A catalytic performance prediction method based on transformer and force-chemical coupling

Abstract

The application discloses a catalytic performance prediction method based on a force-chemical coupling of a Transformer, and belongs to the technical field of the intersection of computational materials science and artificial intelligence; the method comprises the following steps: obtaining a catalytic material sample set, determining key descriptors of four dimensions of surface adsorption force, electronic interaction force, orbital interaction force and chemical bond strength of each sample, and obtaining corresponding kinetic energy barriers, and a training set is formed by the descriptors and the energy barriers; a force-chemical coupling analysis model based on the Transformer is constructed, the model uses a multi-head self-attention mechanism to capture long-range dependence and synergistic effect among the descriptors; the model is trained by using the training set to establish a mapping relationship between the descriptors and the energy barriers; the descriptors of a sample to be predicted are input into the trained model to obtain a predicted energy barrier; the application breaks through the limitation of a single descriptor, and realizes quantitative analysis and accurate prediction of the force-chemical coupling effect.

Description

Technical Field

[0001] This invention relates to the field of computational materials science and artificial intelligence, and in particular to a method for predicting catalytic performance based on Transformer and mechanochemical coupling. Background Technology

[0002] In traditional catalysis theory research, researchers usually focus on finding a single physicochemical descriptor (such as adsorption energy, d-band center, etc.) to describe catalytic activity, and construct a "volcano-shaped curve" by fitting the linear relationship between the descriptor and the catalytic rate or rate-controlling step energy barrier, which serves as the basis for screening catalysts.

[0003] For example, Nørskov's d-band center theory has achieved great success in explaining the catalytic activity of transition metal surfaces. However, with the increasing complexity of research systems (such as single-atom catalysis and heterojunction interface catalysis), the traditional "single descriptor-linear fitting" research paradigm has gradually revealed its limitations: a single descriptor cannot fully capture complex mechanisms. Catalytic reactions are essentially complex kinetic processes involving electronic structure evolution, orbital hybridization, chemical bond breaking and recombination, and changes in surface geometry. Catalytic performance is often influenced by multiple factors (such as lattice stress, electronic interaction forces, orbital overlap, and spin states). A single descriptor often struggles to fully capture these complex "mechanical-chemical" coupling effects, leading to limitations in understanding complex systems. The predicted catalytic activity has a large bias; the universality of linear laws is insufficient. Traditional volcano curve models usually assume a simple linear or quasi-linear relationship between descriptors and energy barriers. However, in actual nanocatalytic systems, the relationship between various physicochemical properties and catalytic performance often exhibits a high degree of nonlinearity; there is a lack of quantitative analysis of the "mechanical-chemical coupling" mechanism. Existing theoretical studies focus more on the electronic structure itself, while there is a lack of systematic theoretical descriptions and quantitative analysis methods for how the mechanical interactions between atoms in the catalytic process couple with the energy barrier of the chemical reaction. This makes it impossible to summarize effective universal laws to guide material design. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting catalytic performance based on Transformer and mechanochemical coupling, which breaks through the limitations of a single descriptor and realizes the quantitative analysis and accurate prediction of mechanochemical coupling effects.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a method for predicting catalytic performance based on Transformer and mechanochemical coupling, comprising the following steps: Obtain a sample set containing different catalytic material models; determine multiple key descriptors for each sample in the sample set, including surface adsorption forces, electronic interaction forces, orbital interaction forces, and chemical bond strength; obtain the kinetic energy barrier corresponding to each sample, and form a training set with multiple key descriptors for each sample; A mechanochemical coupling analytical model based on the Transformer architecture is constructed. The mechanochemical coupling analytical model includes an embedding layer, an encoder layer, and an output layer. The encoder layer adopts a multi-head self-attention mechanism and a feedforward neural network to capture the long-range dependencies and synergistic weights between different descriptors and to weightedly fuse the features corresponding to the descriptors output by the embedding layer. The mechanical-chemical coupling analytical model is trained using a training set to obtain a mechanical-chemical coupling analytical model for predicting kinetic energy barriers.

