A method for predicting MOF chemical warfare agent adsorption performance based on a bidirectional attention mechanism

Abstract

The application relates to the technical field of intelligent prediction of functional materials, in particular to a MOF chemical warfare agent adsorption performance prediction method based on a bidirectional attention mechanism, which comprises the following steps: constructing a standardized data set containing a MOF crystal structure, a guest molecule and an adsorption amount label; using a Matformer self-attention mechanism to extract a MOF atomic-level crystal embedding, and performing topological feature coding on the guest molecule; performing host-guest collaborative interaction modeling through a bidirectional cross-attention mechanism, and dynamically capturing the physical and chemical interactions between the skeleton and the molecule; and inputting the fused features into a regression module to output adsorption prediction values. The application can eliminate the dependence on artificial geometric descriptors, restore the host-guest response in the adsorption process at the atomic scale, significantly improve the prediction accuracy and calculation efficiency, and provide a scientific basis for the rapid screening and directional design of high-performance protective materials.

Description

Technical Field

[0001] This invention belongs to the field of intelligent prediction technology for functional materials, specifically relating to a method for predicting the adsorption performance of MOF chemical warfare agents based on a bidirectional attention mechanism. Background Technology

[0002] Chemical warfare agents, due to their extremely high toxicity, wide-ranging lethality, and long-term environmental residue, have become a major threat in modern warfare and terrorist activities. Their low manufacturing cost and ease of production further increase the difficulty of safety control, posing a direct challenge to human health and environmental safety. In the protection and management of chemical warfare agents, metal-organic frameworks (MOFs), as porous crystalline materials formed by the self-assembly of metal nodes and organic ligands, have demonstrated outstanding application potential due to their advantages such as large specific surface area, tunable pore size, rich framework composition, and strong surface chemical designability. Compared to traditional adsorbent materials, MOFs can achieve selective adsorption and enhancement of specific toxic small molecules through pore size control, functional group modification, and metal site design. Therefore, they have significant scientific research and practical application value in the fields of chemical warfare agent protection, toxic substance removal, and hazardous chemical capture.

[0003] However, due to the vast variety and large structural space of MOF materials, the adsorption performance of MOFs with different structures varies significantly with different warp agent molecules. In a massive material library, relying solely on traditional experimental testing or high-precision molecular simulation methods for individual screening is not only time-consuming and costly, but also unsuitable for large-scale, high-throughput material discovery and design needs. In recent years, artificial intelligence methods, represented by deep learning, have been gradually applied to the prediction of crystalline material properties and the screening of porous materials. By automatically characterizing crystal structures, molecular topologies, and local geometric environments, deep learning models can establish nonlinear mapping relationships between material structures and target properties in a relatively short time, providing a new technical path for predicting MOF adsorption performance. In particular, the introduction of attention mechanisms allows the model to focus on key structural units at the atomic scale, thereby improving the model's characterization capabilities to a certain extent.

[0004] Despite the progress made by deep learning in materials prediction, existing MOF adsorption prediction models still have significant limitations. Most existing methods primarily model single modes within the MOF framework or guest molecules, or employ simplistic and coarse-grained approaches in multimodal information fusion, failing to accurately describe the complex interactions between host and guest molecules during real adsorption. This inadequate characterization of host-guest interactions directly leads to limited prediction accuracy in complex systems. Furthermore, traditional machine learning methods often rely on time-consuming computation of manual, empirical descriptors, resulting in poor model interpretability. Existing graph neural networks also suffer from insufficient accuracy in handling multimodal interactions. Therefore, constructing an interactive deep learning method that can deeply integrate MOF crystal information and chemical warfare agent molecular information, and effectively simulate the dynamic perception relationship between host and guest molecules, to improve the accuracy and interpretability of adsorption performance prediction, is a significant technical challenge currently facing the field of intelligent prediction for functional materials. Summary of the Invention

