CN111611748B - Data-driven material reverse design method and system - Google Patents
Data-driven material reverse design method and system Download PDFInfo
- Publication number
- CN111611748B CN111611748B CN202010446964.4A CN202010446964A CN111611748B CN 111611748 B CN111611748 B CN 111611748B CN 202010446964 A CN202010446964 A CN 202010446964A CN 111611748 B CN111611748 B CN 111611748B
- Authority
- CN
- China
- Prior art keywords
- chromosome
- data
- coding mode
- chromosomes
- design parameters
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/04—Constraint-based CAD
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physiology (AREA)
- Genetics & Genomics (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a data-driven material reverse design method and a data-driven material reverse design system, wherein the method comprises the following steps: firstly, acquiring material data of a molded material sample; then, according to the material data, respectively selecting a machine learning model of the corresponding relation between each performance parameter and the design parameter; searching and selecting parameters of each machine learning model by adopting a cross-validation method based on material data to obtain a modified machine learning model with the corresponding relation between each performance parameter and the design parameter, wherein the modified machine learning model is used as an adaptability function of each performance parameter of a genetic algorithm; finally, the genetic algorithm is adopted for optimization solution, the non-dominant sorting and crowding distance sorting modes are adopted in the genetic algorithm, multi-objective optimization solution is realized, penalty factors are not required to be introduced, the technical defect that the search space in the traditional grid search cannot be exhausted, the design result is inaccurate due to the penalty factor introduction process is overcome, and the comprehensiveness and the accuracy of material reverse design are improved.
Description
Technical Field
The invention relates to the technical field of new material design, in particular to a data-driven material reverse design method and system.
Background
There is a class of material reverse design requirements in material science research, which hope to recommend the next experimental point (the next experimental point is the component or experimental condition which can meet the expected dependent variable by reverse thrust) according to the expected dependent variable, such as the independent variable of material performance reverse thrust component or experimental condition, and then verify whether the design is correct by adopting an experimental mode. Specifically, the reverse material design is to reversely design the components, the structure, the corresponding preparation process and the like of the material from the performance requirement of the material. After the reverse design is completed, whether the design is accurate or not is verified in an experimental mode. The method has the advantages of shortening the period of material research and development and reducing the cost of research and development. Because the traditional material research and development means is to continuously try component combination and preparation process combination to research and develop the material by a trial and error method until the performance index of the material is met, the time and the labor are wasted.
In the past, the next experimental point is generated by expert knowledge, by adopting a trial-and-error method, or by adopting a method such as an orthogonal design method. With the aid of machine learning, there are alternatives in which the more traditional approach of applying machine learning is to use a grid search method to exhaust all possible values of the independent variables.
However, the biggest drawback of the grid search method is the huge amount of computation, which requires a lot of computation resources and computation time. Meanwhile, if the search space is too large, all the possibilities cannot be exhausted, and the defects cause the use field Jing Shouxian to be not widely used.
In addition, a common class of problems in material data machine learning is that the component proportions are back-deduced from the expected material properties, which may be multiple, and the multiple properties may be directly contradictory, and in the material component formulation, besides the main components, there are usually several element dopings, each with an indefinite content but a fixed total amount. Thus, it is a multi-objective optimization problem with linear constraints. For multi-objective optimization problems with linear constraints, a penalty function method is commonly used, which adds penalty terms after the objective function to convert the constrained problem into an unconstrained problem. However, the penalty function method has two drawbacks: first, the search results are only approximately constrained, which is unacceptable in some scenarios. Secondly, the penalty factor is difficult to select, the Hessian matrix of the unconstrained optimization problem is ill-conditioned due to the fact that the penalty factor is too large, solving of the unconstrained optimization problem is affected, and the constraint effect is poor due to the fact that the penalty factor is too small.
How to overcome the technical defects that the search space in the existing machine learning method cannot be exhausted, and the design result is inaccurate due to the introduction of punishment factors in the multi-objective solving process, and the improvement of the comprehensiveness and the accuracy of the reverse material design becomes a technical problem to be solved urgently.
Disclosure of Invention
The invention aims to provide a data-driven material reverse design method and a data-driven material reverse design system, which are used for overcoming the technical defects that the search space in the existing machine learning method cannot be exhausted, and the design result is inaccurate due to the introduction of penalty factors in the multi-objective solving process, and improving the comprehensiveness and the accuracy of material reverse design.
In order to achieve the above object, the present invention provides the following solutions:
a method of reverse engineering a data driven material, the method comprising the steps of:
acquiring material data of a molded material sample; the material data includes performance parameters and design parameters;
according to the material data, respectively selecting a machine learning model of the corresponding relation between each performance parameter and the design parameter;
searching and selecting parameters of each machine learning model based on the material data by adopting a cross validation method to obtain a modified machine learning model of the corresponding relation between each performance parameter and the design parameter, wherein the modified machine learning model is used as an adaptability function of each performance parameter of a genetic algorithm;
Based on the fitness function of each performance parameter, a genetic algorithm based on non-dominant ranking and crowding distance ranking is adopted to determine design parameters of the material which simultaneously meet a plurality of performance indexes.
Optionally, acquiring material data of the molded material sample specifically includes:
collecting material data of the molded material sample from a data source;
data cleaning is carried out on the material data, and cleaned material data are obtained;
and carrying out data transformation on the cleaned material data to obtain transformed material data.
