An expensive multimodal multi-objective optimization method for molecular structure design

By optimizing the partition search and multimodal multi-objective differential evolution algorithm, combined with adaptive resource allocation, the expensive optimization problem in the molecular structure design of catalytic materials is solved. This achieves efficient convergence of the global optimal solution and the diversity mining of local multimodal solutions, thereby improving the model's fitting accuracy and resource utilization efficiency.

CN123116867APending Publication Date: 2026-11-17ZHENGZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611576112.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-12
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies face costly optimization problems in the molecular structure design of catalytic materials. They cannot achieve both efficient convergence of the global optimal solution and the diverse mining of local multimodal solutions with extremely low evaluation budgets. Furthermore, they lack partitioning mechanisms and dedicated surrogate models for multimodal distributions, resulting in resource waste and insufficient model prediction accuracy.

Method used

The decision space is divided into multiple subspaces using a partitioned search strategy. An initial population is generated using Latin hypercube sampling and a dedicated surrogate model is constructed. In each subspace, a multimodal, multi-objective differential evolution algorithm is used for optimization. The surrogate model is updated by combining an adaptive resource allocation filling sampling criterion, thereby achieving efficient convergence of the global Pareto optimal solution and systematic mining of local multimodal solutions.

Benefits of technology

It significantly improves the fitting accuracy and prediction reliability of surrogate models in multimodal optimization scenarios, reduces the waste of computing resources, provides a rich variety of candidate molecular configurations, and improves resource utilization efficiency and model iteration quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116867A_ABST
    Figure CN123116867A_ABST
Patent Text Reader

Abstract

This invention provides an expensive multimodal multi-objective optimization method for molecular structure design, relating to the technical field of molecular structure optimization for catalytic materials. This expensive multimodal multi-objective optimization method for molecular structure design includes the following steps: constructing a multi-objective optimization model with minimizing both molecular formation energy and adsorption Gibbs free energy as dual optimization objectives; determining decision variables and decision space; generating an initial population in each subspace using Latin hypercube sampling; constructing a data archive for each subspace based on the evaluation results; and training surrogate models for the corresponding subspaces using each data archive; repeatedly performing the above evolution and model update operations until the termination condition is met; merging the data archives of each subspace; selecting the set of non-dominated individuals; and outputting the final multimodal Pareto solution set and Pareto front. This invention can significantly improve the efficiency and solution set quality of molecular structure optimization for catalytic materials, effectively reducing R&D computational costs.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the technical field of molecular structure of catalytic materials, and more specifically, to an expensive multimodal multi-objective optimization method for molecular structure design. Background Technology

[0002] In cutting-edge fields such as catalytic materials and drug molecules, high-fidelity simulation calculations based on first principles (represented by VASP software) have become the mainstream method for accurately predicting physicochemical parameters such as molecular formation energy and adsorption free energy. However, the single evaluation of this type of calculation is extremely time-consuming and computationally expensive, making molecular structure optimization problems inherently marked as costly optimizations, which severely restricts the application of evolutionary algorithms in real-world engineering scenarios.

[0003] Currently, for such expensive optimization problems, existing technologies have initially introduced multi-objective evolutionary algorithms and surrogate model-assisted strategies. However, traditional methods still generally suffer from the following limitations: On the one hand, most algorithms rely on tens of thousands of real evaluations to converge, and are prone to premature stagnation under strictly limited computational budgets. Furthermore, they focus only on a single global Pareto front, completely ignoring the multimodal characteristics widely present in molecular structures, i.e., multiple locally optimal solutions with equivalent performance but different configurations within the decision space. This results in a single optimization scheme, failing to provide diverse and high-value alternative configurations for engineering selection. On the other hand, although some schemes introduce surrogate models to reduce the number of evaluations, they mostly use a single global surrogate model, which is difficult to accurately fit the complex local terrain in multimodal scenarios. Moreover, the accompanying filling sampling criteria adopt a fixed resource allocation mechanism, which cannot dynamically adjust the evaluation budget based on the search potential of each sub-region, resulting in severely insufficient prediction accuracy and generalization ability of the surrogate model in key regions.

[0004] When designing expensive molecular structures such as catalytic materials, existing technologies have the following core shortcomings: First, they cannot achieve both efficient convergence of the global optimal solution and the diverse mining of local multimodal solutions with extremely low evaluation budgets. Second, they lack partitioning mechanisms and dedicated proxy models for multimodal distributions, resulting in low model fidelity. Third, the filling sampling strategy lacks adaptive resource allocation capabilities, with scarce computing resources being wasted in low-value regions, making it difficult to drive the proxy model to achieve accurate iteration. Summary of the Invention

[0005] The problem addressed by this invention is how to significantly improve the efficiency and quality of molecular structure optimization of catalytic materials, while effectively reducing the computational cost of research and development.

[0006] To address the above problems, this invention provides an expensive multimodal multi-objective optimization method for molecular structure design, comprising the following steps: Step 1: Construct a multi-objective optimization model with the dual optimization objectives of minimizing molecular formation energy and minimizing adsorption Gibbs free energy, and determine the decision variables and decision space; Step 2: The decision space is divided into multiple non-overlapping subspaces using a partitioned search strategy. In each subspace, an initial population is generated using Latin hypercube sampling. The fitness of individuals in the initial population is evaluated using simulation software. Based on the evaluation results, a data archive for each subspace is constructed, and the surrogate model for the corresponding subspace is trained using each data archive. Step 3: Within each subspace, using the current population as the parent, generate offspring individuals using a multimodal, multi-objective differential evolution algorithm. The fitness value of the offspring individuals is predicted by the surrogate model in their respective subspace. Determine whether to update the surrogate model based on the filling sampling criterion of adaptive resource allocation. If so, select multiple individuals from the candidate solution set of the current subspace for true fitness evaluation, update the data archive using the evaluation results, and retrain the surrogate model. Repeat the above evolution and model update operations until the termination condition is met. Merge the data archives of each subspace, select the set of non-dominated individuals, and output the final multimodal Pareto solution set and Pareto front.

[0007] Optionally, the construction of a multi-objective optimization model with the dual optimization objectives of minimizing molecular formation energy and minimizing adsorption Gibbs free energy, and the determination of decision variables and decision space, includes: The sets of cobalt-site and phosphorus-site substitution elements in the cobalt phosphide catalyst diatomic substitution modification system were determined and encoded by integers to form two-dimensional decision variables. The value range of each dimension of the decision variables was determined by the upper limit of the encoding of the corresponding substitution element set. The molecular formation energy was calculated by the total energy of the doping system, the energy of cobalt element, the energy of cobalt phosphide, and the energy of bulk phase transition metal atoms. The adsorption Gibbs free energy was calculated by the energy of adsorbed nitrate ions, the energy of the catalyst substrate, the energy of gaseous nitrate molecules, and the energy of hydrogen molecules. All of the above energies were obtained by simulation software.

[0008] Optionally, the partition search strategy includes: Select key dimensions from all dimensions of the decision space, and evenly divide the value range of each selected key dimension into a preset number of intervals. By combining the intervals of each key dimension, the entire decision space is deconstructed into multiple non-overlapping subspaces.

[0009] Optionally, both dimensions are selected as key dimensions, and the value range of each dimension is evenly divided into two intervals to form four subspaces.

[0010] Optionally, the Latin hypercube sampling adopts an integer sampling method, generating a preset number of individuals in each subspace. The sampling process includes: uniformly dividing the integer value range of each dimension into equally spaced integer intervals equal to the number of samples; independently generating random permutations for each dimension to ensure that each interval is selected only once; randomly selecting an integer value in each integer interval; and forming a complete integer sample set by random permutations and combinations of each dimension.

[0011] Optionally, the real fitness assessment includes: decoding individuals in the initial population to obtain corresponding molecular structure combinations, calling simulation software to calculate the formation energy and adsorption Gibbs free energy of each molecular structure combination, and storing the calculation results and corresponding codes together in the data archive of each subspace; the surrogate model is a Kriging model, and two Kriging models are trained in each subspace to predict the formation energy and adsorption Gibbs free energy, respectively.

[0012] Optionally, step 3 includes: Step 3-1: In each subspace, using the current population as the parent generation, a new generation of offspring individuals is generated using a cluster-based multimodal multi-objective differential evolution algorithm. The fitness value of the offspring individuals is predicted by the surrogate model of the corresponding subspace. Step 3-2: Based on the filling sampling criterion based on adaptive resource allocation, determine whether the proxy model of each subspace needs to be updated. If so, combine the historical data archive of each subspace with the current population, select multiple excellent individuals through non-dominated sorting and special crowding distance for real fitness evaluation, store the evaluated individuals in the corresponding data archive and retrain the proxy model. Step 3-3: Repeat steps 3-1 and 3-2 until the preset iteration termination condition is reached, and output the updated data archive for each subspace. Steps 3-4: Merge the data archives of all subspaces to obtain the global archive. Perform non-dominated sorting on all individuals in the global archive, filter out the set of non-dominated individuals, and output the final Pareto front and multimodal Pareto solution set.

[0013] Optionally, the cluster-based multimodal multi-objective differential evolution algorithm in step 3-1 includes: performing non-dominated sorting on the current population to obtain the non-dominated levels of each solution; adaptively determining the number of clusters according to the number of individuals in each level and performing hierarchical clustering, calculating the crowding distance in the decision space and the crowding distance in the target space using only individuals in the same cluster as neighbors, and fusing them to obtain the non-dominated and crowding distance indices; sorting the population according to the rule of prioritizing non-dominated levels and descending the indices within the same level; selecting sparsely distributed individuals from the better non-dominated levels as candidate models, and matching a dedicated learning model for each individual through a roulette wheel method with decision space distance weighting; generating offspring individuals based on differential evolution mutation and crossover operations, merging the parent and offspring generations and performing sorting again, and selecting the best individuals to form the next generation parent population.

