A Material Property Prediction Method Based on Automated Machine Learning

Abstract

This invention discloses a material performance prediction method based on automated machine learning, relating to the fields of material performance prediction and machine learning. The method includes: acquiring steel material data and constructing a target dataset; constructing meta-features of the target dataset; constructing an initial population based on the meta-features; using the prediction accuracy of trained individuals as fitness, iteratively optimizing the initial population using a genetic algorithm to obtain the trained individuals with optimal fitness as the material performance prediction model; training the individuals in the population using the target dataset; and predicting the properties of steel materials based on the material performance prediction model. This invention, based on automated machine learning, achieves automatic model selection and automatic hyperparameter optimization, reducing the complexity of model construction in material performance research and improving the accuracy of material performance prediction.

Description

Technical Field

[0001] This invention relates to the field of material property prediction and machine learning, and in particular to a method for predicting material properties based on automated machine learning. Background Technology

[0002] With the rapid development of technologies such as artificial intelligence, data mining, machine learning, and big data, a new research and development model—the data-driven strategy—is gradually emerging. It demonstrates immense potential in revealing the properties of unknown materials and exploring untapped potential materials. The data-driven strategy utilizes large-scale datasets and machine learning algorithms to more effectively analyze material properties and discover correlations and patterns between materials. Through in-depth mining of massive amounts of data, researchers can quickly identify promising new materials, accelerating the materials research and development process. This approach not only reduces research and development costs but also shortens the research and development cycle, providing a completely new avenue for the discovery and application of new materials.

[0003] In the process of predicting material properties using machine learning techniques, decisions regarding the selection of appropriate models, hyperparameter settings, and feature engineering are crucial to model performance. However, these decisions often require researchers to have a deep understanding of machine learning techniques, possess certain experience, and spend a significant amount of time conducting experiments. Furthermore, there are no unified best practices in machine learning modeling. Summary of the Invention

[0004] The purpose of this invention is to provide a material property prediction method based on automatic machine learning, which can realize the automatic construction of material property prediction models and the optimization of hyperparameters, reduce the complexity of model construction in the material property research process, and improve the accuracy of material property prediction in the material property research process.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for predicting material properties based on automated machine learning includes:

[0007] Acquire steel material data and use the steel material data as a sample to construct a target dataset; the steel material data includes the process parameters, composition parameters, and performance parameters of the steel material.

[0008] Construct meta-features of the target dataset; the meta-features include: the number of samples in the target dataset, the number of features, the mean of the kurtosis of each feature, the maximum value of the kurtosis of each feature, the minimum value of the kurtosis of each feature, the standard deviation of the kurtosis of each feature, the mean of the skewness of each feature, the maximum value of the skewness of each feature, the minimum value of the skewness of each feature, and the standard deviation of the skewness of each feature.

[0009] An initial population is constructed based on the meta-features of the target dataset; the individuals in the initial population are different machine learning pipelines, and the machine learning pipelines include a feature preprocessing module, a feature selection module, and a regression module connected in sequence.

[0010] The fitness of the trained individuals is determined by the prediction accuracy of the trained individuals. Based on the initial population, a genetic algorithm is used for iterative optimization to obtain the trained individuals with the best fitness as the material property prediction model. The trained individuals are obtained by training the individuals in the population using the target dataset.

[0011] The properties of steel are predicted based on the aforementioned material property prediction model.

[0012] Optionally, the formula for calculating the characteristic kurtosis is:

[0013]

[0014] Among them, K j Let x be the kurtosis of the j-th characteristic. i,j Let be the value of the j-th feature of the i-th sample, and n be the number of samples. Let be the mean of the j-th feature.

[0015] Optionally, the formula for calculating the characteristic skewness is:

[0016]

[0017] Among them, S j Let x be the skewness of the j-th feature. i,j Let be the value of the j-th feature of the i-th sample, and n be the number of samples. Let σ be the mean of the j-th feature. j Let be the standard deviation of the j-th feature of all samples.