[0005] This technical solution constructs a deep learning architecture for subject-object collaborative perception, utilizes a self-attention mechanism to extract atomic-level structural features of MOF crystals, and combines a bidirectional cross-attention mechanism to dynamically capture multi-scale interactions between the MOF framework and chemical warfare agent molecules. This eliminates the dependence on complex geometric descriptors and enables high-throughput adsorption performance prediction tasks. As a result, it solves the problems of high computational costs of traditional simulation methods and the limitations of existing AI models in characterizing subject-object interactions, achieving efficient screening of chemical warfare agent protective materials.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting the adsorption performance of MOF chemical warfare agents based on a bidirectional attention mechanism, the method comprising the following steps: S1. Construct a standardized dataset for model training and validation. Step S1 specifically involves: First, extensively collecting crystallographic information files of MOF materials from crystal structure databases, and obtaining molecular structure data of chemical warfare agents and their simulants. The chemical warfare agents include sarin, soman, VEX, and mustard gas, and the simulants include dimethyl methylphosphonate and 2-chloroethyl ethyl sulfide. Then, using the grand canonical Monte Carlo simulation method under set temperature and pressure conditions, the adsorption amount of each MOF material for a specific chemical warfare agent molecule is calculated and used as the label value for model learning. During the simulation, the Lennard-Jones potential function is used to describe the van der Waals interactions between atoms, and the MEPO-ML method is applied to perform charge distribution on the MOF framework atoms to accurately describe electrostatic interactions. Finally, the collected adsorption data is standardized, and the dataset is divided into training, validation, and test sets according to a preset ratio, providing a foundation for subsequent neural network parameter optimization.

[0007] S2. Extraction and characterization of subject-guest bimodal features. Step S2 includes atomic-level encoding of the MOF crystal structure and topological feature encoding of the guest molecule. During MOF structure encoding, the CIF file is parsed into an atomic-level crystal diagram representation, where atoms are considered nodes and spatial connections between atoms are considered edges. A Matformer graph neural network based on a self-attention mechanism is used as the feature extractor. The crystal diagram is processed through multiple layers of self-attention blocks to extract node-level crystal embedding representations containing local chemical environment and long-range periodicity information. The Matformer configuration parameters are set as follows: hidden dimension 32, stacked layers 3, and multi-head attention mechanism heads 8. During guest molecule encoding, the Morgan fingerprint vector of the chemical warfare agent molecule is extracted. This fingerprint vector is obtained by sampling the molecular topology with a radius of 2 substructures and mapping it to a 2048-bit bit vector space. Subsequently, a linear mapping layer is used to project the high-dimensional Morgan fingerprint vector into a latent space that matches the dimension of the MOF node representation, generating an object token representation with a dimension of 128, thereby achieving dimensional alignment of subject and object information in the feature space.

[0008] S3. Construct and execute a bidirectional cross-attention interaction model for the host and guest molecules. Step S3 involves introducing a collaborative attention block to deeply fuse the MOF crystal embedding representation with the guest molecule token representation. The interaction module consists of multiple stacked collaborative attention blocks, with each block performing a bidirectional feature update operation. Specifically, the first direction utilizes the MOF structural features as bonds and values, and the guest molecule representation as a query, calculating attention weights to modulate the state of the guest molecule by the MOF structure; the second direction utilizes the guest molecule representation as bonds and values, and the MOF structural representation as a query, calculating attention weights to dynamically update the MOF node features based on the guest molecule information. Through this bidirectional cross-attention calculation, the model simulates the mutual perception process between the host and guest molecules during adsorption, generating a fused feature representation containing interaction information, thus solving the problem that traditional simple feature splicing cannot characterize dynamic physicochemical relationships.

[0009] S4. Perform adsorption performance prediction and model training optimization. Step S4 inputs the fused features obtained through bidirectional interaction into the regression prediction module. The regression prediction module consists of a global pooling layer and a multilayer perceptron. First, it aggregates the variable-length node features into a quantitative global feature vector through global average pooling. Then, it outputs the final adsorption performance prediction value through an MLP containing two fully connected layers with ReLU activation. During model training, mean squared error is used as the loss function, and the AdamW optimizer iteratively updates the model weights. To improve the model's generalization ability and prediction stability, a Dropout mechanism is introduced during training to prevent overfitting, and the mean absolute error and coefficient of determination are used as performance evaluation metrics until the model's loss function converges on the validation set and the prediction accuracy reaches a preset threshold.

[0010] S5. Screening and application of high-performance MOF materials. In step S5, a large-scale MOF candidate library to be screened is input into a trained prediction model. Input features include the CIF files of the candidate MOFs and the molecular structure of the target chemical warfare agent. The model performs forward inference calculations and outputs the predicted adsorption capacity of the batch of materials for the target agent within seconds. The candidate materials are sorted in descending order based on the predicted values, and materials with adsorption capacities higher than a preset benchmark value are selected as high-performance candidate materials.

