Aluminum alloy mechanical property prediction method based on deep learning

Abstract

The invention relates to the technical field of material informatics, in particular to an aluminum alloy mechanical property prediction method based on deep learning, which comprises the following steps: collecting an original data set containing alloy components, process parameters and performance actual measurement data, and cleaning, preprocessing and dividing the original data set; constructing a fusion neural network prediction model, wherein the fusion neural network prediction model integrates a Mama module to extract process sequence features, a KAN network module to extract a component nonlinear relation and a Transform encoder module to perform global feature fusion; training the model by using the training set, and performing hyper-parameter tuning and monitoring by using the verification set; evaluating model precision and engineering applicability through a test set and an external experimental sample; and finally, inputting alloy components to be predicted and process parameters into the trained model, and synchronously outputting predicted values of the yield strength, the tensile strength and the ductility. The method provides an efficient and accurate prediction tool for aluminum alloy design and optimization.

Description

Technical Field

[0001] This invention relates to the field of materials informatics technology, and in particular to a method for predicting the mechanical properties of aluminum alloys based on deep learning. Background Technology

[0002] With the deepening application of artificial intelligence technology in materials science, constructing mapping relationships between composition, processing, and properties using machine learning models has become an important means to accelerate the research and development of new materials. In recent years, traditional machine learning methods such as neural networks, support vector machines, and random forests have been widely used in the performance prediction of various alloys and have shown certain potential. In particular, for aluminum alloy systems, researchers have achieved preliminary predictions of some mechanical property indicators by constructing feature engineering and regression models, providing data-driven auxiliary tools for materials design.

[0003] Most existing prediction models adopt a single classical network architecture, which makes it difficult to simultaneously and effectively capture the complex time-series dependencies of heat treatment process parameters, the highly nonlinear coupling relationship between alloy composition and performance, and achieve efficient fusion of global features. As a result, when dealing with the complex mapping problem of composition-process-performance of multivariable and strongly coupled 7xxx series aluminum alloys, the prediction accuracy and model interpretability are often insufficient, which limits their reliable application in engineering practice. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a deep learning-based method for predicting the mechanical properties of aluminum alloys. It aims to improve the problem that existing single models are unable to accurately capture the sequential dependence of aluminum alloy heat treatment processes, the complex nonlinear relationship between composition and properties, and perform global feature fusion, which leads to insufficient accuracy and poor interpretability in predicting mechanical properties.

[0005] This invention provides the following technical solution: a method for predicting the mechanical properties of aluminum alloys based on deep learning, comprising the following steps: S1. Collect the original dataset containing various aluminum alloy composition data, heat treatment process parameter data and corresponding mechanical property test data. S2. Perform data cleaning and preprocessing on the original dataset to obtain a standard dataset, and divide the standard dataset into a training set, a validation set, and a test set. S3. Construct a fusion neural network model as a prediction model for the mechanical properties of aluminum alloys. The fusion neural network model includes a Mamba module for extracting process parameter sequence-dependent features, a KAN network module for extracting nonlinear interpretable relationships between composition and properties, and a Transformer encoder module for global feature fusion. S4. The aluminum alloy mechanical property prediction model is trained using the training set, and hyperparameters are tuned and training is monitored using the validation set. The goal is to minimize the loss function between the predicted output of the aluminum alloy mechanical property prediction model and the measured mechanical property data, so as to obtain the trained aluminum alloy mechanical property prediction model. S5. Use the test set to evaluate the prediction accuracy and generalization ability of the trained aluminum alloy mechanical property prediction model, and use the measured aluminum alloy sample data that was not involved in the training to conduct external experimental verification of the trained aluminum alloy mechanical property prediction model. S6. Input the composition data and heat treatment process parameter data of the aluminum alloy to be predicted into the trained aluminum alloy mechanical property prediction model, and the trained aluminum alloy mechanical property prediction model outputs the corresponding predicted values ​​of yield strength, tensile strength and elongation.

[0006] Preferably, in step S1, the raw dataset containing various aluminum alloy composition data, heat treatment process parameter data, and corresponding measured mechanical property data specifically includes: Collect alloy composition data, heat treatment process parameter data, and corresponding measured data of yield strength, tensile strength and elongation of various grades or series of aluminum alloys to form an initial data set; Each data record in the initial dataset is standardized into a structured data entry; The structured data entries from different sources are merged, and completely duplicate data entries are removed based on the combined identifier of alloy composition and heat treatment process parameters to obtain the original dataset.

[0007] Preferably, in step S2, the original dataset undergoes data cleaning and preprocessing to obtain a standard dataset. Dividing the standard dataset into a training set, a validation set, and a test set specifically includes: The original dataset is subjected to outlier detection and processing, and data entries with missing values ​​are filled or deleted to obtain a cleaned dataset. The values ​​of alloy element composition data and heat treatment process parameter data in the cleaned dataset are normalized to ensure they fall within a uniform value range. The normalized dataset is divided into the training set, the validation set, and the test set according to a preset ratio.

