blocks and topological structures; the pre-trained performance prediction model is obtained by pre-training the initial performance prediction model using the pre-training data set in combination with the pre-training tasks; the initial performance prediction model includes a crystal graph convolutional neural network and a multimodal Transformer model connected in sequence without the pooling layer; the pre-training data set includes the crystal structure data of the hypothetical metal-organic framework, the 3D energy grid image of the hypothetical metal-organic framework and the corresponding performance labels; the crystal structure data of the hypothetical metal-organic framework is generated using the PORMAKE tool based on the ToBaCCo database, the CoREMOF database, the RCSR database and experimental data; the pre-training tasks include metal-organic framework topology prediction, void fraction prediction and metal cluster / organic connector classification.

[0007] Optionally, the initial performance prediction model is pre-trained using a pre-training data set in combination with a pre-training task to obtain a pre-trained performance prediction model, specifically including: Obtaining the pre-training data set; Preprocessing the crystal structure data of the hypothetical metal organic framework in the pre-training data set to obtain a processed pre-training data set; the processed pre-training data set includes the processed crystal structure data of the hypothetical metal organic framework, a 3D energy grid image of the hypothetical metal organic framework, and a corresponding performance label; Taking the processed crystal structure data of the hypothetical metal organic framework and the 3D energy grid image of the hypothetical metal organic framework as input and the corresponding performance label as output, the initial performance prediction model is pre-trained in combination with the pre-training task to obtain a pre-trained performance prediction model.

[0008] Optionally, preprocessing the crystal structure data of the hypothetical metal organic framework in the pre-training data set to obtain a processed pre-training data set specifically includes: The crystal structure data of the hypothetical metal organic framework are processed by removing solvent molecules to obtain the crystal structure data after removing solvent molecules; Identifying and correcting artifacts in the crystal structure data from which solvent molecules have been removed, to obtain artifact-corrected crystal structure data; Processing the disordered atoms in the crystal structure data after the artifact correction to obtain the crystal structure data after the disorder is eliminated; Supplementing hydrogen atoms to the disorder-eliminated crystal structure data with missing hydrogen atoms to obtain the crystal structure data with supplemented hydrogen atoms; Performing charge balancing processing on the crystal structure data after the charged hydrogen atoms are supplemented to obtain balanced crystal structure data; The balanced crystal structure data is geometrically optimized to obtain processed crystal structure data, thereby forming a processed pre-training data set.

[0009] Optionally, the pre-trained performance prediction model is fine-tuned using a fine-tuning dataset to obtain a metal organic framework performance prediction model, specifically including: Get a fine-tuning dataset; With the crystal structure data of the fine-tuned metal-organic framework and the 3D energy grid image of the fine-tuned metal-organic framework as input and the corresponding performance labels as output, the grid search algorithm is used to fine-tune the model hyperparameters of the pre-trained performance prediction model to obtain the metal-organic framework performance prediction model.

[0010] Optionally, the model hyperparameters include learning rate, regularization parameter, activation function and hidden layer structure.

[0011] Optionally, the crystal graph convolutional neural network with the pooling layer removed includes a first embedding layer and a second embedding layer, a first graph convolutional layer, a second graph convolutional layer, a third graph convolutional layer, a fourth graph convolutional layer, a multi-layer fully connected network, a regression layer and an adaptation layer connected in sequence; the first embedding layer is also connected to the first graph convolutional layer, the second graph convolutional layer, the third graph convolutional layer and the fourth graph convolutional layer respectively; the first graph convolutional layer, the second graph convolutional layer and the third graph convolutional layer are also connected to the multi-layer fully connected network.