[0018] Optionally, an initial population is constructed based on the meta-features of the target dataset, specifically including:

[0019] Obtain multiple public datasets using OpenML;

[0020] A machine learning model is trained using each public dataset.

[0021] Based on the meta-features of each public dataset and the meta-features of the target dataset, the K-nearest neighbor algorithm is used to select a predetermined number of public datasets whose meta-features are adjacent to the meta-features of the target dataset, and the resulting machine learning models are trained to form the initial population.

[0022] Optionally, during the training of individuals in the population using the target dataset, the CMA-ES algorithm is used to distribute and asynchronously optimize the hyperparameters of the machine learning pipeline represented by the individuals in a concurrent manner.

[0023] Optionally, the specific steps for using the CMA-ES algorithm to distribute, asynchronously, and concurrently optimize the hyperparameters of the machine learning pipeline represented by an individual include:

[0024] The mean parameters of the CMA-ES algorithm during the (t+1)th training iteration are determined using the following formula;

[0025]

[0026] Where, m t+1 Let m represent the mean parameter of the CMA-ES algorithm during the (t+1)th training iteration. t Let represent the mean parameter of the CMA-ES algorithm during the t-th training iteration. Indicates the tdth u The mean parameters of the CMA-ES algorithm after training. This represents the set of indexes of the hyperparameters used in this mean update. Let d be the set of hyperparameter indices that caused the delay, where u is the hyperparameter indices that caused the delay due to asynchronous concurrency. The delay is caused by the different computation times required for different hyperparameters and the process scheduling of the operating system. u The hyperparameter u is the algebraic expression for the delay, where the second term represents the compensation for the delay, u. * is the set of hyperparameter indices without delay; π() is sorted in descending order of the accuracy of the machine learning pipeline corresponding to the trained hyperparameters. This indicates that hyperparameters appearing later in the order are less important; Indicates an intermediate function. x u,t Let represent the hyperparameters of the machine learning pipeline with index u sampled during the t-th training process. Indicates the time at td u The hyperparameters of the machine learning pipeline with index u obtained during the training process; Indicates the time at td u' The hyperparameters of the machine learning pipeline with index u' are sampled during the training process, where u' is the index of the hyperparameter without delay.

[0027] Using the mean of the CMA-ES algorithm trained in the (t+1)th iteration, the probability density function is constructed as follows:

[0028]

[0029] Where EI(x) is the probability density function, x is the hyperparameter of the machine learning pipeline, α is the function parameter set to 0.8, and h(x) = N(m t ,σ t 2 C t ) is the sampling distribution function in the CMA-ES algorithm, σ t Let C be the step size of the t-th training iteration in the CMA-ES algorithm. t Let be the covariance matrix of the t-th training iteration in the CMA-ES algorithm, where N is a normal distribution; x m,t For the positive samples in the t-th training iteration, the set of the first M hyperparameters in the hyperparameter sequence obtained by sorting the hyperparameters according to their accuracy in descending order, where M is the number of positive samples and K is the number of positive samples. h () represents a Gaussian kernel with a kernel function bandwidth of h; x n,t For the negative samples of the t-th training, there is the set of the last N hyperparameters in the hyperparameter sequence obtained by sorting the hyperparameters according to the accuracy of the machine learning pipelines corresponding to the trained hyperparameters in descending order, where N is the number of negative samples.

[0030] The hyperparameters of the machine learning pipeline represented by the individual trained in the (t+1)th training iteration are obtained by sampling the probability density function.

[0031] Optionally, the population update steps during the iterative optimization process using a genetic algorithm based on the initial population include:

[0032] The top 10% of individuals in the current population sequence are selected and added to the population in the next iteration; the individual sequence is obtained by sorting the individuals in the current population in descending order of fitness.

[0033] Three individuals are randomly selected from the current population. Two individuals with higher fitness are selected as candidate individuals. The individual with the smaller operator is selected from the two candidate individuals and added to the next generation of the population.