[0008] Preferably, in step S3, constructing a fusion neural network model as a prediction model for the mechanical properties of aluminum alloys specifically includes: The input features of the aluminum alloy mechanical property prediction model are assigned to a subset of alloy element composition features and a subset of heat treatment process parameter features; The Mamba module is constructed and configured to receive a subset of the heat treatment process parameter features as input to extract the sequence dependency features of the process parameters. The KAN network module is constructed and configured to receive the alloy element composition feature subset as input to extract the nonlinear interpretable relationship between composition and performance. The sequence dependency features extracted by the Mamba module are concatenated with the nonlinear interpretable relation features extracted by the KAN network module to form a combined feature vector. The combined feature vector is input into the Transformer encoder module for global feature fusion and encoding. The output of the Transformer encoder module is mapped through at least one fully connected layer to output the prediction result of the aluminum alloy mechanical property prediction model.

[0009] Preferably, in step S4, obtaining the trained aluminum alloy mechanical property prediction model specifically includes: Initialize the network parameters of the aluminum alloy mechanical property prediction model and configure the optimizer and loss function required for training; The data in the training set are input into the aluminum alloy mechanical property prediction model in batches. The loss value between the model prediction output and the corresponding measured mechanical property data is calculated. The network parameters of the aluminum alloy mechanical property prediction model are updated by the optimizer through the backpropagation algorithm. After each round of iterative training or when the preset training period is reached, the performance of the current aluminum alloy mechanical property prediction model is evaluated using the validation set, and the training process is monitored based on the validation set performance, and early stopping judgment or hyperparameter adjustment is performed. When the training meets the preset stopping condition, the iteration stops, and the parameter state of the aluminum alloy mechanical property prediction model with the best validation set performance at this time is determined as the trained aluminum alloy mechanical property prediction model.

[0010] Preferably, the step of using the validation set to evaluate the performance of the current aluminum alloy mechanical property prediction model after each round of training iterations or upon reaching a preset training period, and monitoring the training process based on the validation set performance, and performing early stopping judgment or adjusting hyperparameters specifically includes: When the preset training period arrives, all or part of the data in the validation set is used and input into the current aluminum alloy mechanical property prediction model to calculate the performance evaluation index value of the model on the validation set. The currently calculated performance evaluation index value is compared with the historical performance evaluation index value to monitor the trend of model performance changes; If the performance evaluation index value does not drop to a new lowest value within multiple consecutive training cycles, the early stop mechanism is triggered, and a stop signal is sent to the training process. If the average decrease of the performance evaluation index value over multiple consecutive training cycles is lower than the preset improvement threshold, a hyperparameter adjustment mechanism is triggered to adjust the learning rate of the optimizer or the relevant hyperparameters of the aluminum alloy mechanical property prediction model.

[0011] Preferably, in step S5, evaluating the prediction accuracy and generalization ability of the trained aluminum alloy mechanical property prediction model using the test set specifically includes: The aluminum alloy composition data and heat treatment process parameter data contained in the test set are input into the trained aluminum alloy mechanical property prediction model. The trained aluminum alloy mechanical property prediction model processes the input test set data and outputs the corresponding predicted values ​​of yield strength, tensile strength and elongation. The predicted values ​​are compared one-to-one with the pre-labeled measured mechanical properties data in the test set to calculate a set of performance evaluation indicators for quantifying the prediction error. Based on a set of calculated performance evaluation indicators, the prediction accuracy of the trained aluminum alloy mechanical property prediction model and its generalization ability on the data distribution represented by the test set are comprehensively evaluated.

[0012] Preferably, in step S5, the external experimental verification of the trained aluminum alloy mechanical property prediction model using untrained measured aluminum alloy sample data specifically includes: Obtain one or more measured aluminum alloy samples that are not included in the original dataset to form an external verification sample set; The alloy composition data and heat treatment process parameter data from the external verification sample set are input into the trained aluminum alloy mechanical property prediction model to obtain the model's predicted mechanical properties of the external samples. The predicted mechanical properties of the external samples are compared with the measured mechanical properties obtained by independent physical experiments in the corresponding external verification sample set. Based on the results of the comparative analysis, the accuracy of the trained aluminum alloy mechanical property prediction model in predicting real aluminum alloy samples that did not appear in the training is evaluated, thereby verifying its applicability and generalization ability in actual engineering scenarios.

[0013] Preferably, in step S6, the predicted values ​​of yield strength, tensile strength, and elongation output by the trained aluminum alloy mechanical property prediction model specifically include: Obtain the alloy element composition data of the aluminum alloy to be predicted and the heat treatment process parameter data to be used, and perform the same data preprocessing operation as in S2 to form standardized prediction input data. The standardized prediction input data is input into the trained aluminum alloy mechanical property prediction model. The trained aluminum alloy mechanical property prediction model performs forward propagation calculations on the standardized prediction input data and simultaneously outputs the corresponding predicted values ​​of yield strength, tensile strength, and elongation.

[0014] The present invention has the following beneficial effects: 1. In this invention, by constructing a dedicated neural network model that integrates Mamba, KAN and Transformer, it is possible to collaboratively capture the sequential dependence of heat treatment processes, the complex nonlinear relationship between alloy composition and performance, and the global interaction of multiple features, thereby achieving more accurate multi-task joint prediction of yield strength, tensile strength and elongation of 7xxx series aluminum alloys. Its prediction accuracy is better than that of models using a single traditional architecture.