[0012] In a second aspect, the present application provides a metal organic framework performance prediction system, comprising: A data acquisition module, used to acquire crystal structure data of a metal-organic framework to be predicted; the metal-organic framework to be predicted is a bimetallic metal-organic framework or a monometallic metal-organic framework; A 3D energy grid generation module, used to generate a 3D energy grid image of the metal organic framework to be predicted using a methane molecular probe; A performance prediction module is used to determine the performance of the metal organic framework to be predicted based on the crystal structure data of the metal organic framework to be predicted and the 3D energy grid image, using a metal organic framework performance prediction model; the performance includes gas adsorption capacity, gas diffusivity, electronic properties, stability and thermal stability; wherein the metal organic framework performance prediction model is obtained by fine-tuning the pre-trained performance prediction model using a fine-tuning data set; the fine-tuning data set includes the crystal structure data of the fine-tuned metal organic framework, the 3D energy grid image of the fine-tuned metal organic framework and the corresponding performance label; the crystal structure data of the fine-tuned metal organic framework is the crystal structure data of a hypothetical metal organic framework of a single metal generated by the PORMAKE tool based on the ToBaCCo database, the CoREMOF database and the RCSR database, or based on experimental data, using the PORMA The KE tool generates crystal structure data of a hypothetical metal-organic framework of a bimetallic structure; the experimental data includes bimetallic building blocks and topological structures; the pre-trained performance prediction model is obtained by pre-training the initial performance prediction model using the pre-training data set in combination with the pre-training task; the initial performance prediction model includes a crystal graph convolutional neural network and a multimodal Transformer model connected in sequence without the pooling layer; the pre-training data set includes crystal structure data of the hypothetical metal-organic framework, a 3D energy grid image of the hypothetical metal-organic framework and corresponding performance labels; the crystal structure data of the hypothetical metal-organic framework is generated using the PORMAKE tool based on the ToBaCCo database, the CoREMOF database, the RCSR database and experimental data; the pre-training tasks include metal-organic framework topology prediction, void fraction prediction and metal cluster / organic connector classification.

[0013] In a third aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the metal organic framework performance prediction methods described above.

[0014] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the metal-organic framework performance prediction methods described above.

[0015] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements any of the metal-organic framework performance prediction methods described above.

[0016] According to the specific embodiments provided in this application, this application has the following technical effects: The present application provides a method, system, device, medium and product for predicting the performance of a metal organic framework, obtaining crystal structure data of the metal organic framework to be predicted, and generating a 3D energy grid image of the metal organic framework to be predicted using a methane molecular probe. According to the crystal structure data and the 3D energy grid image, the performance of the metal organic framework to be predicted is determined using the metal organic framework performance prediction model, including gas adsorption capacity, gas diffusivity, electronic properties, stability and thermal stability; the metal organic framework performance prediction model is obtained by fine-tuning the pre-trained performance prediction model using the fine-tuning data set; the fine-tuning data set includes the crystal structure data of the fine-tuned metal organic framework, the 3D energy grid image of the fine-tuned metal organic framework and the corresponding performance labels; the crystal structure data of the fine-tuned metal organic framework is the crystal structure data of a hypothetical metal organic framework of a single metal generated by the PORMAKE tool based on the ToBaCCo database, the CoREMOF database and the RCSR database, or the crystal structure data of a hypothetical metal organic framework of a bimetallic generated by the PORMAKE tool based on experimental data; the experimental data includes bimetallic building blocks and topological structures; the pre-trained performance prediction model is obtained by pre-training the initial performance prediction model using the pre-training data set in combination with the pre-training task; the initial performance prediction model includes a crystal graph convolutional neural network with the pooling layer removed, which are connected in sequence, and a multimodal Transformer model. The metal-organic framework performance prediction model of the present application captures complex structure-property relationships by combining the crystal graph convolutional neural network without the pooling layer to extract structural features and the self-attention mechanism of the multimodal Transformer model, and uses large-scale data sets for training and fine-tuning, thereby achieving high-precision prediction of the performance of single-metal or bimetallic metal-organic frameworks. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 A schematic diagram of a process for predicting metal organic framework performance provided in one embodiment of the present application; Figure 2 A schematic diagram of the architecture of the CGCNN network provided for this application; Figure 3 Schematic diagram of the multimodal Transformer model architecture provided for this application; Figure 4 Schematic diagram for comparison of CO2 adsorption performance of single metal MOF; Figure 5 Schematic diagram for comparison of CO2 adsorption performance of bimetallic MOF; Figure 6 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0020] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0021] For bimetallic MOFs, due to the small data set, this application combines laboratory experimental data to enhance and verify the model. Through transfer learning and data enhancement technology, the problem of scarce bimetallic MOFs data is solved, enabling the model to more accurately predict the performance of bimetallic MOFs. This method combines the artificial intelligence prediction of monometallic MOFs with the experimental data of bimetallic MOFs, improving the generalization ability and prediction accuracy of the model.

