Martensite component prediction method and device

Abstract

The invention provides a martensite component prediction method and device. The method comprises the following steps: training a martensite performance prediction model based on a random forest algorithm; randomly generating a plurality of component vectors, and performing performance prediction on each component vector by using the martensite performance prediction model; sorting the plurality of component vectors according to the performance prediction result through a non-dominated sorting algorithm to determine an excellent component vector; generating a new component vector according to the excellent component vector and a genetic algorithm; and iteratively executing the performance prediction and subsequent steps according to the excellent component vector and the new component vector until a performance prediction result meets a preset value, and outputting a corresponding component vector. According to the method, the component composition of martensite can be quickly determined according to the performance demand, and the probability of generating the component meeting the performance demand can be greatly improved in a mode of randomly generating the primary component, screening the excellent component through the non-dominated sorting algorithm and generating the filial generation according to the excellent component by utilizing the genetic algorithm so as to improve the prediction speed.

Description

Technical Field

[0001] The invention relates to the field of material composition prediction, and in particular to a method and device for predicting martensite composition. Background Art

[0002] At present, there are the following difficulties in the composition design of martensitic steel: the composition of martensitic steel is complex, there are complex synergies between different elements, the influence of elements on material properties is not a simple linear superposition, the combination of these elements will affect the microstructure and performance of the material, and this synergy is often difficult to predict and control; secondly, there are conflicts between properties, for example, increasing strength often reduces plasticity, and increasing hardness may affect corrosion resistance. Application scenarios require martensitic steel to have multiple properties, but in the actual design process, it is necessary to constantly weigh and choose between different properties. For example, in tool manufacturing, martensitic steel needs to have high hardness to keep cutting sharp, and also needs a certain toughness to prevent brittle fracture. In bearing manufacturing, materials not only need high strength and high hardness, but also fatigue resistance and corrosion resistance to ensure a long service life. Faced with the above-mentioned complex game problem of "many to many", the traditional material research method based on trial and error is inefficient and has a long cycle, and can no longer meet the urgent needs of industry for the speed of new material research and development. Summary of the invention

[0003] In order to be able to quickly determine the component composition of martensite according to performance requirements, in one aspect of the present invention, a martensite composition prediction method is proposed, the method comprising: training a martensite performance prediction model based on a random forest algorithm; randomly generating multiple component vectors, and using the martensite performance prediction model to perform performance prediction on each of the component vectors; sorting the multiple component vectors according to the performance prediction results by a non-dominated sorting algorithm to determine an excellent component vector; generating a new component vector based on the excellent component vector and a genetic algorithm; iteratively executing the above-mentioned performance prediction and its subsequent steps based on the excellent component vector and the new component vector until the performance prediction result meets the preset value and outputs the corresponding component vector.

[0004] In one or more implementations, a martensite performance prediction model is trained based on a random forest algorithm, including: constructing a random forest model by configuring the number of decision trees, the maximum depth, the minimum number of sample splits, the minimum number of sample leaf nodes, and the maximum number of features; and training the random forest model using training samples composed of known martensite components and their performance parameters to obtain a martensite performance prediction model.

[0005] In one or more implementations, the martensite composition prediction method of the present invention further includes: adding a martensite transformation temperature prediction function to the martensite property prediction model to predict the martensite transformation temperature.

[0006] In one or more implementations, the martensitic transformation temperature prediction function comprises:

[0007] in, y 3 ( x i ) is the martensitic transformation temperature, () is the mass percentage of the corresponding element.

[0008] In one or more implementations, the multiple component vectors are sorted according to the performance prediction results using a non-dominated sorting algorithm to determine an excellent component vector, including: determining the dominance between any two component vectors according to a preset algorithm; using a triangulation method to determine the congestion of each component vector with the same dominance; sorting the component vectors in order from high to low in terms of the congestion, and determining the component vector with a higher ranking as an excellent component vector.

[0009] In one or more implementations, determining the dominance between any two component vectors according to a preset algorithm includes: if any two component vectors satisfy the following conditions:

[0010] in, x i and x j are any two component vectors, where k,m∈{1,2,3}, then x i Dominate x j ; Determine the dominance degree of each component vector according to the dominance relationship between the component vectors.

[0011] In one or more implementations, triangulation is used for component vectors with the same dominance to determine the crowding of each component vector, including: triangulating each component vector point through the triangulation method to calculate the area of ​​the triangle that encloses the component vector point; and determining the crowding of the corresponding component vector based on the triangle area.

[0012] In one or more implementations, a new component vector is generated based on the excellent component vector and a genetic algorithm, including: generating multiple offspring component vectors by simulating binary crossover for the excellent component vector; introducing a small perturbation in each of the offspring component vectors to perform polynomial mutation to generate multiple new component vectors.

