CN-122024883-A - Parameter optimization method of copper ferrite denitration catalyst based on deep learning

CN122024883ACN 122024883 ACN122024883 ACN 122024883ACN-122024883-A

Abstract

The invention discloses a parameter optimization method of a copper ferrite denitration catalyst based on deep learning, which relates to a neural network model fused with physical simulation priori knowledge, and the method comprises the steps of firstly, constructing a double-channel fused deep learning model, wherein one channel extracts surface active site characteristics from catalyst atomic-level characterization data, and the other channel encodes preparation process parameters; and further, the trained model is used as a high-fidelity proxy model to drive intelligent algorithms such as Bayesian optimization and the like to automatically optimize in a preparation parameter space, and the optimal preparation parameter combination is directly output. The invention realizes the intelligent and accurate design of the catalyst preparation parameters by designing a novel neural network structure and fusing physical simulation knowledge.

Inventors

WANG YUANQING
JI RAN
JIA BAOZHU
JIA XIAOPING
ZHAO XU

Assignees

广东海洋大学

Dates

Publication Date: 20260512
Application Date: 20260210

Claims (10)

1. A method for optimizing parameters of a copper ferrite denitration catalyst based on deep learning, which is characterized in that the copper ferrite denitration catalyst is modified by monoatomic cerium, and the method comprises the following steps: Acquiring atomic-level structure characterization data of the copper ferrite denitration catalyst, and extracting from the atomic-level structure characterization data to obtain surface active site characteristics; inputting the surface active site characteristics into a pre-trained deep learning prediction model, and predicting the anchoring probability distribution of the monoatomic cerium on the surface of the copper ferrite denitration catalyst under a test parameter combination, wherein the test parameter combination is obtained by matching a historical preparation parameter set of the copper ferrite denitration catalyst in a preset preparation parameter space; Simulating an anchoring process of the monoatomic cerium on the surface of the copper ferrite denitration catalyst through molecular dynamics based on the test parameter combination to obtain a running change track, and extracting occupation state data based on the running change track to predict monoatomic distribution evolution trend; And comprehensively analyzing the anchoring probability distribution and the monoatomic distribution evolution trend to obtain an analysis result, and carrying out iterative search in a preparation parameter space by adopting an intelligent optimization algorithm based on the analysis result, wherein an anchoring quality comprehensive evaluation index is used as an optimization target, and optimizing to obtain a target preparation parameter combination.
2. The method for optimizing parameters of a deep learning-based copper ferrite denitration catalyst according to claim 1, wherein the step of inputting the surface active site features into a pre-trained deep learning prediction model to predict an anchoring probability distribution of the monoatomic cerium on the surface of the copper ferrite denitration catalyst under a combination of test parameters comprises the steps of: Inputting the surface active site features in combination with the test parameters to the deep learning predictive model; extracting abstract features in the surface active site features through a feature extraction network in the deep learning prediction model; Mapping the test parameter combination into a parameter feature vector through a parameter coding sub-network in the deep learning prediction model; Fusing the abstract features and the parameter feature vectors through a model fusion layer in the deep learning prediction model to obtain a fusion result; And processing the fusion result through a full-connection layer in the deep learning prediction model to obtain anchoring probability distribution of the monoatomic cerium on different potential active sites on the surface of the copper ferrite denitration catalyst.
3. The parameter optimization method of the copper ferrite denitration catalyst based on deep learning according to claim 2, wherein the construction step of the deep learning prediction model comprises the following steps: Acquiring an atomic-level structure characterization data set of a historical copper ferrite denitration catalyst sample, a corresponding historical preparation parameter set and single-atom cerium anchoring result labeling data; Extracting initial surface active site characteristics of each historical copper ferrite denitration catalyst sample in batches from the atomic-level structure characterization data set; And in the training process, inputting the initial surface active site characteristics into the characteristic extraction network, inputting the corresponding historical preparation parameters into the parameter coding sub-network, performing forward propagation calculation through the model fusion layer and the full connection layer to obtain a predicted anchoring probability, and minimizing a loss function between the predicted anchoring probability and the supervision label through a back propagation algorithm to obtain the pre-trained deep learning prediction model.
4. The method for optimizing parameters of a deep learning-based copper ferrite denitration catalyst according to claim 1, wherein the step of extracting surface active site features from the atomic-level structural characterization data comprises the steps of: Analyzing the atomic-level structure characterization data to obtain surface atomic arrangement, coordination environment and electronic structure information of the copper ferrite denitration catalyst; Identifying potential active sites on the surface of the copper ferrite denitration catalyst based on the surface atomic arrangement, the coordination environment and the electronic structure information, wherein the potential active sites comprise but are not limited to oxygen vacancies, metal sites, step edges and defect sites; Calculating a feature descriptor related to the potential active site, and carrying out normalization and vectorization processing on the feature descriptor to generate the surface active site feature.