[0006] In some optional embodiments, the descriptor includes four dimensions: surface adsorption force, electronic interaction force, orbital interaction force, and chemical bond strength, specifically including: Surface adsorption force descriptors include parameters characterizing the adsorption strength of reactants or intermediates on the catalyst surface; Electron interaction force descriptors include parameters characterizing the charge transfer capability and electron localization degree of active centers; Orbital interaction force descriptors include parameters characterizing the energy level matching and orbital hybridization degree between the catalyst and the adsorbed species; Chemical bond strength and stability class descriptors, including parameters characterizing bond strength and stability.

[0007] In some optional embodiments, the surface adsorption force class descriptor includes at least one of the following parameters: reactant adsorption energy and intermediate adsorption energy.

[0008] In some optional embodiments, the electron interaction force class descriptor includes at least one of the following parameters: The active site charge transfer amount, the charge difference integral value between the reaction intermediate and the active site, and the extremum of the electronic localization function at the active site are based on Bader charge analysis. Among them, the charge transfer amount is used to quantify the redox capability of the active site, and the extremum of the electronic localization function is used to characterize the degree of electronic localization and bonding characteristics.

[0009] In some optional embodiments, the orbital interaction force class descriptor includes at least one of the following parameters: The contribution weight of active atoms to the valence band of the system, the peak overlap energy position of active atom orbitals and adsorbed molecular orbitals in the density of states diagram, the degree of hybridization between active atom orbitals and intermediate orbitals, the d-band center position of the active center, and the energy gap between the highest occupied molecular orbital and the lowest unoccupied molecular orbital; wherein, the d-band center position is used to characterize the average energy level of d electrons, and the energy gap is used to characterize the ease of electronic transition.

[0010] In some optional embodiments, the chemical bond strength and stability class descriptor includes at least one of the following parameters: The crystal orbital Hamiltonian layout integral value, the antibonding state filling level at the Fermi level, and the second-order perturbation stabilization energy based on the natural bond orbital theory are used to quantify the bond strength, the antibonding state filling level is used to characterize the stability of the chemical bond, and the second-order perturbation stabilization energy is used to quantify the charge delocalization stabilization effect between orbitals.

[0011] In some optional embodiments, the method further includes obtaining a mechanochemical coupling analytical model for predicting the kinetic energy barrier, and then using the attention weight matrix and SHAP value in the mechanochemical coupling analytical model to analyze the contribution of each descriptor to the kinetic energy barrier, thereby summarizing the universal mechanochemical coupling rules for this type of catalytic reaction.

[0012] In some optional embodiments, the sample set and the corresponding kinetic energy barrier are obtained based on theoretical calculations. The sample set is constructed by structural optimization and stability screening of different catalytic material models, and the kinetic energy barrier is obtained by simulation calculation of the catalytic reaction pathway.

[0013] In some optional embodiments, the embedding layer normalizes the descriptors and maps them to high-dimensional feature vectors, and introduces positional encoding to preserve the physical attribute sequence information of the descriptors.

[0014] In some optional embodiments, during the training of the mechani-chemical coupling analytical model using the training set, a loss function is used to measure the error between the predicted and true values, the parameters of the mechani-chemical coupling analytical model are updated by an optimizer, and a learning rate scheduling strategy and regularization technique are used to prevent overfitting.

[0015] The catalytic performance prediction method based on Transformer and mechanochemical coupling provided by this invention has at least the following beneficial effects: This invention overcomes the limitations of traditional volcano curves that rely on a single descriptor. By constructing a key descriptor system encompassing four dimensions—surface adsorption forces, electronic interactions, orbital interactions, and chemical bond strength—it achieves a systematic and quantitative description of the "force-chemical coupling" effect in catalytic reactions, comprehensively reflecting the multi-factor synergistic mechanisms in complex catalytic systems. Building upon this, the invention introduces the Transformer architecture into the field of catalytic performance prediction, utilizing a multi-head self-attention mechanism to dynamically capture long-range dependencies and synergistic weights between different descriptors. This effectively solves the problem of high-dimensional nonlinear mapping, which is difficult for traditional linear models to handle, significantly improving the accuracy and reliability of catalytic performance prediction. The method of this invention has broad applicability and can be applied to various catalytic systems and reactions. By replacing a large number of tedious DFT calculations with model prediction, it significantly shortens the development cycle of novel catalytic materials and reduces development costs.