[0034] Crossover and mutation are performed on existing individuals in the next generation population to obtain all individuals of the next generation population.

[0035] Optionally, the mutations include replacement mutations, addition mutations, and deletion mutations.

[0036] A method for predicting material properties based on task chain automated machine learning, comprising the following steps:

[0037] Construct task chains according to the importance of the target performance as predicted;

[0038] Using the above-mentioned material property prediction method based on automated machine learning, a material property prediction model is constructed to predict the performance of the first target in the task chain.

[0039] The output of the material performance prediction model for predicting the performance of the m-th target in the task chain is used as an input to the material performance prediction model for predicting the performance of the (m+1)-th target in the task chain. The above-mentioned material performance prediction method based on automatic machine learning is used to construct the material performance prediction model for predicting the performance of the (m+1)-th target in the task chain.

[0040] Predict the target properties of steel materials based on the material property prediction model used to predict the performance of each target in the task chain.

[0041] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0042] This invention provides a method for predicting material properties based on automated machine learning. The method includes: acquiring steel material data and using the steel material data as samples to construct a target dataset; constructing meta-features of the target dataset; constructing an initial population based on the meta-features of the target dataset; the individuals in the initial population being different machine learning pipelines; using the prediction accuracy of the trained individuals as fitness, iteratively optimizing the initial population using a genetic algorithm to obtain the trained individuals with the optimal fitness as the material property prediction model; training the individuals in the population using the target dataset; and predicting the properties of steel materials based on the material property prediction model. This invention achieves automatic model selection and automatic hyperparameter optimization based on automated machine learning, reducing the complexity of model construction in material property research and improving the accuracy of material property prediction. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A flowchart illustrating a material property prediction method based on automated machine learning, provided as an embodiment of the present invention;

[0045] Figure 2 A schematic diagram of a material property prediction method based on automated machine learning provided in an embodiment of the present invention;

[0046] Figure 3 This is a flowchart of a population initialization process based on meta-learning provided in an embodiment of the present invention;

[0047] Figure 4 A flowchart of a task chain-based automatic machine learning-based material property prediction method provided in an embodiment of the present invention;

[0048] Figure 5 This is a schematic diagram of a machine learning pipeline for predicting fatigue strength obtained through pipeline optimization, provided in an embodiment of the present invention.

[0049] Figure 6 This is a performance comparison chart of different methods for predicting and modeling the properties of steel provided in the embodiments of the present invention;

[0050] Figure 7 This is a scatter plot comparing the actual and predicted values ​​of each prediction target in the automatic machine learning modeling of the task chain provided in this embodiment of the invention.

[0051] Figure 8 This is a comparison diagram of the optimization process using only a genetic algorithm and the optimization process combined with the SMBO algorithm, provided for embodiments of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] The purpose of this invention is to provide a material property prediction method based on automatic machine learning, which can realize the automatic construction of material property prediction models and the optimization of hyperparameters, reduce the complexity of model construction in the material property research process, and improve the accuracy of material property prediction in the material property research process.

[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] In past research, many intelligent systems based on genetic evolution algorithms have demonstrated design capabilities comparable to humans in certain aspects. For example, designing specialized antennas for space missions or identifying and fixing vulnerabilities in large software projects. Inspired by this, embodiments of the present invention will attempt to use genetic evolution algorithms to automatically design targeted machine learning pipelines based on user-provided data.

[0056] The emergence of automated machine learning methods can simplify and improve the efficiency of machine learning processes by automating feature engineering, model selection, and hyperparameter optimization. This helps to address the complexity, cost, and resource challenges of traditional machine learning processes and improve model accuracy.

[0057] Example 1

[0058] This embodiment provides a material property prediction method based on automated machine learning, such as... Figure 1 and Figure 2 As shown, in the process of the material property prediction method based on automatic machine learning, the user receives constraints (time / number of experiments), training data, and indicators. The automatic machine learning program will automatically build a machine learning pipeline, evaluate the machine learning pipeline, and perform hyperparameter optimization to finally obtain a machine learning model with good generalization performance.

