Machine learning-based screening method for dual-effect catalysts of mxenes heterostructures

Abstract

The application discloses a MXenes heterostructure dual-effect catalyst screening method based on machine learning, and specifically comprises the following steps: selecting a MXene structure; selecting a MN4 metal element; constructing a heterostructure; using DFT to calculate the heterostructure as an original data set for machine learning model training data; selecting a regression algorithm to train the machine learning model; using an SR algorithm to describe the internal relationship between the overpotential and the descriptor; obtaining a descriptor formula suitable for ORR and OER overpotential; and constructing a machine learning model by screening feature parameters through feature engineering to predict the overpotential. The application solves the problems of the existing technology, such as a large number of MXenes heterostructure combination styles, slow trial and error, and complicated calculation.

Description

Technical Field

[0001] This invention belongs to the field of catalytic material research and development technology, and relates to a machine learning-based method for screening MXenes heterostructure dual-effect catalysts. Background Technology

[0002] Reduction and oxidation reactions are key to clean energy technologies. The oxygen reduction reaction (ORR) and oxygen evolution reaction (OER) influence the electrochemical conversion process of fuel cells, and their efficiency plays a crucial role. Currently available ORR and OER catalysts, including noble metal (Pt, IrO2) alloys, transition metal carbides / nitrides, and graphene, all exhibit low overpotentials in various devices. Large-area Mo2C / MXene was synthesized on a graphene template. Due to the unique interlayer structure of MXene and the diversity of interactions between transition metals, it is possible to improve ORR and OER performance and is synthetically feasible. However, compared to M2C, each N atom in M2N contains an extra electron, resulting in an excess density of states (DOS) near the Fermi level, providing higher conductivity. This also demonstrates that transition metal nitrides have higher conductivity than transition metal carbides. Furthermore, experimental and computational studies have shown that adding N atoms to the carbon structure can modulate the electronic structure and enhance catalytic activity. Daniel Wilhelm's team used machine learning and first-principles calculations to screen low-dimensional heterostructures in their database, successfully predicting band gaps and interlayer binding energies, thus avoiding brute-force exploration of large combinatorial spaces. Similarly, combining machine learning and density functional theory, they predicted the catalytic activity of 78 unterminated Gra / MXene (V2C) heterostructures, discovering that the CoN4-V2C heterostructure is an ORR / OER bifunctional catalyst, improving design speed and prediction accuracy. However, given the numerous combinations of heterostructures like MN4-Gra and MXene, designing heterostructures through theoretical calculations and experimental trial-and-error is a slow process. Therefore, finding a method for rapidly and accurately designing MXene heterostructure electrocatalysts while effectively reducing computational costs is urgently needed. Summary of the Invention

[0003] The purpose of this invention is to provide a machine learning-based method for screening dual-effect catalysts with heterostructures of MXenes, which solves the problems of multiple combinations of heterostructures of MXenes and complex calculations in the prior art.

[0004] The technical solution adopted in this invention is a machine learning-based screening method for MXenes heterostructure dual-effect catalysts, which is implemented according to the following steps:

[0005] Step 1: Select the MXene structure;

[0006] Step 2: Select MN4 metal elements;

[0007] Step 3: Construct a heterogeneous structure and use DFT to compute the heterogeneous structure as the original dataset for training machine learning models.

[0008] Step 4: Use the SR algorithm to construct descriptors, determine the intrinsic features of MXene heterostructures that affect the overpotential catalytic performance of ORR / OER based on the descriptors, construct a machine learning model with atomic intrinsic features as feature parameters, derive the importance ranking of MXene heterostructure intrinsic features from the machine learning model, and predict the overpotential through the importance ranking of MXene heterostructure intrinsic features to achieve the screening of MXene heterostructure dual-effect catalysts.

[0009] The invention is further characterized in that,

[0010] Step 1: Select 9 MXene structures, including V2NO, Ti2NO, Nb2NO, Zr2NO, Hf2NO, Mo2NO, Ta2NO, W2NO, and Cr2NO.

