Training method, prediction method and device of prediction model of nitrogen reduction reaction

Abstract

A training method, a prediction method and a device of a prediction model of a nitrogen reduction reaction are disclosed. The method comprises: obtaining a sample data set of an electrocatalytic nitrogen reduction reaction; constructing an initial prediction model comprising a first isometric graph attention network encoder, a first single-step path decision maker and a first sequence energy prediction network; updating all network parameters in the first isometric graph attention network encoder based on graph structure data to determine an updated second isometric graph attention network encoder; freezing all network parameters of the second isometric graph attention network encoder, determining a second single-step path decision maker based on reaction step data, and determining a second sequence energy prediction network based on real energy labels; unfreezing all network parameters of the second isometric graph attention network encoder, determining a nitrogen reduction reaction prediction model based on the sample data set, and using the model to predict the input catalyst surface structure. The present application has the effect of improving the prediction efficiency of the nitrogen reduction reaction.

Description

Technical Field

[0001] This disclosure relates to the field of nitrogen reduction reaction prediction technology, and in particular to a training method, prediction method and apparatus for a nitrogen reduction reaction prediction model. Background Technology

[0002] With the continuous development of electrocatalytic ammonia synthesis technology, the computational workload for screening highly active nitrogen reduction catalysts is increasing exponentially. In the post-processing stage of catalyst performance evaluation, the micromechanism analysis module in computational chemistry software can provide researchers with analytical functions for reaction pathways and energy change characteristics by reading the underlying physical model of the catalyst surface structure. However, current processing methods, when faced with massive amounts of catalyst screening data, incur enormous computational and time costs due to the vast volume of microstates involved in the complete nitrogen reduction reaction. With limited computing power, the massive multi-step reaction calculations make it impossible for high-throughput screening processes to proceed normally, or even converge. Therefore, the efficiency of reaction pathway and energy prediction in the performance evaluation process of nitrogen reduction catalysts deserves attention. Summary of the Invention

[0003] In view of this, the present disclosure provides a training method for a prediction model of nitrogen reduction reaction, a prediction method for nitrogen reduction reaction, and an apparatus, in order to improve the prediction efficiency of nitrogen reduction reaction.

[0004] Firstly, a method for training a prediction model for nitrogen reduction reactions is provided, comprising: acquiring a sample dataset of electrocatalytic nitrogen reduction reactions, wherein the sample dataset includes graph structure data of multiple known reaction intermediates on multiple nitrogen reduction reaction pathways, reaction step pairs consisting of adjacent known reaction intermediates, and the true energy label corresponding to each known reaction intermediate; constructing an initial prediction model, the initial prediction model comprising: a first equivalent graph attention network encoder, and a first single-step path decision maker and a first sequence energy prediction network respectively connected to the first equivalent graph attention network encoder, wherein the first equivalent graph attention network encoder is trained based on a preset general material dataset, the first single-step path decision maker is trained based on a classification network structure to be trained, and the first sequence energy prediction network is trained based on a sequence regression network structure to be trained; and based on the graph structure data, training the model for the first nitrogen reduction reaction. All network parameters in the first-order graph attention network encoder are updated to determine the updated second-order graph attention network encoder. All network parameters of the second-order graph attention network encoder are frozen. The first single-step path decision-maker is trained based on reaction step data to determine the second single-step path decision-maker. The first sequence energy prediction network is trained based on real energy labels to determine the second sequence energy prediction network. All network parameters of the second-order graph attention network encoder are unfrozen. The second-order graph attention network encoder, the second single-step path decision-maker, and the second sequence energy prediction network are jointly trained based on the sample dataset to determine the nitrogen reduction reaction prediction model. This nitrogen reduction reaction prediction model is used to predict the surface structure of the input catalyst to generate the reaction path sequence of the nitrogen reduction reaction and the predicted energy distribution corresponding to the reaction path sequence.

[0005] The prediction model obtained through multi-stage joint training can automatically generate a complete reaction pathway sequence and corresponding predicted energy distribution based directly on the input catalyst surface structure. This automated prediction method effectively overcomes the shortcomings of traditional first-principles calculations, which heavily rely on researchers' prior knowledge to manually pre-define reaction pathways, eliminating the arduous task of solving for each adsorption configuration of various reaction intermediates. Simultaneously, the model's direct output of continuous energy distributions reduces the high computational and time costs of traditional density functional theory, thus overcoming the computational bottleneck of existing methods that are limited to case studies. This makes high-throughput screening of thousands of candidate structures truly computationally feasible, accelerating the development of novel, highly efficient nitrogen-fixing catalysts.

[0006] Optionally, the determination of the graph structure data includes: converting the atomic configuration of a known reaction intermediate into a graph structure, wherein each node in the graph structure corresponds to an atom in the atomic configuration, and the feature of the node is configured as the element type of the atom; determining whether the distance between the first atom and the second atom in the atomic configuration is less than a preset cutoff radius; when the distance is less than the cutoff radius, establishing an undirected edge between the first node and the second node corresponding to the first atom and the second atom, respectively, and the feature of the undirected edge is configured as the distance information between the first atom and the second atom.

[0007] Optionally, the first single-step path decision-maker includes a fully connected neural network classifier; the reaction step pair data consists of adjacent preceding and following reaction intermediates; the first single-step path decision-maker is trained based on the reaction step pair data to determine the second single-step path decision-maker, including: inputting the global feature vector corresponding to the preceding reaction intermediate in the reaction step pair data into the fully connected neural network classifier, and outputting the predicted probability of the preceding reaction intermediate being transformed into the next possible reaction intermediate type; updating the network parameters of the fully connected neural network classifier based on the predicted probability, the type of the following reaction intermediate, and the cross-entropy loss function to determine the second single-step path decision-maker.

[0008] Optionally, the first sequence energy prediction network includes a long short-term memory network and a fully connected regression layer. The first sequence energy prediction network is trained based on real energy labels to determine the second sequence energy prediction network, including: inputting a global feature vector sequence composed of multiple known reaction intermediates along the same nitrogen reduction reaction path into the long short-term memory network to determine a hidden state sequence containing historical context information of the reaction process; inputting the hidden state sequence into the fully connected regression layer to output the predicted energy value corresponding to each known reaction intermediate; and updating the network parameters of the long short-term memory network and the fully connected regression layer based on the predicted energy value, the real energy label, and the mean squared error loss function to determine the second sequence energy prediction network.

[0009] Optionally, the acquisition of global feature vectors and global feature vector sequences includes: inputting graph structure data into the second equivalent graph attention network encoder to extract the geometric feature vector corresponding to each atom in the graph structure data; concatenating the geometric feature vectors corresponding to each atom in the atomic dimension to determine the global feature vector corresponding to each known reaction intermediate; and arranging the corresponding multiple global feature vectors in order according to the reaction sequence of multiple known reaction intermediates in the same nitrogen reduction reaction path to determine the global feature vector sequence.