[0014] Optionally, the filling sampling criterion for adaptive resource allocation in step 3-2 includes: initializing the selection probability of each subspace; merging the data archives of all subspaces into a global archive and performing non-dominated sorting, and counting the total number of non-dominated individuals globally; when the selection probability of a certain subspace is greater than a randomly generated probability threshold, performing a filling sampling operation on that subspace, wherein the filling sampling operation includes merging the data archive of the current subspace with the current parent population, performing deduplication and non-dominated sorting, and then using a special crowding distance to perform a secondary sorting of individuals with the same non-dominated level, removing individuals already in the data archive, and selecting those ranked higher. A preset number of individuals are used to evaluate the true fitness of the proxy model and update it. After performing fill sampling, the number of individuals corresponding to the local Pareto front of each subspace in the global Pareto front is counted as the number of contributing individuals in each subspace. The selection probability of each subspace is updated based on the number of contributing individuals in each subspace and the total number of non-dominated individuals in the global subspace: if there are non-dominated individuals in the local data archive of a subspace, the selection probability of that subspace is increased by the product of the constant selection probability and the proportion of the number of contributing individuals in that subspace to the total number of non-dominated individuals in the global subspace; otherwise, the selection probability of that subspace is decreased by the constant selection probability.

[0015] Optionally, after outputting the multimodal Pareto solution set and Pareto front in step 3, the method further includes decoding the encoded individuals in the solution set to obtain the molecular structure substitution element combination corresponding to each non-dominated solution.

[0016] The beneficial effects of the expensive multimodal multi-objective optimization method for molecular structure design of the present invention are: This invention divides the decision space into multiple subspaces using a partitioned search strategy and constructs a dedicated surrogate model for each. This effectively solves the problem that traditional single global surrogate models struggle to accurately fit multiple locally optimal terrain regions in multimodal optimization scenarios, significantly improving the fitting accuracy and prediction reliability of the surrogate model across the entire decision space. Simultaneously, by employing a multimodal, multi-objective differential evolution algorithm to generate offspring, the diversity of the population in both the decision and objective spaces can be maintained during iteration. This allows for efficient convergence of the global Pareto optimal solution and systematic mining of local multimodal solutions under a limited real evaluation budget, providing a rich pool of candidate molecular configurations for engineering applications. Furthermore, by dynamically allocating computational resources based on an adaptive resource allocation filling sampling criterion, scarce real evaluation budgets are precisely allocated to high-value sub-regions, effectively avoiding ineffective consumption of computational resources in low-value regions and significantly improving resource utilization efficiency and the quality of iterative updates of the surrogate model. The three core components mentioned above work together and support each other. The partitioning strategy provides a basis for accurate fitting of the surrogate model, multimodal differential evolution ensures population diversity and convergence, and adaptive filling sampling ensures the continuous fidelity of the surrogate model in key regions. Together, they constitute a complete technical solution for efficiently solving multimodal and multi-objective molecular structure optimization problems under the constraint of expensive computing resources. Attached Figure Description

[0017] Figure 1 This is a flowchart of the expensive multimodal multi-objective optimization method in an embodiment of the present invention; Figure 2 This is a flowchart of the SMMODE_ZS algorithm in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the construction of a multi-agent model in an embodiment of the present invention; Figure 4 (a) is one of the PSs and PF obtained on MMF1 by the five expensive multi-objective algorithms in the embodiments of the present invention; Figure 4 (b) is two of the five expensive multi-objective algorithms obtained on MMF1 in the embodiments of the present invention; Figure 4 (c) is the third of the five expensive multi-objective algorithms obtained on MMF1 in the embodiments of the present invention; Figure 4 (d) is the fourth of the five expensive multi-objective algorithms obtained on MMF1 in the embodiments of the present invention; Figure 4 (e) is the fifth of the five expensive multi-objective algorithms obtained on MMF1 in the embodiments of the present invention; Figure 5This is a flowchart of the experimental design in an embodiment of the present invention; Figure 6 (a) is a set of multimodal solution molecular structures Ru-P molecular structures obtained by SMMODE_ZS in the embodiments of the present invention; Figure 6 (b) is a set of multimodal solution molecular structures Ir-P molecular structures obtained by SMMODE_ZS in the embodiments of the present invention; Figure 7 (a) is a set of multimodal solution molecular structures Tc-P molecular structures obtained by SMMODE_ZS in the embodiments of the present invention.

[0018] Figure 7 (b) is a set of multimodal solution molecular structures Ti-N molecular structures obtained by SMMODE_ZS in the embodiments of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0020] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0021] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0022] like Figure 1 As shown in the figure, an expensive multimodal multi-objective optimization method for molecular structure design provided by an embodiment of the present invention includes the following steps: Step 1: Construct a multi-objective optimization model with the dual optimization objectives of minimizing molecular formation energy and minimizing adsorption Gibbs free energy, and determine the decision variables and decision space; Step 2: The decision space is divided into multiple non-overlapping subspaces using a partitioned search strategy. In each subspace, an initial population is generated using Latin hypercube sampling. The fitness of individuals in the initial population is evaluated using simulation software. Based on the evaluation results, a data archive for each subspace is constructed, and the surrogate model for the corresponding subspace is trained using each data archive. Step 3: Within each subspace, using the current population as the parent, generate offspring individuals using a multimodal, multi-objective differential evolution algorithm. The fitness value of the offspring individuals is predicted by the surrogate model in their respective subspace. Determine whether to update the surrogate model based on the filling sampling criterion of adaptive resource allocation. If so, select multiple individuals from the candidate solution set of the current subspace for true fitness evaluation, update the data archive using the evaluation results, and retrain the surrogate model. Repeat the above evolution and model update operations until the termination condition is met. Merge the data archives of each subspace, select the set of non-dominated individuals, and output the final multimodal Pareto solution set and Pareto front.

[0023] Specifically, "constructing a multi-objective optimization model with minimizing both molecular formation energy and adsorption Gibbs free energy as dual optimization objectives": molecular formation energy ( The thermodynamic stability of cobalt phosphide catalysts is determined by its atomic substitution configuration; a lower value indicates higher thermodynamic stability after substitution. Its functional expression is as follows: In the formula, , , and These represent the energies of transition metal-doped cobalt phosphide, elemental cobalt, cobalt phosphide, and bulk transition metal atoms, respectively.

[0024] Gibbs free energy of nitrate ion adsorption The Gibbs free energy of nitrate ion adsorption is used to measure the spontaneous adsorption capacity of nitrate ions on the surface of a catalyst substrate, which directly determines the initial efficiency of the catalytic reaction. The expression for its Gibbs free energy is shown below:

[0025] In the formula, , ,as well as These represent the Gibbs free energies of nitrate ions adsorbed on the substrate, nitrate molecules in the gas phase, and hydrogen molecules, respectively.

[0026] All the above energy calculations rely on VASP software, and each evaluation is extremely time-consuming, making it a typical expensive optimization problem. There is a typical conflict between the formation energy and the adsorption Gibbs free energy—pursuing extremely low formation energies often leads to decreased adsorption performance, and vice versa—these two together constitute the basic evaluation system for catalyst performance. The optimization objective function is constructed as follows: In the formula, decision variables It is a combination of elements with diatomic substitutions for cobalt phosphide. The problem has a two-dimensional decision space. The solution space of the decision variables is formed by the selection of substitution elements for cobalt and phosphorus atoms in cobalt phosphide, exhibiting typical characteristics of high computational cost multimodal optimization.

[0027] Determine the decision variables and decision space; decision variables The selection of substitution elements for cobalt and phosphorus atoms in cobalt phosphide. Based on relevant research, the candidate substitution elements for Co and P atoms were set to 26 and 29, respectively. Figure 2 As shown, in step 2-1, the elements in the periodic table are arranged according to their column order to ultimately determine the decision space. The upper limit range is [25, 28], and the lower limit range is [0, 0]. That is, the value range of the first dimension (Co-position substitution) is 0~25, and the value range of the second dimension (P-position substitution) is 0~28. Within this optimization solution space, the combination and matching of different substitution elements will form multiple global Pareto optimal solutions and local Pareto optimal solutions, exhibiting complex multimodal optimization characteristics.

[0028] "Dividing the decision space into multiple non-overlapping subspaces using a partitioned search strategy" refers to the following steps: Step 2-1-1, the partitioned search strategy involves selecting d key dimensions from D decision variables (in this embodiment, both dimensions are selected, i.e., d=D=2); Step 2-1-2, uniformly dividing the value range of each selected key dimension into q intervals (in this embodiment, q=2); Step 2-1-3, through the combination of multi-dimensional intervals, deconstructing the entire D-dimensional decision space into K=q. d A number of non-overlapping subspaces.

[0029] "Generate an initial population in each subspace using Latin hypercube sampling" refers to the fact that Latin hypercube sampling is a space-filling design method that can ensure a uniform distribution of samples throughout the sampling space when the number of samples is small. In this embodiment, integer Latin hypercube sampling is used. In step 2-2, NP individuals are generated in each subspace (NP=200 in this embodiment).

[0030] The sampling process includes: First, determining the integer value range of each input variable and the required number of samples, and uniformly dividing the integer value range of each dimension into equally spaced intervals equal to the number of samples; steps 2-3. Then, independently generating random permutations from 1 to NP for each dimension, ensuring that each interval is selected only once; finally, randomly selecting an integer value within each divided integer interval, and randomly combining the values ​​of each dimension according to steps 2-4, generating an integer sample set that satisfies the Latin square property, so that the samples are uniformly distributed and non-overlapping throughout the integer space, denoted as the population. .

[0031] "Evaluating the true fitness of individuals in the initial population using simulation software" refers to steps 2-5: decoding the generated initial population to obtain the corresponding molecular structure combinations (i.e., specific substitution element pairs), and then calling VASP software to calculate the formation energy and Gibbs free energy of nitrate ion adsorption for each corresponding individual. VASP software, based on first-principles calculations and density functional theory, uses plane wave basis sets and projected fused wave methods for high-precision calculations. However, each evaluation is extremely time-consuming, and the computational power required increases exponentially with the scale of the molecular system.

[0032] "Constructing data archives for each subspace based on evaluation results" means storing the formation energy, adsorption Gibbs free energy, and corresponding integer codes of each individual calculated by VASP software into the data database of each subspace. In the process of subsequent iterations, the data archive continuously accumulates real evaluation samples, providing a data foundation for the training and updating of the surrogate model.

