Material mechanical property prediction method and equipment based on transfer learning and ensemble learning, and medium

Abstract

The invention discloses a material mechanical property prediction method and device based on transfer learning and integrated learning and a medium, and belongs to the technical field of crossing of materials, mechanics and artificial intelligence. Aiming at the problems that a traditional material mechanical property test is long in period, high in sample requirement, high in cost and the like, rapid estimation of the material mechanical property of a material small-size sample is achieved through methods such as transfer learning and integrated learning, and the method specifically comprises the steps of preprocessing, modal enhancement, sample enhancement, model migration and combination generation. The method breaks through the limitation of a traditional mechanical test, realizes rapid prediction of the mechanical property of the material, can be widely applied to prototype iterative development and quality inspection of the material in the fields of aviation, automobiles and the like, and remarkably improves the design efficiency and detection efficiency of the material.

Description

Technical Field

[0001] The present invention relates to the fields of materials, mechanics and artificial intelligence, and in particular to a material mechanical property prediction method, device and medium based on transfer learning and ensemble learning. Background Art

[0002] Material mechanical testing is a crucial method for verifying macroscopic material properties during the design and manufacturing stages of the materials industry. Accurate verification of material mechanical properties is a fundamental requirement for product quality in manufacturing industries such as aviation, automotive, and energy, and ensures the safety and reliability of product structural designs. Traditional conventional material mechanical property testing methods (such as static tensile testing, dynamic tensile testing, and fatigue testing) require time-consuming preparation of standard-sized specimens and the scheduling of tests with standard testing machines. These methods present challenges such as long testing cycles, high sample requirements, and high costs, making them particularly challenging for irregularly shaped components and those experiencing material shortages.

[0003] From the perspective of mechanical testing, micro-destructive testing technologies based on small-size samples (such as small punching, nanoindentation, hardness testing, etc.) have made certain developments. However, since there is no direct and explicit correspondence between the original test data of many mechanical experiments and material properties (such as material stress-strain curves, material stress-life curves, material yield strength, material tensile strength, etc.), the conversion from test data to material mechanical properties generally relies on empirical or semi-empirical physical formulas or estimation based on modeling of a small amount of historical experimental data. Due to the complexity of the relationship conversion and the limitation of data scale, it is difficult to achieve accurate prediction of material mechanical properties.

[0004] From a machine learning perspective, traditional data augmentation methods typically rely on data rotation and cropping for sample-level transfer learning. This makes it difficult to effectively apply to data with physical meaning or textual data. Traditional transfer learning methods typically involve fixing model parameters and performing overall or partial parameter updates. Because high-quality data for transfer is often scarce, and the migration path is highly random due to the uncontrollable training process, the effectiveness of transfer learning is often unclear and often results in negative transfer, making it difficult to apply transfer learning in practical engineering applications. Furthermore, traditional ensemble learning, which simply weights the prediction results of single-modal models, lacks quantitative assessment of data reliability and requires further improvement and enhancement.

[0005] Based on this, in order to address the limitations of traditional mechanical testing, we invented a material performance and mechanical properties prediction method based on transfer learning and ensemble learning to achieve accurate and rapid prediction of material performance information. Summary of the Invention

[0006] In order to solve the problems existing in the technical background, the purpose of the present invention is to provide a material performance and mechanical properties prediction method, equipment and medium based on transfer learning and ensemble learning, to solve the limitations of traditional testing methods and the training and prediction problems caused by lack of data, and to achieve accurate and rapid prediction of material mechanical properties.

[0007] The mechanical properties of materials mentioned in the present invention generally refer to common properties that reflect the macroscopic mechanical properties of materials, such as material stress-strain curve, material stress-life curve, material yield strength, material tensile strength, etc.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions: The material mechanical properties prediction method based on transfer learning and ensemble learning includes the following steps: S1) Preprocessing: Preprocess the input data to obtain standardized material test data; S2) Modal enhancement: Modal enhancement is performed on the preprocessed data to obtain material test parameter information, material test curve information, and material test image information, expanding from a single modality to multimodal data; S3) Sample enhancement: Perform sample enhancement on each modality after modality enhancement, expanding from single-fidelity data to multi-fidelity data; S4) Model Transfer: Using transfer learning, we train and transfer models for each modality separately. First, we train a low-fidelity model based on the lowest-fidelity data after sample augmentation. Then, based on the low-fidelity model, we use higher-fidelity data to obtain higher-fidelity models, until all fidelity levels are exhausted. S5) Combination generation: Based on the models of each mode obtained by transfer learning, the mechanical properties data of the target material are generated separately, and the results are combined using an ensemble learning method to obtain the final prediction result.