[0022] In an exemplary embodiment, Figure 1 As shown, a method for predicting the performance of a metal organic framework is provided, comprising the following steps: S1: Acquire crystal structure data of a metal-organic framework to be predicted; the metal-organic framework to be predicted is a bimetallic metal-organic framework or a monometallic metal-organic framework.

[0023] S2: Generate a 3D energy grid image of the metal organic framework to be predicted using a methane molecular probe.

[0024] S3: According to the crystal structure data of the metal organic framework to be predicted and the 3D energy grid image, the performance of the metal organic framework to be predicted is determined using a metal organic framework performance prediction model; the performance includes gas adsorption capacity, gas diffusivity, electronic properties, stability and thermal stability; wherein the metal organic framework performance prediction model is obtained by fine-tuning the pre-trained performance prediction model using a fine-tuning data set; the fine-tuning data set includes the crystal structure data of the fine-tuned metal organic framework, the 3D energy grid image of the fine-tuned metal organic framework and the corresponding performance labels; the crystal structure data of the fine-tuned metal organic framework is the crystal structure data of a hypothetical metal organic framework of a single metal generated by the PORMAKE tool based on the ToBaCCo database, the CoREMOF database and the RCSR database, or the crystal structure data of a hypothetical metal organic framework of a bimetal generated by the PORMAKE tool based on experimental data Crystal structure data of a metal-organic framework; the experimental data include bimetallic building blocks and topological structures; the pre-trained performance prediction model is obtained by pre-training an initial performance prediction model using a pre-training data set in combination with a pre-training task; the initial performance prediction model includes a crystal graph convolutional neural network and a multimodal Transformer model connected in sequence without a pooling layer; the pre-training data set includes crystal structure data of a hypothetical metal-organic framework, a 3D energy grid image of the hypothetical metal-organic framework and corresponding performance labels; the crystal structure data of the hypothetical metal-organic framework is generated using the PORMAKE tool based on the ToBaCCo database, the CoREMOF database, the RCSR database and experimental data; the experimental data include bimetallic building blocks and topological structures; the pre-training tasks include metal-organic framework topology prediction, void fraction prediction and metal cluster / organic linker classification.

[0025] like Figure 2 As shown, the crystal graph convolutional neural network with the pooling layer removed includes a first embedding layer and a second embedding layer, a first graph convolutional layer, a second graph convolutional layer, a third graph convolutional layer, a fourth graph convolutional layer, a multi-layer fully connected network, a regression layer and an adaptation layer connected in sequence; the first embedding layer is also connected to the first graph convolutional layer, the second graph convolutional layer, the third graph convolutional layer and the fourth graph convolutional layer respectively; the first graph convolutional layer, the second graph convolutional layer and the third graph convolutional layer are also connected to the multi-layer fully connected network.

[0026] The predictive performance of MOF is driven by a multimodal Transformer model, which achieves high-precision performance prediction by combining CGCNN to extract structural features and the Transformer's self-attention mechanism to capture complex structure-property relationships, and is trained and fine-tuned using large-scale datasets.

[0027] like Figure 3 As shown, the multimodal Transformer model includes: Input layer: Receives two types of embeddings, atomic-level feature embeddings and energy grid embeddings, as input to the model.

[0028] Transformer encoder layer: The model uses a BERT-based structure, consisting of 12 encoding blocks (L=12), each block has 768 hidden units (H=768), and 12 self-attention heads (A=12).

[0029] [CLS]Token and [SEP]Token: [CLS] and [SEP] are used as learnable embedding layers, respectively, at the beginning and middle of the sequence, mainly used to aggregate features and predict target attributes.

[0030] [VOL]Token: An additional token representing the normalized unit volume, used to compensate for the loss of original energy grid volume information that may occur due to energy grid interpolation.

[0031] In [CLS] and [SEP], atomic features and type embeddings are embedded (type embeddings are set to 0). In [SEP] and [VOL], energy grid embeddings and type embeddings are embedded (type embeddings are set to 1). The role of type embeddings is to tell the model which are atomic graph embeddings and which are energy grid embeddings.

[0032] Position information embedding: Since the position information of the 3D energy grid cannot be directly obtained by the Transformer model after it is divided into patches, position embedding is introduced to provide the position information of each patch in the overall structure.