[0010] Optionally, the second isovariant graph attention network encoder, the second single-step path decision-maker, and the second sequence energy prediction network are jointly trained based on the sample dataset to determine the nitrogen reduction reaction prediction model. This includes: inputting the graph structure data from the sample dataset into the second isovariant graph attention network encoder, and performing forward propagation through the second single-step path decision-maker and the second sequence energy prediction network, outputting the prediction probability for the reaction step pair data and the prediction energy value for the graph structure data, respectively; determining the path decision loss based on the prediction probability, the type of subsequent reaction intermediates in the reaction step pair data, and the cross-entropy loss function; determining the energy prediction loss based on the predicted energy value, the true energy label, and the mean squared error loss function; and using the weighted sum of the path decision loss and the energy prediction loss as the total loss function. The network parameters of the second isovariant graph attention network encoder, the second single-step path decision-maker, and the second sequence energy prediction network are updated simultaneously via backpropagation. An independent validation set for the electrocatalytic nitrogen reduction reaction is obtained. Based on the independent validation set, the path decision accuracy of the second single-step path decision-maker after updating the network parameters and the root mean square error of energy prediction of the second sequence energy prediction network are determined. When the path decision accuracy and the root mean square error of energy prediction meet the preset co-optimization conditions, training is stopped, and the nitrogen reduction reaction prediction model is determined. The preset co-optimization conditions include: the path decision accuracy is greater than or equal to a preset accuracy threshold, and the root mean square error of energy prediction is less than or equal to a preset error threshold; or, the path decision accuracy and the root mean square error of energy prediction reach the optimal Pareto state.

[0011] Secondly, a method for predicting nitrogen reduction reactions is provided, used to predict nitrogen reduction reactions using a nitrogen reduction reaction prediction model trained according to the training method provided in the first aspect above. The method includes: acquiring the catalyst surface structure to be predicted; converting the catalyst surface structure to be predicted into target graph structure data and inputting the target graph structure data into the nitrogen reduction reaction prediction model; iteratively predicting the target graph structure data based on a second isovariant graph attention network encoder and a second single-step path decision-maker to generate a reaction path sequence for the nitrogen reduction reaction, wherein the reaction path sequence includes multiple tandem predicted reaction intermediates generated stepwise according to the reaction sequence, starting from the catalyst surface structure to be predicted; inputting the reaction path sequence into a second sequence energy prediction network to determine the predicted energy distribution corresponding to the reaction path sequence; and outputting the prediction result for the catalyst surface structure to be predicted based on the reaction path sequence and the predicted energy distribution.

[0012] Optionally, based on the second isovariant graph attention network encoder and the second single-step path decision-maker, iterative prediction is performed on the target graph structure data to generate a reaction path sequence for nitrogen reduction reaction. This includes: using the target graph structure data as the current graph structure data and inputting it into the second isovariant graph attention network encoder; executing iterative steps, which include: extracting the geometric feature vector corresponding to each atom in the current graph structure data and concatenating the geometric feature vectors corresponding to each atom along the atomic dimension to determine the global feature vector corresponding to the current graph structure data; inputting the global feature vector into the second single-step path decision-maker to determine the next reaction intermediate type with the highest probability; determining the corresponding predicted reaction intermediate based on the next reaction intermediate type and updating the graph structure data of the predicted reaction intermediate to the current graph structure data; repeatedly executing the iterative steps until a preset termination condition is met; and concatenating the initial structure corresponding to the target graph structure data and the generated multiple predicted reaction intermediates according to the execution order of the iterative steps to determine the reaction path sequence. The preset termination conditions include: the next reaction intermediate type indicates the generation of a preset final product, or the number of iterations reached a preset maximum step threshold.

[0013] Optionally, the reaction path sequence is input into a second sequence energy prediction network to determine the predicted energy distribution corresponding to the reaction path sequence. This includes: inputting the target graph structure data and the graph structure data of multiple predicted reaction intermediates contained in the reaction path sequence into a second isovariant graph attention network encoder to extract and concatenate the corresponding multiple global feature vectors; arranging the multiple global feature vectors according to the reaction order to determine a global feature vector sequence; inputting the global feature vector sequence into a long short-term memory network in the second sequence energy prediction network to determine a hidden state sequence containing contextual history information; inputting the hidden state sequence into a fully connected regression layer in the second sequence energy prediction network to output the predicted energy values ​​corresponding to the target graph structure data and multiple predicted reaction intermediates in the reaction path sequence, thus forming a predicted energy distribution; and outputting the prediction results for the catalyst surface structure to be predicted based on the reaction path sequence and the predicted energy distribution, including: determining the energy change characteristics corresponding to the reaction path sequence based on the predicted energy distribution, wherein the energy change characteristics are used to evaluate the catalytic performance of the catalyst surface structure to be predicted for the nitrogen reduction reaction.

[0014] Thirdly, a nitrogen reduction reaction prediction device is provided, used to predict nitrogen reduction reactions based on a nitrogen reduction reaction prediction model trained according to the training method of the nitrogen reduction reaction prediction model provided in the first aspect above, comprising: an acquisition unit for acquiring the catalyst surface structure to be predicted; a conversion unit for converting the catalyst surface structure to be predicted into target graph structure data and inputting the target graph structure data into the nitrogen reduction reaction prediction model; a reaction path sequence determination unit for iteratively predicting the target graph structure data based on a second isovariant graph attention network encoder and a second single-step path decision-maker to generate a reaction path sequence of the nitrogen reduction reaction, wherein the reaction path sequence includes multiple tandem predicted reaction intermediates generated stepwise according to the reaction occurrence sequence starting from the catalyst surface structure to be predicted; a predicted energy distribution determination unit for inputting the reaction path sequence into a second sequence energy prediction network to determine the predicted energy distribution corresponding to the reaction path sequence; and an output unit for outputting the prediction result for the catalyst surface structure to be predicted based on the reaction path sequence and the predicted energy distribution. Attached Figure Description

[0015] The accompanying drawings used in the description of the embodiments of this disclosure are briefly introduced below: Figure 1 A flowchart illustrating a training method for a prediction model of nitrogen reduction reaction provided in some embodiments of this application is shown. Figure 2 This paper illustrates a flowchart of a method for determining a second single-step path decision-maker provided in some embodiments of this application. Figure 3 This paper illustrates a flowchart of a method for determining a second sequence energy prediction network according to some embodiments of this application. Figure 4 This paper illustrates a flowchart of a method for determining a nitrogen reduction reaction prediction model provided in some embodiments of this application. Figure 5 A flowchart illustrating a method for predicting nitrogen reduction reactions provided in some embodiments of this application is shown. Figure 6 A schematic diagram of the structure of a nitrogen reduction reaction prediction device provided in some embodiments of this application is shown. Detailed Implementation

[0016] To more clearly illustrate the technical solutions in the embodiments of this disclosure, examples of implementation methods of this disclosure will be described below with reference to the accompanying drawings. The accompanying drawings described below are merely some embodiments of this disclosure. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without creative effort. Adjustments and improvements made without departing from the concept of this disclosure are all within the protection scope of this disclosure.