[0016] Furthermore, by conducting interpretability analysis on the model, this invention can quantitatively analyze the contribution of each descriptor to catalytic activity, reveal the physical essence of the dominant catalytic reaction, and provide clear theoretical guidance for catalytic mechanism research and catalyst reverse design. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0018] Figure 1 This is a flowchart of a catalytic performance prediction method based on Transformer and mechanochemical coupling according to an embodiment of the present invention; Figure 2 This is a technical framework diagram of the entire process of rational design of catalytic materials based on Transformer and mechanochemical coupling, provided according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] One embodiment of the present invention relates to a method for predicting catalytic performance based on Transformer and mechanochemical coupling. The implementation details of the method for predicting catalytic performance based on Transformer and mechanochemical coupling in this embodiment are described in detail below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.

[0021] The specific process of the catalytic performance prediction method based on Transformer and mechanochemical coupling in this embodiment can be described as follows: Figure 1 As shown, it includes: Step 101: Obtain a sample set containing different catalytic material models; determine multiple key descriptors for each sample in the sample set, including surface adsorption forces, electronic interaction forces, orbital interaction forces, and chemical bond strength; obtain the kinetic energy barrier corresponding to each sample, and form a training set with multiple key descriptors for each sample. The sample set and the corresponding kinetic energy barrier are obtained based on theoretical calculations. The sample set is constructed by optimizing the structure and screening the stability of different catalytic material models, and the kinetic energy barrier is obtained by simulating the catalytic reaction pathway.

[0022] Surface adsorption force descriptors include parameters characterizing the adsorption strength of reactants or intermediates on the catalyst surface; Electron interaction force descriptors include parameters characterizing the charge transfer capability and electron localization degree of active centers; Orbital interaction force descriptors include parameters characterizing the energy level matching and orbital hybridization degree between the catalyst and the adsorbed species; Chemical bond strength and stability class descriptors, including parameters characterizing bond strength and stability.

[0023] Surface adsorption force descriptors include at least one of the following parameters: reactant adsorption energy and intermediate adsorption energy.

[0024] Electron interaction force class descriptors include at least one of the following parameters: The active site charge transfer amount, the charge difference integral value between the reaction intermediate and the active site, and the extremum of the electronic localization function at the active site are based on Bader charge analysis. Among them, the charge transfer amount is used to quantify the redox capability of the active site, and the extremum of the electronic localization function is used to characterize the degree of electronic localization and bonding characteristics.

[0025] The orbital interaction force class descriptor includes at least one of the following parameters: The contribution weight of active atoms to the valence band of the system, the peak overlap energy position of active atom orbitals and adsorbed molecular orbitals in the density of states diagram, the degree of hybridization between active atom orbitals and intermediate orbitals, the d-band center position of the active center, and the energy gap between the highest occupied molecular orbital and the lowest unoccupied molecular orbital; wherein, the d-band center position is used to characterize the average energy level of d electrons, and the energy gap is used to characterize the ease of electronic transition.

[0026] Chemical bond strength and stability class descriptors include at least one of the following parameters: The crystal orbital Hamiltonian layout integral value, the antibonding state filling level at the Fermi level, and the second-order perturbation stabilization energy based on the natural bond orbital theory are used to quantify the bond strength, the antibonding state filling level is used to characterize the stability of the chemical bond, and the second-order perturbation stabilization energy is used to quantify the charge delocalization stabilization effect between orbitals.