[0009] Furthermore, in step S1), the input data includes data obtained by various mechanical testing methods of small-sized and conventional-sized materials.

[0010] Specifically, the mechanical testing methods include, but are not limited to, small punching, nanoindentation, hardness indentation testing, drop hammer impact, or related mechanical experimental methods. The specific data types obtained include, but are not limited to, time-series or non-time-series scattered point or curve data, such as punch load-displacement curves, indenter pressure-displacement curves, and other response data that can indirectly reflect material properties. Material types include, but are not limited to, metals, plastics, and composite materials.

[0011] Specifically, the standardization process refers to normalizing or regularizing the scattered points or curves of the material test data, using reference values ​​including the mean, peak point or fixed value of the data set.

[0012] Furthermore, in step S2), the material test parameter information refers to one or more characteristic values ​​obtained by feature extraction from the standardized material test data; the material test curve information refers to a curve obtained by upsampling or downsampling from the standardized material test data; and the material test image information refers to an image containing input data information formed by an indirect conversion or direct conversion method based on the standardized material test data.

[0013] Specifically, in the material test parameter information, the extraction method can use an autoregressive model such as a multi-layer perceptron, a convolutional neural network or a combination thereof for extraction, or can use traditional experience or physical formulas for approximate extraction.

[0014] Specifically, in the material test curve information, the downsampling method is generally oriented towards curve data. This is because the amount of data from general material tests is generally large, and the data length is compressed while retaining key features to reduce the input dimension for subsequent machine learning methods; the upsampling method is generally oriented towards scattered data, and the characteristic information of the input scattered data is restored from the scattered data to form a continuous and complete curve.

[0015] Specifically, indirect conversion of material test image information refers to the use of autoregressive models such as multi-layer perceptrons, convolutional neural networks, or a combination thereof to perform structured encoding on the input data to generate image information. However, in this case, the image cannot directly visually represent the scattered point data or curve data information. Direct conversion refers to the conversion of the material test curve into a color or grayscale visual image. During direct conversion, the color space is RGB or CMYK. In this case, the image can directly reflect the geometry or other additional information of the scattered point data or curve data through color and shape.

[0016] Furthermore, in step S3), the material test data is sampled and expanded by using simulation calculation and theoretical calculation methods to obtain theoretical data, simulation data and experimental data of different fidelity.

[0017] Specifically, theoretical data refers to direct estimation of material mechanical properties through physical formulas. For example, it can approximate the static, dynamic, and fatigue mechanical properties of a material using material elements such as carbon, iron, and sulfur content. This allows for the generation of static stress-strain curves, dynamic stress-strain curves, and stress-life curves, establishing an approximate correlation between virtual material test data and actual material mechanical properties. This theoretical data has the lowest fidelity but also the lowest data cost.

[0018] Specifically, simulation data refers to virtual experimental testing using the finite volume method or finite element method, simulating the mechanical response of a specimen under desired experimental conditions (e.g., scattered points or curves), establishing an approximate correlation between input data and the material's mechanical properties. This simulation data offers moderate fidelity and data cost.

[0019] Specifically, experimental data refers to physical testing of materials using actual small-scale or conventional specimens to obtain the mechanical response (e.g., scattered points or curves) under these test conditions. This data is then used to obtain mechanical property data from querying a material database or performing actual experimental testing on reference standard specimens, establishing an accurate correlation between the input data and the material's mechanical properties. This experimental data offers the highest fidelity, but at a higher cost, and serves as a true reference.

[0020] Specifically, the fidelity refers to an indicator that measures the accuracy and reliability of a set of material data. High fidelity means that the data has high accuracy and high reliability, and low fidelity means that the data has low accuracy and low reliability.

[0021] Furthermore, in step S4), the network used by each model includes a multi-layer perceptron, a convolutional neural network, a diffusion model network or a combination thereof; the loss function of each model network includes a semantic loss term, an approximate sample enhancement term, a feature preservation term, a random noise stabilization term and an adversarial enhancement term.