[0015] 2. In this invention, the dual verification system of internal test set evaluation and independent external experimental verification confirms that the proposed method not only performs well on known data distributions, but also has stable predictive ability when facing new grades and new processes of aluminum alloys, demonstrating its strong generalization and applicability in practical engineering scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the deep learning-based method for predicting the mechanical properties of aluminum alloys proposed in this invention. Detailed Implementation

[0017] The technical solutions in 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.

[0018] This invention provides a deep learning-based method for predicting the mechanical properties of aluminum alloys, such as... Figure 1 As shown, it includes the following steps: S1. Collect the original dataset containing various aluminum alloy composition data, heat treatment process parameter data, and corresponding measured mechanical property data.

[0019] Furthermore, in S1, the raw dataset that collects various aluminum alloy composition data, heat treatment process parameter data, and corresponding measured mechanical property data specifically includes: Collect alloy composition data, heat treatment process parameter data, and corresponding measured data of yield strength, tensile strength and elongation of various grades or series of aluminum alloys to form an initial data set; Each data record in the initial dataset is standardized into a structured data entry; Structured data entries from different sources are merged, and completely duplicate data entries are removed based on the combination identifier of alloy composition and heat treatment process parameters to obtain the original dataset.

[0020] Specifically, the first step is to conduct multi-source data collection. Data on various grades or series of aluminum alloys is systematically collected from publicly available materials science databases, published academic literature, and open-source code platforms. For each aluminum alloy, the collected data must include three parts: first, alloy elemental composition data, including at least the content of major elements such as zinc, magnesium, and copper; second, heat treatment process parameter data, which is a chronological sequence of parameters, including at least solution temperature, solution time, aging temperature, and aging time; and third, corresponding measured mechanical property data, specifically laboratory test values ​​for yield strength, tensile strength, and elongation. All collected data items are then compiled to form an initial dataset. Next, data structure normalization is performed. Each unstructured or semi-structured record in the initial dataset is converted into a unified, structured data entry. Each structured data entry contains three distinct field groups: an alloy composition field group to store the content of each element; a process parameter sequence field group to store heat treatment parameters in sequence; and a mechanical property field group to store the three measured values ​​of yield strength, tensile strength, and elongation. Finally, data integration and deduplication are performed. All structured data entries are imported into a unified database. A unique identifier generated from the combination of the alloy composition field group and the process parameter sequence field group is used to compare all entries. If two or more data entries have identical unique identifiers, they are considered completely duplicated; only one is retained, and the remaining duplicate entries are deleted. The dataset obtained after deduplication is the original dataset. The above process ensures the diversity of the original dataset sources, the uniformity of the data structure, and the uniqueness of the samples, thus guaranteeing the reliability of subsequent data processing.

[0021] S2. Perform data cleaning and preprocessing on the original dataset to obtain a standard dataset, and divide the standard dataset into a training set, a validation set, and a test set.

[0022] Furthermore, in S2, the original dataset undergoes data cleaning and preprocessing to obtain a standard dataset, which is then divided into training, validation, and test sets. Specifically: The original dataset is subjected to outlier detection and processing, and data entries with missing values ​​are filled or deleted to obtain a cleaned dataset. The values ​​of alloy element composition data and heat treatment process parameter data in the cleaned dataset are normalized to make them fall into a uniform value range. The normalized dataset is divided into training, validation, and test sets according to a preset ratio.

[0023] Specifically, the first step is to clean the original dataset. Data cleaning includes two main operations: outlier handling and missing value handling. For outlier handling, statistical methods are used to calculate the mean and standard deviation of numerical features such as alloy element content, process parameters, and mechanical properties. Data entries with values ​​outside the range [-3, +3] are identified as outliers and replaced with the mean of that feature. For missing value handling, all data entries are iterated through. If a feature value is found to be missing, it is filled with the mean of that feature among all non-missing entries. If the number of missing features in a data entry exceeds a preset threshold, the data entry is deleted. The dataset obtained after completing the above cleaning operations is called the cleaned dataset. Next, the cleaned dataset is normalized. This normalization process is applied only to the two types of input features: alloy element composition data and heat treatment process parameter data. For any feature to be normalized, its total data value is recorded as... The maximum-minimum normalization method is used, and the calculation is performed according to the following formula: ; in, This represents the i-th original value of the feature. This represents the minimum value among all the original values ​​of this feature. This represents the maximum value among all the original values ​​of this feature. This is the new value after normalization. After this process, the values ​​of all input features are scaled to the [0, 1] interval, forming a normalized dataset. Measured mechanical performance data, used as the prediction target, are not subject to this normalization process. Finally, the normalized dataset is partitioned. A combination of stratified sampling and random sampling is used to ensure consistent distribution of key process parameters across different subsets. Following a pre-defined 8:1:1 ratio, the normalized dataset is randomly divided into three mutually exclusive subsets: a training set, a validation set, and a test set. The training set is used for learning model parameters, the validation set is used for hyperparameter tuning and performance monitoring during training, and the test set is used for final evaluation of the model's generalization ability. After partitioning, the standard dataset consists of the training set, validation set, and test set. Through the above process, the raw data was transformed into a standardized, modular dataset, providing a high-quality data foundation for model training and evaluation.