[0033] "Training the proxy model for the corresponding subspace using each data archive" refers to steps 2-6, based on the data archives of each subspace. Using existing samples, two types of Kriging surrogate models are constructed and trained: one model takes integer codes as input and corresponding formation energy as output, while the other takes integer codes as input and Gibbs free energy of nitrate ion adsorption as output. One set of surrogate models is trained for each subspace. (Including two Kriging models), a total of 4 surrogate models were obtained. ,like Figure 3 As shown. The Kriging model is an interpolation model based on Gaussian processes that can provide both predicted values ​​and estimates of prediction uncertainty, making it suitable as a surrogate model for expensive optimization problems.

[0034] "Within each subspace, using the current population as the parent, a multimodal multi-objective differential evolution algorithm is used to generate offspring individuals" refers to step 3-1: Within each subspace, using the current population as the parent population P, the clustering-based multimodal multi-objective differential evolution algorithm MMODE_CSCD is used as the optimizer to perform evolutionary operations, generating a new generation of offspring individuals through mutation and crossover operations. This algorithm, through adaptive clustering and special crowding distance calculation, can maintain good population diversity and convergence in both the target space and the decision space.

[0035] "The fitness value of offspring individuals is predicted by the surrogate model of their respective subspace" means that in each iteration, the formation energy and adsorption Gibbs free energy of offspring individuals are not directly calculated using the VASP software, but rather through the surrogate model that has been trained in the corresponding subspace. This strategy significantly reduces the number of expensive real-world assessments, allowing limited computational resources to be focused on the accurate evaluation of high-value candidate solutions.

[0036] "Determining whether to update the surrogate model based on the fill sampling criterion of adaptive resource allocation" means that the fill sampling criterion of adaptive resource allocation calculates the ratio of the number of Pareto solutions in each subspace archive to the number of Pareto solutions in the historical archive, assigns different surrogate model update probabilities to each subspace, and then dynamically adjusts the resource allocation rate of each subspace to achieve the tilt of computing resources towards high-value subspaces.

[0037] "Selecting multiple individuals from the candidate solution set of the current subspace for real fitness evaluation" means: if a subspace meets the surrogate model update conditions, then combining the historical data archive of that subspace... Compared with the current population, several superior individuals are selected through non-dominated ordination and special crowding distance (SCD). Non-dominated ordination effectively selects dominant individuals near the first frontier in the target space, while SCD evaluates the crowding distance of an individual in both the target and decision spaces. By prioritizing individuals with larger SCD values ​​in the first frontier for true function evaluation, the individual distribution characteristics of both spaces are comprehensively considered.

[0038] "Update the data archive and retrain the surrogate model using the evaluation results" refers to: using VASP software to evaluate the true fitness of the selected outstanding individuals, and storing the evaluated individuals and their corresponding formation energies and adsorption Gibbs free energies in the corresponding subspace of the data archive. In the process, the Kriging proxy model for that subspace is retrained based on the updated archived data. This is to improve the prediction accuracy and generalization ability of the model in subsequent iterations.

[0039] "Repeating the above evolution and model update operations until the termination condition is met" means: repeatedly performing operations such as offspring generation, surrogate model prediction, imputation sampling judgment, truth evaluation, and model update until the preset maximum number of truth function evaluations (FES) is reached. max (FES in this embodiment) max =400) or the iteration termination condition.

[0040] "Merging the data archives of each subspace, filtering out the set of non-dominated individuals, and outputting the final multimodal Pareto solution set and Pareto front" refers to: archiving the updated data of all subspaces. (i=1,2,3,4) are merged to obtain the global archive DB. A non-dominated sort is performed on all individuals in the global archive DB to filter out the set of non-dominated individuals, outputting the final Pareto front (PF) and multimodal Pareto solution set (PSs). After obtaining the final PSs and PF, the encoded data is decoded to obtain the corresponding combinations of substitution elements, thus providing multiple high-performance candidate schemes for the molecular structure design of catalytic materials.

[0041] In this embodiment, a multi-objective optimization model with the dual objectives of minimizing formation energy and adsorption Gibbs free energy is constructed through step 1. This model can simultaneously take into account the two core indicators of structural stability and catalytic activity in the cobalt phosphide catalyst diatomic substitution modification system. Since there is a typical conflict between formation energy and adsorption Gibbs free energy, the construction of this dual-objective model provides a clear optimization direction and evaluation benchmark for seeking the optimal balance between the two in the future. Step 2 employs a partitioned search strategy to divide the decision space into multiple non-overlapping subspaces, decomposing the complex global multimodal optimization problem into multiple local subproblems and effectively avoiding mutual interference between different modal regions. In each subspace, Latin hypercube sampling is used to generate an initial population, ensuring a uniform distribution of initial samples across the entire subspace even with a limited sample size, providing high-quality training samples for subsequent surrogate model training. VASP software is used to evaluate the true fitness and construct data archives for each subspace, providing accurate labeled data for surrogate model training. Kriging surrogate models for the corresponding subspaces are trained using each data archive, giving each subspace its own dedicated surrogate model capable of accurately fitting the functional terrain of that local region, fundamentally overcoming the technical limitation of a single global surrogate model being unable to adapt to multimodal distribution characteristics. Step 3 generates offspring using a multimodal, multi-objective differential evolution algorithm within each subspace and uses a surrogate model to predict the fitness values ​​of the offspring. This allows for the evaluation of the fitness of a large number of offspring without using expensive VASP software, significantly reducing computational costs. The surrogate model is updated based on a filling sampling criterion using adaptive resource allocation, allowing for the dynamic allocation of scarce computational resources according to the search potential of each subspace. This ensures that the limited real evaluation budget is precisely allocated to high-value regions. By using non-dominated sorting and special crowding distance to screen excellent individuals for real evaluation and update the surrogate model, the model's prediction accuracy in key regions is continuously improved. Finally, by merging the archives from each subspace and screening the set of non-dominated individuals, a uniformly distributed global Pareto front and multiple local multimodal Pareto solutions can be simultaneously output with an extremely low real evaluation budget, providing rich and reliable candidate schemes for the molecular structure design of catalytic materials.

[0042] Optionally, the construction of a multi-objective optimization model with the dual optimization objectives of minimizing molecular formation energy and minimizing adsorption Gibbs free energy, and the determination of decision variables and decision space, includes: The sets of cobalt-site and phosphorus-site substitution elements in the cobalt phosphide catalyst diatomic substitution modification system were determined and encoded by integers to form two-dimensional decision variables. The value range of each dimension of the decision variables was determined by the upper limit of the encoding of the corresponding substitution element set. The molecular formation energy was calculated by the total energy of the doping system, the energy of cobalt element, the energy of cobalt phosphide, and the energy of bulk phase transition metal atoms. The adsorption Gibbs free energy was calculated by the energy of adsorbed nitrate ions, the energy of the catalyst substrate, the energy of gaseous nitrate molecules, and the energy of hydrogen molecules. All of the above energies were obtained by simulation software.

[0043] Specifically, "determining the set of cobalt-substituting elements and the set of phosphorus-substituting elements in the diatomic substitution modification system of cobalt phosphide catalyst" means that cobalt phosphide (CoP) catalysts have a typical crystal lattice structure in which Co atoms and P atoms occupy different lattice positions. Diatomic substitution modification refers to simultaneously replacing some atoms at the Co and P sites with other elements to regulate the electronic structure and surface properties of the catalyst. According to research by relevant scholars, the candidate substitution elements for Co atoms are {Sc, Ti, Zr, Hf, V, Nb, Ta, Cr, Mo, W, Mn, Tc, Re, Fe, Ru, Os, Co, Rh, Ir, Ni, Pd, Pt, Cu, Ag, Au, Zn}, totaling 26; the candidate substitution elements for P atoms are {Sc, Y, Zr, Hf, Nb, Ta, Mo, W, Mn, Os, Co, Ir, Ni, Pt, Cu, Au, Cd, B, C, Si, N, P, As, O, S, Se, Te, Cl, Br}, totaling 29.

[0044] "Encoding them with integers to form two-dimensional decision variables" means: Arranging each candidate substitution element in the periodic table according to column order, mapping each element to a unique integer code. Co-substituents Sc, Ti, Zr, Hf, V, Nb, Ta, Cr, Mo, W, Mn, Tc, Re, Fe, Ru, Os, Co, Rh, Ir, Ni, Pd, Pt, Cu, Ag, Au, and Zn correspond to codes 0-25 respectively; P-substituents Sc, Y, Zr, Hf, Nb, Ta, Mo, W, Mn, Os, Co, Ir, Ni, Pt, Cu, Au, Cd, B, C, Si, N, P, As, O, S, Se, Te, Cl, and Br correspond to codes 0-28 respectively. The two-dimensional code combination (x1, x2) uniquely determines a set of diatomic substitution element combinations.

[0045] "The range of values ​​for each dimension of the decision variable is determined by the upper bound of the encoding of the corresponding set of replacement elements" means that the first dimension (Co bit) has a range of values ​​of [0, 25], corresponding to 26 candidate replacement elements; the second dimension (P bit) has a range of values ​​of [0, 28], corresponding to 29 candidate replacement elements. Therefore, the upper bound of the decision space X is [25, 28], and the lower bound is [0, 0].

[0046] "The above energies were all calculated using simulation software" means that all the above energies... , , , , ,as well as All calculations rely on the VASP software. VASP employs plane wave basis sets and projected embellished wave methods, offering high computational accuracy but with a time complexity of O(N). ³ The extremely long time required for a single structural evaluation makes this a typical expensive optimization problem.