[0033] like Figure 3 As shown, the yellow one is the position information embedding, the red and blue ones are the atomic base graph embedding and the energy grid embedding, and the white grid is to distinguish between the atomic base graph embedding and the energy grid embedding.

[0034] The multimodal Transformer framework is a powerful deep learning model that can process multiple types of data, such as text, images, and structural data. In the study of MOF materials, the multimodal Transformer framework can simultaneously process the structural information and property information of MOF, thereby constructing a more accurate structure-property relationship.

[0035] This application proposes a two-stage prediction method: Single metal prediction: Combining CGCNN (Crystal Graph Convolutional Neural Networks) and the Transformer model, CGCNN is used to extract the crystal structure characteristics of single metal MOFs, and the self-attention mechanism of the Transformer model is used to integrate multiple information to achieve the prediction of the performance of single metal MOFs.

[0036] Bimetallic prediction: Based on the prediction of single metal, the Transformer model is further used to predict the performance of bimetallic MOFs. This model integrates the prediction results of single metal and the structural information of bimetal to capture the complex structure-property relationship.

[0037] By training a large model using the multimodal Transformer framework, researchers can more accurately predict the properties of new MOF materials without having to conduct actual experiments. This approach not only saves a lot of time and resources, but also helps researchers identify the best candidate materials with the desired properties.

[0038] This application builds a high-quality data set through the data accumulated and generated in the laboratory for model training and fine-tuning, and achieves accurate prediction of the performance of monometallic MOFs and bimetallic MOFs. This method improves the efficiency of the laboratory in exploring the MOF structure-performance relationship and can efficiently and accurately predict the performance of MOF based on its structure.

[0039] In practical applications, the architecture of this application that combines the crystal graph convolutional neural network CGCNN and the multimodal Transformer model can be used to evaluate and predict the properties of single and bimetallic MOFs.

[0040] As an optional implementation method, the process of constructing the metal organic framework performance prediction model specifically includes: S31: Obtain the pre-training data set.

[0041] In practical applications, data collection and preprocessing are first performed, and PORMAKE is used to generate 1 million hypothetical single metal MOFs (hMOFs) and 2,000 bimetallic MOFs (two-hMOFs) with different topological structures and building blocks. The single metal building blocks and topological structures are derived from the ToBaCCo database, the CoREMOF (solvent removal) database, and the RCSR database, and the bimetallic building blocks and topological structures are derived from published papers and experimental data obtained in the laboratory, solving the problem of scarce bimetallic data.

[0042] Crystal structure data of MOFs are collected from the ToBaCCo database, CoREMOF database, and RCSR database. These databases provide a large amount of MOFs structural information, which can be used for subsequent analysis and research. Data preprocessing includes removing solvent molecules, identifying and correcting artifacts in the crystal structure, processing disordered atoms, adding missing hydrogen atoms, and performing charge balance and geometry optimization on the crystal structure to construct high-quality hypothetical monometallic and bimetallic MOFs (hMOFs) data sets.

[0043] S32: preprocessing the crystal structure data of the hypothetical metal organic framework in the pre-training data set to obtain a processed pre-training data set; the processed pre-training data set includes the processed crystal structure data of the hypothetical metal organic framework, a 3D energy grid image of the hypothetical metal organic framework and corresponding performance labels.

[0044] As an optional implementation, S32 specifically includes: The crystal structure data of the hypothetical metal organic framework are processed by removing solvent molecules to obtain the crystal structure data without solvent molecules.

[0045] Artifacts in the crystal structure data from which the solvent molecules have been removed are identified and corrected to obtain artifact-corrected crystal structure data.

[0046] The disordered atoms in the crystal structure data after the artifact correction are processed to obtain the crystal structure data after the disorder is eliminated.

[0047] The crystal structure data after the disorder is eliminated and the missing hydrogen atoms are supplemented with hydrogen atoms to obtain the crystal structure data after the hydrogen atoms are supplemented.

[0048] The crystal structure data after the charged hydrogen atoms are supplemented are subjected to charge balancing processing to obtain balanced crystal structure data.

[0049] The balanced crystal structure data is geometrically optimized to obtain processed crystal structure data, thereby forming a processed pre-training data set.

[0050] In practical applications, the crystal structure data is preprocessed as follows: 1. Remove solvent molecules: For MOFs structures in the ToBaCCo and RCSR databases, solvent molecules need to be removed from the crystal structure to improve the quality of the data. This step can be performed in software such as Olex2 or PLATON.