[0059] The material property prediction method based on automated machine learning includes:

[0060] Step 101: Obtain steel material data and use the steel material data as a sample to construct a target dataset; the steel material data includes the process parameters, composition parameters and performance parameters of the steel material.

[0061] Steel material data was collected to form a target dataset, which included the temperatures of three heat treatment processes (normalizing, quenching, and tempering), the content information of nine elements: carbon (C), silicon (Si), manganese (Mn), phosphorus (P), sulfur (S), nickel (Ni), chromium (Cr), copper (Cu), and molybdenum (Mo), one reduction rate, and the area fractions of three non-metallic inclusions: dA (inclusions formed by plastic working), dB (discontinuously arranged inclusions), and dC (isolated inclusions) to predict four target performance characteristics: fatigue strength, tensile strength, fracture strength, and hardness.

[0062] Step 102: Construct the meta-features of the target dataset; the meta-features include: the number of samples in the target dataset, the number of features, the mean of the kurtosis of each feature, the maximum value of the kurtosis of each feature, the minimum value of the kurtosis of each feature, the standard deviation of the kurtosis of each feature, the mean of the skewness of each feature, the maximum value of the skewness of each feature, the minimum value of the skewness of each feature, and the standard deviation of the skewness of each feature.

[0063] The meta-features constructed for the target dataset include: number of samples, number of features, mean of kurtosis of each feature, maximum value of kurtosis of each feature, minimum value of kurtosis of each feature, standard deviation of kurtosis of each feature, mean of skewness of each feature, mean of skewness of each feature, maximum value of skewness of each feature, minimum value of skewness of each feature, and standard deviation of skewness of each feature.

[0064] The formula for calculating the characteristic kurtosis is:

[0065]

[0066] Among them, K j Let x be the kurtosis of the j-th characteristic. i,j Let be the value of the j-th feature of the i-th sample, and n be the number of samples. Let be the mean of the j-th feature.

[0067] The formula for calculating eigenvalue skewness is:

[0068]

[0069] Among them, S j Let x be the skewness of the j-th feature. i,j Let be the value of the j-th feature of the i-th sample, and n be the number of samples. Let σ be the mean of the j-th feature. j Let be the standard deviation of the j-th feature of all samples.

[0070] Step 103: Construct an initial population based on the meta-features of the target dataset; the individuals in the initial population are different machine learning pipelines, and the machine learning pipelines include a feature preprocessing module, a feature selection module, and a regression module connected in sequence.

[0071] Using the OpenML API, target datasets with more than 100 but fewer than 1000 samples were obtained from the platform. All sparse target datasets and those containing time or string-type attributes were removed, resulting in 206 target datasets. Automated machine learning modeling was performed on these target datasets to obtain 206 machine learning pipelines. These pipelines were then evaluated across all machine learning pipelines using all target datasets, and the five pipelines with the best average performance were selected. The machine learning pipelines are structured as a tree, with machine learning operators as nodes and data flows as edges. The root node of the tree is the regression model operator, and the leaf nodes are the input data. Machine learning operators include: feature preprocessing operators (i.e., operators in the feature preprocessing module): Binarizer, FastICA, MaxAbsScaler, Normalizer, Nystroem; feature selection operators (i.e., operators in the feature selection module): SelectFwe (error rate selector), VarianceThreshold (variance threshold selector), SelectPercentile (percentage selector), SelectFromModel (model-based selection); and regression model operators (i.e., operators in the regression module): Ridge, SGDRegressor, ExtraTreesRegressor, GradientBoostingRegressor.

[0072] Based on the meta-features constructed in step 102, the K-nearest neighbor algorithm is used to query the aforementioned pipelines. K-Flod cross-validation is used to evaluate the accuracy of the predictions of all machine learning pipelines in the population, and the evaluation accuracy is used as the fitness of the pipelines.