5. The parameter optimization method of the deep learning-based copper ferrite denitration catalyst according to claim 1, wherein the step of simulating the anchoring process of the monoatomic cerium on the surface of the copper ferrite denitration catalyst based on the test parameter combination through molecular dynamics to obtain a running change track comprises the following steps: Determining initial conditions of molecular dynamics simulation according to the test parameter combination, wherein the initial conditions comprise simulation supercell model construction parameters, simulation system temperature, pressure and initial spatial distribution of anchoring precursor molecules or ions; Based on a preset stress field, performing energy minimization and pre-balancing treatment on a simulation system containing the surface of the copper ferrite denitration catalyst, the monoatomic cerium and an anchoring precursor environment to obtain a treatment result; And under a set constant temperature and constant pressure or constant temperature and constant volume system, performing molecular dynamics simulation on the treatment result, and recording dynamic coordinate changes of adsorption, migration, bonding and desorption of the monoatomic cerium and the precursor on different active sites on the surface of the copper ferrite denitration catalyst to form the running change track.
6. The parameter optimization method of the deep learning-based copper ferrite denitration catalyst according to claim 1, wherein the extraction of occupancy state data based on the running change track to predict a monoatomic distribution evolution trend comprises: Carrying out post-treatment on the running change track to obtain occupation state data of the single-atom cerium on each active site on the surface of the copper ferrite denitration catalyst; Calculating statistical features characterizing the monoatomic distribution state based on the occupancy state data, wherein the statistical features comprise monoatomic occupancy frequency, surface monoatomic density distribution and monoatomic surface diffusion coefficient of each active site; and predicting the single-atom distribution evolution trend under the preset anchoring time scale based on the statistical characteristics.
7. The method for optimizing parameters of a deep learning-based copper ferrite denitration catalyst according to claim 1, wherein the test parameter combination is obtained by spatially matching a historical preparation parameter set of the copper ferrite denitration catalyst in a preset preparation parameter, and comprises the following steps: Determining a preset preparation parameter space, wherein the preparation parameter space consists of preparation process parameters influencing the modification effect of the monoatomic cerium, and the preparation process parameters comprise, but are not limited to, calcination temperature, calcination time, precursor concentration, impregnation time and pH value; Based on the historical preparation parameter set, analyzing the association relation between each preparation parameter and the catalytic performance; And according to the association relation, adopting an optimized sampling algorithm to sample parameters in the preparation parameter space, and generating the test parameter combination.
8. The method for optimizing parameters of a deep learning-based copper ferrite denitration catalyst according to claim 4, wherein the comprehensively analyzing the anchoring probability distribution and the monoatomic distribution evolution trend to obtain an analysis result comprises the following steps: Based on the anchoring probability distribution, obtaining a theoretical anchoring tendency score of the monoatomic cerium at each potential active site; Acquiring dynamic stability scores of the monoatomic cerium at each potential active site based on the monoatomic distribution evolution trend; And carrying out fusion calculation on the theoretical anchoring tendency score and the dynamic stability score to generate an analysis result for evaluating the test parameter combination.
9. The method for optimizing parameters of the copper ferrite denitration catalyst based on deep learning according to claim 1, wherein the iterative search is performed in a preparation parameter space by adopting an intelligent optimization algorithm based on the analysis result, and the optimization is performed by taking an anchoring quality comprehensive evaluation index as an optimization target, so as to obtain a target preparation parameter combination, and the method comprises the following steps: Constructing an objective function based on the analysis result, wherein the anchoring quality comprehensive evaluation index is obtained by fusing the anchoring probability distribution and the monoatomic distribution evolution trend; and in the preparation parameter space, iteratively selecting a preparation parameter combination and evaluating the objective function value of the preparation parameter combination to continuously update the estimation of the objective function distribution until the stopping condition is met, and outputting the preparation parameter combination with the optimal objective function value as the objective preparation parameter combination.
10. The method for optimizing parameters of a deep learning-based copper ferrite denitration catalyst according to claim 9, wherein after the optimizing results in a target preparation parameter combination, the method further comprises: preparing cerium-containing precursor solution according to the precursor concentration, the pH value and the impregnation time in the target preparation parameter combination, and carrying out isovolumetric impregnation on a copper ferrite carrier by utilizing the cerium-containing precursor solution to obtain an impregnated sample; and drying the impregnated sample under preset conditions, placing the dried impregnated sample in a muffle furnace or a tube furnace, and performing heat treatment according to the calcination temperature and the calcination time in the target preparation parameter combination to finally prepare the monoatomic cerium modified copper ferrite denitration catalyst.