[0022] Furthermore, in step S4), specifically, the output of the low-fidelity model serves as an additional input to the high-fidelity model.

[0023] Specifically, in the described model migration, in the iterative migration and update of data with different fidelity, the networks adopted by each model include multi-layer perceptrons, convolutional neural networks, diffusion networks or a combination thereof; in the model migration, the weights of the original migration model remain unchanged or remain variable in subsequent training but are constrained by the original fidelity data, and the weights of the original migration model are used as the initial weights of the current migration model. In particular, unlike traditional methods, the current migration model uses the output of the original migration model as an additional input item, so that learning is more focused on learning the differences between low-fidelity and high-fidelity data, reducing the difficulty of transfer learning and the demand for training data scale, and can improve the prediction accuracy and reliability of the final model. The original transfer model specifically refers to a model pre-trained with lower-fidelity data, while the current transfer model specifically refers to a model to be trained with higher-fidelity data. In particular, unlike traditional methods, the loss function of each model network adds an approximate sample enhancement term, a feature preservation term, a random noise stabilization term, and an adversarial enhancement term to the traditional commonly used loss function, namely the semantic loss term. These additional loss function terms enable transfer learning to naturally generate intermediate approximate sample data that is different from the high-fidelity data and the low-fidelity data during training during gradient backpropagation. Among them, a) the approximate sample enhancement term is used to approximate and temporarily save the approximate sample data. This loss term needs to calculate the gradient in the gradient backpropagation of machine learning to continuously approximate and update the approximate sample data; b) the feature preservation term is used to continuously and dynamically fine-tune the parameters of the low-fidelity model during training while keeping the prediction accuracy of the low-fidelity model basically unchanged; c) the random noise stabilization term is used to stabilize the construction process of the approximate sample data while avoiding entering a local optimal solution; d) the adversarial enhancement term is used to assist in the generation of approximate sample data, making its distribution space closer to the distribution space of the low-fidelity and high-fidelity data.

[0024] The semantic loss term refers to the value obtained by directly comparing the material curve or scatter point predicted by the current migration model with the difference of the real reference value, including the difference of image pixels or actual physical values.

[0025] More specifically, the loss function that is different from traditional transfer learning is described as follows: The approximate sample enhancement item refers to the difference between the current migration model prediction value and the original migration model prediction value, including the difference in image pixels or actual physical values. In particular, the output result of the current migration model prediction is saved as a learnable parameter, which can be dynamically updated during model training and used to replace the output part of the original migration model training data after being updated. The replaced data is then used as the training data for the next round.

[0026] The feature preservation item refers to the difference between the predicted value predicted by the original migration model and the output value of the training data referenced by the original migration model.

[0027] The random noise stabilization term refers to the value obtained by adding a random noise term to the value obtained by comparing the current migration model's predicted value (including image pixels or actual physical values) with the actual reference value. The data type includes image pixels or actual physical values. The mean and variance of the random noise are learnable parameters.

[0028] The adversarial enhancement term refers to a measure that describes the similarity between the current migration model's predicted value and the dataset consisting of the current training data and the original migration model's training data, and is used to determine whether the current migration model's predicted value deviates from the true distribution. The adversarial enhancement term can be implemented through an adversarial neural network, including constructing a discriminator network and calculating adversarial loss. In the discriminator network, the input is the predicted value of the current migration model, the current training data, and the original migration model's training data, and the output is a probability score for the authenticity of the data. In the adversarial loss calculation, the discriminator is used to output a low score (determined as "false") for the current migration model's predicted value and a high score (determined as "true") for the data in the dataset consisting of the current training data and the original migration model's training data.

[0029] Specifically, the learnable parameters in the model migration are based on specific parameter values ​​or settings that can be adjusted, including but not limited to one or more combinations of neuron weights, neuron biases, number of network layers, network structure, learning rate, optimizer, batch size, activation function, and number of iterations.

[0030] Furthermore, in step S5), the ensemble learning method adopts a weighted averaging method, in which the prediction results of each modality are weighted according to their confidence or model accuracy. The result after weighted averaging is used as the input of the combined meta-model, and is trained with a high-fidelity data set to secondary correct the combined meta-model. The final prediction result is the output value of the combined meta-model.