[0073] The process of constructing the initial population, such as Figure 3 As shown, the solid line represents the offline stage of meta-learning, and the dashed line represents the online stage, which specifically includes the following steps:

[0074] A collection of machine learning pipelines that perform well on average across all target datasets.

[0075] A collection of machine learning pipelines that perform well on similar target datasets.

[0076] A randomly generated set of machine learning pipelines.

[0077] Step 104: Using the prediction accuracy of the trained individuals as the fitness, a genetic algorithm is used for iterative optimization based on the initial population to obtain the trained individuals with the best fitness as the material performance prediction model; the trained individuals are obtained by training the individuals in the population using the target dataset.

[0078] The top 10% of the pipelines with the highest fitness in the current population are copied and added to the new population. During the construction of the remaining 90% of the pipelines, three pipelines are randomly selected from the existing population each time. The pipeline with the lowest fitness is eliminated, and then the pipeline with the fewest operators among the remaining two is selected, copied, and added to the new population. Two pipelines are randomly selected, and a random point is chosen in the tree structure to split them, then their contents are swapped. Next, a fixed proportion of the remaining unaffected pipelines are randomly modified. This is equivalent to crossover and mutation in genetic evolution algorithms. Specifically, this invention performs three mutation operations with a one-third probability: replacement mutation, insertion mutation, and deletion mutation. Replacement mutation refers to randomly replacing one random operator in a pipeline with another. Insertion mutation refers to inserting a random operator at a random position in a pipeline. Deletion mutation refers to removing a random subset of operators from a pipeline.

[0079] During hyperparameter optimization, historical hyperparameters are categorized into positive and negative samples. The top 20% of evaluated hyperparameters are used as positive samples, and the remaining hyperparameters are used as negative samples. A Gaussian function is used as the kernel function. The formula for sampling hyperparameters is as follows: Where EI(x) is the probability density function, x is the hyperparameter of the machine learning pipeline, α is the function parameter set to 0.8, and h(x) = N(m t ,σ t 2 C t ) is the sampling distribution function in the CMA-ES algorithm, σ t Let C be the step size of the t-th training iteration in the CMA-ES algorithm. t Let be the covariance matrix of the t-th training iteration in the CMA-ES algorithm, where N is a normal distribution; x m,t For the positive samples in the t-th training iteration, the set of the first M hyperparameters in the hyperparameter sequence obtained by sorting the hyperparameters according to their accuracy in descending order, where M is the number of positive samples and K is the number of positive samples. h () represents a Gaussian kernel with a kernel function bandwidth of h; x n,t The negative samples in the t-th training iteration belong to the set of the last N hyperparameters in the hyperparameter sequence obtained by sorting the hyperparameters according to their accuracy in descending order. N is the number of negative samples.

[0080] Preferably, an asynchronous server / client architecture can be used for distributed hyperparameter optimization during the process of solving the covariance matrix of the CMA-ES algorithm, and the mean in the CMA-ES algorithm can be updated using the following formula:

[0081]

[0082] Where, m t+1 Let m represent the mean parameter of the CMA-ES algorithm during the (t+1)th training iteration. t Let m represent the mean parameters of the CMA-ES algorithm during the t-th training iteration. t-d Let represent the mean parameters of the CMA-ES algorithm during the td-th training iteration. This represents the set of indexes of the hyperparameters used in this mean update. Let d be the set of hyperparameter indices that caused the delay, where u is the hyperparameter indices that caused the delay due to asynchronous concurrency. The delay is caused by the different computation times required for different hyperparameters and the process scheduling of the operating system. u The hyperparameter u is the algebraic expression for the delay, where the second term represents the compensation for the delay, u. * is the set of hyperparameter indices without delay; π() is sorted in descending order of the accuracy of the machine learning pipeline corresponding to the trained hyperparameters. This indicates that hyperparameters appearing later in the order are less important; g(x) u,t-d () represents an intermediate function. x u,t Let represent the hyperparameters of the machine learning pipeline with index u sampled during the t-th training process. Indicates the time at td u The hyperparameters of the machine learning pipeline with index u obtained during the training process; Indicates the time at td u' The hyperparameters of the machine learning pipeline with index u' are sampled during the training process, where u' is the index of the hyperparameter without delay.