[0011] Step 2 selects MN4 metal elements including 3d atoms Ti, V, Cr, Mn, Fe, Co, Ni, Cu, 4d atoms Zr, Nb, Mo, Ru, Rh, Pd, Ag, and 5d atoms Pt, Ir, a total of 17 transition metal atoms.

[0012] In step 3, by calculating the free energy barrier in the four-electron path step of each heterostructure, heterostructures with poor ORR / OER activity, i.e., ORR / OER overpotential greater than 0.43eV / 0.65eV, are removed from the original dataset.

[0013] The descriptor formulas for the ORR and OER overpotentials in step 4 are as follows:

[0014]

[0015]

[0016] In equations (1) and (2), a = R M2 EA M1 b = R M1 -T M1 ·DI M2 c = EA M1 -M M2 In this context, N represents the atomic number, M represents the atomic mass, X represents electronegativity, IE represents the first ionization energy, EA represents the electron affinity, R represents the covalent radius, D represents the density, DI represents the outermost electron, DII represents the outermost D orbital electron, V represents the atomic volume, and T represents the boiling point.

[0017] According to the overpotential descriptors (1) and (2), the characteristic value affecting the ORR overpotential descriptor is R. M2 DII M2 DI M1 EA M1 That is, the intrinsic characteristics R of MXene heterostructure M2 DII M2 DI M1 EA M1 The eigenvalue that affects the ORR overpotential and the OER overpotential descriptor is R. M1 T M1 IE M1 EA M1 That is, the intrinsic characteristics R of MXene heterostructure M1 T M1 IE M1 EA M1 It affects the overpotential of the OER.

[0018] The beneficial effects of this invention are:

[0019] 1) This invention is a machine learning-based method for screening MXene heterostructure dual-effect catalysts. It is a method for rapid screening of MN4-Gra / MXene heterostructure fuel cell cathode catalysts based on machine learning-assisted first-principles calculation and prediction, which overcomes the shortcomings of the traditional "experimental trial and error method" and saves resources and time.

[0020] 2) Machine learning technology was used to predict the catalytic activity of various MXene heterostructures. Model data of 19 heterostructure catalysts were obtained by constructing and calculating by DFT, and the ORR / OER catalytic performance of 117 heterostructure cathode catalysts was predicted. Descriptors that can accurately express atomic features and structural performance were obtained through model data, which simplified the prediction of machine learning models and made the prediction more accurate.

[0021] 3) The entire process does not involve experiments or use of chemicals, does not generate chemical pollution, conforms to the concept of green environmental protection, is low in cost, easy to implement, and is suitable for widespread use. Attached Figure Description

[0022] Figure 1 This is an ML-DFT flowchart;

[0023] Figure 2 shows the heterostructure transition metal diagrams. Figure 2(a) shows the 3d, 4d, and 5d atomic diagrams of the heterostructure transition metal, Figure 2(b) is a top view of the heterostructure transition metal model, and Figure 2(c) is a side view of the heterostructure transition metal model.

[0024] Figure 3The ORR / OER four-electron path diagrams are shown under different water splitting voltages (U = 0V, U = 1.23V).

[0025] Figure 4 is a comparison of the prediction results. Figure 4(a) is the prediction of ORR overpotential by the model, and Figure 4(b) is the prediction of OER overpotential by the model.

[0026] Figure 5 shows the importance analysis of intrinsic features of MXene heterostructures. Figure 5(a) shows the importance analysis of GBR-ORR features, and Figure 5(b) shows the importance analysis of GBR-OER features. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0028] Example 1

[0029] This embodiment provides a machine learning-based method for screening MXenes heterostructure dual-effect catalysts, such as... Figure 1 As shown, please follow these steps:

[0030] Step 1: Select MXene structures, including V2NO, Ti2NO, Nb2NO, Zr2NO, Hf2NO, Mo2NO, Ta2NO, W2NO, and Cr2NO;

[0031] Step 2: Select MN4 metal elements, as shown in Figure 2(a), Figure 2(b), and Figure 2(c), including 17 transition metal atoms: 3d atoms Ti, V, Cr, Mn, Fe, Co, Ni, and Cu; 4d atoms Zr, Nb, Mo, Ru, Rh, Pd, and Ag; and 5d atoms Pt and Ir.