[0027] Step 102: Construct a force-chemical coupling analytical model based on the Transformer architecture. The force-chemical coupling analytical model includes an embedding layer, an encoder layer, and an output layer. The encoder layer uses a multi-head self-attention mechanism and a feedforward neural network to capture the long-range dependencies and synergistic weights between different descriptors and to weightedly fuse the features corresponding to the descriptors output by the embedding layer. The embedding layer normalizes the descriptors and maps them into high-dimensional feature vectors, and introduces positional encoding to preserve the physical attribute sequence information of the descriptors.

[0028] The encoder layer adopts a multi-layer stacked structure, with each layer containing a multi-head self-attention mechanism and a feedforward neural network, and is optimized through residual connections and layer normalization.

[0029] During training, a loss function is used to measure the error between the predicted and true values. The parameters of the mechanistic-chemical coupling analytical model are updated by an optimizer, and a learning rate scheduling strategy and regularization technique are used to prevent overfitting.

[0030] After training, the attention weight matrix and SHAP value in the mechanochemical coupling analytical model are used to analyze the contribution of each descriptor to the kinetic energy barrier, and the universal mechanochemical coupling law for this type of catalytic reaction is summarized.

[0031] Step 103: Train the mechanical-chemical coupling analytical model using the training set to obtain a mechanical-chemical coupling analytical model for predicting the kinetic energy barrier.

[0032] Example 1: Construction of Catalytic Material Sample Set and Descriptor Extraction This embodiment details the construction process of the catalytic material sample set and the extraction method of the multidimensional "mechanical-chemical coupling" descriptor, aiming to provide high-quality, standardized training data for deep learning models.

[0033] The technical framework diagram of the entire process of rational design of catalytic materials based on Transformer and mechanochemical coupling is as follows: Figure 2 As shown in the figure, the closed loop of catalytic material research and development, with theoretical calculation as the core and artificial intelligence technology as the link, is systematically presented. The upper part clarifies the theoretical calculation foundation and multi-scale theoretical simulation system at the thermodynamic and kinetic levels. The middle circle defines the core research objectives of mechanism analysis, law mining, and modification design. The lower part focuses on technology implementation, showing the complete path from extracting 13-dimensional mechanochemical coupling descriptors, constructing a multi-descriptor mechanical coupling model based on deep learning, to analyzing the influencing factors of single-atom catalysis and formulating nonlinear modification strategies. It intuitively reflects the whole chain logic from basic theoretical calculations to AI-assisted mechanism analysis and material optimization.

[0034] Step 1: High-throughput modeling and stability screening of catalytic material systems Initial model construction: Based on first-principles calculation software (such as VASP, CASTEP, etc.), an atomic model of the catalytic material to be studied is constructed. The model is not limited to a specific material type and may include two-dimensional materials (such as MXene, graphene, MoS2), metal surfaces, oxides, or high-entropy alloys. In this embodiment, a 3×3 or 4×4 supercell structure is used as the substrate, and transition metal atoms (such as Fe, Co, Ni, Cu, Zn, etc.) are selected as active centers for doping or loading to construct a single-atom catalyst or a two-atom catalyst model.

[0035] Parameter settings: The projected plane wave method is used to describe the interaction between the ionic core and valence electrons. The exchange correlation functional is the PBE functional under the generalized gradient approximation. To improve computational accuracy, DFT-D2 or DFT-D3 methods are introduced for van der Waals force dispersion correction. The plane wave cutoff energy is set to 450–550 eV, and the energy convergence criterion is 10^2 eV. -5 eV, with a force convergence criterion of 0.01~0.02 eV / Å.

[0036] Stability assessment and screening: Structural relaxation was performed on all initial configurations. Subsequently, stable configurations were selected and stored as valid samples in the database based on the following three indicators:

[0037] Thermodynamic stability: The formation energy is calculated and is required to be less than 0 to ensure that the material is thermodynamically easy to synthesize; Dynamic stability: Calculate the phonon spectrum, requiring no imaginary frequencies across the entire spectrum to ensure that the structure has no dynamic instability; Thermal stability simulation: Perform molecular dynamics simulations (e.g., run at 300K or higher for 10-20 ps), requiring total energy fluctuations to converge and maximum atomic displacement deviations to be less than a threshold (e.g., 0.2 Å), ensuring that the material does not remodel at the operating temperature.