Optionally, the determining, based on the fitness function of each performance parameter, a genetic algorithm based on non-dominant ranking and crowding distance ranking to determine design parameters of a material that simultaneously meets multiple performance indexes specifically includes:
selecting a coding mode according to the type of the design parameters of the material to be determined; the coding mode is a combination optimization coding mode, a real number coding mode or a binary coding mode;
the chromosome of the genetic algorithm is encoded by adopting the encoding mode, so that an initial parent chromosome is obtained;
performing crossover and mutation operations on the parent chromosomes to generate offspring chromosomes;
non-dominant sorting is carried out on the offspring chromosomes and the parent chromosomes according to the order of the fitness function values of all the performance parameters from good to bad; sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small, deleting the chromosome with the smallest crowding distance, repeatedly executing the steps of sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small, deleting the chromosome with the smallest crowding distance until the number of the reserved chromosomes is equal to the number of the chromosomes of a preset proportion; the chromosome is a offspring chromosome or a parent chromosome;
Taking the chromosome with the reserved preset proportion as a parent chromosome, returning to the step of generating a child chromosome by adopting a crossover algorithm and a mutation algorithm, and performing the next iteration until the iteration number reaches an iteration number threshold;
decoding the chromosome with the forefront sequence in the reserved chromosomes by adopting a decoding mode corresponding to the coding mode to obtain the design parameters of the material; the decoding mode is a combination optimization decoding mode, a real number decoding mode or a binary decoding mode.
Optionally, the selecting a coding mode according to the type of the design parameter of the material to be determined specifically includes:
when the design parameters are design parameters with linear constraint, selecting a combined optimized coding mode;
and when the design parameters are design parameters without linear constraint, selecting a real number coding mode or a binary coding mode.
Optionally, when the coding mode is a combination optimization coding mode, the coding mode is used for coding the chromosome of the genetic algorithm to obtain an initial parent chromosome, which specifically includes:
determining the length of the chromosome as the inverse of the precision of the composition of the material; the length of the chromosome is the number of elements contained in the chromosome;
Randomly generating a coded value for each element of the chromosome over the [1, m ] interval; wherein m represents the number of kinds of components of the material, and when the encoded value of an element of the chromosome is i, it represents that the element corresponds to the i-th component.
Optionally, when the decoding mode is a combination optimization decoding mode, the decoding method is used for decoding the chromosome with the optimal fitness function value to obtain the design parameters of the material, and specifically includes:
according to the coding value of each element in the chromosome with optimal fitness function value, the formula is utilized
Calculating the number of elements in the chromosome with the optimal fitness function value corresponding to each component; wherein n is i The number of elements in the chromosome with the optimal fitness function value corresponding to the ith component is represented, L represents the length of the chromosome, sign i (j) A sign function representing the i-th component, +.>
vj represents the encoded value of the j-th element in the chromosome;
according to the number of elements in the chromosome with optimal fitness function value corresponding to each component, using a formula
Calculating the content of each component; where xi represents the content of the i-th component, constant represents a linear constraint value, that is: sigma x i ＝constanct。
Optionally, the sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small specifically includes:
Using crowding distance calculation formulas
Calculating a crowding distance for each chromosome at the same non-dominant level;
wherein, crwding (x j ) Represents the jthNon-dominant grade chromosome x j Is the crowding distance, k represents chromosome x j The number of performance parameters to be met, f i (x j ) Representation of chromosome x j The fitness function value f of the ith performance parameter of (a) i ± (x j ) Fitness function value of the ith performance parameter in all chromosomes for the jth non-dominant class and all non-dominant classes superior to the jth non-dominant class is superior to and closest to chromosome x j Fitness function value f of the ith performance parameter of the chromosome i (max) and f i (min) represents the maximum and minimum values of fitness function values of the ith performance parameter of the jth non-dominant class and all chromosomes of all non-dominant classes superior to the jth non-dominant class, respectively; alpha i A weight representing an ith performance parameter;
according to the calculated crowding distance of each chromosome in the same non-dominant class, the chromosomes in the same non-dominant class are sorted according to the order of the crowding distances from the big to the small.
A data driven material reverse engineering system, the system comprising the steps of:
The material data acquisition module is used for acquiring material data of the molded material sample; the material data includes performance parameters and design parameters;
the machine learning model selection module is used for respectively selecting a machine learning model of the corresponding relation between each performance parameter and the design parameter;
the machine learning model training module is used for searching and selecting the parameters of each machine learning model based on the material data by adopting a cross validation method to obtain a modified machine learning model of the corresponding relation between each performance parameter and the design parameter, and the modified machine learning model is used as an adaptability function of each performance parameter of a genetic algorithm;
and the multi-objective optimization solving module is used for determining design parameters of the material simultaneously meeting a plurality of performance indexes by adopting a genetic algorithm based on non-dominant sorting and crowding distance sorting based on the fitness function of each performance parameter.
Optionally, the material data acquisition module specifically includes:
a material data acquisition sub-module for collecting material data of the molded material sample from a data source;
the data cleaning submodule is used for cleaning the data of the material data to obtain cleaned material data;
And the data conversion sub-module is used for carrying out data conversion on the cleaned material data to obtain converted material data.