[0024] S3. Construct a fusion neural network model as a prediction model for the mechanical properties of aluminum alloys. The fusion neural network model includes a Mamba module for extracting process parameter sequence-dependent features, a KAN network module for extracting nonlinear interpretable relationships between composition and properties, and a Transformer encoder module for global feature fusion.

[0025] Furthermore, in S3, constructing a fusion neural network model as a prediction model for the mechanical properties of aluminum alloys specifically includes: The input features of the aluminum alloy mechanical property prediction model are assigned to a subset of alloy element composition features and a subset of heat treatment process parameter features; Construct a Mamba module and configure it to receive a subset of heat treatment process parameter features as input to extract sequence dependency features of the process parameters; A KAN network module was constructed and configured to receive a subset of alloy element composition features as input to extract a nonlinear interpretable relationship between composition and performance. The sequence dependency features extracted by the Mamba module are concatenated with the nonlinear interpretable relation features extracted by the KAN network module to form a combined feature vector. The combined feature vector is input into the Transformer encoder module for global feature fusion and encoding. The output of the Transformer encoder module is mapped through at least one fully connected layer to output the prediction results of the aluminum alloy mechanical property prediction model.

[0026] Specifically, the preprocessed model input features are explicitly divided into two subsets. The first subset is the alloy element composition feature subset, containing normalized content data for elements such as zinc, magnesium, and copper, denoted as vector C. The second subset is the heat treatment process parameter feature subset, containing normalized parameters such as solution temperature, solution time, aging temperature, and aging time arranged in the process sequence, considered as a sequence, denoted as P. Construct a Mamba module. This module is based on a structured state-space sequence model. Configure its input dimension to match the dimension of the feature subset P of heat treatment process parameters. Use the sequence P as the input to this module. The Mamba module models the sequence P through its internal state-space equations to capture the temporal dependencies between process parameters and outputs a fixed-dimensional feature vector, denoted as . , used to characterize the dependency features of the process sequence; Construct a KAN network module. This module consists of multiple layers of learnable spline functions, instead of traditional fixed activation functions. Configure its input dimension to match the dimension of the alloy element composition feature subset C. Use vector C as the input to this module. The KAN network module performs nonlinear transformations and combinations on the input features through the spline functions of its layers, ultimately outputting a fixed-dimensional feature vector, denoted as . This feature vector It characterizes the complex, nonlinear relationship between components and performance, which can be partially explained by spline function coefficients; The output process sequence dependent feature vector The eigenvectors of the nonlinear relationship between the output components Perform a concatenation operation. Specifically, concatenate the two vectors end-to-end along their feature dimensions to form a higher-dimensional combined feature vector, denoted as [vector name missing]. ,Right now , where Concat represents the vector concatenation function; Construct a Transformer encoder module. This module contains a multi-head self-attention layer and a feedforward neural network layer. Combine the feature vectors... This serves as the input to the Transformer encoder module. This module uses a self-attention mechanism to... The global interaction relationships between all feature elements are modeled and fused, and a globally encoded feature vector is output. Following this, at least one fully connected layer is connected. The feature vector output from the Transformer encoder module is input into this fully connected layer. The fully connected layer maps the high-dimensional features to the three final prediction target dimensions through linear transformation and activation functions. The output of this fully connected layer is the final predicted value of yield strength, tensile strength, and elongation of the aluminum alloy mechanical property prediction model. Through the above steps, a multi-task regression prediction model capable of processing sequential process and component data separately, and fusing features through splicing and global attention, was constructed.

[0027] S4. Use the training set to train the aluminum alloy mechanical property prediction model, and use the validation set to perform hyperparameter tuning and training monitoring. The goal is to minimize the loss function between the predicted output of the aluminum alloy mechanical property prediction model and the measured mechanical property data, so as to obtain the trained aluminum alloy mechanical property prediction model.

[0028] Furthermore, in S4, the trained aluminum alloy mechanical property prediction model specifically includes: Initialize the network parameters of the aluminum alloy mechanical property prediction model and configure the optimizer and loss function required for training; The data in the training set are input into the aluminum alloy mechanical property prediction model in batches. The loss value between the model prediction output and the corresponding measured mechanical property data is calculated. The network parameters of the aluminum alloy mechanical property prediction model are updated by the optimizer through the backpropagation algorithm. After each round of training iterations or when the preset training period is reached, the performance of the current aluminum alloy mechanical property prediction model is evaluated using a validation set, and the training process is monitored based on the validation set performance, and early stopping judgment or hyperparameter adjustment is performed. When the training meets the preset stopping condition, the iteration stops, and the parameter state of the aluminum alloy mechanical property prediction model with the best validation set performance at this time is determined as the trained aluminum alloy mechanical property prediction model.