[0017] To keep the drawings simple, each figure only schematically shows the parts relevant to the embodiment, and they do not represent the actual structure of the product. In addition, for the sake of clarity and ease of understanding, some figures only schematically show parts of components with the same structure or function, and there may actually be more or fewer components with the same structure or function.

[0018] In this disclosure, unless otherwise expressly specified and limited, ordinal numbers, such as “first”, “second”, etc., are used only to distinguish and describe related objects, and should not be construed as indicating or implying the relative importance or order between related objects; furthermore, they do not represent the quantity of related objects. “Multiple” includes two or more, and other quantifiers are similar.

[0019] In the current screening process for nitrogen reduction catalysts, to meet the requirements of path deduction and energy calculation for complex multi-step reactions, it is usually necessary to rely heavily on first-principles calculations based on density functional theory (DFT). Researchers must manually predefine each potential possible path based on prior knowledge and perform static searches one by one to solve for the adsorption configuration of each reaction intermediate on the catalyst surface. However, nitrogen reduction is a complex electrocatalytic process involving multiple proton-electron transfers. Its microscopic evolution data is not only extremely large, but the workload required to determine the optimal reaction path is also enormous. Because high-throughput screening scenarios for electrocatalysis have extremely high requirements for energy prediction accuracy, lossy simplification schemes that deviate from physical laws are often not feasible to reduce computational complexity. If the underlying data layer still relies solely on traditional first-principles calculations or static graph neural network mechanisms, energy analysis for a single complex catalytic surface system will consume extremely high computational resources and time. When facing high-throughput screening of thousands of candidate structures, this traditional strategy cannot flexibly predict and automatically iterate the dynamic path deduction requirements under different business operations, resulting in a large amount of computational resources being wasted on repetitive and inefficient intermediate geometry optimization. This not only limits existing methods to case studies of a limited number of materials, but also makes efficient and systematic exploration of the entire catalyst configuration space computationally infeasible in practice, thus hindering the development of novel high-efficiency catalysts. To address this, this application proposes a training method, prediction method, and apparatus for a prediction model of the nitrogen reduction reaction, aiming to improve the performance prediction efficiency of nitrogen reduction reaction catalyst materials.

[0020] The following description is in conjunction with the accompanying drawings: Figure 1 A flowchart illustrating a training method for a prediction model of nitrogen reduction reaction provided in some embodiments of this application is shown. The training method includes: S110: Obtain a sample dataset of electrocatalytic nitrogen reduction reaction, wherein the sample dataset includes graph structure data of multiple known reaction intermediates on multiple nitrogen reduction reaction pathways, reaction step pairs consisting of adjacent known reaction intermediates, and the true energy label corresponding to each known reaction intermediate. S120: Construct an initial prediction model, which includes: a first isomorphic graph attention network encoder, a first single-step path decision-maker and a first sequence energy prediction network connected to the first isomorphic graph attention network encoder, wherein the first isomorphic graph attention network encoder is trained and established based on a preset general material dataset, the first single-step path decision-maker is established based on a classification network structure to be trained, and the first sequence energy prediction network is established based on a sequence regression network structure to be trained. S130: Based on the graph structure data, update all network parameters in the first equivariant graph attention network encoder to determine the updated second equivariant graph attention network encoder. S140: Freeze all network parameters of the second isomorphic graph attention network encoder, train the first single-step path decision-maker based on the reaction step data, determine the second single-step path decision-maker, and train the first sequence energy prediction network based on the real energy labels, and determine the second sequence energy prediction network. S150: Unfreeze all network parameters of the second isomorphic graph attention network encoder, and jointly train the second isomorphic graph attention network encoder, the second single-step path decision-maker, and the second sequence energy prediction network based on the sample dataset to determine the nitrogen reduction reaction prediction model. The nitrogen reduction reaction prediction model is used to predict the surface structure of the input catalyst to generate the reaction path sequence of the nitrogen reduction reaction and the predicted energy distribution corresponding to the reaction path sequence.

[0021] In the above training method, the sample dataset integrates the evolutionary information of the nitrogen reduction process, which can include various intermediate states on multiple known reaction paths, and constructs them into a graph structure data including atom types, three-dimensional coordinates, and topological connections. Simultaneously, intermediate states with sequential occurrence are paired to construct reaction step pairs, and each intermediate state is labeled with a high-precision calculated true energy tag. Based on the above data source, the network topology of the prediction model can be initialized. The initialized prediction model can include a backbone network responsible for feature extraction and two branch networks responsible for different downstream tasks. The backbone network adopts a first isovariant graph attention network encoder, which is pre-trained on a broad general material database to obtain basic perception capabilities of translational and rotational symmetry in three-dimensional physical space. The two branch networks are a first single-step path decision-maker for handling discrete classification tasks and a first sequence energy prediction network for handling continuous numerical fitting. Firstly, the parameters of the first isovariant graph attention network encoder can be iteratively updated separately using the graph structure data in the samples, guiding its feature extraction preferences to focus from the general material domain to a specific nitrogen reduction reaction system, resulting in a second isovariant graph attention network encoder. Then, all parameters of the second isomorphic graph attention network encoder are frozen to prevent them from participating in gradient calculations. The first single-step path decision-maker and the first sequence energy prediction network are then driven to perform loss descent and parameter updates using reaction steps on the data and true energy labels. This freezing operation shields the untrained downstream networks from the disordered gradients generated in the early stages of backpropagation, preventing them from disrupting the stable feature representation established by the encoder. After training, a second single-step path decision-maker and a second sequence energy prediction network with preliminary prediction capabilities are obtained. Finally, global joint fine-tuning removes the parameter freezing restriction on the encoder. The three trained models are then jointly trained end-to-end using the complete sample dataset, ultimately converging and outputting the trained nitrogen reduction reaction prediction model. During inference deployment, this model can directly receive the surface structure input of an unknown catalyst and autonomously generate a complete reaction path sequence and corresponding energy distribution through internal iterative forward propagation.

[0022] The prediction model obtained through the aforementioned multi-stage joint training can automatically generate a complete reaction pathway sequence and corresponding predicted energy distribution based directly on the input catalyst surface structure. This automated prediction method effectively overcomes the shortcomings of traditional first-principles calculations, which heavily rely on researchers' prior knowledge to manually pre-determine reaction pathways, eliminating the arduous task of solving for each adsorption configuration of various reaction intermediates. Simultaneously, the model's direct output of continuous energy distributions reduces the high computational and time costs of traditional density functional theory, thus overcoming the computational bottleneck of existing methods that are limited to case studies. This makes high-throughput screening of thousands of candidate structures truly computationally feasible, accelerating the development of novel, highly efficient nitrogen-fixing catalysts.