Optionally, the multi-objective optimization solving module specifically includes:
the coding mode selecting sub-module is used for selecting a coding mode according to the type of the design parameters of the material to be determined; the coding mode is a combination optimization coding mode, a real number coding mode or a binary coding mode;
the coding submodule is used for coding the chromosome of the genetic algorithm by adopting the coding mode to obtain an initial parent chromosome;
the crossover mutation operation sub-module is used for carrying out crossover and mutation operation on the parent chromosome to generate a child chromosome;
the sequencing submodule is used for non-dominant sequencing of the child chromosome and the parent chromosome according to the sequence from the good to the bad of the fitness function values of all the performance parameters; sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small, deleting the chromosome with the smallest crowding distance, repeatedly executing the steps of sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small, deleting the chromosome with the smallest crowding distance until the number of the reserved chromosomes is equal to the number of the chromosomes of a preset proportion; the chromosome is a offspring chromosome or a parent chromosome;
Returning to an iteration submodule, wherein the iteration submodule is used for taking the reserved chromosome with the preset proportion as a parent chromosome, and returning to the step of generating a child chromosome by adopting a crossover algorithm and a mutation algorithm to carry out the next iteration until the iteration times reach an iteration times threshold;
the decoding submodule is used for decoding the chromosome with the forefront sequence in the reserved chromosomes by adopting a decoding mode corresponding to the encoding mode to obtain the design parameters of the material; the decoding mode is a combination optimization decoding mode, a real number decoding mode or a binary decoding mode.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
the invention discloses a data-driven material reverse design method and a system, and the method comprises the following steps: firstly, acquiring material data of a molded material sample; according to the material data, a plurality of commonly used machine learning models are used, and each machine learning model with the best fitting effect of the performance parameters and the design parameters is selected through a comparison test; searching and selecting parameters of each machine learning model by adopting a cross verification method based on the material data, so that the machine learning model with the parameters adjusted has a better fitting effect on each performance parameter and the design parameter, and the machine learning model with the parameters adjusted is used as an adaptability function of each performance parameter of a genetic algorithm; based on the fitness function of each performance parameter, a genetic algorithm based on non-dominant ranking and crowding distance ranking is adopted to determine design parameters of the material which simultaneously meet a plurality of performance indexes. The invention adopts genetic algorithm to carry out optimization solution, realizes multi-objective optimization solution by adopting non-dominant sorting and crowding distance sorting modes in the genetic algorithm, does not need to introduce penalty factors, overcomes the technical defects that the search space cannot be exhausted in the traditional grid search and the design result is inaccurate due to the penalty factors introducing process, and improves the comprehensiveness and the accuracy of material reverse design.
Meanwhile, for the multi-objective optimization problem with linear constraint, the method is based on a combined optimization coding and decoding mode in a genetic algorithm, and chromosomes which do not meet the linear constraint are avoided in the crossing and mutation operation of the genetic algorithm.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings that are needed in the embodiments will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a data driven material reverse design method provided by the invention;
FIG. 2 is a schematic diagram of a data driven material reverse design method provided by the present invention;
FIG. 3 is a flow chart of a genetic algorithm provided by the present invention;
FIG. 4 is a schematic diagram of a crossover operation provided by the present invention;
FIG. 5 is a schematic diagram of a mutation operation according to the present invention;
FIG. 6 is a schematic diagram of an optimal chromosome selection strategy provided by the present invention.
Detailed Description
The invention aims to provide a data-driven material reverse design method and a data-driven material reverse design system, which are used for overcoming the technical defects that the search space in a machine learning method cannot be used for exhausting all possible problems and the design result is inaccurate due to the constraint problem conversion process, and improving the comprehensiveness and the accuracy of material reverse design.
In order that the above-recited objects, features and advantages of the present invention will become more readily apparent, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description.
The invention provides a data-driven material reverse design method and a data-driven material reverse design system with a linear constraint multi-objective optimization method for self-adaptive material design, which aim to solve the problem of linear constraint in material design and recommend a next experimental point for the material design more efficiently and rapidly. The multi-objective optimization of the invention means that the material performance requirements in more than or equal to 2 aspects are required to be simultaneously optimized, however, different material performances are mutually restricted at certain time, and all performance indexes are required to be balanced until the requirements on all performances are met, the method of the invention is a multi-objective optimization method with linear constraint based on a genetic algorithm, the constraint of a linear equation can be accurately met, the searching precision can be dynamically adjusted, namely, the longer the chromosome is, the shorter the change step length of each independent variable is, and the obtained result is more accurate. Moreover, due to the natural parallelism of the genetic algorithm, a plurality of candidate solution sets can be obtained through one-time operation, so that a materialist can better select from recommended next experimental points according to field knowledge.
Specifically, as shown in fig. 1, the method for reverse designing a data-driven material according to the present invention comprises the following steps:
The method specifically comprises the following steps:
and (3) data acquisition: the material data of the molded material sample of the present invention, which is collected from a plurality of possible data sources, may be, but is not limited to: material composition data, preparation process data, material microstructure data, material property data and the like collected by material experiments, material calculation, material batch industrial production, publicly published documents and the like.
Data cleaning: removing the data samples suspected of error abnormality, and removing outlier samples deviating from the integral distribution of the samples; and carrying out correlation analysis (such as chi-square test, correlation coefficient calculation, covariance analysis and the like) on the data, and deleting useless features with redundancy or low correlation.
And (3) data transformation: and carrying out smoothing, normalization, standardization, discretization and other transformation on the data according to the subsequent modeling requirement.
Specifically, a plurality of machine learning models such as random forest, support vector regression and the like are used for fitting modeling on material data, and the machine learning model with the best performance is selected as a basic model by indexes such as R square value, mean square error, classification accuracy and the like.
And step 103, searching and selecting the parameters of each machine learning model based on the material data by adopting a cross-validation method to obtain a modified machine learning model of the corresponding relation between each performance parameter and the design parameter, wherein the modified machine learning model is used as an adaptability function of each performance parameter of a genetic algorithm. The method comprises the steps of respectively searching and selecting parameters of each machine learning model based on the material data by adopting a cross validation method, so that the machine learning model after parameter adjustment is better in performance on indexes such as R square value, mean square error, classification accuracy and the like, and the model after parameter adjustment is used as an adaptability function of each performance parameter of a genetic algorithm. Specifically, a particle swarm algorithm or a combined optimization algorithm may be adopted in the process of searching and selecting parameters in each machine learning model, and the distance of the combined optimization algorithm is described as follows: for example, the whole search interval 0-100 is divided into 0-10,10-20, etc., the search of the large interval is first performed, then the optimal large interval is divided into cells, and so on until the optimal solution is found.