[0047] In this optional embodiment, by determining the candidate substitution element set for cobalt and phosphorus sites in the cobalt phosphide catalyst diatomic substitution modification system and encoding it with integers, the complex element combination selection problem is transformed into a standard two-dimensional integer optimization problem, allowing the mutation, crossover, and other operations of the evolutionary algorithm to be performed in an orderly manner within the integer encoding space. By comprehensively calculating the total energy of the doped system, the energy of elemental cobalt, the energy of cobalt phosphide, and the energy of bulk phase transition metal atoms using the formation energy calculation expression, the thermodynamic stability of the catalyst under different substitution combinations can be quantitatively evaluated. Similarly, by comprehensively calculating the energy of adsorbed nitrate ions, the energy of the catalyst substrate, the energy of gaseous nitrate molecules, and the energy of hydrogen molecules using the adsorption Gibbs free energy calculation expression, the initial efficiency of the catalytic reaction under different substitution combinations can be quantitatively evaluated. All of the above energies are calculated using VASP software, ensuring the physical accuracy and reliability of the fitness assessment and providing high-fidelity label data for subsequent training of the surrogate model.

[0048] Optionally, the partition search strategy includes: Select key dimensions from all dimensions of the decision space, and evenly divide the value range of each selected key dimension into a preset number of intervals. By combining the intervals of each key dimension, the entire decision space is deconstructed into multiple non-overlapping subspaces.

[0049] Specifically, "selecting key dimensions from all dimensions of the decision space" means that the decision space has D dimensions (D=2 in this embodiment), and d key dimensions are selected from these D dimensions as the basis for partitioning. The selection of key dimensions can cover all dimensions (d=D), or it can select some dimensions that have a significant impact on the multimodal distribution based on prior knowledge of the problem. In this embodiment, d=D=2, that is, both dimensions are selected as key dimensions.

[0050] "Dividing the value range of each selected key dimension evenly into a preset number of intervals" means that for each selected key dimension j (j=1,2,…,d), its value range [Lj,Uj] is evenly divided into q intervals of equal length. The number of intervals q can be determined based on the number of subspaces required and the complexity of the problem. In this embodiment, q=2, that is, the value range of each dimension is evenly divided into 2 intervals.

[0051] "By combining the intervals of each key dimension, the entire decision space is deconstructed into multiple non-overlapping subspaces" means that the intervals of each key dimension are combined using a Cartesian product to form K=q. ᵈ Each subspace is a non-overlapping subspace. Each subspace is uniquely determined by a specific combination of intervals along each key dimension. For example, when d=2 and q=2, the two intervals of the first dimension and the two intervals of the second dimension combine to form 2×2=4 subspaces. This partitioning method ensures that the physical location of each subspace in the decision space is clear and non-overlapping, laying the foundation for independently constructing surrogate models and performing evolutionary search in each subspace.

[0052] In this optional embodiment, by selecting key dimensions from all dimensions of the decision space and uniformly dividing the value range of each selected dimension into a preset number of intervals, a systematic and regular partitioning of the decision space can be achieved. By combining the intervals of each key dimension, the entire decision space is deconstructed into multiple non-overlapping subspaces, effectively decomposing the original global multimodal optimization problem into multiple local subproblems. This partitioning strategy naturally separates multiple locally optimal regions that were originally scattered in the global decision space into different subspaces. The function terrain within each subspace is relatively simple, and the multimodality is greatly reduced, thereby effectively avoiding mutual interference between different modal regions and significantly reducing the function complexity that the surrogate model needs to fit in each subspace.

[0053] Optionally, both dimensions are selected as key dimensions, and the value range of each dimension is evenly divided into two intervals to form four subspaces.

[0054] Specifically, "selecting both dimensions as key dimensions" means that in this embodiment, the decision space is a two-dimensional space, namely x1 (Co-bit replacement element encoding, value range [0, 25]) and x2 (P-bit replacement element encoding, value range [0, 28]). Selecting both dimensions as key dimensions means that d=D=2.

[0055] "Dividing the value range of each dimension evenly into two intervals" means: for dimension x1, its value range [0, 25] is evenly divided into two intervals: interval 1 is [0, 12], and interval 2 is [13, 25]; for dimension x2, its value range [0, 28] is evenly divided into two intervals: interval 1 is [0, 14], and interval 2 is [15, 28]. During the division, the interval endpoints must be rounded to ensure that each individual value in each subspace is an integer.

[0056] "Combining to form 4 subspaces" means that the Cartesian product of two intervals in each of the two dimensions forms 2×2=4 subspaces, namely: subspace 1 (x1∈[0,12],x2∈[0,14]), subspace 2 (x1∈[0,12],x2∈[15,28]), subspace 3 (x1∈[13,25],x2∈[0,14]), and subspace 4 (x1∈[13,25],x2∈[15,28]). Each subspace covers one quadrant of the decision space, and the four subspaces together completely cover the entire two-dimensional decision space.

[0057] In this optional embodiment, by selecting both dimensions as key dimensions and uniformly dividing each dimension into two intervals to form four subspaces, the most concise and effective partitioning scheme is achieved in this two-dimensional decision space scenario. The number of subspaces matches the problem dimensions, ensuring that each subspace has a sufficiently large area for population evolution search while also achieving sufficient subdivision of the decision space. This partitioning scheme separates multiple local optima that might otherwise be scattered globally into different subspaces. The surrogate model in each subspace only needs to fit the function terrain of that local region, without having to deal with the complex challenges of global multimodal distribution, significantly reducing the fitting difficulty of the surrogate model and improving prediction accuracy.

[0058] Optionally, the Latin hypercube sampling adopts an integer sampling method, generating a preset number of individuals in each subspace. The sampling process includes: uniformly dividing the integer value range of each dimension into equally spaced integer intervals equal to the number of samples; independently generating random permutations for each dimension to ensure that each interval is selected only once; randomly selecting an integer value in each integer interval; and forming a complete integer sample set by random permutations and combinations of each dimension.

[0059] Specifically, "Latin hypercube sampling adopts an integer sampling method" means that since the decision variables of this invention are integer encoded (the encoding of the replacement elements is an integer), Latin hypercube sampling must adopt an integer sampling method, that is, all sampling points are integer coordinates.

[0060] "Generate a preset number of individuals in each subspace" means that within each subspace, NP individuals are generated using integer-type Latin hypercube sampling (NP=200 in this embodiment). Each individual corresponds to a set of two-dimensional integer codes (x1, x2), representing a set of diatomic substitution element combinations.

[0061] "Dividing the integer value range of each dimension into equally spaced integer intervals equal to the number of samples" means that for each dimension j, its integer value range [Lj, Uj] within that subspace is uniformly divided into NP equally spaced integer intervals. Each interval contains at least one integer value, and the lengths of the intervals are equal or as equal as possible.

[0062] "Randomly select an integer value within each integer interval" means that for the k-th interval (k=1,2,...,NP) of each dimension j, randomly select an integer value from all the integers contained in the interval as the value of that dimension at that sample point.

[0063] In this optional embodiment, by employing an integer-type Latin hypercube sampling method, uniform coverage of the integer-type decision space can be achieved with NP samples in each subspace. Since the integer value intervals of each dimension are uniformly divided into equally spaced integer intervals equal to the number of samples, and random permutations are independently generated for each dimension to ensure that each interval is selected only once, the generated sample set has Latin square properties throughout the integer space—each interval in each dimension is selected exactly once, and the samples are highly uniformly distributed in the space without overlap. This sampling strategy, under the condition of a limited number of samples (200 initial samples per subspace in this embodiment), can obtain the most comprehensive information about the subspace function terrain with the minimum sample size, providing an initial training sample set with maximized information for subsequent training of the surrogate model, thereby improving the initial prediction accuracy of the surrogate model from the source.

[0064] Optionally, the real fitness assessment includes: decoding individuals in the initial population to obtain corresponding molecular structure combinations, calling simulation software to calculate the formation energy and adsorption Gibbs free energy of each molecular structure combination, and storing the calculation results and corresponding codes together in the data archive of each subspace; the surrogate model is a Kriging model, and two Kriging models are trained in each subspace to predict the formation energy and adsorption Gibbs free energy, respectively.

[0065] Specifically, the study focused on the structural design of cobalt phosphide diatomic substitution modification, selecting phosphorus and cobalt as doping sites for elemental substitution exploration. Based on the arrangement of elements in the periodic table, standardized numerical codes were assigned to the substitution elements, and the specific codes for each substitution element are shown in Tables 1 and 2.

[0066]

[0067] "Decoding individuals in the initial population to obtain the corresponding molecular structure combination" means that each individual in the initial population is a two-dimensional integer code (x1, x2), where x1 corresponds to the code of the Co-substituted element and x2 corresponds to the code of the P-substituted element. The decoding process is to restore the integer code to the specific element symbol according to the code-element mapping relationship shown in Table 2. For example, the code (7, 21) corresponds to the Cr-P element combination after decoding.

[0068]

[0069] Candidate sample sets are constructed using the element combinations defined in Tables 1 and 2. Under the condition of limited computational resources, the optimal element combination configuration is solved with the formation energy and the Gibbs free energy of nitrate ion adsorption as the dual optimization objectives.

[0070] "Calling simulation software to calculate the formation energy and adsorption Gibbs free energy of each molecular structure combination" means: for each molecular structure combination obtained after decoding (i.e., a specific pair of diatomic substituted elements), calling VASP software to perform first-principles calculations.

[0071] "Storing the calculation results and their corresponding codes together in the data archive of each subspace" means: storing the integer code (x1, x2) of each individual and its corresponding formation energy. The adsorption Gibbs free energy ΔG(NO3) is stored as a data entry in the data archive of the subspace where the individual resides. In (i=1,2,3,4), the data archive is stored in either a database or a file for quick retrieval during subsequent proxy model training and fill sampling.

[0072] "The surrogate model is a Kriging model" means that the Kriging model, also known as a Gaussian process regression model, is a statistical interpolation method. It assumes that the objective function follows a Gaussian process, estimates the function value at unknown points using known sample points, and simultaneously provides an estimate of the variance of the predicted values. Kriging models are particularly suitable for expensive optimization problems with limited sample sizes because they not only provide predicted values ​​but also quantify the uncertainty of the predictions.