[0023] In some embodiments of this application, the determination of graph structure data includes: converting the atomic configuration of a known reaction intermediate into a graph structure, wherein each node in the graph structure corresponds to an atom in the atomic configuration, and the feature of the node is configured as the element type of the atom; determining whether the distance between the first atom and the second atom in the atomic configuration is less than a preset cutoff radius; when the distance is less than the cutoff radius, establishing an undirected edge between the first node and the second node corresponding to the first atom and the second atom respectively, and the feature of the undirected edge is configured as the distance information between the first atom and the second atom.

[0024] Regarding the construction process of graph structure data in the sample dataset, the physical atomic configuration of known reaction intermediates can be converted into a mathematical graph structure that is easy for neural networks to process. Each node in this graph structure corresponds to an actual atom in the original microscopic atomic configuration, and to characterize the basic chemical properties of different atoms, the initial features of these nodes can be directly configured as the element type of the corresponding atom. Based on this, to accurately reconstruct the spatial topology and interaction relationships between atoms in the graph structure, the actual spatial distance, such as the Euclidean distance, between any two atoms in the atomic configuration, for example, labeled as the first and second atoms, can be calculated, and it can be determined whether this distance is less than a preset cutoff radius. This preset cutoff radius can be a physical threshold used to define whether there are effective chemical bonds or van der Waals forces or other interactions between adjacent atoms. When the determination result is yes, that is, the actual distance between the first and second atoms is strictly within the range of the cutoff radius, an undirected edge can be established between the first and second nodes representing these two atoms in the graph structure. Simultaneously, to provide accurate geometric scale input to subsequent network layers, the features of this undirected edge are specifically configured as the actual distance information between the first and second atoms. Through the above mapping and extraction of node element type and edge distance features, the microscopic three-dimensional catalytic reaction intermediates are completely transformed into a standard graph structure data format that can be recognized and processed by graph attention networks.

[0025] In some implementations, the first single-step path decision-maker includes a fully connected neural network classifier; the reaction step pair data consists of adjacent preceding and subsequent reaction intermediates. Figure 2 A flowchart illustrating a method for determining a second single-step path decision-maker according to some embodiments of this application is shown. The method includes: S210: Input the global feature vector corresponding to the preceding reaction intermediate in the reaction step pair data into the fully connected neural network classifier, and output the predicted probability of the preceding reaction intermediate being transformed into the next possible reaction intermediate type. S220: Based on the predicted probability, the type of subsequent reaction intermediates, and the cross-entropy loss function, the network parameters of the fully connected neural network classifier are updated to determine the second single-step path decision-maker.

[0026] The first single-step path decision-maker can be implemented using a fully connected neural network classifier. The reaction step pairs in the aforementioned sample dataset consist of pairs of pre- and post-reaction intermediates that are physically adjacent in the actual reaction sequence. This pairing directly reflects the physical causal relationship of the catalytic system's state evolution. During training, the pre-reaction intermediates in the reaction step pairs are first extracted, and their corresponding global feature vectors are input into the fully connected neural network classifier. The internal network layers calculate and output a probability distribution, representing the predicted probability of the pre-reaction intermediate being transformed into each of the next possible reaction intermediate types. Essentially, this step allows the model to predict the most likely intermediate configuration to be generated in the next step in a discrete classification manner, given the current catalytic microstate. Subsequently, the true types of the actual post-reaction intermediates in the reaction step pairs are extracted as classification labels. Based on the predicted probabilities output by the classifier and the true post-reaction intermediate types, the cross-entropy loss function is used to quantify the difference between the predicted probability distribution and the true categories. By minimizing this cross-entropy loss function, backpropagation can be performed to iteratively update the network parameters of the fully connected neural network classifier until the network converges, thus determining a second single-step path decision-maker with accurate pathfinding capabilities.

[0027] In some implementations, the first sequence energy prediction network includes a long short-term memory network and a fully connected regression layer; Figure 3 A schematic flowchart of a method for determining a second sequence energy prediction network according to some embodiments of this application is shown. The method includes: S310: Input the global feature vector sequence consisting of multiple known reaction intermediates along the same nitrogen reduction reaction pathway into the long short-term memory network to determine the hidden state sequence containing historical information about the reaction process context; S320: Input the hidden state sequence into the fully connected regression layer and output the predicted energy value corresponding to each known reaction intermediate; S330: Based on the predicted energy value, the true energy label, and the mean squared error loss function, the network parameters of the long short-term memory network and the fully connected regression layer are updated to determine the second sequence energy prediction network.

[0028] When training the first sequence energy prediction network, multiple known reaction intermediates arranged in chronological order along the same real nitrogen reduction reaction path can be extracted to obtain their corresponding global feature vectors. These vectors are then concatenated temporally to form a global feature vector sequence, which is input into a Long Short-Term Memory (LSTM) network. Since the real nitrogen reduction reaction is a continuous proton-coupled electron transfer process, the stability energy of a later intermediate is often profoundly affected by the evolution of the preceding geometric configuration and changes in electron configuration. The LTM network, through its internal gating mechanisms (such as forget gates and input gates), can effectively capture and retain this physicochemical dependency across multiple elementary reaction steps, thereby determining a hidden state sequence containing the historical context of the entire reaction process. Subsequently, this hidden state sequence can be input one by one into a fully connected regression layer as a numerical mapping module, fitting the high-dimensional feature space containing the historical evolution context into a continuous scalar value, thus outputting the predicted energy values ​​corresponding to each known reaction intermediate along the reaction path. After obtaining the prediction results, the true energy labels calculated based on first principles from the dataset can be retrieved, and the mean squared error (MSE) loss function can be used to accurately measure the absolute numerical deviation between the predicted energy value and the true physical energy. By backpropagating this MSE loss, the weights and bias parameters within the long short-term memory network and the fully connected regression layer are iteratively updated until the energy fitting error converges, thereby determining the second sequence energy prediction network with high-precision sequence prediction capabilities.

[0029] In some implementations, the acquisition of global feature vectors and global feature vector sequences includes: inputting graph structure data into a second equivalent graph attention network encoder to extract the geometric feature vector corresponding to each atom in the graph structure data; concatenating the geometric feature vectors corresponding to each atom in the atomic dimension to determine the global feature vector corresponding to each known reaction intermediate; and arranging the corresponding multiple global feature vectors in order according to the reaction sequence of multiple known reaction intermediates in the same nitrogen reduction reaction path to determine the global feature vector sequence.

[0030] Figure 4 A schematic flowchart of a method for determining a nitrogen reduction reaction prediction model, provided in some embodiments of this application, is shown. The method includes: S410: Input the graph structure data in the sample dataset into the second isovariant graph attention network encoder, and perform forward propagation through the second single-step path decision-maker and the second sequence energy prediction network to output the predicted probability for the reaction step pair data and the predicted energy value for the graph structure data, respectively. S420: Determine the path decision loss based on the predicted probability, the type of subsequent reaction intermediates in the reaction step pair data, and the cross-entropy loss function; S430: Determine the energy prediction loss based on the predicted energy value, the actual energy label, and the mean squared error loss function; S440: The weighted sum of the path decision loss and the energy prediction loss is used as the total loss function, and the network parameters of the second isovariant graph attention network encoder, the second single-step path decision-maker, and the second sequence energy prediction network are updated simultaneously through backpropagation. S450: Obtain an independent validation set for electrocatalytic nitrogen reduction reactions; S460: Based on the independent validation set, determine the path decision accuracy of the second single-step path decision-maker after updating the network parameters, and the root mean square error of the energy prediction of the second sequence energy prediction network; S470: When the path decision accuracy and the root mean square error of energy prediction meet the preset co-optimization conditions, stop training and determine the nitrogen reduction reaction prediction model; wherein, the preset co-optimization conditions include: the path decision accuracy is greater than or equal to the preset accuracy threshold, and the root mean square error of energy prediction is less than or equal to the preset error threshold; or, the path decision accuracy and the root mean square error of energy prediction reach the optimal Pareto state.