[0011] Furthermore, in step S1, to ensure data accuracy, the GCMC simulation is performed at a temperature of 298 K, with pressure ranges covering multiple sampling points from low-pressure to high-pressure regions. For the MOF framework, a universal force field or DREIDING force field parameters are used; for guest molecules, a TraPPE force field or similarly precise force field parameters are used to ensure that the physical laws learned by the model have cross-system universality.

[0012] Furthermore, in step S2, the construction of the crystal graph employs a nearest neighbor search algorithm under periodic boundary conditions. For each central atom, neighboring atoms are searched within its preset radius and edge connections are established, while the Euclidean distance between atoms is encoded in the edge features. Matformer utilizes this crystal graph with distance information to calculate the interaction weights between nodes through a self-attention mechanism, thereby capturing the geometrically confined effects inside the MOF channels.

[0013] Further, in step S3, the bidirectional cross-attention calculation follows the following mathematical logic: First, the MOF node features are mapped using a linear transformation matrix. Mapping object features to Through calculation Implement MOF to assign features to objects, and simultaneously compute This mechanism enables guest-mediated characterization of MOF nodes. It allows the model to automatically identify the atomic sites that contribute most to adsorption during prediction, such as open metal centers or specific organic functional groups, thereby enhancing the model's interpretability at the molecular level.

[0014] Furthermore, in step S4, to further improve the robustness of the prediction, this scheme employs a multiple training prediction strategy. By changing the random seed of the training set and performing multiple independent training iterations, a set of models with the same structure but slightly different parameters is obtained. In the actual prediction task, the arithmetic mean of the prediction results of this set of models is taken as the final output, thereby offsetting the influence of random noise on the prediction results.

[0015] Furthermore, the multilayer perceptron in the regression prediction module adopts a hierarchical compression structure. The first fully connected layer expands the fusion feature dimension from 128 to 256 to extract higher-order nonlinear combination features, while the second fully connected layer compresses it to 64 dimensions. The final output layer is mapped to a scalar adsorption value. This structural design effectively controls the parameter scale and reduces inference latency while maintaining the model's expressive power.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention introduces a bidirectional cross-attention mechanism for modeling the collaborative interaction between the subject and object. This mechanism can dynamically simulate the complex physicochemical interactions between MOF framework atoms and chemical warfare agent molecules during adsorption, fundamentally solving the problems of insufficient fusion of multimodal information between the subject and object and the inability to accurately characterize dynamic interaction relationships in traditional methods and existing AI models. This method achieves extremely high prediction accuracy on the test set, with predicted values ​​showing a high degree of agreement with high-precision giant canonical Monte Carlo simulation values.

[0017] This approach employs a Matformer graph neural network based on a self-attention mechanism to automatically extract atomic-level crystal embeddings of MOFs and combines this with molecular fingerprint encoding of objects, constructing a complete end-to-end deep learning framework. This framework can autonomously learn the most critical features for adsorption performance directly from the original crystal and molecular structures, avoiding tedious and subjective manual feature design and improving the automation and versatility of the method.

[0018] Compared to giant canonical Monte Carlo simulations, which require significant computational resources for each calculation, this approach can predict the adsorption performance of unknown MOF materials in just seconds after model training. This enables rapid, low-cost, high-throughput virtual screening of a large-scale MOF candidate library containing hundreds of thousands of structures, reducing the workload of traditional simulation methods from several years to several hours, and greatly accelerating the development of high-performance protective materials. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a method for predicting the adsorption performance of MOF chemical warfare agents based on a bidirectional attention mechanism, provided in an embodiment of the present invention. Figure 2 This refers to the correlation between predicted and simulated values ​​in this embodiment of the invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] This invention provides a method for predicting the adsorption performance of MOF chemical warfare agents based on a bidirectional attention mechanism. The core idea is to simulate the complex interactions between metal-organic framework materials and chemical warfare agent molecules during the real physical adsorption process using a deep learning architecture. For example... Figure 1 As shown, the method of this invention covers the entire process from underlying data acquisition, multimodal feature encoding, subject-guest interaction modeling to final performance prediction. In practical applications, such as screening protective materials for sarin gas and its analogue dimethyl methylphosphonate, this method constructs a subject-guest bidirectional sensing neural network, enabling rapid and accurate evaluation of the adsorption capacity of a vast number of candidate MOF materials without the need for time-consuming quantum chemical calculations or giant canonical Monte Carlo simulations.