[0032] Step 3: Construct 19 heterogeneous structures, use DFT to calculate the heterogeneous structures as the original dataset, and use them as training data for machine learning models;

[0033] Step 4: Use the SR algorithm to construct descriptors, determine the intrinsic features of MXene heterostructures that affect the overpotential catalytic performance of ORR / OER based on the descriptors, construct a machine learning model with atomic intrinsic features as feature parameters, derive the importance ranking of MXene heterostructure intrinsic features from the machine learning model, and predict the overpotential through the importance ranking of MXene heterostructure intrinsic features to achieve the screening of MXene heterostructure dual-effect catalysts.

[0034] Example 2

[0035] This embodiment provides a machine learning-based method for screening MXenes heterostructure dual-effect catalysts. Building upon Example 1, step 3 specifically involves calculating the overpotentials of 19 MN4 / Gra-M2NO heterostructures. The structures with the lowest ORR / OER overpotentials are identified as either dual-functional or monofunctional catalysts. Furthermore, Gibbs free energy is used to evaluate ORR and OER activities, and the Gibbs free energy of the adsorption intermediate is calculated to obtain G... *O Since both values ​​are less than 3.52 eV, the ORR / OER reaction pathway only considers the four-electron pathway, using PDS as ΔG. i An important step in the process, ORR / OER activity is also represented by PDS. For example... Figure 3 As shown, by calculating the free energy barrier in the four-electron path steps of each heterostructure, it was found that the combination of TiN4-Gra and MXene exhibits poor ORR / OER activity, making them unsuitable for constructing electrocatalytic heterostructures. ORR / OER calculations for TiN4-Gra / MXene and MN4-Gra / Cr2NO structures also indicate that they are unsuitable as electrocatalyst materials for use in fuel cell cathodes. Therefore, these two heterostructures were not considered during machine learning model training and testing. Heterostructures with poor ORR / OER activity (i.e., ORR / OER overpotential greater than 0.43 eV / 0.65 eV) were removed from the original dataset.

[0036] Example 3

[0037] This embodiment provides a machine learning-based method for screening MXenes heterostructure dual-effect catalysts. Building upon embodiments 1 and 2, step 4, machine learning, is performed in the scikit-learn package within a Python 3 environment. The mechanical performance target is the structural stability target, and the electrical performance target is the structural scalar conductivity target. A proportional conductivity model is established based on gradient boosting regression (GBR), random forest regression (RFR), and support vector regression (SVR). After training and testing the model using GBR, SVR, and RFR, the descriptor SR and the optimal regression algorithm GBR are selected to predict the ORR and OER overpotentials, respectively, to achieve the screening of MXenes heterostructure dual-effect catalysts.

[0038] To verify the learning ability of the machine learning model, 10 heterogeneous structures were selected from the prediction results and compared with the DFT calculation results. In the prediction, the mean absolute error (MAE) was selected to evaluate the prediction accuracy of the descriptor model and the regression model. The MAE of the GBR algorithm was lower than that of the descriptor prediction.

[0039] The SR algorithm is used to construct descriptors. After more than 1000 iterations of the SR loop, a suitable descriptor formula for ORR and OER overpotential is obtained, as follows:

[0040]

[0041]

[0042] In equations (1) and (2), a = R M2 EA M1 b = R M1 -T M1 ·DI M2 c = EA M1 -M M2 In this context, N represents the atomic number, M represents the atomic mass, X represents electronegativity, IE represents the first ionization energy, EA represents the electron affinity, R represents the covalent radius, D represents the density, DI represents the outermost electron, DII represents the outermost D orbital electron, V represents the atomic volume, and T represents the boiling point.