[0031] In the above embodiments, the graph structure data in the sample dataset can be input into the unfrozen second isovariant graph attention network encoder to extract low-level geometric features, and the extracted features are simultaneously distributed to the second single-step path decision-maker and the second sequence energy prediction network for parallel computation. These two branch networks output the predicted probability distribution for the reaction step pair data extrapolation, and the predicted energy scalar value for the input graph structure data, respectively. On one hand, based on the predicted probability and the true subsequent reaction intermediate types in the reaction step pair data, the path decision loss, which focuses on the accuracy of discrete state classification, is calculated using the cross-entropy loss function; on the other hand, based on the predicted energy value and the true energy label obtained from micro-calculation, the energy prediction loss, which focuses on the accuracy of continuous numerical fitting, is calculated using the mean squared error loss function. To address the problem of inconsistent dimensions and mutual interference between classification gradients and regression gradients in multi-task learning, the above-mentioned path decision loss and energy prediction loss can be weighted and summed to construct a global overall loss function. Subsequently, a unified backpropagation algorithm is executed based on the total loss function. Using the calculated global gradient, the end-to-end synchronous updates are performed on all network parameters within the second isomorphic graph attention network encoder, the second single-step path decision-maker, and the second sequence energy prediction network.

[0032] To prevent overfitting, an independent validation set for the electrocatalytic nitrogen reduction reaction, which is not involved in parameter updates, can be periodically acquired during training iterations. Using this validation set, the network performance after the current parameter update is evaluated, and the path decision accuracy (characterizing the decision-maker's pathfinding ability) and the root mean square error of energy prediction (RMSE) (characterizing the predictor's energy fitting bias) are calculated. Based on this, the training process can be terminated according to preset co-optimization conditions. These co-optimization conditions include two parallel triggering mechanisms. For example, one could be an absolute threshold achievement, where training stops directly when the path decision accuracy increases to a level greater than or equal to a preset accuracy threshold and the RMSE of energy prediction decreases to a level less than or equal to a preset error threshold. Alternatively, it could be a relative balance achievement, where training stops when both the path decision accuracy and the RMSE of energy prediction reach the optimal Pareto state. In this multi-task learning scenario, the optimal Pareto state refers to the point where the model parameters have evolved to a dynamic equilibrium critical point in a multi-task game. In this state, any parameter adjustments that attempt to further reduce the root mean square error of energy prediction will inevitably lead to a deterioration in path decision accuracy, and vice versa. When any of the above conditions are met, it can be determined that the prediction model has fully converged, and the current network weights are fixed to determine the final nitrogen reduction reaction prediction model.

[0033] To more clearly demonstrate the complete construction and training process of the nitrogen reduction reaction prediction model described in this application, the following example uses the molybdenum disulfide (MoS2) catalytic system. First, a domain-specific dataset can be constructed (i.e., a sample dataset can be obtained). For the MoS2 target catalytic system, first-principles calculation software (such as VASP) is used for microscopic simulation. When the initial state is determined, the stable structures and energies of N2 molecules adsorbed on the MoS2 surface in different configurations are calculated. Subsequently, based on prior knowledge of chemical reactions, all possible elementary steps of nitrogen reduction are enumerated. For example, in the two competing pathways of the distal and alternating paths, free hydrogen atoms may add to different nitrogen atoms in the intermediate NNH, thereby generating two different intermediates, NNHH and *NHNH, which in turn derive different elementary reaction steps. For each elementary step, the atomic configurations of the reactants and products and their corresponding total system energy (EDFT) can be calculated. That is, each stable intermediate is represented as an atomic system containing a list of coordinates (N×3 matrix) and an atomic number list of all atoms in the material. After obtaining the above calculation results, each stable intermediate configuration is converted into graph structure data, where each atom is treated as a node. The element types (Mo, S, N, H, etc.) are mapped to the initial feature vectors of the nodes. Simultaneously, the Euclidean distance between all atom pairs is calculated. If the distance between two atoms is less than a set 5Å cutoff radius, an undirected edge with distance information is established between the two nodes. Then, the corresponding total system energy EDFT is labeled as the true energy label for each graph structure data. Simultaneously, reaction step pair data representing the evolution logic is constructed: for intermediate Mi in step i and its next possible intermediate Mi+1, a pair (Mi, Mi+1) is formed, and its corresponding path is labeled. The final output sample dataset fully contains the graph structure and true energy labels of intermediates such as M0 (catalyst), M1 (NN adsorbing N2), and M2 (NNH adsorbing H at the distal end), as well as reaction step pair data such as (M0, M1) and (M1, M2). After obtaining the dataset, pre-trained model transfer and full fine-tuning can be performed. First, an EquiformerV2 model pre-trained on millions of general crystal and molecular structures (a pre-defined general materials dataset) is introduced as the first isovariant graph attention network encoder in the initial prediction model. This encoder can convert the three-dimensional coordinates and element types of atoms into high-order geometric feature vectors through the mathematical mechanism of spherical harmonics and graph convolution. Using the graph structure data contained in the nitrogen reduction-specific sample dataset constructed above, the atomic embedding layer, the attention weight matrix in the interaction block, and the spherical harmonic coefficients inside the encoder are fully fine-tuned and updated. Through this step, the model accurately adapts from general materials understanding to the specific characterization of nitrogen reduction reaction, such as being more adaptable to the interaction between N and H adsorbates and the MoS2 surface, thereby determining the updated second isovariant graph attention network encoder.

[0034] After the encoder completes domain adaptation, a first single-step path decision-maker and a first sequence energy prediction network can be constructed, and phased collaborative training can be performed.

[0035] The first step is feature extraction and representation. When the graph structure data of a certain intermediate Mi is input, the fine-tuned second-variable graph attention network encoder outputs the geometric feature vector of each atom, specifically represented by a 49×128 two-dimensional tensor. The tensors of all atoms can be concatenated along the atomic dimension to form a global feature vector Gi representing the macroscopic state of the entire intermediate Mi.