Cross-verifying the machine learning model, adjusting the parameters of the super parameters in the model, and selecting the relatively satisfactory super parameters. Specifically, the built model can be divided into various types according to the needs of the study. For example: if the relation between the material composition, the preparation process and the material surrounding structure and the material performance is to be researched, the model at the moment is a fitting function of the composition, the process, the structure and the performance. If only the relation between the components, the preparation process and the microstructure is to be studied, the model at the moment is the fitting function of the components, the process and the microstructure. Generally, the final properties of the materials need to be studied, so that fitting functions of the model of components, processes, structures and properties are more general.
As shown in fig. 3, step 104 may be implemented by:
Wherein, fitness function selects: taking the machine learning model established in the step 103 as an fitness function; the output of the machine learning model is the target of the study, typically the performance index of the material. For example, the goal of the study is to care about a certain mechanical property of the material, and the output of that model is a certain mechanical property, such as yield strength. The model inputs information such as the composition, preparation process, microstructure and the like of the material.
For the multi-objective optimization problem that the content of each element is variable but the total amount is fixed and the multi-objective optimization problem has obvious linear constraint characteristics, such as material composition formula back-pushing by expected material performance, the following combined optimization coding mode can be used, and the coding mode can be used: the value range of each element of the coding vector V is [1, m ], m is the number of kinds of material components, the length of V is determined according to the percentage content precision of the element, the precision and the length are reciprocal relations, and if the content precision is 0.001 level, the corresponding chromosome length is 1000. This coding scheme is called combinatorial optimization coding.
For other types of materials, such as the design requirements between the preparation process, the structure and the performance, if there is no obvious linear constraint relationship, the conventional coding modes, such as binary coding, real coding, etc., can be used. Binary coding is a process in which each value in the code vector V is a binary number, and each value represents a specific value of a process parameter, and the range of values is between the minimum value and the maximum value that the process parameter can take, and binary coding is commonly used in scenes in which the value is an integer. The real number coding is to make each value in the coding vector V a real number, and the rest is the same as the binary coding. Real number coding is often used for scenes in the range of values that are real. The following description will take as an example design parameters (components) having linear constraints.
crossover operator: and performing crossover operation under a certain probability. Selecting two parent chromosomes, randomly generating a section [ p, q ] so that p is more than or equal to 0, q is less than or equal to V, wherein V is the coding vector in the step 1, V is the length of the coding vector V, p and q are respectively a first random number and a second random number, the values of the p and q are taken from 0 to q, the coding in the coding vector V executes a crossover operator, and the genes of p < q, offspring chromosomes in the intervals of [0, p) and (q, |V| ] are from one chromosome, and the genes of the interval of [ p, q ] are from the other parent chromosome;
mutation operator: and performing mutation operation under a certain probability. Randomly selecting one gene from a parent chromosome, and converting the selected gene into another gene, namely randomly selecting one element in the coding vector V, and changing the element into any integer value in [1, m ];
wherein x is j Representing a solution to rank=j, i.e. chromosome x of the j-th non-dominant class j ，f i ± (x j ) Fitness function value of the ith performance parameter in all chromosomes for the jth non-dominant class and all non-dominant classes superior to the jth non-dominant class is superior to and closest to chromosome x j Fitness function value f of the ith performance parameter of the chromosome i ± (x j ) The meaning represented differs according to the actual problem. For example: when the requirement f i (x j ) The smaller the performance is, the better f i (x j ) All solutions in which the rank value is j and the rank value is less than j are in the objective function f i The value of the upper part is smaller than f i (x j ) Is the maximum value of (2); correspondingly, when the requirement f i (x j ) The larger the performance is, the better f i ± (x j ) All solutions in which the rank value is j and the rank value is less than j are in the objective function f i The value of the upper value is greater than f i (x j ) Minimum value f of (f) i (max) and f i (min) represents that all solutions in rank value j and rank value less than j are in the objective function f i Maximum and minimum values of the values. The crowding distance represents the normalized distance of a solution from a solution that is better than it in each objective function direction, while due to f used i ± (x j )，f i (max) and f i (min) is selected from all solutions in which the rank value is j and the rank value is better than j, and not just at the rank value of j, will causeThe final solution set is focused on the dispersion and non-dominance of the solution on the global, so that the solution set dispersion is more prominent. After calculating the congestion distance, using a round-robin strategy, the smaller the congestion distance, the more similar it behaves in this performance to the solution that it is better than it, the worse the dispersion of the solution, and the better the solution will be deleted. After deleting the solution with the smallest congestion distance, the congestion distance between each solution in the non-dominant level is recalculated, and the above process is iterated until the number of solutions meeting the rank value j and the rank value less than j reaches the population size N set at the time of initialization.
In addition, the congestion distance formula is added with a weight alpha i The method is used for adjusting the bias on the performance of a certain aspect of the material, so that a larger search range can be provided for the performance of the material on the certain aspect of the material through the crowding distance, the solution sets are more dispersed on the performance of the material, and the solution entering the next layer on the same non-supporting stage is more dispersed on the designated material performance through the weighting mode. When specifically setting, if the weight alpha i The smaller the solution is, the more scattered the solution will be in performance in this respect, and the larger the search range of the solution set. If the performance of the plurality of targets is not biased, the weight is set to be 1, and the first 50% of individuals after sorting are reserved to enter the next round of circulation. Repeating the loop from the second step until the algorithm reaches the maximum iteration times;
The process of performing the combination optimization coding is as follows:
step one, determining the length L of a chromosome (code) according to the content precision of the material components, wherein the precision and the length are in reciprocal relation, and if the content precision is 0.001 level, the corresponding chromosome length is 1000.
Generating a chromosome (code) with the length of L, wherein each element in the code is randomly generated, the value range of the element is between [1, m ], m is the number of types of material components, and each element value represents one element.