[0031] Specifically, the ensemble learning method is a stacked generalization based on weighted averaging, which is different from traditional ensemble learning that generally adopts direct weighting or direct parameterized superposition. The weighted averaging is to assign weights to the prediction results of each modality according to its confidence or model accuracy, and the confidence or model accuracy is determined by Dropout based on the Monte Carlo method. In addition, the output of each modality model after weighted averaging also needs to input a combined meta-model with trainable parameters for secondary processing to obtain the final prediction result.

[0032] The weighted average assigns weights based on their confidence or model accuracy. The weights are determined through cross-validation or historical performance data. The calculation formula is:

[0033] in, For the final prediction result, For the i The weight of the modal model, For the i The predicted output of each mode, n is the total number of modes. The weights of each model are calculated using a Monte Carlo Dropout method. Unlike traditional Dropout methods, this method selects the weights of certain hidden layer nodes that need to be closed in the model based on Monte Carlo sampling, and obtains confidence-based weight estimates through statistics.

[0034] The stacked generalization uses a combined meta-model as a combiner; the combined meta-model input is the weighted average output of each modal model, and is trained through a high-fidelity data set to achieve the effect of secondary correction of the model. The meta-model types include models based on linear regression, random forest or neural network to optimize the final prediction accuracy.

[0035] The present invention further provides an electronic device, comprising: one or more processors; a memory for storing one or more computer programs, The one or more processors execute the one or more computer programs to implement the steps of the above method.

[0036] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.

[0037] The present invention also provides a computer program product, comprising a computer program, wherein the computer program implements the steps of the above method when executed by a processor.

[0038] Beneficial effects:

[0039] 1. The present invention innovatively expands the original test data into a multimodal expression of material parameter information, material curve information and material image information, constructs a new multi-fidelity and multimodal transfer learning framework for theoretical, simulation and experimental data, and uses the low-fidelity model as the knowledge basis to gradually transfer physical laws to the high-fidelity model. By constructing special approximate sample enhancement terms and adversarial loss terms in the loss function, approximate intermediate samples are generated to the maximum extent, thereby making full use of various fidelity data, significantly alleviating the bottleneck of the lack of high-fidelity experimental data, and making it possible for the model to accurately predict the mechanical properties of materials under limited real sample conditions, overcoming the problems of negative transfer and transfer instability of traditional transfer learning methods as much as possible, and improving the transfer learning efficiency of the model, breaking through the traditional method's dependence on large-scale labeled data.

[0040] 2. The present invention adopts the idea of ​​ensemble learning to perform combined prediction on the results of different modalities. In ensemble learning, a Monte Carlo weighted mechanism is proposed to dynamically evaluate the confidence of each modality, and a trainable combined meta-model is constructed to correct the weighted results, making full use of the effective information of each modality and compensating for the shortcomings of different types of machine learning models.

[0041] 3. This invention breaks through the limitations of traditional standard material experimental testing methods and proposes a new path for predicting the mechanical properties of materials. It can significantly reduce the requirements for material size and test specifications. It can achieve rapid material mechanical property evaluation in typical scenarios such as special-shaped components in automobiles, energy and aviation, rapid iterative development of materials, and material shortages, providing efficient and reliable technical support for product safety design. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flowchart of the overall design method according to an embodiment of the present invention; Figure 2 The material test data before standardization in the embodiment of the present invention; Figure 3 Schematic diagram of high-fidelity data after modal enhancement in an embodiment of the present invention; Figure 4 Schematic diagram of low-fidelity data after sample enhancement in an embodiment of the present invention; Figure 5 Schematic diagram of the transfer learning framework in an embodiment of the present invention; Figure 6 This is an illustration of approximate sample data obtained by introducing a new loss function in transfer learning in an embodiment of the present invention; Figure 7 4 is a comparison chart of the prediction effects of the embodiment of the present invention and the comparative method. DETAILED DESCRIPTION

[0043] The following examples are provided for a better understanding of the present invention and are not intended to limit the best mode of implementation. They do not limit the content and scope of protection of the present invention. Any product identical or similar to the present invention obtained by anyone under the guidance of the present invention or by combining the features of the present invention with other prior arts shall fall within the scope of protection of the present invention.