[0036] Next, all parameters of the second-order graph attention network encoder are frozen. During the training of the first single-step path decision-maker (fully connected neural network classifier), the global feature vector Gi of the current intermediate is input, and its predicted probability of transformation into each next intermediate type is output. The classifier parameters are updated using the reaction step with the true flow of data and the cross-entropy loss function, thus determining the second single-step path decision-maker. Similarly, during the training of the first sequence energy prediction network, intermediate sequences along the same reaction path, such as M0 to Mk, can be extracted. A global feature vector sequence is obtained using the same encoder and input sequentially into a long short-term memory network. Its hidden states accumulate contextual information about the reaction process. This information is then mapped to the predicted energy values ​​of each intermediate through a fully connected regression layer. The parameters of the LSTM and regression layers are updated only using the true energy label and the mean squared error loss function, thus determining the second sequence energy prediction network.

[0037] Finally, global joint fine-tuning can be performed. The parameters of the second-order graph attention network encoder are unfrozen, and the sample dataset is input into the network. The weighted sum of path decision loss and energy prediction loss is calculated as the total loss function. During joint fine-tuning, the path decision accuracy and root mean square error of energy prediction can be monitored on independent validation sets. When both metrics simultaneously reach preset accuracy and error thresholds, or when improving either metric would harm the other in a Pareto optimal state, the model is considered to have reached the co-optimal condition, and training is stopped. Thus, the final nitrogen reduction reaction prediction model can make reasonable path selections based on historical context and provide accurate and continuous energy predictions.

[0038] Figure 5 A flowchart illustrating a method for predicting nitrogen reduction reactions provided in some embodiments of this application is shown. This prediction method is used to predict nitrogen reduction reactions using a nitrogen reduction reaction prediction model trained according to the training method provided in the above embodiments, and includes: S510: Obtain the surface structure of the catalyst to be predicted; S520: Convert the catalyst surface structure to be predicted into target image structure data, and input the target image structure data into the nitrogen reduction reaction prediction model; S530: Based on the second isovariable graph attention network encoder and the second single-step path decision-maker, the target graph structure data is iteratively predicted to generate a reaction path sequence of nitrogen reduction reaction. The reaction path sequence includes multiple tandem predicted reaction intermediates generated step by step in the order of reaction occurrence, starting from the catalyst surface structure to be predicted. S540: Input the reaction pathway sequence into the second sequence energy prediction network to determine the predicted energy distribution corresponding to the reaction pathway sequence; S550: Based on the reaction pathway sequence and predicted energy distribution, it outputs the prediction results for the surface structure of the catalyst to be predicted.

[0039] In performing actual prediction tasks, the surface structure of the catalyst to be predicted is first acquired, such as the surface of an alloy with a novel composition or a doped and modified material. Since the underlying neural network cannot directly parse macroscopic material chemical formulas or native 3D modeling files, a data preprocessing protocol consistent with the model training phase is required to transform the catalyst surface structure into target graph structure data readable by the model. This target graph structure data, containing the geometric and physicochemical features of the novel catalyst, is formally input into the trained nitrogen reduction reaction prediction model. Once inside the model, the model calls its internal second isomorphic graph attention network encoder to extract high-order geometric features from the currently input target graph structure data and inputs the extracted features into the second single-step path decision-maker. Based on the learned state transition rules, this decision-maker infers the most likely next microscopic reaction action on the catalyst surface, thus generating the next predicted reaction intermediate. To reproduce the real and continuous chemical reaction process, an autoregressive iterative prediction mechanism can be used, where the newly generated predicted reaction intermediate is re-inputted as the current state into the encoder and decision-maker to continue deducing the next reaction path. The above-described deduction process iterates until the preset reaction termination condition is reached, such as the generation of the final product ammonia, thus automatically generating a complete reaction pathway sequence for nitrogen reduction within the model. After successfully identifying the reaction evolution path, an energy assessment can be performed. The complete reaction pathway sequence, composed of multiple intermediates connected in series, is input into the second sequence energy prediction network within the model. Utilizing its internal long short-term memory mechanism, the network fully reads and evaluates the historical context information of the step reactions contained in the sequence, and then directly maps and outputs a predicted energy distribution that strictly corresponds to the reaction pathway sequence. Finally, by combining the automatically generated microscopic reaction pathway sequence and its corresponding macroscopic predicted energy distribution, the final prediction result for the surface structure of the catalyst to be predicted is output. This result intuitively presents the dynamic catalytic process and energy change law of this novel catalyst in the nitrogen fixation process.

[0040] In some embodiments of this application, based on a second isovariant graph attention network encoder and a second single-step path decision-maker, iterative prediction of the target graph structure data is performed to generate a reaction path sequence for nitrogen reduction reaction. This includes: using the target graph structure data as the current graph structure data and inputting it into the second isovariant graph attention network encoder; executing an iterative step, which includes: extracting the geometric feature vector corresponding to each atom in the current graph structure data and concatenating the geometric feature vectors corresponding to each atom along the atomic dimension to determine the global feature vector corresponding to the current graph structure data; inputting the global feature vector into the second single-step path decision-maker to determine the next reaction intermediate type with the highest probability; determining the corresponding predicted reaction intermediate based on the next reaction intermediate type and updating the graph structure data of the predicted reaction intermediate to the current graph structure data; repeatedly executing the iterative step until a preset termination condition is met; and concatenating the initial structure corresponding to the target graph structure data and the generated multiple predicted reaction intermediates according to the execution order of the iterative steps to determine the reaction path sequence. The preset termination condition includes: the next reaction intermediate type indicating the generation of a preset final product, or the number of iterations reached a preset maximum step threshold.

[0041] In some embodiments of this application, the reaction path sequence is input into a second sequence energy prediction network to determine the predicted energy distribution corresponding to the reaction path sequence. This includes: inputting the target graph structure data and the graph structure data of multiple predicted reaction intermediates contained in the reaction path sequence into a second isovariant graph attention network encoder to extract and concatenate the corresponding multiple global feature vectors; arranging the multiple global feature vectors according to the reaction order to determine a global feature vector sequence; inputting the global feature vector sequence into a long short-term memory network in the second sequence energy prediction network to determine a hidden state sequence containing contextual history information; inputting the hidden state sequence into a fully connected regression layer in the second sequence energy prediction network to output the predicted energy values ​​corresponding to the target graph structure data and multiple predicted reaction intermediates in the reaction path sequence, thereby constituting a predicted energy distribution; and outputting a prediction result for the catalyst surface structure to be predicted based on the reaction path sequence and the predicted energy distribution, including: determining the energy change characteristics corresponding to the reaction path sequence based on the predicted energy distribution, wherein the energy change characteristics are used to evaluate the catalytic performance of the catalyst surface structure to be predicted for the nitrogen reduction reaction.