[0043] According to the overpotential descriptors (1) and (2), the eigenvalues ​​affecting the ORR overpotential descriptor (the eigenvalues ​​in the descriptor are intrinsic atomic features) are R. M2 DII M2 DI M1 EA M1 That is, the intrinsic characteristics R of MXene heterostructure M2 DII M2 DI M1 EA M1 The eigenvalue that affects the ORR overpotential and the OER overpotential descriptor is R. M1 T M1 IE M1 EA M1 That is, the intrinsic characteristics R of MXene heterostructure M1 T M1 IE M1 EA M1 It affects the overpotential of the OER.

[0044] The intrinsic features of the MXene heterojunction that affect the overpotential catalytic performance of ORR / OER were determined based on the descriptor. A machine learning model (GBR, RFR, SVR) with atomic intrinsic features as feature parameters was constructed. Three regression algorithms, GBR, RFR, and SVR, were selected to train the machine learning model. The results showed that GBR showed the best fit in the ORR training test and the OER training test.

[0045] The importance ranking of intrinsic features of the MXene heterostructure was derived from the machine learning model, as shown in Figure 4. In the model prediction, the mean absolute error (MAE) was selected as the evaluation metric for both the descriptor and regression models. In descriptor prediction, MAE... ORR0.085V, MAE OER The value is 0.087 in GBR prediction, as shown in Figure 4(a), MAE ORR The voltage is 0.049V, as shown in Figure 4(b), MAE OER The MAE of the GBR algorithm is 0.062, which is lower than that predicted by the descriptor, indicating more accurate prediction results. Data shows that RuN4-Gra / W2NO (GBR vs DFT: 0.39V vs 0.43V), CoN4-Gra / W2NO (GBR vs DFT: 0.39V vs 0.38V), and RhN4-Gra / Ti2NO (GBR vs DFT: 0.39V vs 0.33V) have better ORR overpotentials, while RuN4-Gra / Hf2NO (GBR vs DFT: 0.49V vs 0.40V), RuN4-Gra / W2NO (GBR vs DFT: 0.45V vs 0.43V), and CoN4-Gra / Hf2NO (GBR vs DFT: 0.60V vs 0.50V) have better OER overpotentials.

[0046] As shown in Figure 5(a), the feature value DII in the descriptor M2 (Contribution rate 41.10%), R M2 (Contribution 7.8%), V M2 (Contribution 7.3%) has a strong correlation with ORR overpotential (contribution > 5%, contribution is importance), with a total contribution of 56.20%. As shown in Figure 5(a), the eigenvalue R in the descriptor M2 (Contribution rate 36.70%), V M2 (Contribution rate 17.80%), T M1 (Contribution 14.00%), IE M1 (Contribution 13.5%), V M1 (Contribution 12.7%) has a strong correlation with OER overpotential (contribution > 5%), with a total contribution of 94.6%.

[0047] Overpotential is predicted by ranking the importance of intrinsic features of MXene heterostructures to achieve screening of dual-effect catalysts with MXene heterostructures.

[0048] This embodiment analyzes the stability, catalytic activity, and electronic structure of 19 MN4 / Gra-M2NO catalysts using DFT calculations, and uses machine learning to predict and screen 136 designed MN4 / Gra-MXene heterojunction electrocatalysts. The machine learning approach proposes the SR algorithm for fitting the overpotential descriptor and the GBR (MAE) algorithm. ORR / OERThe predictions (0.049V / 0.062V) successfully showed that MN4-Gra / Ti2NO exhibited good ORR overpotential, and RuN4-Gra / M2NO demonstrated good OER performance. The prediction results also showed that RuN4-Gra / W2NO possessed excellent monofunctional ORR catalytic activity (0.33V). Furthermore, a descriptor without environmental variables was proposed to be used in conjunction with GBR for prediction. The ORR influencing factor focused on the selection of metal atoms in transition metal nitrides (TM1 and IEM1), while the OER influencing factor focused on the selection of transition metals in MXenes (RM2 and DIIM2). These results provide a reliable prediction and screening method for the design of MXenes heterostructures.