[0029] Furthermore, after each round of training iterations or upon reaching a preset training period, the performance of the current aluminum alloy mechanical property prediction model is evaluated using a validation set. Based on the validation set performance, the training process is monitored, and early stopping judgments or hyperparameter adjustments are performed, specifically including: When the preset training period arrives, all or part of the data in the validation set is used to input into the current aluminum alloy mechanical property prediction model, and the performance evaluation index value of the model on the validation set is calculated. The currently calculated performance evaluation index value is compared with the historical performance evaluation index value to monitor the trend of model performance changes; If the performance evaluation metric does not drop to a new minimum value within multiple consecutive training cycles, the early stop mechanism is triggered, sending a stop signal to the training process. If the average decrease in the performance evaluation index value over multiple consecutive training cycles is lower than the preset improvement threshold, the hyperparameter adjustment mechanism is triggered to adjust the learning rate of the optimizer or the relevant hyperparameters of the aluminum alloy mechanical property prediction model.

[0030] Specifically, firstly, the network parameters of the constructed aluminum alloy mechanical property prediction model are initialized. The Xavier initialization method is used to assign initial values ​​to all learnable weight and bias parameters in the model, and the optimizer and loss function required for training are configured. The Adam optimizer is selected as the algorithm for updating model parameters, and an initial learning rate is set for it, for example, = 0.001. Mean squared error is selected as the loss function. This loss function is used to quantify the difference between the model's predicted output and the measured mechanical properties. For a batch of data containing N samples, the formula for calculating the loss value L is: ; in, This represents the model's predicted value for the j-th mechanical performance index of the i-th sample. The values ​​represent the corresponding measured values, where j=1, 2, and 3 correspond to yield strength, tensile strength, and elongation, respectively. The iterative training loop begins. In each iteration, a batch of data is randomly selected from the training set. This batch contains a subset C of alloy composition features and a subset P of heat treatment process parameters, along with the corresponding measured mechanical properties of multiple samples. This batch of data is input into the current state of the aluminum alloy mechanical property prediction model. The model performs forward propagation calculations and outputs the predicted mechanical properties of all samples in this batch. Subsequently, according to the loss function calculation formula defined above, the loss value L on this batch of data is calculated. After obtaining the loss value L, the backpropagation algorithm is executed using the configured Adam optimizer. This algorithm automatically calculates the gradient of the loss value L with respect to each learnable parameter in the model. Then, based on the calculated gradient, its internally maintained momentum, and adaptive learning rate, the Adam optimizer updates all network parameters of the model according to its update rules. A training cycle is defined, for example, one complete iteration of the training set constitutes one training cycle. At the end of each training cycle, the current model's performance is evaluated using the entire validation set. Specifically, the evaluation method involves inputting the input features of all samples in the validation set into the current model to obtain predicted values, and calculating the model's performance evaluation metric V on the entire validation set. In this embodiment, the validation set mean squared error is used as this metric, and a training monitoring mechanism is initiated. The performance evaluation metric V at the end of each training cycle is recorded, and a historical best value is maintained. The initial value is infinity. The core operation for monitoring performance change trends is as follows: First, early stop judgment. Check whether the current performance evaluation index value V is less than the historical best value. If V < Then update V If the performance evaluation index V does not drop to a new minimum value within E consecutive training cycles, it is determined that the model performance no longer improves significantly, triggering the early stopping mechanism and immediately sending a stop signal to the training process. Second, learning rate adjustment. This involves calculating the average decrease in the performance evaluation metric value over the most recent M training epochs. The specific calculation formula is as follows: ,in This represents the index value for the current period t. The indicator value represents the period tM, and an improvement threshold is set. If the average decline... If the improvement rate falls below this threshold, the model performance is deemed to be improving too slowly, triggering the hyperparameter tuning mechanism. In this case, the current learning rate of the Adam optimizer is multiplied by a decay factor, i.e., the updated learning rate is equal to 1, in order to finer-tune the parameters in subsequent training. The training process continues until a preset stopping condition is met. The stopping condition is receiving the aforementioned early stop signal or reaching the preset maximum number of training cycles T. When training stops, the model parameter state corresponding to the training cycle with the optimal performance evaluation index V on the validation set is selected from the entire training history. The model in this parameter state is saved and designated as the final trained aluminum alloy mechanical property prediction model.

[0031] S5. Use the test set to evaluate the prediction accuracy and generalization ability of the trained aluminum alloy mechanical property prediction model, and use the measured aluminum alloy sample data that was not used in the training to conduct external experimental verification of the trained aluminum alloy mechanical property prediction model.

[0032] Furthermore, in S5, the evaluation of the prediction accuracy and generalization ability of the trained aluminum alloy mechanical property prediction model using the test set specifically includes: The aluminum alloy composition data and heat treatment process parameter data contained in the test set are input into the trained aluminum alloy mechanical property prediction model. The trained aluminum alloy mechanical property prediction model processes the input test set data and outputs the corresponding predicted values ​​of yield strength, tensile strength and elongation. The predicted values ​​are compared one-to-one with the pre-labeled measured mechanical properties data in the test set, and a set of performance evaluation indicators for quantifying the prediction error is calculated. Based on a set of performance evaluation indicators obtained through calculation, the prediction accuracy of the trained aluminum alloy mechanical property prediction model and its generalization ability on the data distribution represented by the test set are comprehensively evaluated.