[0013] In one or more implementations, the martensite composition prediction method of the present invention further includes: in response to reaching the prediction iteration number and the performance prediction result still not meeting the preset value, randomly regenerating a plurality of composition vectors.

[0014] In the second aspect of the present invention, a martensitic composition prediction device is proposed, comprising: a composition generation module, configured to generate a plurality of component vectors through a preset random forest model; a performance prediction module, configured to perform performance prediction based on the component vectors through a preset martensitic performance prediction model; a performance sorting module, configured to sort the plurality of component vectors according to the performance prediction results through a non-dominated sorting algorithm to determine an excellent component vector; a performance inheritance module, configured to perform genetic operations based on the excellent component vectors to generate a new component vector; and an iterative decision module, configured to iteratively execute the above-mentioned performance prediction and its subsequent steps based on the excellent component vector and the new component vector until the performance prediction result meets the preset value and outputs the corresponding component vector.

[0015] The beneficial effects of the present invention include: the present invention can quickly determine the composition of martensite according to performance requirements, and by randomly generating primary components, screening excellent components by a non-dominated sorting algorithm, and generating offspring based on excellent components using a genetic algorithm, the probability of generating components that meet performance requirements can be greatly improved to improve the prediction speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other embodiments can be obtained based on these drawings without paying creative work.

[0017] Figure 1 A flowchart of a method for predicting martensite composition according to an embodiment of the present invention; Figure 2 A complete workflow diagram of a method for predicting martensite composition according to an embodiment of the present invention; Figure 3 It is a comparison diagram of component vector distribution before and after iteration based on the method of the present invention; Figure 4 Schematic diagram of a martensite composition prediction device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the embodiments of the present invention are further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0019] It should be noted that all expressions using "first" and "second" in the embodiments of the present invention are for distinguishing two non-identical entities with the same name or non-identical parameters. It can be seen that "first" and "second" are only for the convenience of expression and should not be understood as limitations on the embodiments of the present invention. The subsequent embodiments will not explain this one by one.

[0020] In order to quickly determine the composition of martensite according to performance requirements, in one embodiment of the present invention, a martensite composition prediction method is proposed, such as Figure 1 As shown, it includes: step S1, training a martensite property prediction model based on a random forest algorithm; wherein the random forest algorithm is a powerful and flexible machine learning algorithm, which is commonly used for classification and regression tasks, and is an integrated learning method of the random forest algorithm, which improves the prediction performance and generalization ability of the model by constructing multiple decision trees and integrating them together.

[0021] In one embodiment, a martensitic performance prediction model is trained based on a random forest algorithm, including: constructing a random forest model by configuring the number of decision trees, the maximum depth, the minimum number of sample splits, the minimum number of sample leaf nodes, and the maximum number of features; using known martensitic components and their performance parameters to form training samples to train the random forest model to obtain a martensitic performance prediction model. Among them, the process of obtaining training samples includes: forming component-performance parameters from known martensitic components and their performance parameters; preprocessing the component-performance parameters, including missing value processing, data format unification, and vectorization processing, so as to obtain a sample set; wherein the missing value processing includes deleting missing values, filling missing values, and marking missing values. The process of training a random forest model using training samples includes: dividing the training samples into a training set and a test set in a ratio of 2:1; iteratively training the random forest model using the training set, and judging the performance prediction accuracy of the random forest model using the test set; outputting the random forest model when the performance prediction accuracy of the random forest model reaches a preset value.

[0022] In one embodiment, the martensite performance prediction model trained based on the random forest algorithm is mainly used to predict the two performance parameters of the tensile strength and elongation of the martensite according to the input composition vector. For the performance of the martensite transformation temperature, it is necessary to add a martensite transformation temperature prediction function to the model for prediction, wherein the martensite transformation temperature prediction function is as follows:

[0023] in, y 3 ( x i ) is the component vector x i The martensitic transformation temperature, () is the mass percentage of the corresponding element.

[0024] Step S2, randomly generate multiple component vectors, and use the martensite performance prediction model to predict the performance of each component vector; wherein the random generation algorithm randomly selects elements from the known martensite element composition, and randomly determines the mass percentage of the elements in the component composition. In addition, the martensite performance prediction model in this step is a model with a martensite transformation temperature prediction function added, and its performance prediction results include three performance parameters of martensite: tensile strength, elongation, and transformation temperature.