[0049] As described above, this invention utilizes a machine learning-based method for screening MXene heterostructure dual-effect catalysts. This method constructs a graphene-doped MN4 and MXene (terminated with O) heterostructure electrocatalyst. Replacing the transition metal C with a transition metal N-oxide provides higher conductivity and enhances catalytic activity. The machine learning model was optimized, employing the SR algorithm to elucidate the intrinsic relationship between overpotential and descriptor, obtaining suitable descriptor formulas for ORR and OER overpotentials. Feature engineering was used to screen feature parameters to construct machine learning models (GBR, RFR, SVR) for overpotential prediction. A combination of ML-DFT was used, employing the simplest eigenvalue fitting descriptor and machine learning model to liberate first-principles calculations, achieving rapid prediction of MXene heterostructure catalyst performance. This provides a new research approach for MXene heterostructure design, innovatively using environmental variable-free feature fitting descriptors, reducing trial-and-error costs from the outset and providing a reliable theoretical basis for experiments.

Claims

1. A machine learning-based method for screening MXenes heterostructure dual-effect catalysts, characterized in that, The specific steps are as follows: Step 1: Select the MXene structure; Step 2: Select MN4 metal elements; Step 3: Construct a heterogeneous structure and use DFT to compute the heterogeneous structure as the original dataset for training machine learning models. Step 4: Use the SR algorithm to construct descriptors, and determine the intrinsic features of the MXene heterostructure that affect the ORR / OER overpotential catalytic performance based on the descriptors. The descriptor formulas for ORR and OER overpotential are as follows: （1） (2) In equations (1) and (2), a = b = c = In this context, N represents atomic number, X represents electronegativity, IE represents first ionization energy, EA represents electron affinity, R represents covalent radius, D represents density, DI represents outermost electron, DII represents outermost D orbital electron, V represents atomic volume, and T represents boiling point. According to the overpotential descriptors Equations (1) and (2), the eigenvalues ​​affecting the ORR overpotential descriptor are RM2, DIIM2, DIM1, and EAM1, that is, the intrinsic features of the MXene heterostructure RM2, DIIM2, DIM1, and EAM1 affect the ORR overpotential, and the eigenvalues ​​affecting the OER overpotential descriptor are RM1, TM1, IEM1, and EAM1, that is, the intrinsic features of the MXene heterostructure RM1, TM1, IEM1, and EAM1 affect the OER overpotential. A machine learning model with atomic intrinsic features as feature parameters is constructed. The importance ranking of intrinsic features of MXene heterostructure is obtained from the machine learning model. The overpotential is predicted by the importance ranking of intrinsic features of MXene heterostructure, so as to achieve the screening of dual-effect catalysts of MXene heterostructure.

2. The machine learning-based screening method for MXenes heterostructure dual-effect catalysts according to claim 1, characterized in that, Step 1 selects nine MXene structures, including V2NO, Ti2NO, Nb2NO, Zr2NO, Hf2NO, Mo2NO, Ta2NO, W2NO, and Cr2NO.

3. The machine learning-based screening method for MXenes heterostructure dual-effect catalysts according to claim 1, characterized in that, Step 2 selects MN4 metal elements including 3d atoms Ti, V, Cr, Mn, Fe, Co, Ni, Cu, 4d atoms Zr, Nb, Mo, Ru, Rh, Pd, Ag, and 5d atoms Pt, Ir, a total of 17 transition metal atoms.

4. The machine learning-based screening method for MXenes heterostructure dual-effect catalysts according to claim 1, characterized in that, In step 3, by calculating the free energy barrier in the four-electron path step of each heterostructure, heterostructures with poor ORR / OER activity, i.e., ORR / OER overpotential greater than 0.43eV / 0.65eV, are removed from the original dataset.

Images