[0033] Furthermore, in S5, the trained aluminum alloy mechanical property prediction model is validated externally using measured aluminum alloy sample data that was not used in the training process. This specifically includes: Obtain one or more measured aluminum alloy samples that are not included in the original dataset to form an external validation sample set; The alloy composition data and heat treatment process parameter data from the external validation sample set are input into the trained aluminum alloy mechanical property prediction model to obtain the model's predicted mechanical properties of the external samples. The predicted mechanical properties of the external samples are compared with the measured mechanical properties obtained by independent physical experiments in the corresponding external validation sample set. Based on the results of the comparative analysis, the accuracy of the trained aluminum alloy mechanical property prediction model in predicting real aluminum alloy samples that did not appear in the training was evaluated, thereby verifying its applicability and generalization ability in actual engineering scenarios.

[0034] Specifically, the predefined and retained test set is extracted. This test set contains multiple sets of pre-processed aluminum alloy composition data and heat treatment process parameter data, along with corresponding unnormalized measured mechanical property data. The input features of all samples in this test set are then input into a trained aluminum alloy mechanical property prediction model. The model performs forward propagation calculations and outputs the predicted yield strength, tensile strength, and elongation values ​​for each test sample. For each sample in the test set, the predicted values ​​of its three mechanical performance indicators are compared one-to-one with the corresponding measured values. Based on this, a set of standard performance evaluation indicators for quantifying prediction errors is calculated, specifically including the following three indicators: Mean absolute error: ; in, This represents the total number of samples in the test set. This represents the model's predicted value for a certain mechanical performance index of the i-th sample. This represents the corresponding measured value. This calculation requires separate calculations for the three indices: yield strength, tensile strength, and elongation. Root mean square error: ; This indicator also needs to be calculated separately for the three mechanical performance indicators; Coefficient of determination: ; in, This represents the average value of all measured values ​​of the mechanical performance index on the test set. The closer the value is to 1, the better the model's predictions fit the actual values. The MAE, RMSE, and sum values ​​of the three mechanical performance indicators were calculated and observed. These quantitative indicators were used to evaluate the prediction accuracy of the trained aluminum alloy mechanical property prediction model on unseen data represented by the test set, which are distributed similarly to the training data. High prediction accuracy and stable low errors for all three indicators indicate that the model has good generalization ability and can reliably handle new, similarly distributed data. Prepare a new set of measured aluminum alloy samples. The grade, composition, or heat treatment parameters of these samples should be ensured to be not included in the original dataset constructed above. Perform the same heat treatment process and mechanical property tests as in the previous steps on these samples, accurately recording the alloy composition, heat treatment parameters, and measured yield strength, tensile strength, and elongation values ​​for each sample. All samples constitute an external validation sample set. For each sample in the external validation sample set, perform data cleaning and normalization preprocessing on its alloy composition and heat treatment parameter data in the same manner as described above to form standardized input data. Input this standardized input data into the trained aluminum alloy mechanical property prediction model to obtain the model's predicted values ​​for three mechanical properties for each external sample. The predicted values ​​are compared with the actual values ​​measured through independent physical experiments on a sample-by-sample and index-by-index basis. The mean absolute error on this external validation sample set is calculated. Simultaneously, the correlation coefficient between the predicted and measured values ​​is calculated, based on... The absolute size of the external validation error and its relative difference from the evaluation metrics on the internal test set are used to comprehensively evaluate the model's prediction accuracy for real aluminum alloy samples that may have new distribution characteristics and were not present in the training. If the external validation error is similar to the internal test error and remains within the engineering-acceptable error range, it proves that the model not only has good data generalization ability but also practical engineering applicability, and can provide effective performance prediction support for the research and development of new materials or processes.

[0035] S6. Input the composition data and heat treatment process parameter data of the aluminum alloy to be predicted into the trained aluminum alloy mechanical property prediction model, and the trained aluminum alloy mechanical property prediction model outputs the corresponding predicted values ​​of yield strength, tensile strength and elongation.

[0036] Furthermore, in S6, the predicted values ​​of yield strength, tensile strength, and elongation output by the trained aluminum alloy mechanical property prediction model specifically include: Obtain the alloy element composition data of the aluminum alloy to be predicted and the heat treatment process parameter data to be used, and perform the same data preprocessing operation as in S2 to form standardized prediction input data. Standardized prediction input data is input into the trained aluminum alloy mechanical property prediction model; The trained aluminum alloy mechanical property prediction model performs forward propagation calculations on the standardized prediction input data and simultaneously outputs the corresponding predicted values ​​of yield strength, tensile strength, and elongation.