[0025] In an optional embodiment, the method of randomly generating multiple component vectors includes: sorting the known martensitic constituent elements, and randomly generating a set of random numbers using a random number generation algorithm, and selecting elements with corresponding serial numbers according to the random numbers and determining their mass percentages to generate a component vector. For example, if a set of randomly generated random numbers is 356189, then an element ranked as 3 is selected and its element mass percentage is determined to be 5%, an element ranked as 6 is selected and its element mass percentage is determined to be 1%, and an element ranked as 8 is selected and its element mass percentage is determined to be 9%, thereby forming a component vector. In an optional embodiment, the random number generation algorithm can be configured to generate a range of random numbers, thereby generating random numbers of varying lengths to generate component vectors composed of different numbers of elements.

[0026] Step S3, sorting the multiple component vectors according to the performance prediction results by a non-dominated sorting algorithm to determine the excellent component vector; wherein the non-dominated sorting determines which components perform better in the multi-objective optimization by comparing multiple objective values.

[0027] In an optional embodiment, the multiple target values ​​are three parameters: tensile strength, elongation and transformation temperature of martensite. The multiple component vectors are sorted according to the performance prediction results by a non-dominated sorting algorithm to determine an excellent component vector, including: Step S3.1, determine the dominance between any two component vectors according to the following algorithm;

[0028] Among them, k,m∈{1,2,3}, represent three performance parameters, and Represent two different performance parameters respectively. The above formula indicates that when the component vector x i The three performance parameters are less than or equal to x j , and they cannot all be equal, then determine x i Dominate xj .

[0029] Step S3.2, use triangulation method to determine the crowding degree of each component vector for component vectors with the same dominance; specifically, first project the component vectors into two-dimensional space, and then use Delaunay triangulation method to triangulate each projection point and calculate the area of ​​the triangle that wraps the projection point. The larger the area, the lower the density around the projection point, and thus the smaller the crowding degree and the higher the priority. Among them, Delaunay triangulation method is an algorithm in computational geometry, which is used to divide a set of points on a plane into a series of non-overlapping triangles so that the set of these triangles can completely cover the convex hull of the original point set.

[0030] Step S3.3: Sort the component vectors from high to low according to the congestion degree, and determine the component vectors with higher ranking as excellent component vectors.

[0031] Step S4, generating a new component vector according to the excellent component vector and the genetic algorithm; wherein the genetic algorithm includes a crossover algorithm and a mutation algorithm; wherein generating a new component vector according to the excellent component vector and the genetic algorithm includes: For the excellent component vector, multiple offspring component vectors are generated by simulating binary crossover through the following formula; ; Among them, β is a crossover factor generated based on random numbers, x p and x q are two parent component vectors, k represents the number of generations; In each offspring component vector, a small perturbation is introduced by the following formula to perform polynomial mutation to generate multiple new component vectors; .

[0032] Step S5, iteratively executing the above performance prediction and subsequent steps according to the excellent component vector and the new component vector until the performance prediction result meets the preset value and outputs the corresponding component vector; In an optional embodiment, in response to reaching the prediction iteration number, the performance prediction result still does not meet the preset value and a plurality of random component vectors are regenerated. Figure 2 shown.

[0033] Example Use the above steps to design a composition with a tensile strength of >= 2.5 GPa, an elongation of >= 5%, and a martensitic transformation temperature of >= 300°C, such as Figure 3As shown in the figure. In the 0th generation without algorithm optimization, the component points are randomly generated by the program; these randomly generated component points show a relatively scattered distribution in the target space, reflecting that their performance indicators have a large fluctuation range; as the program iterates to the 30th generation, through multiple optimizations of non-dominated sorting and crowding calculation, the component points gradually concentrate in the range where the performance is close to the target area; it can be observed that these component points begin to show obvious local aggregation characteristics, indicating that under the drive of multi-objective optimization, the performance of the components gradually approaches the set target performance, and finally reaches a balance between multiple objectives.

[0034] Based on the above-mentioned method embodiments, the present invention can quickly determine the composition of martensite according to performance requirements, and by randomly generating primary components, screening excellent components using a non-dominated sorting algorithm, and generating offspring based on excellent components using a genetic algorithm, the probability of generating components that meet performance requirements can be greatly improved to improve the prediction speed.

[0035] In a second aspect of the present invention, a martensite composition prediction device is provided, such as Figure 4 As shown, it includes: a component generation module 10, configured to generate multiple component vectors through a preset random forest model; a performance prediction module 20, configured to perform performance prediction based on the component vectors through a preset martensite performance prediction model; a performance sorting module 30, configured to sort multiple component vectors according to the performance prediction results through a non-dominated sorting algorithm to determine an excellent component vector; a performance inheritance module 40, configured to perform genetic operations based on the excellent component vectors to generate new component vectors; an iterative decision module 50, configured to iteratively execute the above-mentioned performance prediction and its subsequent steps based on the excellent component vectors and the new component vectors until the performance prediction result meets the preset value and outputs the corresponding component vector.