[0038] Step 2: Extraction and Quantification of Multidimensional "Mechanical-Chemical Coupling" Descriptors This step is one of the core features of this invention. Unlike traditional methods that only use d-band centers, this invention extracts 13 key descriptors (X1~X13) from four dimensions: mechanics, electronics, orbitology, and chemical bonding, to holographically map the microscopic physical picture of the catalytic reaction. All descriptors are obtained through DFT post-processing:

[0039] First dimension: Surface adsorption force parameters (reflecting the direct interaction between geometry and energy) X1 (Reactant Adsorption Energy): Calculates the adsorption energy of the substrate on the initial reactant molecules (such as H2O2, H2O, N2, etc.), characterizing the affinity of the reaction in its initial state.

[0040] X2 (intermediate adsorption energy): Calculates the adsorption energy of key reaction intermediates (such as *OH, *O, *OOH, etc.). This parameter is usually closely related to the linear proportional relationship.

[0041] The second dimension: electron interaction force parameters (reflecting charge transfer and localization) X3 (Bader charge transfer): Using the Bader charge analysis method, the number of electrons gained or lost by the active center atom during the catalytic process is calculated to quantify the redox capability.

[0042] X4 (Differential Charge Density Integral): Calculates the differential charge density between the active site and the adsorbed molecule, and integrates its spatial distribution to characterize the intensity of interfacial charge rearrangement.

[0043] X9 (Electron Localization Function (ELF) Extrema): Calculates the electron localization function at the active site and takes its extreme value (0~1). High values ​​represent covalent bond characteristics (high electron localization), while low values ​​represent metallic or ionic bond characteristics (electron delocalization).

[0044] The third dimension: orbital interaction parameters (reflecting energy level matching and hybridization mechanisms). X5 (Valence Band Contribution Weight): Calculates the contribution ratio of active atomic orbitals to electronic states near the top of the valence band or the Fermi level of the system through density of states (DOS) analysis.

[0045] X6 (DOS peak overlap energy): Identifies the main peak positions of the d orbitals of active atoms and the p orbitals of adsorbed molecules in the DOS diagram, calculates the energy difference between the two, and characterizes the orbital resonance matching degree.

[0046] X7 (Orbital Hybridization): Calculates the integral overlap area between the bonding orbitals of the active center and the adsorbed species, quantifying the strength of hybridization.

[0047] X8 (d-band center): The position of the d-band center of the active metal atom is calculated according to the formula. This is a classic electronic structure descriptor.

[0048] X12 (Frontier band gap): Calculates the energy difference between the highest occupied molecular orbital (HOMO) and the lowest unoccupied molecular orbital (LUMO) of the system, characterizing the ease or difficulty of electronic transitions.

[0049] Fourth dimension: Chemical bond strength and stability parameters (reflecting the nature of bonding and the energy barrier for breaking bonds) X10 (Crystal Orbital Hamiltonian Layout (ICOHP): Calculates the crystal orbital Hamiltonian layout (COHP) of active bonds (such as metal-oxygen bonds) using the LOBSTER program and integrates it to the Fermi level (ICOHP). The more negative the value, the stronger the bond.

[0050] X11 (Antibonding state filling degree): The number of electrons filling the antibonding orbitals below the Fermi level. The more antibonding states filled, the more unstable the chemical bond, and the easier it is to break or desorb.

[0051] X13 (NBO Second-Order Perturbation Energy): Based on the Natural Bond Orbital (NBO) theory, this method calculates the second-order perturbation interaction energy between electron donor and acceptor orbitals, accurately quantifying the charge delocalization stabilization effect between orbitals.