The process of performing the combinatorial optimization decoding is:
step one, calculating the number n of values of various elements in the chromosome to be decoded 1 ,n 2 ,…,n m 。
Step two, calculating the proportion x of each component represented by the chromosome coding i . The calculation formula is as follows:
defining a sign function: i represents the ith component, j is used to count the jth, V of the encoded vector V j Representing the code value of the j-th element in the code vector V;
at this time, the ratio x of each component can be calculated by the following formula i ：
Wherein constant represents a linear constraint value, namely: sigma x i ＝constanct。
The invention also carries out material experiments on the experimental points obtained in the step 104, evaluates the experimental results, and ends if the experimental results meet the actual requirements, otherwise, newly generated experimental data are added into the original material data, and iterates again.
The specific steps of step 104 of the present invention are described below in connection with specific embodiments:
selecting a coding mode according to the type of the design parameters of the material to be determined, wherein the coding mode specifically comprises the following steps: when the design parameters are design parameters with linear constraint, selecting a combined optimized coding mode; and when the design parameters are design parameters without linear constraint, selecting a real number coding mode or a binary coding mode. The invention is described by taking design parameters-components with linear constraint as examples, and at the moment, the coding mode is a combined optimized coding mode, and the decoding mode corresponding to the coding mode is a combined optimized decoding mode, which is specifically as follows:
Coding the chromosome of the genetic algorithm by adopting a combined optimized coding mode, initializing the chromosome of the genetic algorithm, and obtaining an initial parent chromosome; for example, the population size of the multi-objective genetic algorithm is set to be 50, the iteration number is 500, the fitness function selects a model and super parameters in the modeling process, and the material to be designed has three element components.
Performing crossover and mutation operations on the parent chromosomes to generate offspring chromosomes;
referring to FIG. 4, crossover operations are performed with crossover intervals of [2,4], where the individual chromosomes produced by crossover are in the [0,2 ] and (2, 4] intervals with the genes from one parent chromosome and the [2,4] interval genes from the other parent chromosome.
Referring to fig. 5, a mutation operation is performed such that the mutation position generated at random is 2, and the gene B at the second position is randomly changed to any one of the two elements a and C other than B, and here, the element C is randomly changed.
Non-dominant sorting is carried out on the offspring chromosomes and the parent chromosomes according to the order of the fitness function values of all the performance parameters from good to bad; sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small, deleting the chromosome with the smallest crowding distance, repeatedly executing the steps of sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small, deleting the chromosome with the smallest crowding distance until the number of the reserved chromosomes is equal to the number of the chromosomes of a preset proportion; the chromosome is a offspring chromosome or a parent chromosome; taking the preserved chromosomes with preset proportions as parent chromosomes, returning The next iteration is carried out by adopting a crossover algorithm and a mutation algorithm to generate a child chromosome until the iteration number reaches an iteration number threshold; referring to fig. 6, selection operator and elite strategies are performed, individuals with father and cross variation are selected together, all individuals to be selected are subjected to non-dominant ranking, and the ranking is the forefront. Assuming that the total number of individuals with rank value 1 and individuals with rank value 0 is greater than population size 50 at non-dominant rank value 1, then the individuals at rank value 1 are further crowded distance ordered (where a greater search range for the first performance is desired in terms of performance, so the first performance weight α is determined 1 Set to 0.5 and the rest set to 1, calculate the crowding distance by formula
It is desirable here that all properties are smaller and better, therefore
Taking the current non-dominant level and less than the current non-dominant level, all solutions are in the objective function f i The value of the upper part is smaller than f i (x 1) after calculating the congestion distance of the solutions, deleting the smallest solution, then re-calculating the congestion distance, continuing the iteration until the number of solutions entering the next generation is equal to the population size 50 set in the step one), entering the next generation with a large congestion distance, and selecting individuals entering the next generation through non-dominant ranking and congestion distance ranking.
And decoding the chromosome with the forefront sequence in the reserved chromosomes in a combined optimized decoding mode to obtain the content of the components of the material.
The method for coding the chromosome of the genetic algorithm by adopting the combination optimization coding mode specifically comprises the following steps:
determining the length of the chromosome as the inverse of the precision of the composition of the material; the length of the chromosome is the number of elements contained in the chromosome; for example, the chromosome (code) length L is determined according to the material composition content accuracy, where the content accuracy is 0.2, so the corresponding chromosome length l=1/0.2=5.
Randomly generating a coded value for each element of the chromosome over the [1, m ] interval; wherein m represents the number of kinds of components of the material, and when the encoded value of an element of the chromosome is i, it represents that the element corresponds to the i-th component. For example, a length 5 chromosomal code is generated, where there are three elements a, B, C, and a chromosomal code [1, 2,3] can be generated assuming that the corresponding constituent content percentages are 0.4,0.4,0.2, respectively, where 1,2,3 represent elements a, B, C, respectively.
The method for decoding the chromosome with the optimal fitness function value by adopting the combination optimization decoding mode to obtain the content of the components of the material comprises the following steps:
According to the coding value of each element in the chromosome with optimal fitness function value, the formula is utilized
Calculating the number of elements in the chromosome with the optimal fitness function value corresponding to each component; wherein n is i The number of elements in the chromosome with the optimal fitness function value corresponding to the ith component is represented, L represents the length of the chromosome, sign i (j) A sign function representing the i-th component, +.>
v j A coded value representing a j-th element in the chromosome; according to the number of elements in the chromosome with optimal fitness function value corresponding to each component, using the formula +.>
Calculating the content of each component; wherein x is i Representing the content of the i-th component, constant represents a linear constraint value, that is: sigma x i ＝constanct。
For example, assume that the constraint is x 1 +x 2 +x 3 =3, 3 is total content. Para [1, 2,3]The process of performing the combinatorial optimization decoding is:
calculation of the chromosome to be decodedNumber n of values of various elements 1 ＝2,n 2 ＝2,n 3 =1. Calculating the ratio x of each component represented by the chromosome coding i To obtain x 1 ＝2/5*3＝1.2,x 2 ＝2/5*3＝1.2,x 3 =1/5*3 =0.6, i.e. the three components content is 1.2 units, 0.6 units, respectively.