[0042] Continuing with the example of using the surface structure of a molybdenum disulfide catalyst as the input to be predicted, we can first obtain a clean initial surface structure of the MoS2 catalyst, convert it into target graph structure data, and input it as the current graph structure data into the second isovariant graph attention network encoder within the model to initiate the iterative step. In a single iteration, the geometric feature vector corresponding to each atom in the current graph structure data (assumed to be the initial state M0) is extracted and non-destructively stitched along the atomic dimension to determine the global feature vector corresponding to the current graph structure data. Subsequently, this global feature vector is input into the second single-step path decision-maker, which determines the type of the next reaction intermediate with the highest probability based on the internal probability distribution. For example, the model infers that M0 is most likely to adsorb nitrogen, thus predicting the first-step reaction intermediate M1 (such as the NN configuration). Based on this prediction result, the predicted reaction intermediate corresponding to M1 is constructed, and its graph structure data is updated and replaced with the new current graph structure data, and then sent to the next iteration to predict M2 (such as the NNH configuration with distal hydrogen adsorption). The above iterative steps will be executed cyclically within the model until the preset termination condition is met. The termination condition can be set as follows: the type of the next reaction intermediate predicted by the decision-maker indicates the generation of a preset final product (e.g., the prediction of *NH3 desorption), or, to prevent the model from entering an infinite loop, the number of iterations is set to reach a preset maximum step threshold. After iteration terminates, the initial structure corresponding to M0 and the multiple predicted reaction intermediates (M1, M2…) generated by the model are concatenated according to the actual order of the iteration steps, thus determining a complete reaction path sequence [M0, M1, M2, …, Mk]. After obtaining the complete reaction path sequence, it can be input into the second sequence energy prediction network for macroscopic energy calculations. The target graph structure data (M0) and the graph structure data of the multiple predicted reaction intermediates (M1 to Mk) contained in the sequence are respectively fed into the same second isovariant graph attention network encoder for extraction and concatenation, resulting in multiple corresponding global feature vectors. These vectors are arranged according to the actual order of the microscopic reactions, forming a temporally continuous global feature vector sequence. Next, the global feature vector sequence is sequentially input into the Long Short-Term Memory (LSTM) network within the second sequence energy prediction network. The LSM network determines a hidden state sequence containing the entire reaction process's contextual history by accumulating structural evolution information from preceding intermediates (e.g., from M0 to Mi-1). This hidden state sequence is then input into a fully connected regression layer, directly mapping the predicted energy values ​​(E0, E1, E2, ..., Ek) corresponding to each graph structure data in the reaction path sequence, thus forming the predicted energy distribution of the sequence. Finally, based on the determined reaction path sequence and predicted energy distribution, the predicted results for the surface structure of the MoS2 catalyst are output.

[0043] Figure 6A schematic diagram of a nitrogen reduction reaction prediction device provided in some embodiments of this application is shown. The nitrogen reduction reaction prediction model, trained according to the training method of the nitrogen reduction reaction prediction model provided in the above embodiments, includes: an acquisition unit 610 for acquiring the catalyst surface structure to be predicted; a conversion unit 620 for converting the catalyst surface structure to be predicted into target graph structure data and inputting the target graph structure data into the nitrogen reduction reaction prediction model; a reaction path sequence determination unit 630 for iteratively predicting the target graph structure data based on a second isovariant graph attention network encoder and a second single-step path decision-maker to generate a reaction path sequence for the nitrogen reduction reaction, wherein the reaction path sequence includes multiple tandem predicted reaction intermediates generated stepwise according to the reaction sequence, starting from the catalyst surface structure to be predicted; a predicted energy distribution determination unit 640 for inputting the reaction path sequence into a second sequence energy prediction network to determine the predicted energy distribution corresponding to the reaction path sequence; and an output unit 650 for outputting the prediction result for the catalyst surface structure to be predicted based on the reaction path sequence and the predicted energy distribution.

[0044] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not described in detail or in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Furthermore, the above embodiments can be freely combined as needed.

Claims

1. A training method for a predictive model of nitrogen reduction reaction, characterized in that, include: Obtain a sample dataset of electrocatalytic nitrogen reduction reaction, wherein the sample dataset includes graph structure data of multiple known reaction intermediates on multiple nitrogen reduction reaction pathways, reaction step pairs consisting of adjacent known reaction intermediates, and the actual energy label corresponding to each known reaction intermediate; An initial prediction model is constructed, comprising: a first isomorphic graph attention network encoder, and a first single-step path decision-maker and a first sequence energy prediction network respectively connected to the first isomorphic graph attention network encoder, wherein the first isomorphic graph attention network encoder is trained and established based on a preset general material dataset, the first single-step path decision-maker is established based on a classification network structure to be trained, and the first sequence energy prediction network is established based on a sequence regression network structure to be trained. Based on the graph structure data, all network parameters in the first equivariant graph attention network encoder are updated to determine the updated second equivariant graph attention network encoder. Freeze all network parameters of the second isomorphic graph attention network encoder, train the first single-step path decision-maker based on the reaction step data to determine the second single-step path decision-maker, and train the first sequence energy prediction network based on the real energy label to determine the second sequence energy prediction network. Unfreeze all network parameters of the second isomorphic graph attention network encoder, and jointly train the second isomorphic graph attention network encoder, the second single-step path decision-maker, and the second sequence energy prediction network based on the sample dataset to determine the nitrogen reduction reaction prediction model. The nitrogen reduction reaction prediction model is used to predict the surface structure of the input catalyst to generate the reaction path sequence of the nitrogen reduction reaction and the predicted energy distribution corresponding to the reaction path sequence.

2. The training method for the prediction model of nitrogen reduction reaction according to claim 1, characterized in that, The determination of the graph structure data includes: The atomic configuration of the known reaction intermediate is converted into a graph structure, wherein each node in the graph structure corresponds to an atom in the atomic configuration, and the features of the node are configured as the element type of the atom; Determine whether the distance between the first atom and the second atom in the atomic configuration is less than a preset cutoff radius; When the spacing is less than the cutoff radius, an undirected edge is established between the first node and the second node corresponding to the first atom and the second atom, respectively. The feature of the undirected edge is configured as the distance information between the first atom and the second atom.

3. The training method for the prediction model of nitrogen reduction reaction according to claim 2, characterized in that, The first single-step path decision maker includes a fully connected neural network classification. utensil; The reaction step pair data consists of adjacent preceding and subsequent reaction intermediates; The step of training the first single-step path decision-maker based on the reaction step pair data to determine the second single-step path decision-maker includes: The global feature vector corresponding to the preceding reaction intermediate in the reaction step pair data is input into the fully connected neural network classifier, and the predicted probability of the preceding reaction intermediate being transformed into the next possible reaction intermediate type is output. Based on the predicted probability, the type of the subsequent reaction intermediate, and the cross-entropy loss function, the network parameters of the fully connected neural network classifier are updated to determine the second single-step path decision-maker.

4. The training method for the prediction model of nitrogen reduction reaction according to claim 3, characterized in that, The first sequence energy prediction network includes a long short-term memory network and a fully connected regression layer; The step of training the first sequence energy prediction network based on the real energy labels to determine the second sequence energy prediction network includes: The global feature vector sequence consisting of multiple known reaction intermediates along the same nitrogen reduction reaction pathway is input into the long short-term memory network to determine the hidden state sequence containing historical information about the reaction process context. The hidden state sequence is input into the fully connected regression layer, and the predicted energy value corresponding to each of the known reaction intermediates is output. Based on the predicted energy value, the actual energy label, and the mean squared error loss function, the network parameters of the long short-term memory network and the fully connected regression layer are updated to determine the second sequence energy prediction network.