[0036] Based on the above-mentioned device embodiment, the present invention can quickly determine the composition of martensite according to performance requirements, and by randomly generating primary components, screening excellent components using a non-dominated sorting algorithm, and generating offspring based on excellent components using a genetic algorithm, the probability of generating components that meet performance requirements can be greatly improved to improve the prediction speed.

[0037] The above are exemplary embodiments disclosed in the present invention, but it should be noted that various changes and modifications can be made without departing from the scope of the disclosure of the embodiments of the present invention as defined in the claims. The functions, steps and / or actions of the method claims according to the disclosed embodiments described herein do not need to be performed in any particular order. A person of ordinary skill in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope disclosed in the embodiments of the present invention (including the claims) is limited to these examples; under the idea of ​​the embodiments of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and there are many other changes in different aspects of the embodiments of the present invention as above, which are not provided in detail for the sake of simplicity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the scope of protection of the embodiments of the present invention.

Claims

1. A method for predicting martensite composition, characterized in that: The method comprises: Training martensite property prediction model based on random forest algorithm; Randomly generate a plurality of component vectors, and use the martensite performance prediction model to perform performance prediction on each of the component vectors; Sorting the plurality of component vectors according to the performance prediction results by a non-dominated sorting algorithm to determine an excellent component vector; Generate a new component vector according to the excellent component vector and the genetic algorithm; The above performance prediction and subsequent steps are iteratively performed according to the excellent component vector and the new component vector until the performance prediction result meets the preset value and the corresponding component vector is output.

2. The method for predicting martensite composition according to claim 1, characterized in that: The martensitic properties prediction model is trained based on the random forest algorithm, including: Build a random forest model by configuring the number of decision trees, maximum depth, minimum number of sample splits, minimum number of sample leaf nodes, and maximum number of features; The random forest model is trained using training samples composed of known martensite components and their performance parameters to obtain a martensite performance prediction model.

3. The method for predicting martensite composition according to claim 2, characterized in that: The method further comprises: A martensitic transformation temperature prediction function is added to the martensitic property prediction model to predict the martensitic transformation temperature.

4. The method for predicting martensite composition according to claim 3, characterized in that: The martensitic transformation temperature prediction function comprises: in, y 3( x i ) is the martensitic transformation temperature, () is the mass percentage of the corresponding element.

5. The method for predicting martensite composition according to claim 1, characterized in that: Sorting the plurality of component vectors according to the performance prediction result by a non-dominated sorting algorithm to determine an excellent component vector includes: Determine the dominance between any two component vectors according to a preset algorithm; Using triangulation method to determine the crowding degree of each component vector with the same dominance degree; The component vectors are sorted in descending order of the congestion degree, and component vectors with higher sorting are determined as excellent component vectors.

6. The method for predicting martensite composition according to claim 5, characterized in that: The dominance between any two component vectors is determined according to a preset algorithm, including: If any two component vectors satisfy the following conditions: in, x i and x j are any two component vectors, where k, m represent two different performance parameters k, m∈{1,2,3}, then x i Dominate x j ; The dominance degree of each component vector is determined according to the dominance relationship between the component vectors.

7. The method for predicting martensite composition according to claim 5, characterized in that: The triangulation method is used to determine the crowding degree of each component vector having the same dominance, including: By triangulating each component vector point, the triangulation method is used to calculate the area of ​​the triangle that encloses the component vector point; The congestion degree of the corresponding component vector is determined according to the area of ​​the triangle.

8. The method for predicting martensite composition according to claim 1, characterized in that: Generating a new component vector according to the excellent component vector and the genetic algorithm includes: Generating a plurality of offspring component vectors by simulating binary crossover for the excellent component vector; A small disturbance is introduced into each of the child component vectors to perform polynomial mutation to generate a plurality of the new component vectors.

9. The method for predicting martensite composition according to claim 1, characterized in that: The method further comprises: In response to reaching the prediction iteration number, the performance prediction result still does not meet the preset value and a plurality of component vectors are randomly regenerated.

10. A martensite composition prediction device, characterized in that: include: A component generation module, configured to generate a plurality of component vectors through a preset random forest model; A performance prediction module, configured to perform performance prediction according to the composition vector using a preset martensite performance prediction model; A performance sorting module, configured to sort the plurality of component vectors according to the performance prediction results by using a non-dominated sorting algorithm to determine an excellent component vector; A performance inheritance module configured to perform a genetic operation based on the excellent component vector to generate a new component vector; The iterative decision module is configured to iteratively execute the above performance prediction and subsequent steps according to the excellent component vector and the new component vector until the performance prediction result meets the preset value and outputs the corresponding component vector.

Images