The invention also provides a data-driven material reverse design system, which comprises the following steps:
The material data acquisition module is used for acquiring material data of the molded material sample; the material data includes performance parameters and design parameters; the material data acquisition module specifically comprises: a material data acquisition sub-module for collecting material data of the molded material sample from a data source; the data cleaning submodule is used for cleaning the data of the material data to obtain cleaned material data; and the data conversion sub-module is used for carrying out data conversion on the cleaned material data to obtain converted material data.
The machine learning model selection module is used for respectively selecting machine learning models of the corresponding relation between each performance parameter and the design parameter, namely, according to the material data, selecting a model with the best fitting effect of the corresponding relation between each performance parameter and the design parameter by using a plurality of common machine learning models through a comparison experiment;
the machine learning model training module is used for searching and selecting the parameters of each machine learning model by adopting a cross validation method based on the material data to obtain a modified machine learning model of the corresponding relation between each performance parameter and the design parameter as an adaptability function of each performance parameter of a genetic algorithm, namely searching and selecting the parameters of each machine learning model by adopting the cross validation method based on the material data to enable the machine learning model with the adjusted parameters to have better fitting effect on each performance parameter and the design parameter, and taking the model after the parameter adjustment as the adaptability function of each performance parameter of the genetic algorithm;
And the multi-objective optimization solving module is used for determining design parameters of the material simultaneously meeting a plurality of performance indexes by adopting a genetic algorithm based on non-dominant sorting and crowding distance sorting based on the fitness function of each performance parameter. The multi-objective optimization solving module specifically comprises: the multi-objective optimization solving module specifically comprises:
the coding mode selecting sub-module is used for selecting a coding mode according to the type of the design parameters of the material to be determined; the coding mode is a combination optimization coding mode, a real number coding mode or a binary coding mode; the coding submodule is used for coding the chromosome of the genetic algorithm by adopting the coding mode to obtain an initial parent chromosome; the crossover mutation operation sub-module is used for carrying out crossover and mutation operation on the parent chromosome to generate a child chromosome; the sequencing submodule is used for non-dominantly sequencing the offspring chromosomes and the parent chromosomes according to the order of the fitness function values of all the performance parameters from good to bad; sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small, deleting the chromosome with the smallest crowding distance, repeatedly executing the steps of sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small, deleting the chromosome with the smallest crowding distance until the number of the reserved chromosomes is equal to the number of the chromosomes of a preset proportion; the chromosome is a offspring chromosome or a parent chromosome; returning to an iteration submodule, wherein the iteration submodule is used for taking the reserved chromosome with the preset proportion as a parent chromosome, and returning to the step of generating a child chromosome by adopting a crossover algorithm and a mutation algorithm to carry out the next iteration until the iteration times reach an iteration times threshold; the decoding submodule is used for decoding the chromosome with the forefront sequence in the reserved chromosomes by adopting a decoding mode corresponding to the encoding mode to obtain the design parameters of the material; the decoding mode is a combination optimization decoding mode, a real number decoding mode or a binary decoding mode.
The invention has the advantages that: the basic model of the invention can select a conventional machine learning model for predicting material performance and structure, and takes the model as a fitness function of a genetic algorithm to guide the algorithm to carry out multi-objective optimization. Secondly, in order to solve the problem of linear constraint in the optimizing process, the chromosomes are encoded in a combined optimizing mode in the population initialization, so that solutions which do not meet the linear constraint are avoided in the crossing and mutation operation of the genetic algorithm, and the practicability of understanding is greatly improved. Meanwhile, the method is a heuristic optimization method, so that the time performance of searching is greatly improved.
In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other.
The principles and embodiments of the present invention have been described herein with reference to specific examples, which are intended to be only illustrative of the methods and concepts underlying the invention, and not all examples are intended to be within the scope of the invention as defined by the appended claims.
Claims (4)
1. A method of reverse engineering a data driven material, the method comprising the steps of:
acquiring material data of a molded material sample; the material data includes performance parameters and design parameters;
according to the material data, respectively selecting a machine learning model of the corresponding relation between each performance parameter and the design parameter;
searching and selecting parameters of each machine learning model based on the material data by adopting a cross validation method to obtain a modified machine learning model of the corresponding relation between each performance parameter and the design parameter, wherein the modified machine learning model is used as an adaptability function of each performance parameter of a genetic algorithm;
determining design parameters of materials simultaneously meeting a plurality of performance indexes by adopting a genetic algorithm based on non-dominant sorting and crowding distance sorting based on the fitness function of each performance parameter;
the fitness function based on each performance parameter adopts a genetic algorithm based on non-dominant sorting and crowding distance sorting to determine design parameters of materials meeting a plurality of performance indexes at the same time, and the method specifically comprises the following steps:
selecting a coding mode according to the type of the design parameters of the material to be determined; the coding mode is a combination optimization coding mode, a real number coding mode or a binary coding mode;
The chromosome of the genetic algorithm is encoded by adopting the encoding mode, so that an initial parent chromosome is obtained;
performing crossover and mutation operations on the parent chromosomes to generate offspring chromosomes;