[0044] This implementation case includes the following steps: Step 1: Preprocessing. In this embodiment, fixed values ​​are used to standardize the material test data from the experiment. The experimental method adopts the standard test method for small punching experiments on metal materials. The materials are aluminum-magnesium alloy, low-alloy steel, and duplex steel. The material thickness is 1mm. The material data type is the load-displacement curve of the material obtained in the experiment. The standardization uses a fixed value of 30kN for the load and a fixed value of 5mm for the displacement. The fixed values ​​of standardization are related to the maximum load range of the test machine and the expected subsequent prediction range.

[0045] The load displacement curve is a curve of the relationship between the punching impact load and the punch displacement obtained after the standard test of the small punching experiment of the metal material. The original curve data is as follows Figure 2 As shown, Figure 2 a) is the material test parameter information, Figure 2 b) is the material test curve information, Figure 2 c) Material test image information.

[0046] Step 2: Modal Enhancement. In this embodiment, modal enhancement is performed on the load-displacement curves obtained after the three material tests described in Step 1 to obtain material test parameter information (parameter mode), material test curve information (curve mode), and material test image information (image mode), a total of three modes. Among them, the material test parameter information uses the empirical method to extract key information, and the slope of the initial straight line segment of the load-displacement curve is E, the peak load value is P, and the displacement corresponding to the peak load is S, a total of three key empirical parameters; the material test curve information, through the downsampling method to compress the original load-displacement curve, only retaining the points starting from the load and displacement 0, with a displacement interval of 0.025mm (the maximum displacement is 5mm) as the sampling distance, a total of 201 data points, such as (0,0), (0.2KN, 0.025mm)... (10.2KN, 5mm); the material test image information, based on the closed area enclosed by the projection of the last coordinate point in the original load-displacement curve (displacement is the X-axis, load is the Y-axis) to the X-axis, forms a picture, the picture color uses the RGB color channel, in this embodiment it is black (255,255,255), and different color channel values ​​can also be assigned according to the need for additional information expansion. Extended modal information such as Figure 3 As shown, Figure 3 a) is the material test parameter information, Figure 3 b) is the material test curve information, Figure 3 c) Material test image information.

[0047] Step 3: Sample enhancement. Perform sample enhancement on each mode after modal enhancement, i.e., material test parameter information, material test curve information, and material test image information, to form multi-fidelity data. This embodiment includes two fidelities, i.e., low-fidelity data and high-fidelity data. High-fidelity data is paired data of the test data (load-displacement curve) obtained from step 2 and its corresponding real material mechanical properties (stress-strain curve). For this embodiment, the sample enhancement is performed by simulation to obtain low-fidelity data. Assume that the stress-strain curve of the material can be described by the following formula:

[0048] in is stress, It's strain. H is the material hardening coefficient, n is the material hardening index. Specifically, according to the approximate range of common metal material properties, 2000 sets of data are generated through finite element simulation, wherein the material hardening index varies from 0.02 to 0.5, and the material hardening coefficient varies from 100 to 2000. 2000 sets of data are randomly sampled from the combination of the material hardening index and the material hardening coefficient to form a parameter group that approximately covers all approximate material properties. According to these 2000 sets of parameters, the strain is defined to vary from 0 to 1, and a point is taken at intervals of 0.02 to obtain the final 2000 sets of stress-strain relationship curves. According to these 2000 sets of stress-strain relationship curves, 2000 different virtual materials are defined, and standard small punching experimental tests are simulated in finite element simulation to obtain the corresponding load-displacement curves, and finally form pairs of low-fidelity data including load-displacement curves and stress-strain relationships. According to the method described in step two, the obtained low-fidelity data is expanded into three modal data, namely material test parameter information, material test curve information, and material test image information, such as Figure 4 As shown, Figure 4 a) is the material test parameter information, Figure 4 b) is the material test curve information, Figure 4 c) Material test image information.