[0022] S1. Constructing a standardized dataset for model training and validation. In this stage, crystal structure information files for metal-organic framework materials are extensively retrieved and downloaded from the CoRE MOF2019 and hMOF databases. These files contain unit cell parameters, atomic coordinates, and space group information, forming the basis for subsequent crystal diagram construction. For chemical warfare agents, this embodiment selects sarin as the target contaminant and its common chemical simulant DMMP as an auxiliary research object to obtain its corresponding molecular structure information. To obtain high-quality training labels, GCMC simulations are performed using the RASPA2 software package to calculate the adsorption amount of the above molecules by the MOF under specific environmental conditions. In terms of simulation parameter settings, the van der Waals interactions of the main MOF atoms are described by the Lennard-Jones potential parameters provided by a universal force field, while the force field parameters for sarin and DMMP molecules are based on the TraPPE force field to ensure accurate expression of molecular flexibility and electrostatic properties. For the partial charge of MOF atoms, this invention uses the MEPO-ML method for rapid estimation. This method predicts the local charge distribution of atoms based on a machine learning model, thereby significantly improving data preparation efficiency while ensuring accuracy. After the simulation was completed, a triplet dataset containing thousands of sets of "MOF structure, guest molecule, and adsorption amount" was obtained. This dataset was then labeled and standardized, meaning the adsorption values ​​were mapped to a standard normal distribution range by subtracting the mean and dividing by the standard deviation. Finally, the dataset was randomly divided into three parts: 70% training set, 15% validation set, and 15% test set.

[0023] S2. Extraction and characterization of subject-object bimodal features. This step is responsible for converting the original physicochemical information into a tensor form that can be processed by a neural network. For MOF crystal structure input, this embodiment uses the Matformer encoding module for processing. First, the CIF file is parsed into an atomic-level crystal diagram, where nodes represent atoms, and their initial feature vectors contain basic attributes such as atomic number, electronegativity, and atomic radius; edges represent spatial connections between atoms, and their features encode the interatomic spacing under periodic boundary conditions. The Matformer encoding module processes the crystal diagram through a multi-layer self-attention mechanism. Each layer of self-attention calculation dynamically allocates weights according to the relative positions of atoms and the chemical environment, thereby extracting a node-level crystal embedding representation containing long-range periodicity.

[0024] In this embodiment, the hidden layer dimension of the Matformer is set to 32, the stacking layer is 3, and the number of heads in the multi-head attention mechanism is 8. This configuration effectively controls computational overhead while ensuring the model's expressive power. For the input of the object molecular structure, this invention extracts the molecule's Morgan fingerprint as its original feature. The Morgan fingerprint generates a 2048-bit sparse bit vector by sampling substructures with a radius of 2 from the molecular topology. Since the dimension of this vector is inconsistent with the feature dimension of the MOF node, this invention introduces a linear mapping layer responsible for projecting the 2048-bit fingerprint vector into a 128-dimensional latent space to generate the object token representation. This linear mapping process achieves dimensional alignment between the subject features and the object features, laying the foundation for subsequent cross-modal interaction.

[0025] S3. Construct and execute bidirectional cross-attention interaction modeling between the host and guest. This is the core technical step of the invention, designed to simulate the dynamic response between the host and guest during adsorption. The MOF node embedding representation obtained in S2 and the guest token representation are input into the collaborative attention block. The collaborative attention block performs bidirectional cross-attention calculation. The first interaction is "guest-perceived host," where the guest token is used as the query vector, and all MOF node embeddings are used as the key and value vectors. By calculating the dot product similarity between the query and the key, and then normalizing it using Softmax, an attention weight matrix is ​​obtained. This matrix reflects the guest molecule's preference for different atomic sites in the MOF. Subsequently, the value is weighted and summed using this weight, integrating the MOF's structural information into the guest representation. The second interaction is "host-perceived guest," where each MOF node embedding is used as the query, and the guest token is used as the key and value. This process allows each atom in the MOF to adjust its feature expression according to the characteristics of the current guest molecule, simulating the inductive effect of the guest molecule on the electron cloud distribution of the framework after entering the pore. Through the iterative alternation of multiple layers of collaborative attention blocks, the model generates deeply fused subject-object interaction features. Compared to traditional feature stitching, this mechanism achieves dynamic coupling at the feature level, enabling the capture of subtle local interactions such as hydrogen bonding and π-π stacking.