[0051] 2. Identify and correct artifacts in the crystal structure: Identify and correct various artifacts in the MOFs crystal structure, including incorrectly bonded atoms, isolated atoms, overlapping atoms, etc., to ensure the accuracy of the data.

[0052] 3. Dealing with disordered atoms: For disordered atoms, they need to be processed to retain a single set of structures, which may involve deleting some atoms or rearranging them to eliminate the disorder.

[0053] 4. Add missing hydrogen atoms: Supplement the missing hydrogen atoms to ensure the integrity and accuracy of the structure.

[0054] 5. Charge balance treatment: For MOFs structures with charges, charge-balancing ions need to be added or adjusted to maintain the electrical neutrality of the structure.

[0055] 6. Structural optimization: Use packages such as LAMMPS and UFF force field to perform geometric optimization on the MOFs structure to obtain a more accurate structural model.

[0056] 7. Dataset construction: The optimized MOFs structures are constructed as datasets for machine learning model training. These datasets will contain features and labels for predicting MOFs performance.

[0057] Features that predict MOFs performance include atomic-level characteristics and 3D energy grids.

[0058] Atom-based Graph Embedding: Atoms (nodes), bonds (edges) and their features (such as interatomic distances) are converted into features in vector space through the crystal graph convolutional neural network CGCNN, representing the local features of MOFs. These local features can capture the atomic-level arrangement and interaction in the material's microstructure.

[0059] 3D Energy-grid Embedding: The 3D energy grids calculated by the methane molecular probe are used to describe the van der Waals interactions between MOFs and methane molecules. These energy grids are viewed as 3D images, where each grid point and energy value are used as pixels and single-channel colors, respectively, representing the global characteristics of MOFs and reflecting the overall behavior and energy distribution of MOFs during gas adsorption.

[0060] Labels are used to represent the various physical and chemical properties of MOFs, including: Gas adsorption capacity: For example, the amount of hydrogen (H2) adsorption is used to measure the adsorption and storage capacity of MOFs for gas molecules.

[0061] Gas diffusivity: For example, the diffusion coefficients of carbon dioxide (CO2) and hydrogen (H2) characterize the diffusion rate of gases in MOFs and reflect the efficiency of the material in gas transport.

[0062] Electronic properties: such as band gap, which is used to describe the electronic structure of MOFs, especially their potential in catalysis or electronic conductivity.

[0063] Solvent Removal Stability and Thermal Stability: These data were obtained through text mining techniques and were used to evaluate the stability of MOFs under different environmental conditions, especially during solvent removal and thermal treatment.

[0064] S33: Taking the processed crystal structure data of the hypothetical metal-organic framework and the 3D energy grid image of the hypothetical metal-organic framework as input and the corresponding performance label as output, the initial performance prediction model is pre-trained in combination with the pre-training task to obtain a pre-trained performance prediction model.

[0065] S34: Get a fine-tuning dataset.

[0066] S35: Using the fine-tuned crystal structure data of the metal organic framework and the fine-tuned 3D energy grid image of the metal organic framework as input and the corresponding performance label as output, a grid search algorithm is used to fine-tune the model hyperparameters of the pre-trained performance prediction model to obtain a metal organic framework performance prediction model. The model hyperparameters include a learning rate, a regularization parameter, an activation function, and a hidden layer structure.

[0067] In practical applications, the crystal graph convolutional neural network CGCNN is used to convert crystal structure data (i.e. atoms (nodes), bonds (edges) and their features (distances between atoms)) into vector space. CGCNN contains convolutional layers and pooling layers, but in the metal organic framework performance prediction model of this application, the CGCNN output vector without pooling layer is used to retain atomic-level features (i.e. crystal structure features) to ensure the integrity of the structural information.

[0068] The energy grid calculated using the methane molecular probe is used as one of the inputs of the multimodal Transformer model to describe the van der Waals interactions between MOFs and methane molecules and construct multimodal input data. The 3D energy grid is considered as a 3D image, where the grid points and energy values ​​are used as pixels and single-channel colors, respectively.

[0069] In terms of model design, a multimodal framework is constructed by combining CGCNN and the multimodal Transformer model. The crystal structure characteristics and energy grid information are integrated through the self-attention mechanism of the Transformer to achieve efficient prediction of MOFs performance.