[0052] Step 3: Building and training a deep learning model based on the Transformer architecture This step utilizes advanced AI algorithms to uncover the nonlinear "mechanical-chemical" coupling between the aforementioned 13 descriptors and catalytic activity (such as the reaction barrier Ea).

[0053] Data preprocessing: The 13-dimensional descriptor data extracted in step 2 is organized into a feature matrix and Z-score standardized to eliminate dimensional differences.

[0054] Model architecture setup: Build a Transformer regression model based on the Self-Attention mechanism.

[0055] Embedding layer: Maps the 13-dimensional input vector to a high-dimensional latent space (e.g., d_model=64 or 128) and superimposes positional encoding to preserve the logical information of the physical parameters.

[0056] Encoder: Stacked L layers (L=4~6 recommended) of Transformer Encoder Blocks. Each layer contains two sub-layers:

[0057] Multi-head self-attention layer: Set the number of heads h = 4 or 8. By calculating the query, key, and value matrix, the model can automatically capture long-range dependencies and collaborative weights between different descriptors (such as d-band centers and ICOHPs).

[0058] Feedforward neural network layer: contains two fully connected layers and ReLU or GELU activation functions to enhance nonlinear fitting capabilities.

[0059] Residual connection and normalization: Each sub-layer is followed by Add&Norm (Layer Normalization) operations to prevent gradient vanishing and accelerate convergence.

[0060] Output layer: The high-dimensional features output by the encoder are compressed into a one-dimensional scalar through a fully connected layer, namely the predicted catalytic reaction energy barrier Ea.

[0061] Model training: Divide the model into training, validation, and test sets (e.g., 8:1:1). Set the loss function to mean squared error (MSE). Use the Adam optimizer with an initial learning rate of 10^6. -3 ~10^ -4 The learning rate is dynamically adjusted using a cosine annealing strategy. Dropout (ratio 0.1~0.2) and L2 regularization are introduced to prevent overfitting.

[0062] Model Interpretation and Pattern Extraction: After training, the attention weight matrix in the Transformer model is extracted. By analyzing the weight heatmap, the key descriptor combinations that contribute most to the prediction results are identified, thereby summarizing the physical laws governing this type of catalytic reaction.

[0063] Example 2: Study on the release pattern of reactive oxygen species (ROS) in transition metal single-atom catalysts using the method of the present invention. This embodiment uses a TiNO2-type MXene substrate-supported transition metal single-atom catalytic system as an example to demonstrate how to use the above-mentioned general method to resolve the catalytic mechanism of H2O2 decomposition to generate ·OH radicals.

[0064] 1. Dataset Construction and Computation Model system: TiNO2-type MXene was selected as the substrate, and more than 20 transition metal atoms, such as Sc, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Mo, Ru, Rh, Pd, Ag, W, Os, Ir, Pt, and Au, were selected and anchored to different vacancies (such as Ti vacancies) on the MXene surface to construct single-atom catalysts (SACs).

[0065] First-principles calculations: Calculations were performed using VASP software. The plane wave cutoff energy was set to 500 eV, and the k-point grid was 5×5×1. Formation energy calculations and molecular dynamics simulations (300 K, 10 ps) were performed on all systems, and 28 structurally stable doped systems were ultimately selected as valid samples.

[0066] Reaction pathway simulation: For the reaction pathway H2O2→2OH→·OH+OH*, the CI-NEB method was used to search for transition states, and the kinetic energy barrier (Ea) of the rate-determining step (usually OO bond breaking or ·OH desorption) was calculated as the target label for the model. The calculation results show that the Ea of different metal-doped systems ranges from 0.3 eV to 1.5 eV.

[0067] 2. Descriptor Extraction For the 28 systems mentioned above, 13-dimensional feature vectors were extracted. The calculation details of some key data are as follows:

[0068] d-band center (X8): Integral calculation of the d-orbital density of states of a single metal atom reveals that the d-band center of metals such as Fe and Co is closer to the Fermi level (around -1.5 eV), while that of metals such as Zn and Cu is farther away (below -6.0 eV).