[0073] In this optional embodiment, by decoding the initial population individuals to obtain the corresponding molecular structure combinations and calling VASP software for real fitness evaluation, the physical accuracy and reliability of the initial training samples are ensured, providing high-fidelity label data for the surrogate model. By storing the calculation results and corresponding codes together in the data archives of each subspace, a complete mapping database of code-performance is established, providing a unified data management foundation for subsequent training and iterative updates of the surrogate model. By employing a Kriging model as the surrogate model and training two independent Kriging models in each subspace to predict the formation energy and adsorption Gibbs free energy respectively, each objective function has a dedicated surrogate model for accurate fitting in each subspace. The Kriging model can still provide good prediction accuracy and uncertainty estimation even with a limited number of samples. Modeling the two objective functions separately avoids the mutual interference that may occur when a single model simultaneously fits two conflicting objectives. This multi-surrogate model architecture allows the two objective functions in each subspace to obtain independent, high-precision surrogate predictions, providing accurate and reliable fitness estimates for subsequent multi-objective differential evolution algorithms.

[0074] Optionally, step 3 includes: Step 3-1: In each subspace, using the current population as the parent generation, a new generation of offspring individuals is generated using a cluster-based multimodal multi-objective differential evolution algorithm. The fitness value of the offspring individuals is predicted by the surrogate model of the corresponding subspace. Step 3-2: Based on the filling sampling criterion based on adaptive resource allocation, determine whether the proxy model of each subspace needs to be updated. If so, combine the historical data archive of each subspace with the current population, select multiple excellent individuals through non-dominated sorting and special crowding distance for real fitness evaluation, store the evaluated individuals in the corresponding data archive and retrain the proxy model. Step 3-3: Repeat steps 3-1 and 3-2 until the preset iteration termination condition is reached, and output the updated data archive for each subspace. Steps 3-4: Merge the data archives of all subspaces to obtain the global archive. Perform non-dominated sorting on all individuals in the global archive, filter out the set of non-dominated individuals, and output the final Pareto front and multimodal Pareto solution set.

[0075] Specifically, "Step 3-1, within each subspace, using the current population as the parent, and employing a cluster-based multimodal multi-objective differential evolution algorithm to generate a new generation of offspring individuals" means that in each subspace, the current population is used as the parent population P, and the MMODE_CSCD algorithm (a cluster-based multimodal multi-objective differential evolution algorithm) is used as the optimizer. This algorithm, through adaptive clustering and special crowding distance calculation, can maintain both the diversity of the decision space and the objective space while preserving population convergence.

[0076] "The fitness value of offspring individuals is predicted by the surrogate model of the corresponding subspace" means that in each iteration of step 3-1, the formation energy and adsorption Gibbs free energy of the newly generated offspring individuals are not directly calculated by calling the VASP software, but are calculated by the surrogate model that has been trained in the subspace. It enables rapid prediction. This prediction process has extremely low computational cost, allowing the algorithm to evaluate and select a large number of offspring within a limited computational budget.

[0077] Step 3-2, "Determining whether the surrogate model of each subspace needs to be updated based on the fill sampling criterion of adaptive resource allocation," refers to: The fill sampling criterion based on adaptive resource allocation dynamically allocates different surrogate model update probabilities to each subspace by calculating the ratio of the number of Pareto solutions in the archive of each subspace to the number of Pareto solutions in the historical archive. When the selection probability of a certain subspace... When the probability value is greater than the randomly generated probability threshold rand (rand∈[0,1]), the proxy model update operation of that subspace is triggered, and the filling sampling criterion is executed.

[0078] "Combining historical data archives of each subspace with the current population, and selecting multiple superior individuals for true fitness evaluation through non-dominated ranking and special crowding distance" refers to: when a subspace meets the update conditions, the data of that subspace is archived. The individuals are merged with all individuals in the current parent population Pᵢ. After deduplication, a non-dominated ranking is performed on the individuals in the merged set. For individuals with the same non-dominated level, a special crowding distance (SCD) is used for further ranking. After ranking, individuals already in the data archive are removed from the candidate set, and the top η individuals (η=6 in this example) from the remaining individuals are selected as candidate solutions. The VASP software is then used to evaluate their true fitness.

[0079] The Special Crowding Distance (SCD) is calculated as follows: In the formula, This indicates the crowding distance of an individual in the decision-making space. This indicates the crowding distance of an individual in the target space. This represents the average crowding distance in the decision-making space. This represents the average crowding distance in the target space.

[0080] "Storing the evaluated individuals into the corresponding data archive and retraining the surrogate model" means storing the encoding, formation energy, and adsorption Gibbs free energy of the η individuals that have undergone real evaluation by VASP into the data archive of this subspace. In this process, two Kriging surrogate models for the subspace are retrained based on the updated data archive, thereby gradually improving the prediction accuracy of the surrogate models in subsequent iterations.

[0081] "Step 3-3, repeat steps 3-1 and 3-2 until the preset iteration termination condition is reached, and output the updated data archives for each subspace" means that in each iteration cycle, step 3-1 is executed first (evolutionary generation of offspring and fitness prediction by the surrogate model), then step 3-2 is executed (determining whether the surrogate model needs to be updated and performing a true evaluation), and this process is repeated until the preset maximum number of true function evaluations (FES) is reached. max (FES in this embodiment) max =400). After the iteration is complete, output the final updated data archive for each subspace. .

[0082] "Steps 3-4, merging the data archives of all subspaces to obtain the global archive, performing a non-dominated sort on all individuals in the global archive, filtering out the set of non-dominated individuals, and outputting the final Pareto front and multimodal Pareto solution set" refers to: archiving the updated data archives of all subspaces. (i=1,2,3,4) are merged into a global archive DB. A non-dominated sort is performed on all individuals in the DB to obtain the set of non-dominated individuals (i.e., the Pareto optimal solution set). The non-dominated front in the target space is the Pareto front (PF), and the corresponding non-dominated solution set in the decision space is the multimodal Pareto solution set (PSs). The final output PSs contains multiple molecular structure schemes with equivalent or similar performance in the target space but different distributions in the decision space.

[0083] In this optional embodiment, step 3-1 generates offspring using a clustering-based multimodal multi-objective differential evolution algorithm within each subspace, and uses a surrogate model to predict the fitness values ​​of the offspring. This allows for the evaluation and selection of a large number of offspring without triggering expensive VASP computations. Step 3-2 dynamically determines whether the surrogate model in each subspace needs updating based on an adaptive resource allocation-based filling sampling criterion, enabling the scarce real evaluation budget to be allocated differentially according to the search potential of each subspace—subspaces contributing more non-dominated solutions receive a higher update probability. Excellent individuals are selected for real evaluation through non-dominated ranking and special crowding distance, ensuring that the selected candidate solutions are both near the Pareto front in the target space and sparsely distributed in both the decision and target spaces. This simultaneously improves the convergence accuracy of the surrogate model near the Pareto front and its generalization ability across the entire space. The evaluated individuals are stored in the data archive, and the surrogate model is retrained, allowing it to gradually approach the real function terrain with iterative progress. Step 3-3 iterates until the termination condition is met, ensuring that the algorithm converges sufficiently within a limited computational budget. By merging the data archives of all subspaces and filtering the set of non-dominated individuals in steps 3-4, the final Pareto front has good distribution uniformity. The multimodal Pareto solution set contains multiple candidate molecular structures that are equivalent in performance in the target space but have different positions in the decision space, providing a variety of selection schemes for the design of catalytic materials.

[0084] Optionally, the cluster-based multimodal multi-objective differential evolution algorithm in step 3-1 includes: performing non-dominated sorting on the current population to obtain the non-dominated levels of each solution; adaptively determining the number of clusters according to the number of individuals in each level and performing hierarchical clustering, calculating the crowding distance in the decision space and the crowding distance in the target space using only individuals in the same cluster as neighbors, and fusing them to obtain the non-dominated and crowding distance indices; sorting the population according to the rule of prioritizing non-dominated levels and descending the indices within the same level; selecting sparsely distributed individuals from the better non-dominated levels as candidate models, and matching a dedicated learning model for each individual through a roulette wheel method with decision space distance weighting; generating offspring individuals based on differential evolution mutation and crossover operations, merging the parent and offspring generations and performing sorting again, and selecting the best individuals to form the next generation parent population.

[0085] Specifically, "performing non-dominated sorting on the current population to obtain the non-dominated levels of each solution" means that non-dominated sorting is a classic operation in multi-objective optimization, which divides individuals in the population into different frontier levels based on Pareto dominance. Individuals in the first frontier are not dominated by any other individuals and represent the optimal solution in the current population; individuals in the second frontier are dominated only by individuals in the first frontier, and so on.

[0086] "Adaptively determine the number of clusters based on the number of individuals in each layer and perform hierarchical clustering" means that for each non-dominated layer, the number of k-means clusters k is adaptively determined based on the number of individuals in that layer, and k-means clustering is performed independently within each layer to divide individuals in the same layer into several clusters.

[0087] "Calculating crowding distance in the decision space and target space using only individuals within the same cluster as neighbors" means that crowding distance is typically used to measure the spatial distribution density of individuals. Traditional methods calculate crowding distance across the entire population, but this easily overlooks inter-cluster distances in multimodal optimization. In this algorithm, the calculation of an individual's crowding distance is restricted to within the same cluster; that is, the crowding distance CD of that individual in the decision space is calculated only using other individuals in the same cluster as neighbors. i,x and the crowding distance CD in the target space i,f .

[0088] "The fusion of non-dominated and crowding distance indicators" refers to the fusion of an individual's non-dominated hierarchy information with the two crowding distances mentioned above to obtain a comprehensive non-dominated and crowding distance (CSCD) indicator. The CSCD indicator comprehensively considers the convergence of an individual in the target space (reflected by the non-dominated hierarchy) and the sparsity of its distribution in the decision space and target space (reflected by the intra-cluster crowding distance), providing a multi-dimensional evaluation basis for population ranking.

[0089] Step 3-1-1: Initialize the algorithm for each subspace by setting the population size, maximum number of iterations, and difference scaling factor. Hyperparameters such as crossover probability are defined; NP individuals are randomly generated in the decision space to form an initial population P; the multi-objective function values ​​of all individuals are calculated to complete population evaluation, and the iterative loop is started. The scaling factor F=0.8 and the crossover probability Cr=1 are used in the differential evolution algorithm; the number of individuals NP is set to 100 and 50 respectively according to different dimensions of the problem.