[0026] S4. Perform adsorption performance prediction and model training optimization. The features, after bidirectional interaction, contain rich physicochemical information and are input into the subsequent prediction chain. First, a global pooling layer is used to aggregate the node features of the MOF. Since different MOFs have different numbers of atoms, the global pooling layer compresses the variable-length node feature matrix into a fixed-length global feature vector through average pooling or summation pooling operations. Subsequently, this vector is fed into a multilayer perceptron. The MLP regressor consists of two fully connected layers with a ReLU activation function inserted in between to introduce a nonlinear transformation. Finally, the output layer of the MLP regressor produces a scalar value, which is the adsorption performance prediction result. During training, the model uses the mean squared error as the optimization objective and the AdamW optimizer is used for parameter updates.

[0027] To prevent overfitting during training, this embodiment sets a dropout rate of 0.1 in the fully connected layers and employs a weight decay strategy. Furthermore, to eliminate the randomness introduced by model initialization, each prediction task is trained five times independently, and the average of the predicted values ​​is used as the final result to ensure the stability of the prediction performance.

[0028] S5. Screening and Application of High-Performance MOF Materials. In the actual materials discovery process, researchers face a candidate MOF library containing hundreds of thousands of structures. Using a trained model, the structure files of all MOFs in the library are input into the model in batches. The model uses pre-learned bidirectional attention weights to automatically identify materials with suitable pore sizes and strong interaction sites, and quickly provides predicted adsorption amounts for Sarin or DMMP. Based on the prediction results, the materials are ranked, and the top 1% of high-quality candidate materials are selected.

[0029] like Figure 2 As shown, the method of this invention exhibits extremely high prediction accuracy on the test set, with a coefficient of determination R² reaching 0.928. The predicted values ​​closely match the GCMC simulated values, demonstrating the reliability of this technical solution in practical screening tasks. The screened high-performance materials can be further used to manufacture chemical protective suit filter layers, industrial waste gas purification devices, or environmental emergency detection sensors, thereby playing a substantial role in the fields of public safety and environmental protection.

[0030] This invention achieves a precise mapping from static structure to dynamic performance through the complete execution process of S1 to S5 described above. In terms of operational principle, this invention utilizes a self-attention mechanism to extract the intrinsic structural regularities of the host material and a cross-attention mechanism to capture the external interaction regularities between the host and guest components. This two-layer attention architecture reconstructs the essence of the adsorption phenomenon at the atomic scale. During operation, the Matformer encoding module is responsible for perceiving the spatial geometric environment, while the collaborative attention block is responsible for perceiving chemical affinity. Their collaborative work enables the model to maintain extremely high generalization ability when dealing with MOF systems with vastly different structures. Compared to traditional methods, this invention no longer requires manually defining complex geometric descriptors; instead, the neural network autonomously learns the features that play a decisive role in adsorption. This not only improves prediction accuracy but also provides a new perspective for understanding the MOF adsorption mechanism. For example, by visualizing the weight distribution in the collaborative attention block, researchers can clearly observe that the model focuses its attention on open metal sites or specific functional groups in the MOF, which is highly consistent with common sense in physicochemical knowledge, demonstrating the interpretability of this invention in scientific discovery.

[0031] Furthermore, this invention exhibits unique advantages in handling highly hazardous substances such as chemical warfare agents. Since experimentally determining the adsorption data of chemical warfare agents is extremely dangerous and costly, this method provides a safe, low-cost, and highly efficient alternative by combining simulation data-driven approaches with deep learning interactive modeling. In screening large-scale hypothetical MOF databases, this method can complete in just a few hours the workload that traditional simulation methods would take years, significantly accelerating the development of new chemical defense materials. Simultaneously, because the model uses a standardized input interface, it is not only applicable to Sarin and DMMP, but can also be quickly transferred to prediction tasks for other chemical warfare agents such as Soman, VEX, or mustard gas through fine-tuning, demonstrating strong technical scalability.

[0032] In the design of the regression prediction module, the multilayer perceptron, through deep nonlinear mapping of fused features, is able to capture the complex nonlinear trend of adsorption amount changing with structure. For example... Figure 2 The correlation curves shown demonstrate that the model maintains minimal prediction bias even across adsorption amounts spanning multiple orders of magnitude. This is primarily attributed to the fine modulation of subject-guest characteristics by the bidirectional cross-attention mechanism, resulting in a highly environmentally sensitive feature representation. This highly sensitive feature representation is particularly crucial for identifying confined adsorption effects in microporous materials. Through dimensionality reduction processing using a global pooling layer, the model transforms microscopic atomic interaction information into macroscopic adsorption indices, achieving cross-scale modeling from microscopic mechanisms to macroscopic properties. This provides a complete and efficient technical path for the intelligent design of functional porous materials.