Description

Parameter optimization method of copper ferrite denitration catalyst based on deep learning Technical Field The invention relates to the field of machine learning, in particular to a parameter optimization method of a copper ferrite denitration catalyst based on deep learning. Background The preparation parameters of the copper ferrite denitration catalyst are optimized, and the core aim is to solve the key challenges in the rational design of the single-atom catalyst. By anchoring modification such as monoatomic cerium (Ce) on the copper ferrite carrier, the denitration activity of the catalyst can be remarkably improved. However, the existing optimization method based on machine learning often only depends on limited experimental data to construct a black box model, and has the obvious defects that firstly, the models generally lack fusion of atomic-level physical and chemical mechanisms of a material system, the physical interpretability of a prediction result is poor, the extrapolation capability is weak, secondly, the input characteristics of the model are macroscopic process parameters, high-dimensional atomic-level structure information cannot be fully utilized, the prediction precision is limited, and finally, a single static prediction model cannot reflect the dynamic effect of atomic migration in the preparation process, so that the optimization result may fail in the actual dynamic heat treatment process. Therefore, developing a special deep learning model capable of deeply fusing atomic-level structural characteristics, technological parameters and physical process dynamic information is an urgent need for realizing the rational design and parameter accurate optimization of the single-atom catalyst. Disclosure of Invention The invention provides a parameter optimization method of a copper ferrite denitration catalyst based on deep learning, which can realize the intelligent and accurate design of single-atom catalyst preparation parameters. The invention provides a parameter optimization method of a copper ferrite denitration catalyst based on deep learning, which comprises the following steps of: Acquiring atomic-level structure characterization data of the copper ferrite denitration catalyst, and extracting from the atomic-level structure characterization data to obtain surface active site characteristics; inputting the surface active site characteristics into a pre-trained deep learning prediction model, and predicting the anchoring probability distribution of the monoatomic cerium on the surface of the copper ferrite denitration catalyst under a test parameter combination, wherein the test parameter combination is obtained by matching a historical preparation parameter set of the copper ferrite denitration catalyst in a preset preparation parameter space; Simulating an anchoring process of the monoatomic cerium on the surface of the copper ferrite denitration catalyst through molecular dynamics based on the test parameter combination to obtain a running change track, and extracting occupation state data based on the running change track to predict monoatomic distribution evolution trend; And comprehensively analyzing the anchoring probability distribution and the monoatomic distribution evolution trend to obtain an analysis result, and carrying out iterative search in a preparation parameter space by adopting an intelligent optimization algorithm based on the analysis result, wherein an anchoring quality comprehensive evaluation index is used as an optimization target, and optimizing to obtain a target preparation parameter combination. According to the embodiment of the invention, through obtaining and extracting the surface active site characteristics of the atomic-level structure of the catalyst, the physical target point of stable anchoring of single-atom cerium is accurately locked from a microscopic level, an accurate physical basis is provided for subsequent quantitative prediction and simulation, secondly, the static anchoring probability distribution of single-atom cerium is rapidly predicted based on the surface characteristics and test parameter combination by utilizing a pre-trained deep learning model, the rapid and preliminary evaluation of the possible distribution situation of single-atom before preparation is realized, the defect of high blindness of the traditional trial-and-error method is overcome, then, the anchoring process is dynamically repeated on an atomic scale by molecular dynamics simulation, and the evolution trend is extracted, so that the migration and stable dynamic behavior rule of single-atom in the preparation process is captured, the neglect of the influence of static prediction on the process dynamics is made up, and finally, the result of the static prediction and the dynamic simulation is comprehensively analyzed, and an intelligent optimization algorithm is driven to be automatically and iteratively searched in a parameter space by using the pre-train