[0090] Step 3-1-2: First, the population is divided into layers to obtain the non-dominated levels of each solution; then, the number of k-means clusters is adaptively determined according to the number of individuals in each layer, and the layered clustering is completed. Only individuals in the same cluster are considered as neighbors to calculate the decision and target space crowding, and then the non-dominated and crowding distance (CSCD) indices are fused; finally, the population is sorted according to the rule of non-dominated level priority and CSCD descending order in the same layer, and the ordered population is output.

[0091] "Ranking the population according to the rule of non-dominated hierarchy priority and descending order of the indicators within the same hierarchy" means that: the first criterion for population ranking is the non-dominated hierarchy, and the lower the hierarchy (the earlier the individual), the higher the priority; for individuals in the same non-dominated hierarchy, they are arranged in descending order according to the CSCD index, and the higher the CSCD value (representing the sparser distribution in that hierarchy), the higher the priority.

[0092] Step 3-1-3: Traverse all individuals in the population layer by layer according to the non-dominated ranking index. First, randomly select the model source layer from the upper layers with better quality. Then, select high-quality individuals with sparse distribution in this layer as candidate models. Finally, match a unique learning model for each individual using a roulette wheel method with decision space distance weighting. .

[0093] "Selecting sparsely distributed individuals from the better non-dominated hierarchy as candidate models" means: model individuals ( () is the source of guiding direction in differential evolution mutation operations. From the better (front) non-dominated levels, individuals with larger CSCD values ​​(sparse distribution) are selected as the candidate model set. These individuals represent the better quality solutions in the current population and the solutions with sparse distribution around them.

[0094] "Matching a unique learning model to each individual using a roulette wheel approach with decision space distance weighting" means that for each individual in the population, the Euclidean distance between that individual and each candidate model in the decision space is calculated from the candidate model set. The reciprocal of the distance is used as the weight for roulette wheel selection. The closer the candidate model is, the higher the probability of it being selected, thus matching each individual with a high-quality model that is close to it in the decision space as a learning direction.

[0095] "Generating offspring individuals based on differential evolution mutation and crossover operations" refers to: Step 3-1-4, for each target individual in the population... Utilizing its exclusive learning role models and two distinct individuals randomly selected from the population , According to the mutation formula: In the formula Two distinct individuals are randomly selected from the population, and F is the difference scaling factor (F=0.8 in this example). Then, a binomial crossover operation of differential evolution is performed to generate offspring individuals.

[0096] The differential scaling factor F and the crossover probability Cr are the core control parameters of the differential evolution algorithm. In this embodiment, F=0.8 is the scaling factor for the mutation amplitude. A larger value (close to 1) can enhance the algorithm's global exploration ability and avoid getting trapped in local optima too early. Cr=1 is the probability of binomial crossover. A value of 1 indicates that offspring individuals completely inherit the genes of the mutation vector and do not retain gene fragments from the parents. This extreme setting aims to maximize the mutation diversity of offspring individuals, which is beneficial for exploring more local optima in multimodal optimization scenarios. The value of the population size NP is adjusted according to the problem dimension—NP=100 is set for the CEC2019 test set (high dimension), and NP=50 is set for the IDMP test set (low dimension) to ensure that the population size matches the complexity of the decision space.

[0097] By setting the differential scaling factor to F=0.8, a larger search step size can be maintained during mutation operations, enhancing the algorithm's global exploration capability in the decision space and helping to cross the low-fitness "valleys" between different modal regions. By setting the crossover probability to Cr=1, offspring individuals adopt the mutation vector entirely instead of the parent gene fragment, further improving population diversity and avoiding the loss of multimodal solutions caused by excessive similarity between offspring and parents. By adjusting the population size NP according to the problem dimension, the algorithm can maintain sufficient population diversity in molecular design problems of different complexities, preventing the omission of multimodal solutions due to an excessively small population or the waste of computational resources due to an excessively large population.

[0098] "After merging the parent and offspring generations, perform another sorting and select the best individuals to form the next generation of parent population" means: merging all parent individuals PP and the generated offspring individuals OP into a temporary mixed population POP = P∪OP. Then, performing a complete CSCD-based non-dominated sorting on POP again yields an ordered mixed population POP*. Finally, selecting NP best individuals from POP* based on sorting priority as the parent population for the next iteration.

[0099] In this optional embodiment, by performing non-dominated sorting on the current population to obtain the non-dominated levels of each solution, the quality of individuals in the population can be clearly distinguished, making convergence the primary criterion for population sorting. By adaptively determining the number of clusters based on the number of individuals in each level and performing hierarchical clustering, and calculating the crowding distance in the decision space and the crowding distance in the target space using only individuals in the same cluster as neighbors, the calculation of the crowding distance is not affected by individuals in other clusters, and can truly reflect the sparseness of the distribution of individuals in their local regions. By integrating the non-dominated levels and the two crowding distances into a CSCD index, and sorting the population according to the rule of prioritizing non-dominated levels and descending CSCD within the same level, a sorting strategy of "convergence first, diversity second" is realized, ensuring that the population maintains a trend of approaching the Pareto front and a wide distribution in both the target space and the decision space during the evolutionary process. By selecting sparsely distributed individuals from the better non-dominated levels as candidate role models, and matching each individual with a dedicated learning role model through a roulette wheel approach weighted by decision space distance, the mutation direction of each individual is directed towards the high-quality sparse solutions in its vicinity in the decision space. This guides the individuals in the population to evolve towards the optimal solution in their local region, effectively maintaining the multimodal distribution characteristics of the population. Offspring are generated through mutation and crossover operations based on differential evolution. After merging the parent and offspring generations, a second sorting is performed to select the best individuals to form the next generation of parent population, achieving an elite preservation mechanism and ensuring that the population quality does not degrade with iteration.

[0100] Optionally, the filling sampling criterion for adaptive resource allocation in step 3-2 includes: initializing the selection probability of each subspace; merging the data archives of all subspaces into a global archive and performing non-dominated sorting, and counting the total number of non-dominated individuals globally; when the selection probability of a certain subspace is greater than a randomly generated probability threshold, performing a filling sampling operation on that subspace, wherein the filling sampling operation includes merging the data archive of the current subspace with the current parent population, performing deduplication and non-dominated sorting, and then using a special crowding distance to perform a secondary sorting of individuals with the same non-dominated level, removing individuals already in the data archive, and selecting those ranked higher. A preset number of individuals are used to evaluate the true fitness of the proxy model and update it. After performing fill sampling, the number of individuals corresponding to the local Pareto front of each subspace in the global Pareto front is counted as the number of contributing individuals in each subspace. The selection probability of each subspace is updated based on the number of contributing individuals in each subspace and the total number of non-dominated individuals in the global subspace: if there are non-dominated individuals in the local data archive of a subspace, the selection probability of that subspace is increased by the product of the constant selection probability and the proportion of the number of contributing individuals in that subspace to the total number of non-dominated individuals in the global subspace; otherwise, the selection probability of that subspace is decreased by the constant selection probability.

[0101] Specifically, "initializing the selection probability of each subspace" means: in step 3-2-1, at the beginning of the algorithm, assigning an initial selection probability to each subspace i (i=1,2,3,4). In this embodiment, the initial selection probability of all subspaces is set to the same value (e.g., ...). =0.5), ensuring that each subspace gets an equal chance to be updated in the early stages of the algorithm.

[0102] "Merge the data archives of all subspaces into a global archive and perform non-dominated sorting, then count the total number of globally non-dominated individuals" refers to: Step 3-2-2, which involves archiving the data archives of all subspaces. (i=1,2,3,4) are merged into a global archive. Step 3-2-3, for Perform a non-dominated sort on all individuals to obtain the global non-dominated solution set PSs and the global Pareto front PF. Count the total number of globally non-dominated individuals, denoted as B.

[0103] "When the selection probability of a certain subspace is greater than the randomly generated probability threshold, perform a filling sampling operation on that subspace" means: Step 3-2-4, for each subspace i, generate a random number rand∈[0,1]. If the probability threshold... If the value is greater than or equal to rand, then the filling sampling operation of the subspace is triggered; otherwise, the subspace will not be evaluated in this iteration.

[0104] "The filling sampling operation includes merging the data archive of the current subspace with the current parent population" means: The data archive of the i-th subspace... Compared with the current parent population P i All individuals are merged and stored in a temporary set T1.

[0105] "After deduplication and non-dominated sorting, a secondary sorting is performed on individuals with the same non-dominated level using a special crowding distance" means: performing a deduplication operation to remove duplicate individuals from the temporary set T1, and then performing a non-dominated sort on all deduplicated individuals. For individuals with the same non-dominated level, a further descending sort is performed using a special crowding distance (SCD).

[0106] "Removing existing individuals from the data archive, selecting a predetermined number of top-ranked individuals for true fitness evaluation, and updating the surrogate model" refers to: removing individuals from the data archive... The individuals already in the set are removed from the sorted set T2, and the top η individuals (η=6 in this example) from the remaining individuals are selected as candidate solutions. The VASP software is used to evaluate the true fitness of the candidate solutions, and the evaluated individuals are stored in the data archive. In, and based on the updated training data Retrain the agent model .

[0107] "After performing padding sampling, the number of individuals corresponding to the local Pareto fronts of each subspace included in the global Pareto front is counted as the number of contributing individuals in each subspace" refers to: Step 3-2-5, archiving the local data of each subspace i. Perform non-dominated sorting on each solution and obtain its local non-dominated solution set. With local Pareto front The statistical global Pareto front (PF) includes local Pareto fronts. The number of corresponding individuals in the middle, and this number is denoted as , which is the number of individuals contributing to the global Pareto front from the i-th subspace.

[0108] "Updating the selection probability of each subspace based on the number of contributing individuals in each subspace and the total number of globally non-dominated individuals" refers to: based on statistically obtained... The selection probability of each subspace is updated through an adaptive resource allocation mechanism, based on the total number B of globally non-dominated individuals. .

[0109] "If a non-dominant individual exists in the local data archive of a subspace, then increase the selection probability of that subspace by the product of the constant selection probability and the proportion of the number of contributing individuals in that subspace to the total number of non-dominant individuals globally" refers to step 3-2-6, based on the statistically obtained data. Total number of globally non-dominated individuals The selection probability corresponding to this subspace is calculated through an adaptive resource allocation mechanism. Choice probability The calculation formula is as follows: In the formula, This represents the constant selection probability in each update. Indicates the first The number of non-dominated solutions in each archive. This represents the number of non-dominated solutions in the total archive. As can be seen from the formula, the more non-dominated individuals in a sub-archive, the higher the selection probability. The larger. Furthermore, if or ,but or .

[0110] Step 3-2-7 involves repeating the above-described filling sampling criterion and selection probability update iterative process for each subspace, ultimately returning the updated surrogate model. With data archive .

[0111] In this optional embodiment, adaptive allocation of computing resources among subspaces is achieved by initializing the selection probability of each subspace and dynamically updating it during iteration. A global benchmark for evaluating the contribution of each subspace is established by merging the data archives of all subspaces into a global archive, performing non-dominated sorting, and counting the total number of globally non-dominated individuals. When the selection probability of a subspace exceeds a randomly generated probability threshold, a filling sampling operation is triggered. The data archive of the current subspace is merged with the current parent population, deduplicated, and then non-dominated. A secondary sorting is performed on individuals with the same non-dominated level using a special crowding distance, ensuring that the selected candidate solutions are located near the Pareto front in the target space and sparsely distributed in both the decision and target spaces. This balances the convergence accuracy of the surrogate model near the Pareto front with its generalization ability across the entire space. Existing individuals in the data archive are removed, and a predetermined number of top-ranked individuals are selected for real evaluation, avoiding the waste of computing resources caused by repeatedly evaluating labeled samples. The number of individuals corresponding to the local Pareto fronts of each subspace within the global Pareto front is counted as the number of contributing individuals in each subspace, and the selection probability of each subspace is updated accordingly. Subspaces that contribute more non-dominated solutions receive a greater increase in selection probability, so that in subsequent iterations, subspaces with higher search value receive more real evaluation budgets.

[0112] Tables 3 to 5 show the Wilcoxon rank-sum test results of SMMODE_ZS and its comparative algorithms rPSP, IGDX, and IGDF under the same conditions. Figure 4 (a) Figure 4 (b) Figure 4 (c) Figure 4 (d) and Figure 4 (e) shows the final PSs and PF distributions obtained for each algorithm of the MMF1 function.

[0113] To verify the performance of the proposed algorithm in the decision space, Tables 3 and 4 show the mean and standard deviation of five expensive optimization algorithms under the same conditions for rPSP and IGDX metrics. Table 3 shows that, in 28 test functions, the SMMODE_ZS algorithm significantly outperforms the ParEGO, PB-NSGAIII, KTA2, and K-RVEA algorithms in rPSP metrics (20, 19, 20, and 24 test functions, respectively). Furthermore, this algorithm achieves the optimal rPSP value in 16 of the 28 test functions. Table 4 shows that, in 28 test functions, the SMMODE_ZS algorithm significantly outperforms the ParEGO, PB-NSGAIII, KTA2, and K-RVEA algorithms in IGDX metrics (23, 20, 20, and 24 test functions, respectively); and this algorithm achieves the optimal IGDX value in 17 of the 28 test functions. Experimental results using rPSP and IGDX metrics show that, compared to the four comparison algorithms, the SMMODE_ZS algorithm achieves better diversity and convergence performance in the decision space, further validating its performance advantage in the decision space.

[0114] To verify the performance of the proposed algorithm in the target space, Table 5 shows the mean and standard deviation of the five expensive optimization algorithms on the IGDF index under the same conditions. As shown in Table 5, among the 28 test functions, the SMMODE_ZS algorithm significantly outperforms the ParEGO, PB-NSGAIII, KTA2, and K-RVEA algorithms on the IGDF index, with 27, 16, 16, and 23 test functions respectively. Compared with the PB-NSGAIII and KTA2 algorithms, the proposed SMMODE_ZS algorithm performs poorly on some test functions containing local optimum (PSs) (such as MMF13 and MMF14). This is mainly because the PB-NSGAIII and KTA2 algorithms both employ a double-archive mechanism, which effectively avoids the algorithm getting trapped in local optima. However, the SMMODE_ZS algorithm, in order to improve population diversity, divides the entire decision space, which to some extent increases the risk of the algorithm getting trapped in local optima. Nevertheless, from the overall experimental results, the SMMODE_ZS algorithm still outperforms the four comparative algorithms in the target space. It should be noted that the four expensive multi-objective comparison algorithms mentioned above focus on solving the Pareto front, while the SMMODE_ZS algorithm not only needs to improve the diversity and convergence of the Pareto front, but also needs to search for multiple uniformly distributed PSs. Experimental results fully demonstrate that the proposed SMMODE_ZS algorithm can find well-distributed Pareto fronts while simultaneously searching for multiple uniformly distributed PSs.

[0115] The experimental results and analysis above fully verify the superiority of the proposed SMMODE_ZS algorithm. On one hand, the SMMODE_ZS algorithm employs a multi-agent model construction method based on partition search, which effectively reduces the adverse effects of multimodal characteristics on algorithm performance while improving population diversity. This algorithm obtains a well-distributed power factor (PF) and multiple evenly distributed power factors (PSs), an advantage not possessed by traditional expensive multi-objective algorithms. On the other hand, the filling sampling criterion based on an adaptive resource allocation mechanism can rationally allocate computational resources to each subspace, significantly improving resource utilization efficiency. Simultaneously, this criterion fully considers the distribution characteristics of individuals in the target and decision spaces, selecting excellent individuals to update the agent model in real time, thereby enabling the algorithm to obtain evenly distributed PSs and PF.

[0116] To more intuitively verify the comprehensive performance of the SMMODE_ZS algorithm in both the decision and objective spaces, this section further compares the PF and PSs generated by five expensive optimization algorithms on the MMF1 test function. Figure 4 As shown, in the target space, both the PB-NSGAIII algorithm and the SMMODE_ZS algorithm obtained a uniformly distributed power distribution. However, in the decision space, the power distribution obtained by the SMMODE_ZS algorithm is significantly better than the other four comparison algorithms, further verifying the superiority of the algorithm.

[0117]

[0118]

[0119]

[0120] To address the expensive multimodal, multi-objective optimization problem of minimizing the diatomic substitution configuration formation energy and the Gibbs free energy of nitrate ion adsorption in cobalt phosphide catalysts, the proposed SMMODE_ZS algorithm was used in optimization experiments. The aim was to verify the algorithm's solution performance and engineering application value in practical scenarios of catalytic material molecular design. Figure 5 As shown. To comprehensively evaluate the overall performance of the SMMODE_ZS algorithm, its variant MMODE_ZS was introduced as a comparison algorithm in the experimental design. This comparison algorithm, except for omitting the surrogate model module, maintains complete consistency with SMMODE_ZS in all other core strategies and parameter configurations to ensure the fairness and rationality of the experimental comparison. Regarding experimental parameter settings, the population size for both algorithms was set to NP=200, and the maximum number of evaluations of the true function was limited to FESmax=400. Table 6 shows the experimental results of the SMMODE_ZS algorithm and its variant. To more intuitively demonstrate the multimodal solutions found by the algorithm, Figure 6 (a) Figure 6 (b) and Figure 7(a) Figure 7 (b) The molecular structures of the two sets of multi-objective, multi-modal solutions are presented respectively. To ensure the reliability of the experimental data, the true fitness values ​​of all experimental samples were calculated using VASP software, providing accurate basic data support for algorithm performance evaluation.

[0121]

[0122] As shown in Table 6, under the same conditions, the SMMODE_ZS algorithm successfully searched for 6 global Pareto solutions (X1-X6), while the comparative algorithm MMODE_ZS only found 3 global Pareto solutions (Y1-Y3). Within a limited number of real evaluations, the SMMODE_ZS algorithm demonstrates better solution space exploration capability and population diversity maintenance performance, and can more efficiently locate a set of global Pareto solutions with good diversity. In addition to global Pareto solutions, the SMMODE_ZS algorithm also successfully explored two sets of multimodal solutions (X7, X8 and X9, X6) located on the local Pareto front. 10 The MMODE_ZS algorithm, however, did not detect any multimodal solutions. Although such multimodal solutions are not at the global Pareto front, they characterize the catalytic performance limits under specific structures, providing important local information for revealing the structure-activity relationship of cobalt phosphide catalysts. In contrast, the MMODE_ZS algorithm, due to the lack of a surrogate model, cannot effectively mine such valuable multimodal solutions under limited computational resource constraints.

[0123] From a computational efficiency perspective, the performance advantage of the SMMODE_ZS algorithm essentially stems from the strong predictive capability of the surrogate model. In computationally expensive optimization problems such as cobalt phosphide catalyst design, a single evaluation of the true function consumes significant computational resources and time. The SMMODE_ZS algorithm uses a surrogate model to predictively evaluate individual solutions in the population. With limited computational resources, it can obtain a greater number and higher quality optimization solutions by reducing the number of true evaluations, effectively avoiding the ineffective consumption of computational resources in non-optimal regions and significantly improving optimization efficiency. Figure 6 and Figure 7 The study demonstrates the different molecular structures corresponding to two sets of multimodal solutions with the same formation energy and Gibbs free energy of nitrate ion adsorption.

[0124] The experimental results above fully demonstrate that the proposed algorithm can provide decision-makers with more options. When adapting to high-stability, high-activity high-end scenarios, individuals with the lowest formation energy and moderate Gibbs free energy for nitrate adsorption should be prioritized, but the manufacturing cost of such solutions is usually high. If manufacturing cost is considered and performance requirements are moderate, individuals with moderate formation energy and superior Gibbs free energy for nitrate adsorption can be selected to balance performance and economic cost. The discovery of multimodal solutions further enriches the selection of solutions and effectively saves design costs. For example, when the objective is formation energy... When the Gibbs free energy of nitrate adsorption is −0.8 and ΔG(NO3) = 0, the lower-cost Ru-P combination can be used instead of the higher-cost Ir-P combination, minimizing cost while meeting performance requirements. In summary, the SMMODE_ZS algorithm demonstrates significant performance in solving the typical EMMOPs of cobalt phosphide catalyst diatomic substitution modification design. This algorithm not only searches for more and better global Pareto solutions within limited computational resources but also effectively mines multimodal solutions with research value, thus verifying the effectiveness and practicality of the proposed algorithm in complex engineering optimization problems.

[0125] Optionally, after outputting the multimodal Pareto solution set and Pareto front in step 3, the method further includes decoding the encoded individuals in the solution set to obtain the molecular structure substitution element combination corresponding to each non-dominated solution.

[0126] Specifically, "after outputting the multimodal Pareto solution set and Pareto front in step 3" means that after performing the global non-dominated sorting in steps 3-4 and filtering out the set of non-dominated individuals, the algorithm has obtained the final multimodal Pareto solution set (PSs) and Pareto front (PF). At this time, each individual in PSs is still in integer encoding form (x1, x2).

[0127] "Decoding the encoded individuals in the solution set" means that decoding is the reverse process of encoding. For each encoded individual (x1, x2) in PSs, according to the encoding-element mapping relationship shown in Table 2, x1 is restored to the corresponding Co-bit substitution element symbol, and x2 is restored to the corresponding P-bit substitution element symbol.

[0128] "Obtaining the molecular structure substitution element combinations corresponding to each non-dominated solution" means that after decoding, each encoded individual (x1, x2) is converted into a specific molecular structure description, namely, the element combination pair of "first substitution atom - second substitution atom" (such as "Ru-P", "Ir-P", "Tc-P", "Ti-N", etc.). These element combinations directly correspond to the actual molecular structure configuration and can be directly used by materials scientists for subsequent synthesis experiments or further analysis.

[0129] In this optional embodiment, by decoding the encoded individuals in the solution set after outputting the multimodal Pareto solution set and Pareto front, the abstract integer codes are transformed into specific combinations of molecular structure substitution elements, enabling the optimization results output by the algorithm to be directly understood by materials scientists and applied to the design of practical catalytic materials. For example... Figure 6 and Figure 7 As shown, the multimodal solutions obtained after decoding reveal different molecular structures with the same formation energy and adsorption Gibbs free energy—for example, when the target is the formation energy. When the Gibbs free energy of nitrate adsorption is −0.8 and ΔG(NO3) = 0, the Ru-P combination can be used instead of the more expensive Ir-P combination; when the objective is When ΔG(NO3) = −1 and ΔG(NO3) = 0.2, the Tc-P or Ti-N combination can be selected. These decoded multimodal molecular structures provide decision-makers with a wide range of options: when adapting to high-stability and high-activity high-end scenarios, individuals with the lowest formation energy and moderate adsorption Gibbs free energy can be prioritized; when considering manufacturing costs and moderate performance requirements, individuals with moderate formation energy and superior adsorption Gibbs free energy can be selected; when multiple candidate solutions with equivalent performance exist, lower-cost element combinations can be used to replace higher-cost combinations, minimizing costs while meeting performance requirements.

[0130] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. An expensive multimodal multi-objective optimization method for molecular structure design, characterized in that, Includes the following steps: Step 1: Construct a multi-objective optimization model with the dual optimization objectives of minimizing molecular formation energy and minimizing adsorption Gibbs free energy, and determine the decision variables and decision space; Step 2: The decision space is divided into multiple non-overlapping subspaces using a partitioned search strategy. In each subspace, an initial population is generated using Latin hypercube sampling. The fitness of individuals in the initial population is evaluated using simulation software. Based on the evaluation results, a data archive for each subspace is constructed, and the surrogate model for the corresponding subspace is trained using each data archive. Step 3: In each subspace, using the current population as the parent generation, the multimodal multi-objective differential evolution algorithm is used to generate offspring individuals. The fitness value of the offspring individuals is predicted by the surrogate model of their respective subspace. Based on the filling sampling criterion based on adaptive resource allocation, it is determined whether to update the surrogate model. If so, multiple individuals are selected from the candidate solution set of the current subspace for real fitness evaluation. The evaluation results are used to update the data archive and retrain the surrogate model. The above evolution and model update operations are repeated until the termination condition is met. The data archives of each subspace are merged, the set of non-dominated individuals is selected, and the final multimodal Pareto solution set and Pareto front are output.

2. The expensive multimodal multi-objective optimization method for molecular structure design according to claim 1, characterized in that, The construction of a multi-objective optimization model with minimizing both molecular formation energy and adsorption Gibbs free energy as dual optimization objectives, and the determination of decision variables and decision space, include: The sets of cobalt-site and phosphorus-site substitution elements in the cobalt phosphide catalyst diatomic substitution modification system were determined and encoded by integers to form two-dimensional decision variables. The value range of each dimension of the decision variables was determined by the upper limit of the encoding of the corresponding substitution element set. The molecular formation energy was calculated by the total energy of the doping system, the energy of cobalt element, the energy of cobalt phosphide, and the energy of bulk phase transition metal atoms. The adsorption Gibbs free energy was calculated by the energy of adsorbed nitrate ions, the energy of the catalyst substrate, the energy of gaseous nitrate molecules, and the energy of hydrogen molecules. All of the above energies were obtained by simulation software.

3. The expensive multimodal multi-objective optimization method for molecular structure design according to claim 1, characterized in that, The partition search strategy includes: Select key dimensions from all dimensions of the decision space, and evenly divide the value range of each selected key dimension into a preset number of intervals. By combining the intervals of each key dimension, the entire decision space is deconstructed into multiple non-overlapping subspaces.

4. The expensive multimodal multi-objective optimization method for molecular structure design according to claim 3, characterized in that, Two dimensions were selected as key dimensions, and the value range of each dimension was evenly divided into two intervals to form four subspaces.

5. The expensive multimodal multi-objective optimization method for molecular structure design according to claim 1, characterized in that, The Latin hypercube sampling adopts an integer sampling method, generating a preset number of individuals in each subspace. The sampling process includes: uniformly dividing the integer value range of each dimension into equally spaced integer intervals equal to the number of samples; independently generating random permutations for each dimension to ensure that each interval is selected only once; randomly selecting an integer value in each integer interval; and forming a complete integer sample set by random permutations and combinations of each dimension.

6. The expensive multimodal multi-objective optimization method for molecular structure design according to claim 5, characterized in that, The true fitness assessment includes: decoding individuals in the initial population to obtain corresponding molecular structure combinations, calling simulation software to calculate the formation energy and adsorption Gibbs free energy of each molecular structure combination, and storing the calculation results and corresponding codes in the data archive of each subspace; the surrogate model is a Kriging model, and two Kriging models are trained in each subspace to predict the formation energy and adsorption Gibbs free energy, respectively.

7. The expensive multimodal multi-objective optimization method for molecular structure design according to claim 1, characterized in that, Step 3 includes: Step 3-1: In each subspace, using the current population as the parent generation, a new generation of offspring individuals is generated using a cluster-based multimodal multi-objective differential evolution algorithm. The fitness value of the offspring individuals is predicted by the surrogate model of the corresponding subspace. Step 3-2: Based on the filling sampling criterion based on adaptive resource allocation, determine whether the proxy model of each subspace needs to be updated. If so, combine the historical data archive of each subspace with the current population, select multiple excellent individuals through non-dominated sorting and special crowding distance for real fitness evaluation, store the evaluated individuals in the corresponding data archive and retrain the proxy model. Step 3-3: Repeat steps 3-1 and 3-2 until the preset iteration termination condition is reached, and output the updated data archive for each subspace. Steps 3-4: Merge the data archives of all subspaces to obtain the global archive. Perform non-dominated sorting on all individuals in the global archive, filter out the set of non-dominated individuals, and output the final Pareto front and multimodal Pareto solution set.

8. The expensive multimodal multi-objective optimization method for molecular structure design according to claim 7, characterized in that, The clustering-based multimodal multi-objective differential evolution algorithm in step 3-1 includes: performing non-dominated sorting on the current population to obtain the non-dominated levels of each solution; adaptively determining the number of clusters according to the number of individuals in each level and performing hierarchical clustering, calculating the crowding distance in the decision space and the crowding distance in the target space using only individuals in the same cluster as neighbors, and fusing them to obtain the non-dominated and crowding distance indices; sorting the population according to the rule of prioritizing non-dominated levels and descending the indices within the same level; selecting sparsely distributed individuals from the better non-dominated levels as candidate models, and matching a dedicated learning model for each individual through a roulette wheel method with decision space distance weighting; generating offspring individuals based on differential evolution mutation and crossover operations, merging the parent and offspring generations and performing sorting again, and selecting the best individuals to form the next generation parent population.

9. The expensive multimodal multi-objective optimization method for molecular structure design according to claim 8, characterized in that, The filling sampling criterion for adaptive resource allocation in step 3-2 includes: initializing the selection probability of each subspace; merging the data archives of all subspaces into a global archive and performing non-dominated sorting, and counting the total number of non-dominated individuals globally; when the selection probability of a subspace is greater than a randomly generated probability threshold, performing a filling sampling operation on that subspace, the filling sampling operation including merging the data archive of the current subspace with the current parent population, performing deduplication and non-dominated sorting, using a special crowding distance to perform a secondary sorting of individuals with the same non-dominated level, removing individuals already in the data archive, and selecting those ranked higher. Assume a number of individuals are used to evaluate the true fitness of the surrogate model and update it. After performing fill sampling, the number of individuals corresponding to the local Pareto fronts of each subspace in the global Pareto front is counted as the number of contributing individuals in each subspace. The selection probability of each subspace is updated based on the number of contributing individuals in each subspace and the total number of non-dominated individuals in the global dataset: if there are non-dominated individuals in the local dataset of a subspace, the selection probability of that subspace is increased by the product of the constant selection probability and the proportion of the number of contributing individuals in that subspace to the total number of non-dominated individuals in the global dataset; otherwise, the selection probability of that subspace is decreased by the constant selection probability.

10. The expensive multimodal multi-objective optimization method for molecular structure design according to claim 1, characterized in that, After outputting the multimodal Pareto solution set and Pareto front in step 3, the method further includes decoding the encoded individuals in the solution set to obtain the molecular structure substitution element combination corresponding to each non-dominated solution.