[0070] In the model training, three pre-training tasks were designed: MOF topology prediction (MTP), void fraction prediction (VFP), and metal cluster / organic connector classification (MOC). The pre-training model used 1 million single metal hMOFs to perform 100 rounds of pre-training on a multimodal framework combining CGCNN and a multimodal Transformer model, using the AdamW optimizer with a learning rate of 10 -4 , weight decay is 10 -2 .

[0071] Fine-tuning model uses a high-quality dataset of 5,000 to 20,000 MOFs (fine-tuning dataset) to fine-tune the pre-trained multimodal framework combining CGCNN and multimodal Transformer models. The model input parameters are Input=m×n, where m=1, n=406. The activation function is GELU, the optimizer is AdamW, the learning rate is 0.0001, the regularization term is 0.65, the number of hidden layers is 2, and the number of neurons is (18, 12). Model hyperparameters are optimized by grid search.

[0072] The dataset was split into training set, validation set, and test set with a ratio of 8:1:1. The training set was used to learn the model, the validation set was used to verify the model accuracy and hyperparameter configuration, and the test set was used to evaluate the prediction results of MOFs properties.

[0073] Compare the performance of our model with other machine learning models, including R2 score and Mean Absolute Error (MAE).

[0074] The attention scores of the self-attention layer are analyzed to determine the importance of the input features and provide the possibility of a deeper understanding of the relationship between the structure and properties of MOFs.

[0075] The model performance evaluation results show that Figure 4 As shown in the figure, after fine-tuning on the single-metal MOF dataset, the model's predicted values ​​(blue) are compared with the true values ​​(red) to show high accuracy of CO2 adsorption performance, and the data points are close to a straight line. The prediction accuracy of single-metal MOFs reaches R²=89%, and the validation set contains 2136 samples; Figure 5 As shown in Figure 2, after fine-tuning on the bimetallic MOF dataset, although the dispersion between the predicted values ​​and the true values ​​is high, which is caused by the small size of the validation set, the R 2The value is still high, and it can still effectively predict the overall trend of carbon dioxide adsorption performance. The prediction accuracy of bimetallic MOFs reached R²=82%, and the validation set contained 53 samples. Through comparative analysis with other machine learning models, the R² score and mean absolute error MAE of this application in predictive performance are significantly better than traditional models. In addition, by analyzing the attention score of the self-attention layer in the multimodal Transformer model, the complex relationship between the structure and properties of MOFs is further revealed, providing a scientific basis and technical support for the optimization and design of MOFs materials. This method effectively solves the problem of a large number of MOFs but scarce property data, and realizes efficient and accurate performance prediction. In particular, it conducts innovative exploration of the complex structure-property relationship of bimetallic MOFs, providing new ideas and tools for MOFs performance prediction and material design.

[0076] The multimodal architecture of this application, which combines the crystal graph convolutional neural network CGCNN and the multimodal Transformer model, is used to predict the various properties of bimetallic MOFs. It solves the problem that researchers have to spend a lot of time and effort to predict the properties of new MOF materials based on the material structure, as well as the prediction limitations caused by the scarcity of bimetallic MOFs data. By integrating laboratory experimental data and transfer learning technology, the generalization ability of the model is enhanced, and efficient and accurate prediction of the performance of monometallic and bimetallic MOFs can be achieved.

[0077] This application constructs a high-quality dataset through feature engineering, uses atom-based graphs and energy grid embeddings as model inputs, optimizes the multimodal Transformer model using a grid search optimization algorithm, and uses the attention scores in the self-attention layer to analyze the feature importance. The resulting model can be used to accurately predict the properties of unknown MOFs outside the training dataset.

[0078] Based on the same inventive concept, the embodiment of the present application also provides a system for implementing the metal organic framework performance prediction involved above. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme recorded in the above method, so the specific limitations in the metal organic framework performance prediction system embodiment provided below can refer to the limitations of the metal organic framework performance prediction method above, and will not be repeated here.

[0079] In an exemplary embodiment, a metal organic framework performance prediction system is provided, comprising: The data acquisition module is used to acquire the crystal structure data of the metal-organic framework to be predicted; the metal-organic framework to be predicted is a bimetallic metal-organic framework or a monometallic metal-organic framework.