[0083] In a distributed environment, it is divided into computing nodes and sampling nodes. The computing nodes are responsible for training and verifying the performance of a certain hyperparameter. The sampling nodes are responsible for sampling a series of hyperparameters based on the current probability density function, and updating the probability density function after a certain number of hyperparameters have been trained and verified. That is, the timing of the m update is when the number of verified hyperparameters reaches a certain number.

[0084] When updating the sampling distribution, it is updated based on the previous hyperparameter validation results. During this process, due to a series of software and hardware limitations, not all hyperparameters are sampled from generation t when updating the sampling distribution from generation t to generation t+1; some are sampled from generation td. u In the case of algebraic distributions, these delay hyperparameters constitute a set.

[0085] Step 105: Predict the properties of steel based on the material property prediction model.

[0086] Example 2

[0087] This embodiment uses a random forest to rank feature importance based on the reduction of node impurity. After ranking, the resulting target ranking is: tensile strength, hardness, fracture strength, and fatigue strength. Automatic machine learning is then performed on each predicted target sequentially, and the output of the previous automatic machine learning model is concatenated with the target dataset features, which is then used as feature input to the next model.

[0088] This embodiment provides a method for predicting material properties based on task chain-based automated machine learning, which includes the following steps:

[0089] Construct task chains based on the importance of the target performance predicted according to the needs.

[0090] Using the aforementioned material property prediction method based on automated machine learning, a material property prediction model is constructed to predict the performance of the first target in the task chain.

[0091] Using the output of the material performance prediction model for predicting the performance of the m-th target in the task chain as an input to the material performance prediction model for predicting the performance of the (m+1)-th target in the task chain, the above-mentioned material performance prediction method based on automatic machine learning is used to construct the material performance prediction model for predicting the performance of the (m+1)-th target in the task chain.

[0092] Predict the target properties of steel materials based on the material property prediction model used to predict the performance of each target in the task chain.

[0093] This embodiment arranges the machine learning pipeline in reverse order according to the task chain: serialization encoding starts with the final output and ends with the input data. Every operator must precede all its input operators. This applies to both single-input and multi-input operators.

[0094] like Figure 4As shown, this embodiment divides the prediction targets into primary prediction targets and auxiliary prediction targets. The auxiliary prediction targets are used to rank the feature importance of the primary prediction targets, and a chain-like automatic machine learning model is performed based on the ranking results. In this embodiment, a random forest is used to rank feature importance based on the reduction of node impurity. After ranking, automatic machine learning is performed on each prediction target sequentially, and the output of the previous automatic machine learning model is concatenated with the target dataset features, then used as feature input to the next model.

[0095] Prioritize processing the longest sub-pipe: In the case of multiple or parallel sub-pipes, process the longest sub-pipe first. If multiple parallel sub-pipes are the same, the priority order is randomly determined.

[0096] Equal Encoding Length: To ensure that all pipelines have the same encoding length, this embodiment of the invention defines a fixed maximum number of operators L = 15 for all pipelines. For smaller pipelines, blank operators are used to fill the gaps, and the XGBRanker model is used to predict the performance ranking of the pipelines.

[0097] After incorporating a meta-learning-based population initialization algorithm, targeted population initialization is performed on the target dataset provided by the user, significantly improving the speed of automated machine learning, especially under time constraints of 1 minute and 10 minutes. The population pre-screening method further enhances optimization speed by filtering out potentially poor-performing machine learning pipelines. After improvements using the two methods proposed in this paper, the average MSE of the machine learning pipeline is only 69.21% of the original value at 1 minute and 83.72% at 10 minutes.