[0069] ICOHP (X10): The focus is on analyzing the bonding strength between metal atoms and oxygen atoms in H2O2. The results show that the Co-O bond has a larger ICOHP integral value, indicating a stronger bond, which is conducive to activating the OO bond.

[0070] HOMO-LUMO band gap (X12): Calculations show that the band gap of highly efficient catalysts is usually small (<1.0 eV), which is conducive to electron transitions between the metal and the substrate.

[0071] 3. Transformer Model Training and Prediction Model parameters: Build a Transformer model with 6 encoder layers, feature dimension d_model=64, number of attention heads 8, and fully connected layer dimension 256.

[0072] Training process: Training was performed using 2000 epochs with a batch size of 4. During training, the training set loss rapidly decreased from an initial 1.5 to 0.02 and then converged.

[0073] Predictive Performance: On the test set, the model's predicted energy barrier value achieves a determination coefficient (R²) of 0.92 compared to the actual value calculated by the DFT, with a root mean square error (RMSE) of only 0.08 eV. In contrast, the traditional univariate linear fitting model based on d-band centers has an R² of only 0.65. This demonstrates the significant advantage of the Transformer architecture of this invention in handling complex nonlinear relationships.

[0074] 4. Analysis of patterns and revelation of the "mechanical-chemical" mechanism By analyzing the Attention Map and SHAP values ​​of the trained model, this embodiment derives the following physically significant catalytic laws: Dominant Factor Identification: Among all 13 descriptors, the d-band center (X8), ICOHP (X10), and HOMO-LUMO band gap (X12) had the highest attention weights, with the sum of the three exceeding 45%. This indicates that the synergistic effect of electronic structure and chemical bond strength is the core factor determining ROS release efficiency.

[0075] Nonlinear synergistic effect: The model reveals a strong cross-attention between X8 (d-band center) and X10 (ICOHP). Specifically, when the d-band center is in a specific range (e.g., -1.0 eV to -2.0 eV), the influence weight of ICOHP on the energy barrier increases significantly. This explains why, in certain metallic systems, simply adjusting the d-band center cannot achieve optimal activity; bonding strength must also be optimized simultaneously.

[0076] Optimization Strategy: Based on model feedback, a material design strategy of "prioritizing d-orbital modulation while also considering antibonding state filling" was proposed. For example, it was predicted that if the d-band center of the Ni-doped system could be shifted upward by 0.3 eV, and the antibonding state filling degree could be increased by 10% through substrate stress modulation, its ROS catalytic activity would be improved by 40%. This prediction was subsequently verified by DFT.

[0077] In summary, this embodiment not only verifies the high-precision predictive capability of the method of the present invention, but more importantly, demonstrates its powerful function in revealing deep physical mechanisms and guiding reverse material design. This method can be fully extended to other heterogeneous catalysis fields such as hydrogen evolution, oxygen evolution, and nitrogen reduction, and has extremely broad application prospects.

[0078] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the protection scope of this invention. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the protection scope of this invention.

[0079] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0080] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0081] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing the present invention, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A method for predicting catalytic performance based on Transformer and mechanochemical coupling, characterized in that, The method includes: Obtain a sample set containing different catalytic material models; determine multiple key descriptors for each sample in the sample set, including surface adsorption forces, electronic interaction forces, orbital interaction forces, and chemical bond strength; obtain the kinetic energy barrier corresponding to each sample, and form a training set with multiple key descriptors for each sample; A mechanochemical coupling analytical model based on the Transformer architecture is constructed. The mechanochemical coupling analytical model includes an embedding layer, an encoder layer, and an output layer. The encoder layer adopts a multi-head self-attention mechanism and a feedforward neural network to capture the long-range dependencies and synergistic weights between different descriptors and to weightedly fuse the features corresponding to the descriptors output by the embedding layer. The mechanical-chemical coupling analytical model is trained using a training set to obtain a mechanical-chemical coupling analytical model for predicting kinetic energy barriers.