[0080] The 3D energy grid generation module is used to generate a 3D energy grid image of the metal organic framework to be predicted using a methane molecular probe.

[0081] A performance prediction module is used to determine the performance of the metal organic framework to be predicted based on the crystal structure data of the metal organic framework to be predicted and the 3D energy grid image, using a metal organic framework performance prediction model; the performance includes gas adsorption capacity, gas diffusivity, electronic properties, stability and thermal stability; wherein the metal organic framework performance prediction model is obtained by fine-tuning a pre-trained performance prediction model using a fine-tuning data set; the fine-tuning data set includes the crystal structure data of the fine-tuned metal organic framework, the 3D energy grid image of the fine-tuned metal organic framework and the corresponding performance label; the crystal structure data of the fine-tuned metal organic framework is the crystal structure data of a hypothetical metal organic framework of a single metal generated by the PORMAKE tool based on the ToBaCCo database, the CoREMOF database and the RCSR database, or the crystal structure data of a bimetallic metal organic framework generated by the PORMAKE tool based on experimental data The crystal structure data of a hypothetical metal-organic framework; the experimental data include bimetallic building blocks and topological structures; the pre-trained performance prediction model is obtained by pre-training the initial performance prediction model using the pre-training data set in combination with the pre-training task; the initial performance prediction model includes a crystal graph convolutional neural network and a multimodal Transformer model connected in sequence without the pooling layer; the pre-training data set includes the crystal structure data of the hypothetical metal-organic framework, the 3D energy grid image of the hypothetical metal-organic framework and the corresponding performance labels; the crystal structure data of the hypothetical metal-organic framework is generated using the PORMAKE tool based on the ToBaCCo database, the CoREMOF database, the RCSR database and the experimental data; the experimental data include bimetallic building blocks and topological structures; the pre-training tasks include metal-organic framework topology prediction, void fraction prediction and metal cluster / organic linker classification.

[0082] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the above-mentioned metal organic framework performance prediction method when executing the computer program.

[0083] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, which implements the above-mentioned metal-organic framework performance prediction method when executed by a processor.

[0084] In an exemplary embodiment, a computer program product is provided, including a computer program, which implements the above-mentioned metal organic framework performance prediction method when executed by a processor.

[0085] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 6 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a metal organic framework performance prediction method is implemented.

[0086] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0087] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0088] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0089] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0090] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0091] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for predicting the performance of a metal organic framework, characterized in that: include: Obtaining crystal structure data of the metal-organic framework to be predicted; Using a methane molecular probe to generate a 3D energy grid image of the metal-organic framework to be predicted; According to the crystal structure data of the metal organic framework to be predicted and the 3D energy grid image, the performance of the metal organic framework to be predicted is determined using a metal organic framework performance prediction model; the metal organic framework performance prediction model is obtained by fine-tuning a pre-trained performance prediction model using a fine-tuning data set; the fine-tuning data set includes the crystal structure data of the fine-tuned metal organic framework, the 3D energy grid image of the fine-tuned metal organic framework and the corresponding performance labels; the pre-trained performance prediction model is obtained by pre-training an initial performance prediction model using a pre-training data set in combination with a pre-training task; the initial performance prediction model includes a crystal graph convolutional neural network with the pooling layer removed and a multimodal Transformer model connected in sequence.

2. The metal organic framework performance prediction method according to claim 1, characterized in that: The initial performance prediction model is pre-trained using the pre-training data set and combined with the pre-training task to obtain a pre-trained performance prediction model, which specifically includes: Obtaining the pre-training data set; Preprocessing the crystal structure data of the hypothetical metal organic framework in the pre-training data set to obtain a processed pre-training data set; the processed pre-training data set includes the processed crystal structure data of the hypothetical metal organic framework, a 3D energy grid image of the hypothetical metal organic framework, and a corresponding performance label; Taking the processed crystal structure data of the hypothetical metal organic framework and the 3D energy grid image of the hypothetical metal organic framework as input and the corresponding performance label as output, the initial performance prediction model is pre-trained in combination with the pre-training task to obtain a pre-trained performance prediction model.