[0037] Specifically, the raw data of the aluminum alloy to be predicted is obtained. This data includes alloy element composition data and heat treatment process parameter data. The alloy element composition data specifically refers to the mass percentage content of each major element in the aluminum alloy to be predicted, forming a vector. The specific heat treatment process parameters include the planned heat treatment process steps and parameters, such as the solution treatment temperature. Solution time Aging temperature Time limit These parameters should be arranged in the order of the processes to form a sequence. To ensure that the model input and training data are in the same feature space, it is necessary to process the new data... And perform preprocessing operations. Specifically, check... and Check for any outliers that clearly exceed the scope of common engineering knowledge. If so, correct or remove them based on materials science principles. Check for any missing values. If missing values ​​are found, fill them in using professional knowledge or similar processes, or declare the data unusable. and For each feature in the dataset, normalization is performed using the minimum value min(x) and maximum value max(x) of the corresponding feature, calculated and saved in the steps above. The calculation formula is: ; in, Here, x represents the original value of a feature in the new data, and min(x) and max(x) are the minimum and maximum values ​​of that feature calculated from the training set during the training phase. It is the new value after normalization; After the above processing, a standardized subset of alloy composition characteristics is obtained. and heat treatment process parameter characteristic subset ; The standardized prediction input data, i.e. and The data is input into a pre-trained aluminum alloy mechanical property prediction model, and the model automatically performs forward propagation calculations according to the defined architecture. Specifically, Enter the Mamba module to extract sequence features. The KAN network module extracts nonlinear relationship features, the two are concatenated and then globally fused by the Transformer encoder module, and finally mapped through a fully connected layer. After the model's forward propagation calculation is completed, three values ​​are directly generated in its output layer, representing the model's prediction results of the mechanical properties of aluminum alloy under the input conditions, denoted as the predicted yield strength value. Predicted tensile strength and elongation prediction value ; These predicted values ​​are calculated by the model based on the learned complex mapping relationship between composition, process, and performance. They can serve as a rapid and quantitative reference for evaluating the mechanical properties of the aluminum alloy design scheme to be predicted. This completes the entire process of predicting the mechanical properties of new aluminum alloy compositions and process schemes using a trained deep learning model. When applying this model, simply ensure that the input data follows the same preprocessing specifications as the training data to obtain the corresponding multi-objective performance predictions.

[0038] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the mechanical properties of aluminum alloys based on deep learning, characterized in that, Includes the following steps: S1. Collect the original dataset containing various aluminum alloy composition data, heat treatment process parameter data and corresponding mechanical property test data. S2. Perform data cleaning and preprocessing on the original dataset to obtain a standard dataset, and divide the standard dataset into a training set, a validation set, and a test set. S3. Construct a fusion neural network model as a prediction model for the mechanical properties of aluminum alloys. The fusion neural network model includes a Mamba module for extracting process parameter sequence-dependent features, a KAN network module for extracting nonlinear interpretable relationships between composition and properties, and a Transformer encoder module for global feature fusion. S4. The aluminum alloy mechanical property prediction model is trained using the training set, and hyperparameters are tuned and training is monitored using the validation set. The goal is to minimize the loss function between the predicted output of the aluminum alloy mechanical property prediction model and the measured mechanical property data, so as to obtain the trained aluminum alloy mechanical property prediction model. S5. Use the test set to evaluate the prediction accuracy and generalization ability of the trained aluminum alloy mechanical property prediction model, and use the measured aluminum alloy sample data that was not involved in the training to conduct external experimental verification of the trained aluminum alloy mechanical property prediction model. S6. Input the composition data and heat treatment process parameter data of the aluminum alloy to be predicted into the trained aluminum alloy mechanical property prediction model, and the trained aluminum alloy mechanical property prediction model outputs the corresponding predicted values ​​of yield strength, tensile strength and elongation.

2. The method for predicting the mechanical properties of aluminum alloys based on deep learning according to claim 1, characterized in that, In S1, the raw dataset containing various aluminum alloy composition data, heat treatment process parameter data, and corresponding measured mechanical property data specifically includes: Collect alloy composition data, heat treatment process parameter data, and corresponding measured data of yield strength, tensile strength and elongation of various grades or series of aluminum alloys to form an initial data set; Each data record in the initial dataset is standardized into a structured data entry; The structured data entries from different sources are merged, and completely duplicate data entries are removed based on the combined identifier of alloy composition and heat treatment process parameters to obtain the original dataset.

3. The method for predicting the mechanical properties of aluminum alloys based on deep learning according to claim 1, characterized in that, In step S2, the original dataset undergoes data cleaning and preprocessing to obtain a standard dataset. The standard dataset is then divided into a training set, a validation set, and a test set. Specifically, this includes: The original dataset is subjected to outlier detection and processing, and data entries with missing values ​​are filled or deleted to obtain a cleaned dataset. The values ​​of alloy element composition data and heat treatment process parameter data in the cleaned dataset are normalized to ensure they fall within a uniform value range. The normalized dataset is divided into the training set, the validation set, and the test set according to a preset ratio.