[0098] The automated machine learning pipeline ultimately discovered through the task chain approach revealed four feature preprocessing methods: MinMaxScalar, RobustScaler, StandardScaler, and PolynomialFeatures, and five regression models: RandomForest, ExtraTree, XGBoost, GradientBoosting, and RidgeCV. Figure 5 As shown.

[0099] Across all prediction tasks, hyperparameter optimization methods, automated machine learning methods, and task-chain automated machine learning methods all achieved some degree of performance improvement. Among these, the task-chain automated machine learning method achieved the highest prediction accuracy across all tasks. For example... Figure 6As shown, compared to the Random Forest algorithm with default parameters and an average R² of 0.919, Task Chain Automatic Machine Learning improves the R² to 0.967, achieving a significant performance improvement. The improvement is particularly noticeable in the tensile strength prediction task, increasing the R² from 0.923 to 0.980. Figure 7 and Figure 8 As shown, Figure 7 Images (a), (b), (c), and (d) in the figure show the comparison between the actual and predicted values ​​of the task chain automated machine learning method for fatigue strength, tensile strength, fracture strength, and hardness. Figure 8 (a), (b), (c), and (d) in the figure show the comparison results of the optimization process using only the genetic algorithm versus the process of optimizing fatigue strength, tensile strength, fracture strength, and hardness by combining the genetic algorithm with the SMBO algorithm.

[0100] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting material properties based on automated machine learning, characterized in that, The material property prediction method based on automated machine learning includes: Acquire steel material data and use the steel material data as a sample to construct a target dataset; the steel material data includes the process parameters, composition parameters, and performance parameters of the steel material. Construct meta-features of the target dataset; the meta-features include: the number of samples in the target dataset, the number of features, the mean of the kurtosis of each feature, the maximum value of the kurtosis of each feature, the minimum value of the kurtosis of each feature, the standard deviation of the kurtosis of each feature, the mean of the skewness of each feature, the maximum value of the skewness of each feature, the minimum value of the skewness of each feature, and the standard deviation of the skewness of each feature. An initial population is constructed based on the meta-features of the target dataset; the individuals in the initial population are different machine learning pipelines, and the machine learning pipelines include a feature preprocessing module, a feature selection module, and a regression module connected in sequence. The fitness of the trained individuals is determined by the prediction accuracy of the trained individuals. Based on the initial population, a genetic algorithm is used for iterative optimization to obtain the trained individuals with the best fitness as the material property prediction model. The trained individuals are obtained by training the individuals in the population using the target dataset. The properties of steel are predicted based on the aforementioned material property prediction model; During the training of individuals in the population using the target dataset, the CMA-ES algorithm is used to optimize the hyperparameters of the machine learning pipeline represented by the individuals in a distributed, asynchronous, and concurrent manner. The specific steps for using the CMA-ES algorithm to distribute, asynchronously, and concurrently optimize the hyperparameters of the machine learning pipeline represented by individuals include: The mean parameters of the CMA-ES algorithm during the (t+1)th training iteration are determined using the following formula; in, Let represent the mean parameters of the CMA-ES algorithm during the (t+1)th training iteration. Let represent the mean parameter of the CMA-ES algorithm during the t-th training iteration. Indicates the first The mean parameters of the CMA-ES algorithm after training. This represents the set of indexes of the hyperparameters used in this mean update. For the set of hyperparameter indices that generated the delay, These are the hyperparameter numbers that indicate latency caused by asynchronous concurrency. The latency varies depending on the computation time required for different hyperparameters and the operating system's process scheduling. Hyperparameters for generating delay Delayed algebra, This is the set of hyperparameter indices without delay; To sort the machine learning pipelines from highest to lowest accuracy based on the hyperparameters obtained during training. = This indicates that hyperparameters ranked later in the order are less important; Indicates an intermediate function. , Let represent the hyperparameters of the machine learning pipeline with index u sampled during the t-th training process. Indicates the t-th The hyperparameters of the machine learning pipeline with index u obtained during the training process; Indicates the t-th The sequence number obtained during this training process is Hyperparameters of a machine learning pipeline The hyperparameter index without delay; Using the mean of the CMA-ES algorithm trained in the (t+1)th iteration, the probability density function is constructed as follows: ； in, Let be the probability density function. For hyperparameters of the machine learning pipeline, As a function parameter, set it to 0.