5. The training method for the prediction model of nitrogen reduction reaction according to claim 4, characterized in that, The acquisition of the global feature vector and the global feature vector sequence includes: The graph structure data is input into the second equivariant graph attention network encoder to extract the geometric feature vector corresponding to each atom in the graph structure data; The geometric feature vectors corresponding to each atom are concatenated at the atomic dimension to determine the global feature vector corresponding to each known reaction intermediate. According to the reaction sequence of the known reaction intermediates in the same nitrogen reduction reaction path, the corresponding global feature vectors are arranged in order to determine the global feature vector sequence.

6. The training method for the prediction model of nitrogen reduction reaction according to claim 5, characterized in that, The step of jointly training the second isomorphic graph attention network encoder, the second single-step path decision-maker, and the second sequence energy prediction network based on the sample dataset to determine the nitrogen reduction reaction prediction model includes: The graph structure data in the sample dataset is input into the second isovariant graph attention network encoder, and forward propagated through the second single-step path decision-maker and the second sequence energy prediction network to output the predicted probability for the reaction step pair data and the predicted energy value for the graph structure data, respectively. Based on the predicted probability, the type of subsequent reaction intermediates in the reaction step pair data, and the cross-entropy loss function, the path decision loss is determined. Based on the predicted energy value, the actual energy label, and the mean squared error loss function, the energy prediction loss is determined; The weighted sum of the path decision loss and the energy prediction loss is used as the total loss function, and the network parameters of the second isomorphic graph attention network encoder, the second single-step path decision-maker, and the second sequence energy prediction network are updated simultaneously through backpropagation. Obtain an independent validation set for the electrocatalytic nitrogen reduction reaction; Based on the independent validation set, the path decision accuracy of the second single-step path decision-maker after updating the network parameters and the root mean square error of energy prediction of the second sequence energy prediction network are determined. When the path decision accuracy and the root mean square error of energy prediction meet the preset co-optimization conditions, training is stopped and the nitrogen reduction reaction prediction model is determined. The preset collaborative optimization conditions include: the path decision accuracy is greater than or equal to a preset accuracy threshold, and the root mean square error of energy prediction is less than or equal to a preset error threshold; or, the path decision accuracy and the root mean square error of energy prediction reach the optimal Pareto state.

7. A method for predicting nitrogen reduction reactions, characterized in that, The nitrogen reduction reaction prediction model trained by the training method for the nitrogen reduction reaction prediction model according to any one of claims 1 to 6 performs predictions, including: Obtain the surface structure of the catalyst to be predicted; The catalyst surface structure to be predicted is converted into target image structure data, and the target image structure data is input into the nitrogen reduction reaction prediction model; Based on the second isomorphic graph attention network encoder and the second single-step path decision-maker, the target graph structure data is iteratively predicted to generate the reaction path sequence of the nitrogen reduction reaction, wherein the reaction path sequence includes multiple tandem predicted reaction intermediates generated step by step in the order of reaction occurrence, starting from the catalyst surface structure to be predicted. The reaction path sequence is input into the second sequence energy prediction network to determine the predicted energy distribution corresponding to the reaction path sequence; Based on the reaction path sequence and the predicted energy distribution, a prediction result for the surface structure of the catalyst to be predicted is output.

8. The method for predicting nitrogen reduction reactions according to claim 7, characterized in that, The step of iteratively predicting the target graph structure data based on the second isomorphic graph attention network encoder and the second single-step path decision-maker to generate the reaction path sequence of the nitrogen reduction reaction includes: The target graph structure data is used as the current graph structure data and input into the second equivariant graph attention network encoder; The iterative steps include: extracting the geometric feature vector corresponding to each atom in the current graph structure data, and concatenating the geometric feature vectors corresponding to each atom in the atomic dimension to determine the global feature vector corresponding to the current graph structure data; The global feature vector is input into the second single-step path decision-maker to determine the next reaction intermediate type with the highest probability; the corresponding predicted reaction intermediate is determined based on the next reaction intermediate type, and the graph structure data of the predicted reaction intermediate is updated to the current graph structure data; The iterative steps are repeated until a preset termination condition is met; According to the execution order of the iterative steps, the initial structure containing the target graph structure data and the generated multiple predicted reaction intermediates are connected in series to determine the reaction path sequence; The preset termination conditions include: the next reaction intermediate type indicates the generation of a preset final product, or the number of times the iterative steps are executed reaches a preset maximum step threshold.

9. The method for predicting nitrogen reduction reactions according to claim 8, characterized in that, The step of inputting the reaction path sequence into the second sequence energy prediction network to determine the predicted energy distribution corresponding to the reaction path sequence includes: The target graph structure data contained in the reaction path sequence and the graph structure data of multiple predicted reaction intermediates are respectively input into the second equivariant graph attention network encoder to extract and concatenate the corresponding multiple global feature vectors. Arrange the multiple global feature vectors according to the order in which the reactions occur to determine the global feature vector sequence; The global feature vector sequence is input into the long short-term memory network in the second sequence energy prediction network to determine the hidden state sequence containing contextual history information; The hidden state sequence is input into the fully connected regression layer in the second sequence energy prediction network, and the target graph structure data in the reaction path sequence and the predicted energy values ​​corresponding to the multiple predicted reaction intermediates are output to form the predicted energy distribution. Based on the reaction pathway sequence and the predicted energy distribution, the predicted results for the catalyst surface structure to be predicted are output, including: Based on the predicted energy distribution, the energy change characteristics corresponding to the reaction path sequence are determined, wherein the energy change characteristics are used to evaluate the catalytic performance of the catalyst surface structure to be predicted for the nitrogen reduction reaction.

10. A device for predicting nitrogen reduction reactions, characterized in that, The nitrogen reduction reaction prediction model trained by the training method for the nitrogen reduction reaction prediction model according to any one of claims 1 to 6 performs predictions, including: Acquisition unit, used to acquire the surface structure of the catalyst to be predicted; A conversion unit is used to convert the catalyst surface structure to be predicted into target image structure data, and input the target image structure data into the nitrogen reduction reaction prediction model; The reaction path sequence determination unit is used to iteratively predict the target graph structure data based on the second isovariant graph attention network encoder and the second single-step path decision-maker to generate the reaction path sequence of the nitrogen reduction reaction, wherein the reaction path sequence includes multiple tandem predicted reaction intermediates generated step by step in the order of reaction occurrence, starting from the catalyst surface structure to be predicted. A predicted energy distribution determination unit is used to input the reaction path sequence into the second sequence energy prediction network and determine the predicted energy distribution corresponding to the reaction path sequence. The output unit is used to output a prediction result for the surface structure of the catalyst to be predicted based on the reaction path sequence and the predicted energy distribution.

Images