3. The metal organic framework performance prediction method according to claim 2, characterized in that: Preprocessing the crystal structure data of the hypothetical metal organic framework in the pre-training data set to obtain a processed pre-training data set specifically includes: The crystal structure data of the hypothetical metal organic framework are processed by removing solvent molecules to obtain the crystal structure data after removing solvent molecules; Identifying and correcting artifacts in the crystal structure data from which solvent molecules have been removed, to obtain artifact-corrected crystal structure data; Processing the disordered atoms in the crystal structure data after the artifact correction to obtain the crystal structure data after the disorder is eliminated; Supplementing hydrogen atoms to the disorder-eliminated crystal structure data with missing hydrogen atoms to obtain the crystal structure data with supplemented hydrogen atoms; Performing charge balancing processing on the crystal structure data after the charged hydrogen atoms are supplemented to obtain balanced crystal structure data; The balanced crystal structure data is geometrically optimized to obtain processed crystal structure data, thereby forming a processed pre-training data set.

4. The metal organic framework performance prediction method according to claim 1, characterized in that: The pre-trained performance prediction model is fine-tuned using the fine-tuning dataset to obtain a metal organic framework performance prediction model, which specifically includes: Get a fine-tuning dataset; With the crystal structure data of the fine-tuned metal-organic framework and the 3D energy grid image of the fine-tuned metal-organic framework as input and the corresponding performance labels as output, the grid search algorithm is used to fine-tune the model hyperparameters of the pre-trained performance prediction model to obtain the metal-organic framework performance prediction model.

5. The metal organic framework performance prediction method according to claim 4, characterized in that: The model hyperparameters include learning rate, regularization parameter, activation function and hidden layer structure.

6. The metal organic framework performance prediction method according to claim 1, characterized in that: The crystal graph convolutional neural network without the pooling layer includes a first embedding layer and a second embedding layer, a first graph convolutional layer, a second graph convolutional layer, a third graph convolutional layer, a fourth graph convolutional layer, a multi-layer fully connected network, a regression layer and an adaptation layer connected in sequence; the first embedding layer is also connected to the first graph convolutional layer, the second graph convolutional layer, the third graph convolutional layer and the fourth graph convolutional layer respectively; the first graph convolutional layer, the second graph convolutional layer and the third graph convolutional layer are also connected to the multi-layer fully connected network.

7. A metal organic framework performance prediction system, characterized in that: include: A data acquisition module, used to acquire crystal structure data of a metal-organic framework to be predicted; the metal-organic framework to be predicted is a bimetallic metal-organic framework or a monometallic metal-organic framework; A 3D energy grid generation module, used to generate a 3D energy grid image of the metal organic framework to be predicted using a methane molecular probe; A performance prediction module is used to determine the performance of the metal organic framework to be predicted based on the crystal structure data of the metal organic framework to be predicted and the 3D energy grid image, using a metal organic framework performance prediction model; the performance includes gas adsorption capacity, gas diffusivity, electronic properties, stability and thermal stability; wherein the metal organic framework performance prediction model is obtained by fine-tuning the pre-trained performance prediction model using a fine-tuning data set; the fine-tuning data set includes the crystal structure data of the fine-tuned metal organic framework, the 3D energy grid image of the fine-tuned metal organic framework and the corresponding performance label; the crystal structure data of the fine-tuned metal organic framework is the crystal structure data of a hypothetical metal organic framework of a single metal generated by the PORMAKE tool based on the ToBaCCo database, the CoREMOF database and the RCSR database, or based on experimental data, using the PORMA The KE tool generates crystal structure data of a hypothetical metal-organic framework of a bimetallic structure; the experimental data includes bimetallic building blocks and topological structures; the pre-trained performance prediction model is obtained by pre-training the initial performance prediction model using the pre-training data set in combination with the pre-training task; the initial performance prediction model includes a crystal graph convolutional neural network and a multimodal Transformer model connected in sequence without the pooling layer; the pre-training data set includes crystal structure data of the hypothetical metal-organic framework, a 3D energy grid image of the hypothetical metal-organic framework and corresponding performance labels; the crystal structure data of the hypothetical metal-organic framework is generated using the PORMAKE tool based on the ToBaCCo database, the CoREMOF database, the RCSR database and experimental data; the pre-training tasks include metal-organic framework topology prediction, void fraction prediction and metal cluster / organic connector classification.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the metal organic framework performance prediction method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the metal organic framework performance prediction method described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the metal organic framework performance prediction method described in any one of claims 1 to 6 is implemented.

Images