[0049] Step 4: Model migration. The basic framework of model migration is as follows Figure 5As shown. In this embodiment, low-fidelity model training is first performed, that is, the low-fidelity data obtained after sample enhancement in step three is used to train parameter models, curve models, and image models based on neural networks, Seq2Seq networks, and diffusion networks, respectively. Among them, the neural network has 4 hidden layers, the number of neurons in each layer is 100, and the activation function uses ReLU; the Seq2Seq network is composed of 2 convolutional neural networks, namely a convolutional network and a deconvolutional network. The convolutional network adopts a 3-layer one-dimensional convolution structure, the convolution kernel size of each layer is 5, the step size is 2, the padding is 2, the number of channels is 32, 64, and 128 respectively, and the activation function uses ReLU. The deconvolution network adopts a three-layer one-dimensional deconvolution structure. The convolution kernel size of each layer is 5, the stride is 2, the padding is 2, the number of channels is 128, 64, and 32 respectively, and the activation function uses ReLU; finally, a one-dimensional convolution layer (kernel size 1, stride 1) is added to reduce the number of channels to 1 and output the target sequence; the diffusion network adopts a noise prediction network based on the U-Net architecture, which includes 4 stages of downsampling and upsampling. Each stage uses a residual block (including 2 convolution layers and group normalization), and the middle layer is equipped with a self-attention mechanism. The time step information is injected into each residual block through sinusoidal position embedding.

[0050] After the low-fidelity model is trained, transfer learning is performed on the low-fidelity model using higher-fidelity data to obtain a high-fidelity model. In this embodiment, all learnable parameters of the low-fidelity model can also be dynamically adjusted during high-fidelity model training, and the output of the low-fidelity model is used as additional input to the high-fidelity model. The loss function defines a semantic loss term, an approximate sample enhancement term, a feature preservation term, a random noise stabilization term, and an adversarial enhancement term. The total loss function is the weighted sum of each component.

[0051] The semantic loss term, with a weight coefficient of 1, is the sum of the squares of the difference between the current high-fidelity model prediction value and the actual material properties. The approximate sample enhancement term, with a weight coefficient of 0.5, is the absolute value of the difference between the current high-fidelity model prediction value and the low-fidelity model prediction value. The feature preservation term, with a weight coefficient of 0.1, refers to the difference between the low-fidelity model prediction value and the output value of the low-fidelity original training data. The random noise stabilization term adds a random noise term based on the difference between the high-fidelity model prediction value and the actual material properties. The mean and variance of the random noise are learnable parameters. The adversarial enhancement term, with a weight coefficient of 1, refers to a measure of the similarity between the high-fidelity model prediction value and the data set consisting of the actual material properties data and the low-fidelity data. In this embodiment, it is implemented through an adversarial neural network, including constructing a discriminator network and calculating the adversarial loss. In the discriminator network, the input is the prediction value of the current transfer model, the current training data and the original transfer model training data, and the output is a probability score of the data authenticity. In the adversarial loss calculation, the discriminator is used to output a low score (judged as "false") for the current migration model's prediction value, and a high score (judged as "true") for the data set consisting of the current training data and the original migration model training data.

[0052] Figure 6 This is an illustration of approximate sample data obtained by introducing a new loss function in transfer learning in an embodiment of the present invention, where: Figure 6 a) is the material test parameter information, Figure 6 b) is the material test curve information, Figure 6 c) Material test image information.

[0053] Step 5: Combination Generation: In this embodiment, high-fidelity models of three modalities are obtained through training according to step 4, and the predicted values ​​of the mechanical properties of the material are output by fusing the prediction results of each modal model.

[0054] First, the model results are weighted, and the weighting coefficient is determined by Dropout based on the Monte Carlo method: multiple forward propagations are performed on each high-fidelity model (100 times in this embodiment), and the Dropout layer is kept activated during each forward propagation to introduce randomness, and the prediction results of each time are recorded; the variance of each model prediction result is calculated as its uncertainty measure, and the weighting coefficient is inversely proportional to the uncertainty (that is, the smaller the variance, the larger the weight), and finally the weight is normalized to a probability distribution through the Softmax function.

[0055] The weighted results are then stacked and generalized, feeding them into a stacking generalization module. This stacking generalization uses a combined metamodel as a combiner, trained on high-fidelity data. In this example, the metamodel employs a two-layer fully connected neural network with 50 hidden layer neurons and a Reluctant Unit (ReLU) activation function. This minimizes the mean squared error (MSE) on the high-fidelity data to optimize the final prediction accuracy. The final material mechanical property predictions are the output of the combined metamodel.

[0056] In this embodiment, for the convenience of comparing the results, the stress-strain curve is used as a reference to illustrate the effect of the embodiment. The predicted effect of the material stress-strain curve can be seen in Table 1 and Figure 7 .