2. The catalytic performance prediction method based on Transformer and mechanochemical coupling as described in claim 1, characterized in that, The descriptor includes four dimensions: surface adsorption force, electronic interaction force, orbital interaction force, and chemical bond strength, specifically including: Surface adsorption force descriptors include parameters characterizing the adsorption strength of reactants or intermediates on the catalyst surface; Electron interaction force descriptors include parameters characterizing the charge transfer capability and electron localization degree of active centers; Orbital interaction force descriptors include parameters characterizing the energy level matching and orbital hybridization degree between the catalyst and the adsorbed species; Chemical bond strength and stability class descriptors, including parameters characterizing bond strength and stability.

3. The catalytic performance prediction method based on Transformer and mechanochemical coupling as described in claim 2, characterized in that, The surface adsorption force descriptor includes at least one of the following parameters: reactant adsorption energy and intermediate adsorption energy.

4. The catalytic performance prediction method based on Transformer and mechanochemical coupling as described in claim 2, characterized in that, The electron interaction force class descriptor includes at least one of the following parameters: The active site charge transfer amount, the charge difference integral value between the reaction intermediate and the active site, and the extremum of the electronic localization function at the active site are based on Bader charge analysis. Among them, the charge transfer amount is used to quantify the redox capability of the active site, and the extremum of the electronic localization function is used to characterize the degree of electronic localization and bonding characteristics.

5. The catalytic performance prediction method based on Transformer and mechanochemical coupling as described in claim 2, characterized in that, The orbital interaction force class descriptor includes at least one of the following parameters: The contribution weight of active atoms to the valence band of the system, the peak overlap energy position of active atom orbitals and adsorbed molecular orbitals in the density of states diagram, the degree of hybridization between active atom orbitals and intermediate orbitals, the d-band center position of the active center, and the energy gap between the highest occupied molecular orbital and the lowest unoccupied molecular orbital; wherein, the d-band center position is used to characterize the average energy level of d electrons, and the energy gap is used to characterize the ease of electronic transition.

6. The catalytic performance prediction method based on Transformer and mechanochemical coupling as described in claim 2, characterized in that, The chemical bond strength and stability class descriptor includes at least one of the following parameters: The crystal orbital Hamiltonian layout integral value, the antibonding state filling level at the Fermi level, and the second-order perturbation stabilization energy based on the natural bond orbital theory are used to quantify the bond strength, the antibonding state filling level is used to characterize the stability of the chemical bond, and the second-order perturbation stabilization energy is used to quantify the charge delocalization stabilization effect between orbitals.

7. The catalytic performance prediction method based on Transformer and mechanochemical coupling as described in claim 1, characterized in that, It also includes obtaining a mechanochemical coupling analytical model for predicting the kinetic energy barrier, using the attention weight matrix and SHAP value in the mechanochemical coupling analytical model to analyze the contribution of each descriptor to the kinetic energy barrier, and summarizing the universal mechanochemical coupling rules for this type of catalytic reaction.

8. The catalytic performance prediction method based on Transformer and mechanochemical coupling as described in claim 1, characterized in that, The sample set and the corresponding kinetic energy barrier are obtained based on theoretical calculations. The sample set is constructed by optimizing the structure and screening the stability of different catalytic material models, and the kinetic energy barrier is obtained by simulating the catalytic reaction pathway.

9. The catalytic performance prediction method based on Transformer and mechanochemical coupling as described in claim 1, characterized in that, The embedding layer normalizes the descriptors and maps them into high-dimensional feature vectors, and introduces positional encoding to preserve the physical attribute sequence information of the descriptors.

10. The catalytic performance prediction method based on Transformer and mechanochemical coupling as described in claim 1, characterized in that, During the training process of the mechani-chemical coupling analytical model using the training set, a loss function is used to measure the error between the predicted and the true values. The parameters of the mechani-chemical coupling analytical model are updated by an optimizer, and a learning rate scheduling strategy and regularization technique are used to prevent overfitting.

Images