4. The method for predicting the mechanical properties of aluminum alloys based on deep learning according to claim 1, characterized in that, In S3, constructing a fusion neural network model as a prediction model for the mechanical properties of aluminum alloys specifically includes: The input features of the aluminum alloy mechanical property prediction model are assigned to a subset of alloy element composition features and a subset of heat treatment process parameter features; The Mamba module is constructed and configured to receive a subset of the heat treatment process parameter features as input to extract the sequence dependency features of the process parameters. The KAN network module is constructed and configured to receive the alloy element composition feature subset as input to extract the nonlinear interpretable relationship between composition and performance. The sequence dependency features extracted by the Mamba module are concatenated with the nonlinear interpretable relation features extracted by the KAN network module to form a combined feature vector. The combined feature vector is input into the Transformer encoder module for global feature fusion and encoding. The output of the Transformer encoder module is mapped through at least one fully connected layer to output the prediction result of the aluminum alloy mechanical property prediction model.

5. The method for predicting the mechanical properties of aluminum alloys based on deep learning according to claim 1, characterized in that, In S4, the specifically obtained trained aluminum alloy mechanical property prediction model includes: Initialize the network parameters of the aluminum alloy mechanical property prediction model and configure the optimizer and loss function required for training; The data in the training set are input into the aluminum alloy mechanical property prediction model in batches. The loss value between the model prediction output and the corresponding measured mechanical property data is calculated. The network parameters of the aluminum alloy mechanical property prediction model are updated by the optimizer through the backpropagation algorithm. After each round of iterative training or when the preset training cycle is reached, the performance of the current aluminum alloy mechanical property prediction model is evaluated using the validation set, and the training process is monitored based on the validation set performance, and early stopping judgment or hyperparameter adjustment is performed. When the training meets the preset stopping condition, the iteration stops, and the parameter state of the aluminum alloy mechanical property prediction model with the best validation set performance at this time is determined as the trained aluminum alloy mechanical property prediction model.

6. The method for predicting the mechanical properties of aluminum alloys based on deep learning according to claim 5, characterized in that, The step of evaluating the performance of the current aluminum alloy mechanical property prediction model using the validation set after each round of training iterations or upon reaching a preset training period, and monitoring the training process based on the validation set performance, and performing early stopping judgment or adjusting hyperparameters specifically includes: When the preset training period arrives, all or part of the data in the validation set is used and input into the current aluminum alloy mechanical property prediction model to calculate the performance evaluation index value of the model on the validation set. The currently calculated performance evaluation index value is compared with the historical performance evaluation index value to monitor the trend of model performance changes; If the performance evaluation index value does not drop to a new lowest value within multiple consecutive training cycles, the early stop mechanism is triggered, and a stop signal is sent to the training process. If the average decrease of the performance evaluation index value over multiple consecutive training cycles is lower than the preset improvement threshold, the hyperparameter adjustment mechanism is triggered to adjust the learning rate of the optimizer or the relevant hyperparameters of the aluminum alloy mechanical property prediction model.

7. The method for predicting the mechanical properties of aluminum alloys based on deep learning according to claim 1, characterized in that, In step S5, evaluating the prediction accuracy and generalization ability of the trained aluminum alloy mechanical property prediction model using the test set specifically includes: The aluminum alloy composition data and heat treatment process parameter data contained in the test set are input into the trained aluminum alloy mechanical property prediction model. The trained aluminum alloy mechanical property prediction model processes the input test set data and outputs the corresponding predicted values ​​of yield strength, tensile strength and elongation. The predicted values ​​are compared one-to-one with the pre-labeled measured mechanical properties data in the test set to calculate a set of performance evaluation indicators for quantifying the prediction error. Based on a set of calculated performance evaluation indicators, the prediction accuracy of the trained aluminum alloy mechanical property prediction model and its generalization ability on the data distribution represented by the test set are comprehensively evaluated.

8. The method for predicting the mechanical properties of aluminum alloys based on deep learning according to claim 1, characterized in that, In step S5, the external experimental verification of the trained aluminum alloy mechanical property prediction model using untrained measured aluminum alloy sample data specifically includes: Obtain one or more measured aluminum alloy samples that are not included in the original dataset to form an external verification sample set; The alloy composition data and heat treatment process parameter data from the external verification sample set are input into the trained aluminum alloy mechanical property prediction model to obtain the model's predicted mechanical properties of the external samples. The predicted mechanical properties of the external samples are compared with the measured mechanical properties obtained by independent physical experiments in the corresponding external verification sample set. Based on the results of the comparative analysis, the accuracy of the trained aluminum alloy mechanical property prediction model in predicting real aluminum alloy samples that did not appear in the training is evaluated, thereby verifying its applicability and generalization ability in actual engineering scenarios.

9. The method for predicting the mechanical properties of aluminum alloys based on deep learning according to claim 1, characterized in that, In step S6, the predicted values ​​of yield strength, tensile strength, and elongation output by the trained aluminum alloy mechanical property prediction model specifically include: Obtain the alloy element composition data of the aluminum alloy to be predicted and the heat treatment process parameter data to be used, and perform the same data preprocessing operation as in S2 to form standardized prediction input data. The standardized prediction input data is input into the trained aluminum alloy mechanical property prediction model. The trained aluminum alloy mechanical property prediction model performs forward propagation calculations on the standardized prediction input data and simultaneously outputs the corresponding predicted values ​​of yield strength, tensile strength, and elongation.

Images