[0057] Table 1 Comparison of the prediction effect of the embodiment of the present invention on key quantities with comparison method 1, comparison method 2 and real reference

[0058] Among them, the prediction result of the method of the embodiment of the present invention is the output value of the combined meta-model finally obtained through steps one to five, using the model framework and loss function of the embodiment of the present invention; the prediction result of comparison method 1 is the use of a traditional transfer learning method, that is, the loss function uses general semantic loss, and the model framework does not use the output of the low-fidelity model as an additional input of the high-fidelity model; the prediction result of comparison method 2 is the use of a transfer learning method with an improved model framework, that is, the loss function uses general semantic loss, but the model framework is consistent with the framework described in the embodiment of the present invention, that is, the output of the low-fidelity model is used as an additional input of the high-fidelity model; the real reference value comes from the experiment, belongs to high-fidelity data, and is the standard for benchmark verification.

[0059] Figure 7 This is a comparison chart of the prediction effects of the embodiment of the present invention, where Figure 7 a) is a comparison chart of the embodiment of the present invention and the real reference results, Figure 7 b) is a comparison chart of comparison method 1 and the real reference results. Figure 7 c) is a comparison chart of comparison method 2 and the real reference results.

[0060] From Table 1 and Figure 7 As can be seen, the embodiment of the present invention, due to its use of a new model architecture and loss function, achieves a low final prediction error of less than 5%. In contrast, the average prediction error of Comparison Method 1, based on traditional transfer learning, generally exceeds 10%. While Comparison Method 2 shows a slight improvement over Comparison Method 1, its average prediction error remains around 10% due to the lack of transfer learning aided by similar samples. To ensure a consistent comparison benchmark, all of the aforementioned methods utilize the ensemble learning method described in this invention.

[0061] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principles and purpose of the present invention. The scope of protection of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. All implementation schemes within its scope shall be subject to the constraints of the present invention.

Claims

1. A material mechanical properties prediction method based on transfer learning and ensemble learning, characterized in that: The steps include: S1) Preprocessing: Preprocess the input data to obtain standardized material test data; S2) Modal enhancement: Modal enhancement is performed on the preprocessed data to obtain material test parameter information, material test curve information, and material test image information, expanding from a single modality to multimodal data; S3) Sample enhancement: Perform sample enhancement on each modality after modality enhancement, expanding from single-fidelity data to multi-fidelity data; S4) Model transfer: Using transfer learning, the model of each modality is trained and transferred separately; First, a low-fidelity model is trained based on the lowest-fidelity data after sample augmentation. Then, based on the low-fidelity model, higher-fidelity data is used in turn to obtain a higher-fidelity model until all fidelity levels are traversed. S5) Combination generation: Based on the models of each mode obtained by transfer learning, the mechanical properties data of the target material are generated separately, and the results are combined using an ensemble learning method to obtain the final prediction result.

2. The method according to claim 1, characterized in that In step S1), the input data includes data obtained by various mechanical testing methods of small-sized and conventional-sized materials.

3. The method according to claim 1, characterized in that In step S2), the material test parameter information refers to one or more characteristic values ​​obtained by feature extraction from the standardized material test data; the material test curve information refers to a curve obtained by upsampling or downsampling from the standardized material test data; and the material test image information refers to an image containing input data information formed by an indirect conversion or direct conversion method based on the standardized material test data.

4. The method according to claim 1, wherein In step S3), the material test data is sampled and expanded by using simulation calculation and theoretical calculation methods to obtain theoretical data, simulation data and experimental data of different fidelity.

5. The method according to claim 1, wherein In step S4), the network used by each model includes a multi-layer perceptron, a convolutional neural network, a diffusion model network, or a combination thereof; the loss function of each model network includes a semantic loss term, an approximate sample enhancement term, a feature preservation term, a random noise stabilization term, and an adversarial enhancement term.

6. The method according to claim 1, characterized in that In step S4), specifically, the output of the low-fidelity model serves as an additional input to the high-fidelity model.

7. The method according to claim 1, characterized in that In step S5), the ensemble learning method adopts a weighted averaging method, in which the prediction results of each modality are weighted according to their confidence or model accuracy. The result after weighted averaging is used as the input of the combined meta-model, and is trained with a high-fidelity data set to secondary correct the combined meta-model. The final prediction result is the output value of the combined meta-model.

8. An electronic device comprising: one or more processors; a memory for storing one or more computer programs, The method further comprises: executing the one or more processors to execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images