[0033] In summary, this invention presents a method for predicting the adsorption performance of MOF chemical warfare agents based on a bidirectional attention mechanism. By constructing a complete pipeline including Matformer encoding, linear mapping alignment, bidirectional cross-attention interaction, and MLP regression prediction, it achieves high-precision and high-efficiency evaluation of the adsorption performance of MOF materials. This invention not only solves the problem of missing subject-object interaction representation in traditional machine learning methods, but also eliminates the dependence on artificial features through an end-to-end deep learning architecture, providing strong technical support for chemical warfare agent protection and toxic and harmful gas capture. In the construction of future chemical defense and security systems, this method can be integrated as a core algorithm into intelligent materials design platforms, helping researchers to quickly respond to sudden chemical threats and develop more efficient filtration and protection equipment.

Claims

1. A method for predicting the adsorption performance of MOF chemical warfare agents based on a bidirectional attention mechanism, characterized in that, The method includes the following steps: S1. Construct a standardized dataset for model training and validation: Collect crystallographic information files of MOF materials from crystal structure databases, obtain molecular structure data of chemical warfare agents and their simulants, use the giant canonical Monte Carlo simulation method to calculate the adsorption amount of each MOF material for a specific chemical warfare agent molecule as the label value, and divide the data into training set, validation set and test set after standardization. S2. Extraction and representation of subject-object dual-modal features: The MOF CIF file is parsed into an atomic-level crystal diagram, and a node-level crystal embedding representation is extracted using a graph neural network Matformer based on a self-attention mechanism; for the object molecule, its Morgan fingerprint vector is extracted and projected to a latent space that matches the MOF node representation dimension through a linear mapping layer to generate the object token representation; S3. Construct and execute bidirectional cross-attention interaction modeling between subject and object: Input the MOF crystal embedding representation and the object molecule token representation into an interaction module composed of multiple layers of collaborative attention blocks. Bidirectional cross-attention calculation is performed within each layer. In the first direction, the object molecule representation is used as the query and the MOF structural features are used as the key and value. In the second direction, the MOF node features are used as the query and the object molecule representation is used as the key and value, generating a fusion feature representation containing interaction information. S4. Perform adsorption performance prediction and model training optimization: Input the fused features into the regression prediction module, which includes a global pooling layer and a multilayer perceptron, and output the adsorption performance prediction value; during model training, mean square error is used as the loss function, and an optimizer is used for iterative updates; S5. Screening and application of high-performance MOF materials: Input the MOF candidate library to be screened into the trained prediction model, output the prediction results of the adsorption amount of the target chemical warfare agent, and screen high-performance candidate materials based on the prediction values.

2. The method for predicting the adsorption performance of MOF chemical warfare agents based on a bidirectional attention mechanism according to claim 1, characterized in that, In step S1, the giant canonical Monte Carlo simulation is performed at a temperature of 298 K and covers multiple pressure sampling points from low pressure to high pressure regions; a general force field or DREIDING force field parameter is used for the MOF framework, and a TraPPE force field parameter is used for the guest molecules.

3. The method for predicting the adsorption performance of MOF chemical warfare agents based on a bidirectional attention mechanism according to claim 1, characterized in that, In step S2, the configuration parameters of the Matformer include: a hidden dimension of 32, a stacking layer of 3 layers, and 8 heads for the multi-head attention mechanism; the dimension of the object molecule token is 128.

4. The method for predicting the adsorption performance of MOF chemical warfare agents based on a bidirectional attention mechanism according to claim 1, characterized in that, In step S3, the bidirectional cross-attention calculation specifically involves mapping the MOF node features to... Mapping object features to Through calculation Implement MOF to assign features to objects, and simultaneously compute Implement feature modification of MOF nodes by objects.

5. The method for predicting the adsorption performance of MOF chemical warfare agents based on a bidirectional attention mechanism according to claim 1, characterized in that, In step S4, a multiple training prediction strategy is adopted. By changing the random seed of the training set, multiple independent trainings are performed to obtain a set of models. In actual prediction, the arithmetic mean of the prediction results of this set of models is taken as the final output.

6. The method for predicting the adsorption performance of MOF chemical warfare agents based on a bidirectional attention mechanism according to claim 1, characterized in that, In step S4, the multilayer perceptron in the regression prediction module adopts a hierarchical compression structure design. Its first fully connected layer expands the fusion feature dimension to 256 dimensions, the second fully connected layer compresses it to 64 dimensions, and the final output layer is mapped to a scalar adsorption value.

Images