8. This is the sampling distribution function in the CMA-ES algorithm. Let be the step size for the t-th training iteration in the CMA-ES algorithm. Let be the covariance matrix of the t-th training iteration in the CMA-ES algorithm. It follows a normal distribution; , Let M be the positive samples from the t-th training iteration. M represents the set of the first M hyperparameters in the hyperparameter sequence, ordered from largest to smallest by the accuracy of the machine learning pipelines corresponding to the trained hyperparameters. M is the number of positive samples. This represents a Gaussian kernel with a kernel function bandwidth of h; , For the negative samples of the t-th training, there is the set of the last N hyperparameters in the hyperparameter sequence obtained by sorting the hyperparameters according to the accuracy of the machine learning pipelines corresponding to the trained hyperparameters in descending order, where N is the number of negative samples. The hyperparameters of the machine learning pipeline represented by the individual trained in the (t+1)th training iteration are obtained by sampling the probability density function.

2. The material property prediction method based on automated machine learning according to claim 1, characterized in that, The formula for calculating the characteristic kurtosis is: ； in, For the first j One characteristic kurtosis, For the first i The first sample j The values ​​of each feature, n For the number of samples, For the first j The mean of each feature.

3. The material property prediction method based on automated machine learning according to claim 1, characterized in that, The formula for calculating eigenvalue skewness is: ; in, For the first j Each feature skewness, For the first i The first sample j The values ​​of each feature, n For the number of samples, For the first j The mean of each feature, Let be the standard deviation of the j-th feature of all samples.

4. The material property prediction method based on automated machine learning according to claim 1, characterized in that, Based on the meta-features of the target dataset, an initial population is constructed, specifically including: Obtain multiple public datasets using OpenML; A machine learning model is trained using each public dataset. Based on the meta-features of each public dataset and the meta-features of the target dataset, the K-nearest neighbor algorithm is used to select a predetermined number of public datasets whose meta-features are adjacent to the meta-features of the target dataset, and the resulting machine learning models are trained to form the initial population.

5. The material property prediction method based on automated machine learning according to claim 1, characterized in that, The population update steps during iterative optimization using a genetic algorithm based on the initial population include: The top 10% of individuals in the current population sequence are selected and added to the population in the next iteration; the individual sequence is obtained by sorting the individuals in the current population in descending order of fitness. Three individuals are randomly selected from the current population. Two individuals with higher fitness are selected as candidate individuals. The individual with the smaller operator is selected from the two candidate individuals and added to the next generation of the population. Crossover and mutation are performed on existing individuals in the next generation population to obtain all individuals of the next generation population.

6. The material property prediction method based on automated machine learning according to claim 5, characterized in that, The mutations include replacement mutations, addition mutations, and deletion mutations.

7. A method for predicting material properties based on task chain-based automated machine learning, characterized in that, The method for predicting material properties based on task chain-based automated machine learning includes the following steps: Construct task chains according to the importance of the target performance as predicted; Using the material property prediction method based on automatic machine learning as described in any one of claims 1-6, a material property prediction model for predicting the performance of the first target in the task chain is constructed. Using the output of the material performance prediction model for predicting the performance of the m-th target in the task chain as an input to the material performance prediction model for predicting the performance of the (m+1)-th target in the task chain, the material performance prediction model for predicting the performance of the (m+1)-th target in the task chain is constructed using the material performance prediction method based on automatic machine learning as described in any one of claims 1-6. Predict the target properties of steel materials based on the material property prediction model used to predict the performance of each target in the task chain.

Images