non-dominant sorting is carried out on the offspring chromosomes and the parent chromosomes according to the order of the fitness function values of all the performance parameters from good to bad; sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small, deleting the chromosome with the smallest crowding distance, repeatedly executing the steps of sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small, deleting the chromosome with the smallest crowding distance until the number of the reserved chromosomes is equal to the number of the chromosomes of a preset proportion; the chromosome is a offspring chromosome or a parent chromosome;
taking the chromosome with the reserved preset proportion as a parent chromosome, returning to the step of generating a child chromosome by adopting a crossover algorithm and a mutation algorithm, and performing the next iteration until the iteration number reaches an iteration number threshold;
decoding the chromosome with the forefront sequence in the reserved chromosomes by adopting a decoding mode corresponding to the coding mode to obtain the design parameters of the material; the decoding mode is a combination optimization decoding mode, a real number decoding mode or a binary decoding mode;
The coding mode is selected according to the type of the design parameters of the material to be determined, and specifically comprises the following steps:
when the design parameters are design parameters with linear constraint, selecting a combined optimized coding mode;
when the design parameters are design parameters without linear constraint, selecting a real number coding mode or a binary coding mode;
when the coding mode is a combination optimization coding mode, the coding mode is adopted to code the chromosome of the genetic algorithm to obtain an initial parent chromosome, and the method specifically comprises the following steps:
determining the length of the chromosome as the inverse of the precision of the composition of the material; the length of the chromosome is the number of elements contained in the chromosome;
randomly generating a coded value for each element of the chromosome over the [1, m ] interval; wherein m represents the number of kinds of components of the material, and when the coding value of an element of the chromosome is i, the element corresponds to the i-th component;
when the decoding mode is a combination optimization decoding mode, decoding a chromosome with the optimal fitness function value by adopting the decoding mode to obtain design parameters of the material, wherein the method specifically comprises the following steps:
according to the coding value of each element in the chromosome with optimal fitness function value, the formula is utilized 
Calculating the number of elements in the chromosome with the optimal fitness function value corresponding to each component; wherein n is i The number of elements in the chromosome with the optimal fitness function value corresponding to the ith component is represented, L represents the length of the chromosome, sign i (j) A sign function representing the i-th component, +.>
vj represents the encoded value of the j-th element in the chromosome;
according to the number of elements in the chromosome with optimal fitness function value corresponding to each component, using a formula
Calculating the content of each component; wherein xi represents the content of the ith component, constant represents linearityConstraint values, namely: sigma x i ＝constanct。
2. The method of claim 1, wherein obtaining material data for a molded material sample, comprises:
collecting material data of the molded material sample from a data source;
data cleaning is carried out on the material data, and cleaned material data are obtained;
and carrying out data transformation on the cleaned material data to obtain transformed material data.
3. A data driven material reverse engineering system, the system comprising the following modules:
the material data acquisition module is used for acquiring material data of the molded material sample; the material data includes performance parameters and design parameters;
The machine learning model selection module is used for respectively selecting a machine learning model of the corresponding relation between each performance parameter and the design parameter;
the machine learning model training module is used for searching and selecting the parameters of each machine learning model based on the material data by adopting a cross validation method to obtain a modified machine learning model of the corresponding relation between each performance parameter and the design parameter, and the modified machine learning model is used as an adaptability function of each performance parameter of a genetic algorithm;
the multi-objective optimization solving module is used for determining design parameters of materials which simultaneously meet a plurality of performance indexes by adopting a genetic algorithm based on non-dominant sorting and crowding distance sorting based on the fitness function of each performance parameter;
the multi-objective optimization solving module specifically comprises:
the coding mode selecting sub-module is used for selecting a coding mode according to the type of the design parameters of the material to be determined; the coding mode is a combination optimization coding mode, a real number coding mode or a binary coding mode;
the coding submodule is used for coding the chromosome of the genetic algorithm by adopting the coding mode to obtain an initial parent chromosome;
The crossover mutation operation sub-module is used for carrying out crossover and mutation operation on the parent chromosome to generate a child chromosome;
the sequencing submodule is used for non-dominant sequencing of the child chromosome and the parent chromosome according to the sequence from the good to the bad of the fitness function values of all the performance parameters; sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small, deleting the chromosome with the smallest crowding distance, repeatedly executing the steps of sorting the chromosomes in the same non-dominant class according to the order of the crowding distance from large to small, deleting the chromosome with the smallest crowding distance until the number of the reserved chromosomes is equal to the number of the chromosomes of a preset proportion; the chromosome is a offspring chromosome or a parent chromosome;
returning to an iteration submodule, wherein the iteration submodule is used for taking the reserved chromosome with the preset proportion as a parent chromosome, and returning to the step of generating a child chromosome by adopting a crossover algorithm and a mutation algorithm to carry out the next iteration until the iteration times reach an iteration times threshold;
the decoding submodule is used for decoding the chromosome with the forefront sequence in the reserved chromosomes by adopting a decoding mode corresponding to the encoding mode to obtain the design parameters of the material; the decoding mode is a combination optimization decoding mode, a real number decoding mode or a binary decoding mode;
The coding mode is selected according to the type of the design parameters of the material to be determined, and specifically comprises the following steps:
when the design parameters are design parameters with linear constraint, selecting a combined optimized coding mode;
when the design parameters are design parameters without linear constraint, selecting a real number coding mode or a binary coding mode;
when the coding mode is a combination optimization coding mode, the coding mode is adopted to code the chromosome of the genetic algorithm to obtain an initial parent chromosome, and the method specifically comprises the following steps:
determining the length of the chromosome as the inverse of the precision of the composition of the material; the length of the chromosome is the number of elements contained in the chromosome;
randomly generating a coded value for each element of the chromosome over the [1, m ] interval; wherein m represents the number of kinds of components of the material, and when the coding value of an element of the chromosome is i, the element corresponds to the i-th component;
when the decoding mode is a combination optimization decoding mode, decoding a chromosome with the optimal fitness function value by adopting the decoding mode to obtain design parameters of the material, wherein the method specifically comprises the following steps:
according to the coding value of each element in the chromosome with optimal fitness function value, the formula is utilized 
Calculating the number of elements in the chromosome with the optimal fitness function value corresponding to each component; wherein n is i The number of elements in the chromosome with the optimal fitness function value corresponding to the ith component is represented, L represents the length of the chromosome, sign i (j) A sign function representing the i-th component, +.>
vj represents the encoded value of the j-th element in the chromosome;
according to the number of elements in the chromosome with optimal fitness function value corresponding to each component, using a formula
Calculating the content of each component; where xi represents the content of the i-th component, constant represents a linear constraint value, that is: sigma x i ＝constanct。
4. The data driven material reverse engineering system of claim 3 wherein the material data acquisition module comprises in particular:
a material data acquisition sub-module for collecting material data of the molded material sample from a data source;
the data cleaning submodule is used for cleaning the data of the material data to obtain cleaned material data;
and the data conversion sub-module is used for carrying out data conversion on the cleaned material data to obtain converted material data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010446964.4A CN111611748B (en) | 2020-05-25 | 2020-05-25 | Data-driven material reverse design method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010446964.4A CN111611748B (en) | 2020-05-25 | 2020-05-25 | Data-driven material reverse design method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111611748A CN111611748A (en) | 2020-09-01 |
| CN111611748B true CN111611748B (en) | 2023-05-26 |
Family
ID=72205160
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010446964.4A Active CN111611748B (en) | 2020-05-25 | 2020-05-25 | Data-driven material reverse design method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111611748B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112488871A (en) * | 2020-10-23 | 2021-03-12 | 广西电网有限责任公司电力科学研究院 | Method and system for eliminating redundant data of original input features of power grid |
| JP7088399B1 (en) | 2021-12-17 | 2022-06-21 | Dic株式会社 | Novolac-type phenolic resin search method, information processing device, and program |
| CN114822698B (en) * | 2022-06-21 | 2022-09-13 | 华中农业大学 | Knowledge reasoning-based biological large sample data set analysis method and system |
| US11837333B1 (en) | 2022-12-20 | 2023-12-05 | Dow Global Technologies Llc | Simulation guided inverse design for material formulations |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102817088A (en) * | 2012-08-09 | 2012-12-12 | 东华大学 | Fusant conveying process of fusant directly-spun terylene filament |
| WO2016165392A1 (en) * | 2015-04-17 | 2016-10-20 | 华南理工大学 | Genetic algorithm-based cloud computing resource scheduling method |
| CN108880663A (en) * | 2018-07-20 | 2018-11-23 | 大连大学 | Incorporate network resource allocation method based on improved adaptive GA-IAGA |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10474953B2 (en) * | 2016-09-19 | 2019-11-12 | The Aerospace Corporation | Systems and methods for multi-objective optimizations with decision variable perturbations |
-
2020
- 2020-05-25 CN CN202010446964.4A patent/CN111611748B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102817088A (en) * | 2012-08-09 | 2012-12-12 | 东华大学 | Fusant conveying process of fusant directly-spun terylene filament |
| WO2016165392A1 (en) * | 2015-04-17 | 2016-10-20 | 华南理工大学 | Genetic algorithm-based cloud computing resource scheduling method |
| CN108880663A (en) * | 2018-07-20 | 2018-11-23 | 大连大学 | Incorporate network resource allocation method based on improved adaptive GA-IAGA |
Non-Patent Citations (1)
| Title |
|---|
| 基于RQGA和非支配排序的多目标混沌量子遗传算法;王瑞琪;李珂;张承慧;裴文卉;;电机与控制学报;第16卷(第04期);91-99 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111611748A (en) | 2020-09-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111611748B (en) | Data-driven material reverse design method and system | |
| CN108334949B (en) | Image classifier construction method based on optimized deep convolutional neural network structure fast evolution | |
| CN110046378B (en) | Selective hierarchical integration Gaussian process regression soft measurement modeling method based on evolutionary multi-objective optimization | |
| Köksalan et al. | An interactive territory defining evolutionary algorithm: iTDEA | |
| CN106647262A (en) | Differential evolution method facing agile satellite multi-object task planning | |
| US20040167721A1 (en) | Optimal fitting parameter determining method and device, and optimal fitting parameter determining program | |
| CN111986737B (en) | Durable concrete mixing proportion optimization method based on RF-NSGA-II | |
| García-Martínez et al. | Role differentiation and malleable mating for differential evolution: an analysis on large-scale optimisation | |
| CN112561039A (en) | Improved search method of evolutionary neural network architecture based on hyper-network | |
| CN115481727A (en) | Intention recognition neural network generation and optimization method based on evolutionary computation | |
| CN114004153A (en) | Penetration depth prediction method based on multi-source data fusion | |
| CN114792126A (en) | Convolutional neural network design method based on genetic algorithm | |
| Pan et al. | Neural architecture search based on evolutionary algorithms with fitness approximation | |
| CN114241267A (en) | Structural entropy sampling-based multi-target architecture search osteoporosis image identification method | |
| Song et al. | Importance weighted expectation-maximization for protein sequence design | |
| CN116010291A (en) | Multipath coverage test method based on equalization optimization theory and gray prediction model | |
| CN117149615A (en) | Method and corresponding device for generating test case execution path | |
| CN115620046A (en) | Multi-target neural architecture searching method based on semi-supervised performance predictor | |
| CN116611504A (en) | Neural architecture searching method based on evolution | |
| CN114117917B (en) | Multi-objective optimization ship magnetic dipole array modeling method | |
| CN109934344A (en) | A kind of multiple target Estimation of Distribution Algorithm of improved rule-based model | |
| El-Dahshan et al. | Genetic programing modeling for nucleus–nucleus collisions | |
| CN115661546A (en) | Multi-objective optimization classification method based on feature selection and classifier joint design | |
| KR102103578B1 (en) | Method for Selecting Asset Portfolio | |
| CN110162704B (en) | Multi-scale key user extraction method based on multi-factor genetic algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |