Titanium alloy performance prediction model construction method, application method and device

Abstract

This invention provides a method, application method, and apparatus for constructing a titanium alloy performance prediction model. The construction method includes: constructing a multimodal performance dataset based on titanium alloy samples from different heat treatment batches; constructing auxiliary image data based on microstructure images of titanium alloys acquired under different heat treatment conditions and their corresponding microstructure category labels; training a visual backbone network on the auxiliary image data to perform microstructure category classification tasks, obtaining a visual feature extractor for extracting morphological features of titanium alloys; transferring the visual feature extractor to a pre-constructed initial neural network model, and iteratively training the initial neural network model based on the multimodal performance dataset until the loss function converges or a preset number of iterations is reached, determining that the iterative training is complete, and obtaining a titanium alloy performance prediction model for outputting mechanical property prediction results. This invention can improve the accuracy and stability of titanium alloy performance prediction under small sample conditions.

Description

Technical Field

[0001] This invention relates to the fields of material performance evaluation and artificial intelligence technology, and in particular to a method, application method and apparatus for constructing a performance prediction model for titanium alloys. Background Technology

[0002] Titanium alloys are a class of metallic structural materials possessing ultra-high specific strength, excellent corrosion resistance, and stable high-temperature mechanical properties. They are irreplaceable core materials in aerospace and high-end equipment manufacturing, and are widely used in critical load-bearing components of aircraft. Plane strain fracture toughness is a core indicator for evaluating the service safety of high-strength titanium alloys, directly reflecting the material's ability to resist crack instability and propagation. Its value is jointly determined by alloy composition, heat treatment regime, and microstructure. Different solution treatment and aging processes alter the morphology, size, and phase interface distribution of the α phase, β matrix, grain boundary α, and secondary α phase within the material, directly controlling the crack initiation location and propagation path. Therefore, establishing a mechanical property prediction model that can simultaneously integrate heat treatment process and microstructure information, clarifying the intrinsic evolution law between process, microstructure, and fracture toughness, has extremely high engineering value for optimizing titanium alloy heat treatment processes, assessing component service safety, and shortening the development cycle of new materials.

[0003] Current performance evaluation of titanium alloys mainly falls into two categories: traditional experimental testing and data-driven machine learning. While traditional mechanical testing yields reliable results, it requires the mass production of standard specimens and multiple sets of repeated fracture toughness tests, resulting in drawbacks such as high material consumption, long testing cycles, and high overall R&D costs. Data-driven machine learning, which uses deep learning to predict material properties, employs two single-modal modeling approaches: one relies on training convolutional networks using microscopic images such as SEM and EBSD to extract visual features of microstructure and phase distribution, but this method cannot incorporate the historical information of microstructure evolution carried by heat treatment processes; the other uses structured process parameters such as temperature and duration for modeling, which, while quantifying the impact of process variables, lacks an intuitive representation of the true microstructure and local microscale differences. A single data modality is insufficient to fully characterize the coupling relationship between process, microstructure, and fracture toughness. When sample size is insufficient, heat treatment regimes vary significantly, and image magnification scales are inconsistent, problems such as overfitting, low prediction accuracy, and poor generalization stability easily arise, making it difficult to meet the engineering requirements of rapid iteration of heat treatment regimes and accurate prediction of fracture toughness for titanium alloys. Summary of the Invention

[0004] The present invention aims to provide a method for constructing, applying, and using a titanium alloy performance prediction model to solve the above-mentioned technical problems and improve the accuracy and stability of titanium alloy performance prediction under small sample conditions.

[0005] To address the aforementioned technical problems, this invention provides a method for constructing a titanium alloy performance prediction model, comprising: A multimodal performance dataset was constructed based on titanium alloy samples from different heat treatment batches; each titanium alloy sample included associated heat treatment process parameters, mechanical property test labels, microstructure images, and / or microstructure image sub-blocks obtained by random cropping based on the microstructure images. Auxiliary image data was constructed based on the microstructure images of titanium alloys acquired under different heat treatment conditions and their corresponding microstructure category labels; among them, the microstructure images of titanium alloys include a variety of typical microstructure morphologies; The visual backbone network was trained on the auxiliary image data to perform an organizational category classification task, resulting in a visual feature extractor for extracting the morphological features of titanium alloys. The visual feature extractor is transferred to a pre-built initial neural network model. The initial neural network model is iteratively trained based on a multimodal performance dataset until the loss function converges or the preset number of iterations is reached. The iterative training is then considered complete, resulting in a titanium alloy performance prediction model for outputting mechanical performance prediction results. The initial neural network model includes a visual feature mapping module, a FiLM conditional modulation fusion module, and a regression prediction network connected in sequence. The visual feature mapping module maps the microstructure image feature vectors output by the visual feature extractor to a preset dimension to obtain image modal features; the FiLM conditional modulation fusion module recalibrates the image modal features based on conditional modulation parameters generated from heat treatment process parameters; and the regression prediction network makes predictions based on the recalibrated image modal features, outputting normalized mechanical property prediction results. In the above scheme, a multimodal dataset binding heat treatment process, mechanical labels, and microscopic images is constructed to address the problem of incomplete representation of single-class data. Randomly cropping image sub-blocks can expand the sample base. Simultaneously, auxiliary image data with microstructure category labels is constructed. Based on this auxiliary data, the visual backbone classification pre-training is completed, allowing the visual backbone to learn the specific microstructure texture and morphological features of titanium alloys in advance, forming a transferable feature extractor. This eliminates the need to train deep networks from scratch, alleviating overfitting problems with small samples. Furthermore, the initial model adopts a hierarchical module design, unifying the feature dimension of visual feature mapping to ensure matching of subsequent modulation operations. Unlike simple splicing, the FiLM module recalibrates the image modal features based on conditional modulation parameters generated from heat treatment process parameters, allowing the process to dynamically correct morphological features as a constraint, fully exploring the intrinsic relationship between process, microstructure, and performance. Finally, a regression network completes the mechanical performance fitting output. By leveraging pre-trained transfer learning to reduce the difficulty of training with small samples, and by utilizing multimodal complementary information to improve representation capabilities, deep fusion of two types of data is achieved through FiLM conditional modulation. Compared with single-modal and simple fusion schemes, this can effectively improve the accuracy and generalization stability of predicting the mechanical properties of titanium alloys, and reduce experimental testing costs.

[0006] In one implementation, a multimodal performance dataset is constructed based on titanium alloy samples from different heat-treated batches, specifically including: Obtain titanium alloy samples from different heat-treated batches; Microstructure images of each titanium alloy sample were obtained, and the effective microstructure areas in the microstructure images were randomly cropped to obtain several microstructure image sub-blocks. Obtain the heat treatment process parameters for each titanium alloy sample; wherein, the heat treatment process parameters include at least one of the following: solution type, solution temperature, solution time, solution cooling method, aging temperature, aging time, and aging cooling method; Obtain the mechanical property test label for each titanium alloy sample; wherein the mechanical property test label includes at least one of fracture toughness, tensile strength, yield strength, elongation after fracture, reduction of area, impact toughness, fatigue life and fatigue crack propagation performance; Based on the unique sample number of each titanium alloy sample, the heat treatment process parameters, mechanical property test labels, microstructure images, and / or microstructure image sub-blocks of the same titanium alloy sample are associated as a multimodal sample; wherein, the data expression of the multimodal sample is: ; In the formula, For the first Data records for one titanium alloy sample; For the first The first titanium alloy sample corresponding to the first Zhang microscopic tissue images or image sub-blocks; For the first Number of microstructure images or image sub-blocks corresponding to each titanium alloy sample; For the first Heat treatment process parameters for one titanium alloy sample; For the first Mechanical property test labels for individual titanium alloy samples; A multimodal performance dataset is constructed based on multimodal samples; the multimodal samples undergo integrity checks and consistency verification.

[0007] In the above scheme, multiple batches of heat-treated titanium alloy samples are collected, microscopic images are acquired simultaneously, and sub-blocks are cropped to expand the image sample size. All-dimensional heat treatment parameters and various mechanical property test labels are collected. Based on the unique number, three types of data are bound to form standardized multimodal samples, which are then integrated into a dataset after integrity and consistency verification. Multiple batches of samples cover differentiated processes, allowing the model to fully learn the evolution laws of processes, microstructures, and properties. Random image cropping expands the data and enriches the microscopic observation field when the number of physical samples is limited. Multiple types of processes and performance parameters adapt to diverse prediction needs. The multimodal samples bound from the same source simultaneously carry process evolution information and microscopic morphological features, making up for the lack of information in single-modal data.

[0008] In one implementation, auxiliary image data is constructed based on microstructure images of titanium alloys acquired under different heat treatment conditions and their corresponding microstructure category labels, specifically including: Collect images of the microstructure of titanium alloys under different heat treatment conditions and their corresponding microstructure category labels; The cracked areas in the microstructure image of the titanium alloy are marked, and the microstructure image of the titanium alloy is cropped by a pixel sliding window of a preset size to obtain several image sub-blocks that do not contain the cracked areas. Image sub-blocks derived from the same titanium alloy microstructure image are divided into the same data subset, and auxiliary image data is constructed based on the data subset.

[0009] In the above scheme, labeled microscopic images of titanium alloys with multiple heat treatment processes and typical morphologies are collected as basic materials. First, image sub-blocks containing crack interference areas are labeled and removed to avoid crack-irrelevant features interfering with the network's learning of the intrinsic microstructure and morphology of titanium alloys. Then, all sub-blocks generated from the same original image are grouped into the same subset to prevent data leakage caused by splitting images from the same source. The auxiliary image data constructed in this way is free of noise interference and has a rigorous grouping logic. It can be used for pre-training of the visual backbone network for classification, enabling the network to learn in advance the specific features of various microscopic textures, phase interfaces, and morphologies of titanium alloys. This allows the network to obtain transferable visual extraction capabilities that are suitable for downstream mechanical prediction tasks, effectively alleviating the problem of overfitting in small-sample modeling of titanium alloys in downstream applications, while ensuring the objectivity and reliability of model evaluation indicators in the pre-training stage.

[0010] In one implementation, the visual backbone network is trained on auxiliary image data to perform an organizational category classification task, resulting in a visual feature extractor for extracting the morphological features of titanium alloys. Specifically, this includes: A visual backbone network is constructed based on a pre-defined neural network model; wherein the neural network includes at least one of ResNet, MobileNet, EfficientNet, ConvNeXt, VisionTransformer, or improved networks thereof; The visual backbone network is iteratively trained based on auxiliary image data until its performance indicators reach the preset requirements. The training of the visual backbone network is then considered complete, resulting in a visual feature extractor capable of extracting texture, boundaries, phase distribution, tissue scale differences, and morphological differences from microscopic tissue images. Among these features, when training the visual backbone network for tissue category classification, a category-weighted cross-entropy loss is used.

[0011] In the above scheme, various mainstream networks, including convolutional and visual Transformer networks, are flexibly selected to build the visual backbone, adapting to different computing power and image extraction needs. Pre-training for tissue classification is conducted using dedicated auxiliary image data for titanium alloys. This is combined with category-weighted cross-entropy loss to balance the training bias caused by the uneven number of samples from different tissues, allowing the network to learn the microscopic textures, phase boundaries, phase distributions, tissue scales, and morphological features of various titanium alloys in a balanced manner. Once the classification performance meets the standards, a dedicated visual feature extractor is generated. With this pre-training transfer learning method, downstream small-sample titanium alloy mechanical prediction tasks do not require training deep networks from scratch, significantly reducing the risk of model overfitting in small-sample scenarios. The microscopic features output by the extractor closely match the physical properties of titanium alloy materials, exhibiting stronger feature transfer adaptability compared to general image pre-training weights.

[0012] In one implementation, before iteratively training the initial neural network model based on the multimodal performance dataset, the method further includes model input preprocessing of the multimodal performance dataset, including: A standardized micro-tissue image is obtained by performing preprocessing and pixel normalization operations on the micro-tissue image. The preprocessing operations include at least one of the following: grayscale conversion, denoising, brightness correction, contrast enhancement, histogram equalization, adaptive histogram equalization, scale unification, and target region cropping. The expression for the standardized micro-tissue image is: ; In the formula, To standardize microscopic tissue images; Pixel normalization processing; For optional data augmentation operations; For image preprocessing operations; For the first The first titanium alloy sample corresponding to the first Zhang microscopic tissue images or image sub-blocks; Standardization is performed based on the type of heat treatment process parameters to generate standardized process features, which are then assembled to obtain a structured heat treatment process parameter vector. The expression for this structured heat treatment process parameter vector is as follows: ; In the formula, For the first A structured heat treatment process parameter vector for a titanium alloy sample; It is a numerical process feature; Indicates categorical process characteristics; This indicates the number of numerical heat treatment process parameters; Indicates the number of category-type heat treatment process parameters; Indicates feature concatenation operation; The mechanical property test labels are normalized to obtain normalized performance labels; the expression for the normalized performance labels is as follows: ; In the formula, For the first Normalized mechanical property test labels for individual titanium alloy samples; For the first Mechanical property test labels for individual titanium alloy samples; and These are the minimum and maximum values ​​of the corresponding mechanical performance test labels in the current training subset, respectively; To prevent extremely small constants with a denominator of zero.

[0013] In the above scheme, hierarchical standardization preprocessing is performed on microscopic images, heat treatment process parameters, and mechanical performance test labels, respectively. Various image preprocessing operations such as grayscale conversion, denoising, and scale unification are used to eliminate imaging noise and size and brightness differences. Pixel normalization and training-specific data enhancement are combined to unify the image input distribution and expand sample diversity. Numerical process parameters are standardized and categorical process parameters are encoded and concatenated into a unified structured vector to eliminate differences in the dimensions of different processes. The mechanical labels are normalized using the Min-Max formula based only on the maximum and minimum values ​​of the current training subset, and a minimum constant is added to avoid computational failure. Training and validation data are isolated throughout the process to prevent information leakage. After standardized preprocessing, the three types of data have unified dimensionality and numerical distribution, eliminating training interference caused by scale, dimension, and imaging conditions between multimodal data, reducing the difficulty of model fitting, accelerating network convergence speed, and reducing prediction errors caused by noise and data leakage.

[0014] In one implementation, a visual feature mapping module maps the microstructure image feature vector output by the visual feature extractor to a preset dimension to obtain image modal features; a FiLM conditional modulation fusion module recalibrates the image modal features based on conditional modulation parameters generated from heat treatment process parameters; and a regression prediction network makes predictions based on the recalibrated image modal features, outputting normalized mechanical performance prediction results, specifically including: A standardized microstructure image is input into a visual feature extractor, which extracts texture, boundary, phase distribution, tissue scale differences, or morphological differences from the standardized microstructure image to obtain an image feature vector, which is then input into the visual feature mapping module. The expression for the image feature vector is as follows: ; In the formula, The feature vector of the microscopic tissue image; The parameter is Visual feature extractor; To standardize microscopic tissue images; The visual feature mapping space maps the received image feature vectors to a fusion feature space that matches the conditional modulation parameters, obtaining image modal features, which are then input into the FiLM conditional modulation fusion module; wherein, the expression for the image modal features is: ; In the formula, The mapped image modal features; For visual feature mapping module; The FiLM conditional modulation fusion module generates FiLM conditional modulation parameters of a preset dimension based on the structured heat treatment process parameter vector, and recalibrates the image modal features based on the FiLM conditional modulation parameters to obtain multimodal fusion features constrained by heat treatment process conditions, which are then input into the regression prediction network. The regression prediction network performs nonlinear mapping on the received multimodal fusion features to obtain the normalized prediction results of the mechanical properties of titanium alloy.

[0015] In the above scheme, a pre-trained visual feature extractor first extracts specific features such as texture, phase boundary, and microstructure of standardized micro-images to obtain high-dimensional feature vectors. Then, a visual feature mapping module maps these vectors to a unified dimension that matches the process modulation parameters, solving the feature dimension mismatch problem and helping to suppress overfitting in small samples. Relying on the FiLM conditional modulation fusion module, scaling and offset parameters generated by the heat treatment process are used to recalibrate the image modal features element by element. This allows the process to act as an external constraint to dynamically control the visual features. Unlike simple feature splicing, this approach can accurately characterize the performance differences of the same microstructure under different heat treatment regimes, fully integrating the complementary information of process evolution and microstructure. Finally, a regression network completes nonlinear fitting based on the fused comprehensive features to output normalized mechanical prediction values. The entire hierarchical network structure fully explores the intrinsic relationship between process, microstructure, and mechanical properties, effectively making up for the lack of single-modal information and significantly improving the accuracy and generalization stability of small-sample titanium alloy mechanical property prediction.

[0016] In one implementation, the FiLM conditional modulation fusion module generates FiLM conditional modulation parameters of a preset dimension based on the structured heat treatment process parameter vector, and recalibrates the image modal features based on the FiLM conditional modulation parameters to obtain multimodal fusion features constrained by the heat treatment process conditions, which are then input into the regression prediction network. Specifically, this includes: The structured heat treatment process parameter vector is input into the process condition parameter generation network to generate conditional modulation parameters for modulating image modal features; wherein, the conditional modulation parameters include at least one of scaling parameters, offset parameters, gating coefficients, attention weights or combinations thereof; Image modal features are modulated element-wise based on conditional modulation parameters to generate multimodal fusion features constrained by heat treatment process conditions; the expression for the multimodal fusion features is as follows: ; In the formula, This is a multimodal fusion feature; and These represent the scaling parameter and offset parameter generated based on the structured heat treatment process parameter vector, respectively; This is the identifier for element-wise multiplication; These are the mapped image modal features.

[0017] In the above scheme, a network-analyzed structured heat treatment process vector is generated using independent process condition parameters. This vector outputs adjustable condition modulation parameters such as scaling and offset. The image modal features are recalibrated through FiLM operations that multiply and superimpose offsets element-wise, transforming the heat treatment process into a dynamic control factor rather than a simple set of parallel features. This approach can match differentiated feature scaling and offset degrees for different processes, accurately reflecting the constraint effect of heat treatment regimes on the mechanical response of microstructures. Compared to feature stitching and simple gating fusion methods, it can couple process and morphology information more deeply, fully depicting the intrinsic relationship between process and microstructure. The generated multimodal fusion features simultaneously carry a complete process history and microstructure details, effectively improving the accuracy of titanium alloy mechanical property prediction.

[0018] In one implementation, the visual feature extractor is transferred to a pre-built initial neural network model. The initial neural network model is iteratively trained based on a multimodal performance dataset until the loss function converges or a preset number of iterations is reached. The iterative training is then considered complete, resulting in a titanium alloy performance prediction model used to output mechanical performance prediction results. Specifically, this includes: Leave-one-out cross-validation training is performed based on the unique sample number of each titanium alloy sample; wherein, each leave-one-out cross-validation training involves selecting one titanium alloy sample as the validation sample and the remaining titanium alloy samples as training samples. In each training iteration, the parameters of the initial neural network model are trained based on the training sample parameters in the current training iteration, and the training loss function between the output mechanical property prediction result and the mechanical property prediction result of the titanium alloy sample is calculated; wherein, when training the initial neural network model, the feature extraction part of the visual feature extractor is transferred to the initial neural network model, and the parameters of the feature extraction part are frozen according to the data scale of the titanium alloy sample. The parameters of the initial neural network model in the next training session are optimized based on the training loss function output from the previous training session, until leave-one-out cross-validation training is completed, resulting in a titanium alloy performance prediction model.

[0019] In the above scheme, the pre-trained visual feature extractor is transferred to the initial neural network, and the parameters of the feature extraction layer are flexibly frozen according to the number of multimodal samples. By leveraging transfer learning to reuse the microstructure characterization capability of titanium alloys, the overfitting problem of the model in small sample scenarios is greatly alleviated. A leave-one-out cross-validation method is adopted, which uses the unique sample number as the basis for partitioning. Only a single complete sample is selected as the validation set each time, and all remaining samples are used as the training set. This avoids the problem of leakage of the same sample image sub-block across datasets and makes full use of the limited titanium alloy samples to complete training and unbiased evaluation. Each fold updates the network parameters and calculates the loss function based on the current training subset. The model weights are continuously optimized iteratively with loss feedback. After all folds of training are completed, the final prediction model is obtained. This approach maximizes the mining of the correlation between process, microstructure and mechanical properties within the small sample dataset, and can objectively and stably verify the model's generalization ability, ensuring that the obtained titanium alloy performance prediction model has reliable prediction accuracy and generalization stability.

[0020] Secondly, this application also provides methods for applying the titanium alloy performance prediction model, including: Collect titanium alloy samples to be evaluated; wherein, the titanium alloy samples include associated heat treatment process parameters, microstructure images and / or microstructure image sub-blocks obtained by random cropping based on the microstructure images; The titanium alloy sample is input into the titanium alloy performance prediction model so that the titanium alloy performance prediction model outputs the predicted results of the titanium alloy sample in terms of the target mechanical properties; wherein, the titanium alloy performance prediction model is constructed using the construction method of the titanium alloy performance prediction model described above.

[0021] In the above scheme, the heat treatment process parameters, microstructure images, and cut sub-blocks of the titanium alloy sample to be evaluated are first collected as model inputs, and then fed into the titanium alloy performance prediction model to complete the mechanical property prediction. This model, built upon a pre-trained visual feature extractor and a FiLM conditional modulation fusion structure, can simultaneously utilize process evolution information and microstructure features. Compared to evaluation methods relying solely on a single process or image, it provides more complete information dimensions and higher prediction accuracy. The entire application process only requires sample characterization and process information recording, eliminating the need for destructive mechanical tests such as fracture toughness tests, significantly shortening the sample testing cycle and reducing the cost of test consumables and equipment wear. Simultaneously, the microstructure image cut sub-blocks provide multiple perspectives, reducing prediction bias caused by local microstructure. Combined with a model optimized through multi-fold cross-validation, prediction stability is ensured, enabling rapid non-destructive prediction of the mechanical properties of new titanium alloy samples. This provides a convenient and feasible technical means for rapid screening of heat treatment processes and early evaluation of material properties.

[0022] Thirdly, this application also provides a device for constructing a titanium alloy performance prediction model, including: a first data module, a second data module, a network training module, and a model construction module; The first data module constructs a multimodal performance dataset based on titanium alloy samples from different heat treatment batches; each titanium alloy sample includes associated heat treatment process parameters, mechanical property test labels, microstructure images and / or microstructure image sub-blocks obtained by randomly cropping the microstructure images; The second data module constructs auxiliary image data based on the microstructure images of titanium alloys acquired under different heat treatment conditions and their corresponding microstructure category labels; among them, the microstructure images of titanium alloys include a variety of typical microstructure morphologies; The network training module trains the visual backbone network on the auxiliary image data to perform a category classification task, resulting in a visual feature extractor for extracting the morphological features of titanium alloys. The model building module is used to transfer the visual feature extractor to the pre-built initial neural network model. Iterative training of the initial neural network model is performed based on the multimodal performance dataset until the loss function converges or the preset number of iterations is reached. The iteration training is then determined to be complete, and a titanium alloy performance prediction model for outputting mechanical performance prediction results is obtained. The initial neural network model includes a visual feature mapping module, a FiLM conditional modulation fusion module, and a regression prediction network connected in sequence. The visual feature mapping module maps the microstructure image feature vectors output by the visual feature extractor to a preset dimension to obtain image modal features; the FiLM conditional modulation fusion module recalibrates the image modal features based on the conditional modulation parameters generated by the heat treatment process parameters; and the regression prediction network makes predictions based on the recalibrated image modal features and outputs normalized mechanical performance prediction results. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the method for constructing a titanium alloy performance prediction model according to an embodiment of the present invention. Figure 2 This is a topological schematic diagram of a titanium alloy performance prediction model provided in one embodiment of the present invention; Figure 3 This is a flowchart illustrating the application method of the titanium alloy performance prediction model provided in one embodiment of the present invention. Figure 4 This is a schematic diagram of the module of the device for constructing a titanium alloy performance prediction model according to an embodiment of the present invention. Detailed Implementation

[0024] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0025] The terms "first" and "second," etc., in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0026] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0027] Example 1 See Figure 1 , Figure 1 This is a flowchart illustrating a method for constructing a titanium alloy performance prediction model according to an embodiment of the present invention. The method for constructing a titanium alloy performance prediction model according to this embodiment includes steps 101 to 104, each step being as follows: Step 101: Construct a multimodal performance dataset based on titanium alloy samples from different heat treatment batches; wherein each titanium alloy sample includes associated heat treatment process parameters, mechanical property test labels, microstructure images and / or microstructure image sub-blocks obtained by randomly cropping microstructure images.

[0028] In this embodiment of the invention, titanium alloy samples from different heat-treated batches are collected to construct a multimodal performance dataset. Each sample is associated with heat treatment process parameters, mechanical property test labels, and microstructure images or image cropping sub-blocks. By binding multiple types of data, including process, performance, and images, a complete multimodal sample foundation is formed. Image sub-blocks are used to expand the data volume and overcome the deficiency of insufficient data under small sample conditions.

[0029] In one embodiment, a multimodal performance dataset is constructed based on titanium alloy samples from different heat treatment batches, specifically including: Obtain titanium alloy samples from different heat-treated batches; Microstructure images of each titanium alloy sample were obtained, and the effective microstructure areas in the microstructure images were randomly cropped to obtain several microstructure image sub-blocks. Obtain the heat treatment process parameters for each titanium alloy sample; wherein, the heat treatment process parameters include at least one of the following: solution type, solution temperature, solution time, solution cooling method, aging temperature, aging time, and aging cooling method; Obtain the mechanical property test label for each titanium alloy sample; wherein the mechanical property test label includes at least one of fracture toughness, tensile strength, yield strength, elongation after fracture, reduction of area, impact toughness, fatigue life and fatigue crack propagation performance; Based on the unique sample number of each titanium alloy sample, the heat treatment process parameters, mechanical property test labels, microstructure images, and / or microstructure image sub-blocks of the same titanium alloy sample are associated as a multimodal sample; wherein, the data expression of the multimodal sample is: ; In the formula, For the first Data records for one titanium alloy sample; For the first The first titanium alloy sample corresponding to the first Zhang microscopic tissue images or image sub-blocks; For the first Number of microstructure images or image sub-blocks corresponding to each titanium alloy sample; For the first Heat treatment process parameters for one titanium alloy sample; For the first Mechanical property test labels for individual titanium alloy samples; A multimodal performance dataset is constructed based on multimodal samples; the multimodal samples undergo integrity checks and consistency verification.

[0030] In this embodiment of the invention, titanium alloy samples from different heat treatment batches are obtained, and multiple groups of titanium alloy samples are prepared in batches. Each group adopts different solution treatment and aging heat treatment regimes to form a multi-heat treatment batch sample library. High-strength and high-toughness titanium alloys such as TB18 and TC18 are selected to ensure that the samples cover multiple heat treatment process ranges, providing a physical basis for subsequent construction of multimodal data containing process differences. Original microscopic images of each titanium alloy sample are acquired using SEM, TEM, OM, and EBSD equipment. The original images contain invalid areas such as scales, black borders, and annotations. Irrelevant areas are first removed, leaving only the effective microstructure areas. Then, a random cropping operation is performed on the effective areas, and image sub-blocks of fixed size are uniformly output. Multiple image sub-blocks can be generated from the same sample, expanding the total number of image samples under the condition of a limited number of titanium alloy physical samples, enriching the field of view for microscopic morphology observation, and allowing the model to fully learn various microscopic details such as α phase, β phase, and phase interface. Complete heat treatment records were collected for each sample. Parameters were categorized into numerical and categorical types, covering seven key information categories: solution type, solution temperature, solution time, solution cooling method, aging temperature, aging time, and aging cooling method. One or more combinations of these categories could be used for recording. These structured parameters record the entire history of the formation of the titanium alloy microstructure, reflecting the regulatory effect of heat treatment on microstructure evolution. This is key structured modal information that distinguishes it from microscopic images. Standard mechanical testing machines were used to conduct standard mechanical property tests on the corresponding samples. At least one of the following indicators—fracture toughness, tensile strength, yield strength, and fatigue crack propagation performance—was collected as a true label. The label represents the actual service mechanical level of the sample and serves as the baseline true value that the model ultimately needs to fit and predict, used to establish a regression mapping relationship from images and processes to mechanical properties. A unique, non-repeating sample number was assigned to each titanium alloy sample. Using this number as an association index, all image sub-blocks, the complete set of heat treatment process parameters, and the measured mechanical labels of that sample were bound into a complete multimodal sample. It should be noted that all images under the same sample share the same set of process and performance labels, ensuring a one-to-one correspondence between modal data and eliminating sample matching errors. A dual verification of completeness and consistency is performed on all associated multimodal samples. Completeness checks remove incomplete samples with missing images, missing process parameters, or no performance labels; consistency checks identify anomalous data such as mismatched sample numbers, incorrect image-sample correspondence, and blurry or faulty images. All verified multimodal samples are compiled and integrated to form a standardized titanium alloy multimodal performance dataset.

[0031] For example, TB18 titanium alloy samples and heat treatment batches were acquired. Each sample had a unique sample number, used to associate the corresponding heat treatment process parameters, microstructure image, and mechanical property test label. SEM microstructure images of the TB18 titanium alloy samples were acquired. SEM images were acquired using a FEI Helios Nanolab G3 UC device. Before SEM characterization, the samples underwent mechanical grinding and polishing, followed by etching with Kroll etchant to reveal the microstructure. Each TB18 sample corresponded to 2–6 SEM microstructure images. The original SEM images had varying magnifications, including 2000× and 5000×. Before constructing the model input images, non-microstructure areas such as rulers, annotation text, and black borders were removed from the original images. The effective microstructure areas were then cropped to obtain 224×224 pixel SEM image sub-blocks. The cropping process could use random cropping to expand the number of images in the target multimodal performance dataset and increase the visible differences in microstructure areas during model training. Mechanical property test labels for the TB18 titanium alloy samples were also acquired. The target performance label in this embodiment of the invention is plane strain fracture toughness K. IC K IC The tests were conducted using an MTS Landmark 370.25 electro-hydraulic servo fatigue testing machine. The specimens were compact tensile specimens with a thickness of 25 mm and an effective width of 50 mm. Pre-cracks were introduced using the decreasing K-method, with a pre-crack frequency of 18 Hz, and a final pre-crack stress intensity factor of 22. The obtained K-method... IC Numerical values ​​were used as model training labels; a sample-level multimodal association record was established. SEM microstructure image sub-blocks, heat treatment process parameters, and K0 were used to associate the same TB18 titanium alloy sample with the corresponding heat treatment process parameters. IC The test label is associated with a multimodal sample. See Table 1, which is a summary example of a TB18 titanium alloy multimodal dataset provided in this embodiment of the invention.

[0032] Table 1

[0033] Step 102: Construct auxiliary image data based on the microstructure images of titanium alloys acquired under different heat treatment process conditions and their corresponding microstructure category labels; wherein, the microstructure images of titanium alloys include a variety of typical microstructure morphologies.

[0034] In this embodiment of the invention, microscopic images of titanium alloys with different heat treatments and various morphologies, along with corresponding tissue category labels, are used to construct an auxiliary image dataset. Dedicated image data with tissue classification labels is provided to offer exclusive training materials for the pre-training of visual networks.

[0035] In one embodiment, auxiliary image data is constructed based on microstructure images of titanium alloys acquired under different heat treatment conditions and their corresponding microstructure category labels, specifically including: Collect images of the microstructure of titanium alloys under different heat treatment conditions and their corresponding microstructure category labels; The cracked areas in the microstructure image of the titanium alloy are marked, and the microstructure image of the titanium alloy is cropped by a pixel sliding window of a preset size to obtain several image sub-blocks that do not contain the cracked areas. Image sub-blocks derived from the same titanium alloy microstructure image are divided into the same data subset, and auxiliary image data is constructed based on the data subset.

[0036] In this embodiment of the invention, titanium alloy samples that have undergone differentiated solution treatment and aging heat treatment are selected, and microstructure images are captured using characterization equipment such as SEM. The microstructure types are classified according to the overall morphology of the α-phase, β-phase, and phase interface within the images, and a corresponding microstructure category label is matched to each original microstructure image. Different heat treatment processes generate differentiated microstructures such as lamellar and biphasic states. Diverse images and labels allow the subsequent visual network to learn the unique textures and boundary features of various typical titanium alloy microstructures, providing basic materials for pre-training tasks. Auxiliary image data may include self-built titanium alloy microstructure image data, publicly available material image data, material microstructure images extracted from literature, or combinations thereof. Annotation tools are used to annotate non-microstructure interference areas such as fatigue cracks in the original microstructure images. Then, a fixed-size pixel sliding window is set, and the entire image is slid-cropped region by region according to a set step size, generating a large number of image sub-blocks in batches. After cropping, sub-blocks with crack area ratios exceeding a threshold are filtered out, retaining only valid sub-blocks containing only pure microstructures without crack interference. Cracks are not intrinsic microstructural features of materials and can interfere with the network's learning of phase distribution and grain morphology. Removing crack-containing sub-blocks eliminates irrelevant noise, ensuring that pre-training focuses solely on the intrinsic microstructural features. All sub-blocks obtained by sliding cropping from the same original microscopic image are grouped into the same data subset, without splitting them into different groups for training, validation, and testing. After grouping, all qualified and clearly grouped image sub-blocks are integrated with their corresponding microstructural category labels to form a complete auxiliary image dataset. Restricting homogeneous sub-blocks to different subsets avoids image information leakage and prevents morphological features from the same original image from appearing simultaneously in both training and evaluation groups, ensuring the objectivity of subsequent visual network pre-training results and the reliability of evaluation metrics.

[0037] For example, in this embodiment of the invention, a self-built titanium alloy SEM microstructure image classification dataset is used as the auxiliary image data to obtain the original auxiliary image data. This original auxiliary image data originates from SEM microstructure images after fatigue crack propagation experiments, mainly including microstructure images or sub-images and their corresponding tissue category labels. The original auxiliary image data includes 221 original SEM images, of which 139 are lamellar SEM images and 82 are bimorphic SEM images. The original SEM images are categorized and stored according to tissue category, and corresponding binary classification labels of "lamellae / bimorphic" are established. Since some original SEM images contain fatigue crack regions, to avoid crack regions interfering with the tissue category classification pre-training, this embodiment uses Labelme to annotate the crack regions in the original SEM images. Subsequently, a 224×224 pixel sliding window is used to obtain microstructure sub-images from the original SEM images, with a sliding step size of 112 pixels, allowing overlap between adjacent cropped blocks. Regions with edge areas less than 224×224 pixels are discarded; when the area of ​​a cracked region in a cropped block exceeds 1%, the cropped block is removed. After quality screening, target region selection, sliding window cropping, and crack removal, approximately 10,000 224×224 pixel microscopic tissue sub-images are obtained for tissue category classification pre-training. All sub-blocks cropped from a single original image are uniformly assigned to one of the training, validation, or test subsets without splitting; all qualified sub-blocks are divided into training, validation, and test subsets in an 8:1:1 ratio. Each subset contains only sub-blocks from the same source image. The sub-blocks are integrated with their corresponding tissue classification labels to obtain complete auxiliary image data for visual main intervention training. The training and validation sets are used for image classification model training and model selection, while the test set is only used for independent evaluation after the image classification model training is completed.

[0038] Step 103: Train the visual backbone network on the auxiliary image data to perform an organizational category classification task, and obtain a visual feature extractor for extracting the morphological features of titanium alloys.

[0039] In this embodiment of the invention, auxiliary image data is used to perform tissue classification training on the visual backbone, resulting in a dedicated visual feature extractor for the microstructure of titanium alloys. Through classification pre-training, the network autonomously learns the microstructure features of titanium alloys, acquiring transferable image feature extraction capabilities.

[0040] In one embodiment, a visual backbone network is trained on auxiliary image data to perform an organizational category classification task, resulting in a visual feature extractor for extracting morphological features of titanium alloys, specifically including: A visual backbone network is constructed based on a pre-defined neural network model; wherein the neural network includes at least one of ResNet, MobileNet, EfficientNet, ConvNeXt, VisionTransformer, or improved networks thereof; The visual backbone network is iteratively trained based on auxiliary image data until its performance indicators reach the preset requirements. The training of the visual backbone network is then considered complete, resulting in a visual feature extractor capable of extracting texture, boundaries, phase distribution, tissue scale differences, and morphological differences from microscopic tissue images. Among these features, when training the visual backbone network for tissue category classification, a category-weighted cross-entropy loss is used.

[0041] In this embodiment of the invention, the network can include convolutional networks such as ResNet, MobileNet, EfficientNet, ConvNeXt, and the visual Transformer network VisionTransformer. Improved versions of these networks can also be used, and any one or a combination of them can be selected as the visual backbone. Considering the scenario of single-channel, low-computational-power training of titanium alloy microscopic images, a lightweight variant network is preferred. The network input is adapted to the size of the microscopic tissue image. The main network is responsible for extracting global features from shallow texture to deep tissue layer by layer. A classification output head is temporarily added at the end to adapt to the tissue classification pre-training task. The preprocessed auxiliary image sub-blocks are input into the constructed visual backbone network to carry out iterative training for tissue binary classification. The training process uses the class-weighted cross-entropy loss function. This loss function addresses the problem of imbalanced sample numbers for slices and bimorphic tissues in the auxiliary data. It assigns higher loss weights to the class with fewer samples, avoiding network bias towards the tissue type with a larger sample proportion, and ensuring that both types of microscopic morphology features are fully learned. The expression for the class weight is: In the formula, Let k be the class weights of the training samples of class k. C represents the total number of training samples and the number of classes. denoted as the number of training samples for the k-th class. The network parameters are continuously iterated and optimized, while performance metrics such as classification accuracy and F1 score on the validation set are monitored simultaneously. Training stops once these metrics reach a preset passing standard. After training, the final classification output layer of the network is removed, retaining only the intermediate convolutional, pooling, and feature mapping structures. This retained portion constitutes the visual feature extractor, which can autonomously extract transferable deep features such as microscopic image texture, grain boundary, two-phase distribution, microstructure size, and morphological differences in titanium alloys.

[0042] For example, this embodiment of the invention selects the lightweight improved model TinyEfficientNet as the backbone, without loading ImageNet general pre-trained weights, and completes training from scratch using microscopic images of titanium alloys, avoiding interference from general natural image features on metal structure recognition. Combining the characteristics of microscopic SEM images, the network input is set to a 224×224 single-channel grayscale image; the first layer of the network is configured with a convolution + BN + SiLU activation module with 1 input channel and 32 output channels, and the main body stacks 5 sets of lightweight MBConv modules. Each MBConv integrates extended convolution, depthwise separable convolution, SE attention module, and projection convolution. The SE module achieves feature adaptive heavy texture feature weights through channel compression and activation; the end of the network is equipped with a dedicated classification head, which sequentially passes through 1×1 convolution, global average pooling calibration, strengthens the titanium alloy phase interface, refines, flattens, dropsout, and fully connected layers, and finally sets 2 classification channels to correspond to the two types of titanium alloy structures: layer sheets and bistates. The entire network is highly lightweight with low computational overhead, specifically designed for extracting local texture, boundary, and microstructure distribution features from microscopic images of titanium alloys, providing feature extraction capabilities for downstream small-sample fracture toughness prediction. The batch size for image classification pre-training is set to 32, the number of training epochs to 30, the initial learning rate to 0.001, the optimizer to Adam, the weight decay to 0.0001, and the learning rate scheduling to a cosine annealing scheduler. During training, model weights are selected based on the validation set accuracy, and the model weights with the highest validation set accuracy are saved to prevent overfitting. After training, unbiased performance is evaluated using an independent test set. When the accuracy, weighted F1, and other metrics reach the preset passing standards, the backbone network training is considered complete. In this embodiment, the overall accuracy of the classification model on the test set is 99.17%, and the weighted average F1 value is 0.9917. This result demonstrates that the visual backbone can identify the differences in texture, phase interface, phase distribution, and morphology between lamellar and bimodal structures in titanium alloys. This visual backbone network can serve as a subsequent multimodal K-mode model for titanium alloys. IC A pre-trained visual feature extractor for prediction tasks.

[0043] Step 104: Transfer the visual feature extractor to the pre-built initial neural network model, and iteratively train the initial neural network model based on the multimodal performance dataset until the loss function converges or the preset number of iterations is reached. Determine that the iterative training is complete, and obtain the titanium alloy performance prediction model for outputting mechanical performance prediction results. The initial neural network model includes a visual feature mapping module, a FiLM conditional modulation fusion module, and a regression prediction network connected in sequence. The visual feature mapping module is used to map the microstructure image feature vector output by the visual feature extractor to a preset dimension to obtain image modal features; the FiLM (Feature-wise Linear Modulation) conditional modulation fusion module is used to recalibrate the image modal features based on the conditional modulation parameters generated by the heat treatment process parameters; the regression prediction network is used to make predictions based on the recalibrated image modal features and output normalized mechanical performance prediction results.

[0044] In this embodiment of the invention, the visual feature extractor is transferred into the initial neural network. The transfer of the pre-trained extractor combined with iterative training can quickly complete the model construction and realize the automatic prediction of the mechanical properties of titanium alloys.

[0045] In one embodiment, before iteratively training the initial neural network model based on the multimodal performance dataset, the method further includes preprocessing the multimodal performance dataset for model input, including: A standardized micro-tissue image is obtained by performing preprocessing and pixel normalization operations on the micro-tissue image. The preprocessing operations include at least one of the following: grayscale conversion, denoising, brightness correction, contrast enhancement, histogram equalization, adaptive histogram equalization, scale unification, and target region cropping. The expression for the standardized micro-tissue image is: ; In the formula, To standardize microscopic tissue images; Pixel normalization processing; For optional data augmentation operations; For image preprocessing operations; For the first The first titanium alloy sample corresponding to the first Zhang microscopic tissue images or image sub-blocks; Standardization is performed based on the type of heat treatment process parameters to generate standardized process features, which are then assembled to obtain a structured heat treatment process parameter vector. The expression for this structured heat treatment process parameter vector is as follows: ; In the formula, For the first A structured heat treatment process parameter vector for a titanium alloy sample; It is a numerical process feature; Indicates categorical process characteristics; This indicates the number of numerical heat treatment process parameters; Indicates the number of category-type heat treatment process parameters; Indicates feature concatenation operation; The mechanical property test labels are normalized to obtain normalized performance labels; the expression for the normalized performance labels is as follows: ; In the formula, For the first Normalized mechanical property test labels for individual titanium alloy samples; For the first Mechanical property test labels for individual titanium alloy samples; and These are the minimum and maximum values ​​of the corresponding mechanical performance test labels in the current training subset, respectively; To prevent extremely small constants with a denominator of zero.

[0046] In this embodiment of the invention, after reading the image sub-blocks, they are first converted into single-channel grayscale images, then the image size is scaled to 256×256 pixels, and center-cropped to obtain a 224×224 pixel model input. Subsequently, global histogram equalization is performed on the center-cropped grayscale image. Imaging differences are eliminated and input specifications are standardized for the original microscopic image, and different processing logics are executed for the training and validation / prediction stages. Preprocessing can use one or more operations such as grayscale conversion, denoising, histogram equalization, size unification, and target cropping to eliminate interference such as scale differences, uneven brightness, and inconsistent sizes. During the training stage, to improve the model's robustness to changes in image orientation and local pose, random data augmentation is further performed after center cropping. Random data augmentation includes random horizontal flipping, random vertical flipping, and random angle rotation, where the random angle rotation range is ±10°. The enhanced image is first converted into a tensor, scaling the pixel values ​​from 0-255 to the range of 0-1. Then, it is standardized using the mean and standard deviation, further linearly mapping the pixel values ​​to approximately -1 to 1, to obtain model input with consistent numerical scale. It should be noted that data augmentation is only enabled during the model training phase, expanding image diversity and improving model robustness through flipping and rotation. Augmented data is not used in the validation and prediction phases. Pixel normalization unifies the pixel value distribution, eliminating numerical bias caused by image brightness. Pixel normalization uses a mean of 0.5 and a standard deviation of 0.5, which can be expressed as: . This represents the image tensor after grayscale conversion, global histogram equalization, and tensor transformation. The image processing during the training phase can be represented as: The image processing procedure in the verification or prediction stage can be represented as follows: .in, This represents the operation of converting an image into a tensor and scaling pixel values ​​to the [0,1] range; This indicates a center clipping operation; This indicates grayscale conversion and global histogram equalization processing; This indicates the random data augmentation operation used during the training phase; This indicates an image scaling operation.

[0047] Heat treatment process parameters are divided into two categories: numerical and categorical. Numerical parameters include solution temperature, solution time, aging temperature, and aging time, while categorical parameters include solution type, solution cooling method, and aging cooling method. After standardization and encoding according to parameter type, the parameters are concatenated into a one-dimensional vector. Numerical parameters are standardized using Z-score, eliminating dimensions based on the mean and standard deviation of the current training subset. Categorical parameters such as solution type and cooling method are converted into numerical features using sequential / one-heat encoding. Specifically, for numerical heat treatment process parameters, z-score standardization is performed using the mean and standard deviation of the training set. ;in, Represents the standardized first Numerical process parameters; Indicates the first Numerical process parameters, and These represent the mean and standard deviation of the parameter in the current training subset, respectively. For categorical heat treatment process parameters, sequential encoding is used for numerical representation, and the encoding process can be expressed as follows: ;in, Indicates the first Each type of heat treatment process parameter; Indicates a category encoder; The encoded categorical process features are represented; the standardized numerical process features and the encoded categorical process features are concatenated to obtain a 7-dimensional structured heat treatment process parameter vector: ;in, to These represent the standardized numerical characteristics corresponding to solution temperature, solution time, aging temperature, and aging time, respectively. to These represent the coding features corresponding to solution type, solution cooling method, and aging cooling method, respectively. Indicates feature concatenation operation; Indicates the first A structured heat treatment process parameter vector for each titanium alloy sample.

[0048] The original fracture toughness, strength, and other mechanical indicators have large differences in physical dimensions and numerical ranges, which would cause difficulties in the convergence of the regression network if trained directly. Therefore, a Min-Max scaling mapping is used to map them to the 0~1 range. The mechanical performance test labels are sorted and normalized to obtain normalized performance labels for model training. This is the true value fitted for the model regression task, which is then converted back to the actual mechanical value after prediction. It should be noted that the mean and standard deviation of numerical process parameters, the encoder of categorical process parameters, and K... IC The minimum and maximum values ​​of the labels are determined solely by the current training subset. Validation samples do not participate in the fitting process of the above parameters; they are only transformed using the processing rules determined by the current training subset. This approach avoids leakage of validation sample information into the training process and ensures the consistency of data processing rules across the training, validation, testing, and prediction phases.

[0049] In one embodiment, the visual feature mapping module is used to map the microstructure image feature vector output by the visual feature extractor to a preset dimension to obtain image modal features; the FiLM conditional modulation fusion module is used to recalibrate the image modal features according to the conditional modulation parameters generated by the heat treatment process parameters; the regression prediction network is used to make predictions based on the recalibrated image modal features and output normalized mechanical performance prediction results, specifically including: A standardized microstructure image is input into a visual feature extractor, which extracts texture, boundary, phase distribution, tissue scale differences, or morphological differences from the standardized microstructure image to obtain an image feature vector, which is then input into the visual feature mapping module. The expression for the image feature vector is as follows: ; In the formula, The feature vector of the microscopic tissue image; The parameter is Visual feature extractor; To standardize microscopic tissue images; The visual feature mapping space maps the received image feature vectors to a fusion feature space that matches the conditional modulation parameters, obtaining image modal features, which are then input into the FiLM conditional modulation fusion module; wherein, the expression for the image modal features is: ; In the formula, The mapped image modal features; For visual feature mapping module; The FiLM conditional modulation fusion module generates FiLM conditional modulation parameters of a preset dimension based on the structured heat treatment process parameter vector, and recalibrates the image modal features based on the FiLM conditional modulation parameters to obtain multimodal fusion features constrained by heat treatment process conditions, which are then input into the regression prediction network. The regression prediction network performs nonlinear mapping on the received multimodal fusion features to obtain the normalized prediction results of the mechanical properties of titanium alloy.

[0050] See Figure 2 , Figure 2 This is a topological schematic diagram of a titanium alloy performance prediction model provided in one embodiment of the present invention. In this embodiment, a standardized microstructure image, after complete preprocessing such as grayscale and normalization, is input into a pre-trained visual feature extractor with fixed parameters. Fixed parameters The visual extractor, based on the representational capabilities learned from previous tissue classification tasks, automatically mines microscopic information specific to titanium alloys within images, such as texture, phase interfaces, two-phase distribution, microstructure size, and morphology, and outputs a high-dimensional original image feature vector. This vector fully carries all the visual information of a single SEM image, but its high dimensionality makes it impossible to directly match the modulation parameters generated by the process. Therefore, it is fed down to the visual feature mapping module. For example, a standardized SEM image is input into a frozen visual feature extractor to obtain a 320-dimensional microstructure image feature vector: , This represents the 320-dimensional microscopic tissue image feature vector corresponding to the image. The visual feature mapping module is an intermediate transformation unit in the image branch, receiving the high-dimensional feature vector. The core function is to map the original image features to a preset fusion dimension consistent with the FiLM modulation parameters. The module integrates a fully connected layer Linear(320,32), a layer normalization layer LayerNorm, and a Dropout layer. On the one hand, it compresses the feature dimension to achieve dimension matching; on the other hand, it suppresses overfitting in small sample scenes through normalization and regularization terms. After processing, it generates a 32-dimensional image modality feature with uniform dimensions. The data is then directly fed into the FiLM conditional modulation fusion module to await process parameter adjustment. The process parameters are branched off into a separate branch, with the structured heat treatment process vector independently input into the process parameter generation network, outputting FiLM modulation parameters. The FiLM conditional modulation fusion module uses the generated FiLM conditional modulation parameters to globally recalibrate the image modal features, obtaining multimodal fusion features. This design differs from simple feature stitching; the heat treatment process is no longer an independent parallel input, but rather a control condition to correct microscopic image features, allowing images with the same morphology to generate differentiated fusion features under different processes. This accurately expresses the coupling relationship between the process and the tissue, ultimately generating process-constrained multimodal fusion features that are fed into the regression network. The regression prediction network is the model's final output module, receiving the multimodal fusion features output by FiLM. It constructs a multi-layer nonlinear mapping structure through multiple fully connected layers, nonlinear activation layers, and Dropout layers, with a Sigmoid activation function at the tail. The output numerical range perfectly matches the mechanical labels after Min-Max normalization, yielding normalized fracture toughness prediction values ​​within the range of 0 to 1. During the training phase, the MSE loss can be directly calculated with the normalized true labels to update the parameters. During the prediction phase, the maximum and minimum values ​​stored in the training set are used to convert the values ​​to K with real physical units. ICNumerical values. For example, the regression prediction network structure is as follows: The Dropout ratio is 0.3, and the Sigmoid output is used to keep the prediction results within the normalized label range.

[0051] In one embodiment, the FiLM conditional modulation fusion module generates FiLM conditional modulation parameters of a preset dimension based on the structured heat treatment process parameter vector, and recalibrates the image modal features based on the FiLM conditional modulation parameters to obtain multimodal fusion features constrained by the heat treatment process conditions, which are then input into the regression prediction network. Specifically, this includes: The structured heat treatment process parameter vector is input into the process condition parameter generation network to generate conditional modulation parameters for modulating image modal features; wherein, the conditional modulation parameters include at least one of scaling parameters, offset parameters, gating coefficients, attention weights or combinations thereof; Image modal features are modulated element-wise based on conditional modulation parameters to generate multimodal fusion features constrained by heat treatment process conditions; the expression for the multimodal fusion features is as follows: ; In the formula, This is a multimodal fusion feature; and These represent the scaling parameter and offset parameter generated based on the structured heat treatment process parameter vector, respectively; This is the identifier for element-wise multiplication; These are the mapped image modal features.

[0052] See Figure 2 , Figure 2 This is a topological diagram of a titanium alloy performance prediction model provided in one embodiment of the present invention. In this embodiment, the structured heat treatment process parameter vector is input as an independent branch into the process condition parameter generation network. This network is constructed from a multilayer perceptron, a normalization layer, and a nonlinear activation layer. The network receives all encoded process information related to solid solution and aging, learns the intrinsic relationship between the process and microstructure and mechanical properties, and outputs condition modulation parameters of a preset dimension. The modulation parameters can be a single type or a combination of scaling parameters, offset parameters, gating coefficients, and attention weights. For example, scaling parameters and offset parameters are selected as the core parameters of FiLM modulation. The 7-dimensional structured heat treatment process parameter vector is input into the process condition parameter generation network, outputting 64-dimensional condition parameters, which are then split into 32-dimensional scaling parameters and 32-dimensional offset parameters to obtain the condition modulation parameters: (12); among which, The parameter is The process condition parameter generation network; Indicates the first A 7-dimensional structured heat treatment process parameter vector for a titanium alloy sample. This represents the scaling parameter generated from the heat treatment process parameters; This represents the offset parameter generated by the heat treatment process parameters. Then, the scaling parameter generated by the process will be... Image modal features Element-wise multiplication is performed to scale the image features across each dimension; then the offset parameter is added. Feature translation correction was completed. After dual recalibration using scaling and offset, the result was obtained. This refers to multimodal fusion features. This mechanism differs from simple splicing and fusion; it doesn't simply merge process and image features. Instead, it uses the heat treatment process as an external constraint to dynamically adjust the expression of microscopic image features. The same type of microscopic morphology image will generate completely different fusion features under different process parameters, accurately reflecting the physical law that "heat treatment processes change microstructure evolution and ultimately affect mechanical properties." The fusion features simultaneously carry microscopic morphology information and heat treatment history information, and are finally fed into a regression prediction network to complete performance fitting.

[0053] In one embodiment, the visual feature extractor is transferred to a pre-constructed initial neural network model. The initial neural network model is iteratively trained based on a multimodal performance dataset until the loss function converges or a preset number of iterations is reached. The iterative training is then considered complete, resulting in a titanium alloy performance prediction model for outputting mechanical performance prediction results. Specifically, this includes: Leave-one-out cross-validation training is performed based on the unique sample number of each titanium alloy sample; wherein, each leave-one-out cross-validation training involves selecting one titanium alloy sample as the validation sample and the remaining titanium alloy samples as training samples. In each training iteration, the parameters of the initial neural network model are trained based on the training sample parameters in the current training iteration, and the training loss function between the output mechanical property prediction result and the mechanical property prediction result of the titanium alloy sample is calculated; wherein, when training the initial neural network model, the feature extraction part of the visual feature extractor is transferred to the initial neural network model, and the parameters of the feature extraction part are frozen according to the data scale of the titanium alloy sample. The parameters of the initial neural network model in the next training session are optimized based on the training loss function output from the previous training session, until leave-one-out cross-validation training is completed, resulting in a titanium alloy performance prediction model.

[0054] In this embodiment of the invention, the dataset is divided according to the unique sample number of the titanium alloy before training, and a leave-one-out cross-validation method is adopted. In each round of validation, one complete heat-treated titanium alloy sample is extracted as the validation sample, and all remaining heat-treated samples are used as the training samples for that round. At the same time, it is stipulated that all SEM images, image sub-blocks, process parameters, and performance labels corresponding to the same titanium alloy sample are all classified into the same subset and will not be split into the training set and validation set. This division method can avoid the problem of data leakage across rounds of the same sample. In the scenario of a small sample size with very few titanium alloy physical samples, the limited samples are fully utilized to complete the model training and objective evaluation of the effect. The pre-trained visual feature extractor is transferred to the initial neural network model, and the extractor parameters are frozen based on the total number of titanium alloy samples, i.e., the total number of multimodal samples. In this scheme, there are only 12 samples, which is a typical small sample scenario. Therefore, all weights of the visual feature extractor are frozen throughout the process, and only the parameters of the visual feature mapping module, process parameter generation network, FiLM modulation module, and regression prediction network are iteratively updated. During each training fold, the network parameters of the trainable modules are updated by backpropagation using optimizers such as Adam, based on the loss function value. Regularization techniques such as early stopping, weight decay, and Dropout are used to suppress overfitting of small samples. If the validation loss does not decrease for several consecutive rounds, the single-fold training is terminated early.

[0055] For example, normalized K is used. IC Labels and Normalized K IC The mean squared error between the predicted values ​​is used as the training loss function: ;in, This indicates the number of prediction results used in loss calculation during training. Indicates the first Normalized K IC Predicted value Indicates the corresponding normalized K IC Real labels. The Adam optimizer is used to update the parameters of the visual feature mapping module, the process condition parameter generation network, the FiLM conditional modulation fusion module, and the regression prediction network. The training epochs are 50, the batch size is 16, the initial learning rate is 0.001, and the weight decay coefficient is 0.0001. During training, freezing the pre-trained visual backbone, Dropout regularization, weight decay, and early stopping strategies are used to reduce the risk of overfitting. The early stopping patience value is set to 15, meaning training stops when the validation loss does not decrease for 15 consecutive epochs. During each fold, the model weights with the lowest validation loss are saved and used for the K-values ​​of the validation samples in that fold. IC predict.

[0056] After single-fold training is completed, the next set of validation samples is switched, and the process of partitioning, training, and parameter optimization is repeated until all samples have undergone leave-one-fold testing. After all cross-validation processes are completed, the network weights with the best generalization effect are selected by combining the model evaluation metrics of each fold, and a complete titanium alloy mechanical property prediction model is obtained, which can be used to predict the fracture toughness of novel titanium alloy samples.

[0057] For example, let the set of all TB18 sample IDs be... , No. The sample number retained is Then the first The training and validation sets are represented as follows: ; ;in, Indicates the first Fold the training sample set, Indicates the first Verification sample set. All SEM images, heat treatment process parameters, and K0 of the same sample. IC Labels only appear in the same fold of the training or validation set. During the validation phase, the image-level K corresponding to each SEM image in the validation samples is first obtained. IC The predicted value is then averaged across all image-level predicted values ​​for the same sample to obtain the sample-level K. IC Predicted value: ;in, Indicates the first Sample level K of TB18 samples IC Predicted value Indicates the first The TB18 sample K corresponding to Zhang SEM image IC Predicted value This indicates the number of SEM images corresponding to this sample; After the 12-fold leave-one cross-validation, the sample-level K is based on 12 samples. IC The root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R²), and mean absolute percentage error (MAPE) are calculated based on the predicted and measured values ​​to evaluate the model's prediction accuracy and stability. The definitions of each indicator are as follows: ; ; ; ; in, Indicates the number of samples participating in the evaluation. Indicates the first Measured K for each sample IC value, Indicates the first K of samples IC Predicted value Indicates all measured K IC The average value To prevent constants with a denominator of zero.

[0058] Preferably, embodiments of the present invention also include several control models to verify the impact of input modality and feature fusion method on K. IC Impact on Predictive Performance. The comparison models include Tab-only, Image-only, Concat, GMU, CrossAttn, and FiLM. The Tab-only model is a single-modal model that uses only heat treatment process parameters for prediction; the Image-only model is a single-modal model that uses only SEM microstructure images for prediction; the Concat model is a multi-modal model that directly concatenates image features and process parameter features for prediction; the GMU model is a multi-modal model that combines image features and process parameter features through gated fusion; and the CrossAttn model is a multi-modal model that achieves interaction between image features and process parameter features through cross-attention. Each comparison model uses the same data partitioning method, training and evaluation process, and evaluation metrics as the titanium alloy performance prediction model of this embodiment, namely the TinyEfficientNet-FiLM model. The TinyEfficientNet-FiLM multi-modal performance prediction model is compared with the Tab-only, Image-only, Concat, GMU, and CrossAttn models. See Table 2, which shows the performance of different models provided in this embodiment on TB18 titanium alloy K... IC A comparison table of leave-one-out cross-validation results in the prediction task.

[0059] Table 2

[0060] Referring to Table 2, the comparison model includes two types of single-modal networks and three mainstream multimodal fusion networks: Tab-only only inputs heat treatment process parameters and has no feature fusion operation; Image-only only inputs SEM microstructure images and does not introduce process information; Concat uses a simple fusion method of directly concatenating image features and process parameters; GMU completes the fusion of two types of features through a gating mechanism; CrossAttn relies on cross-attention to realize the interaction between image and process features. The TinyEfficientNet-FiLM model provided in this embodiment of the invention has the best overall performance, with an RMSE of 4.6685 and a MAE of 3.1156. The RMSE reached 0.9201, with a MAPE of only 5.64%, making it the lowest indices and highest goodness of fit among all models, demonstrating superior overall prediction accuracy and stability. A horizontal comparison of the two single-modal models shows that, compared to the Tab-only model relying solely on process parameters, this approach reduces RMSE by approximately 30.8073%. Improved by 10.1916%; compared to the Image-only model that only uses microscopic images, the RMSE reduction was as high as 58.9889%. The improvement was 73.4729%. This comparison fully demonstrates that the microstructure evolution information recorded by the heat treatment process and the real morphological information carried by the SEM image are highly complementary. The fusion of dual-modal inputs can compensate for the lack of information from a single data source and significantly improve the accuracy of fracture toughness prediction. Furthermore, compared with three multimodal fusion schemes—Concat, GMU, and CrossAttn—the TinyEfficientNet-FiLM model still has lower RMSE and MAE. In the small-sample prediction scenario for TB18 titanium alloy, the FiLM conditional modulation fusion approach, which uses the heat treatment process as a conditional factor to dynamically modulate the microscopic image features, is superior to traditional feature interaction methods such as simple stitching, gated fusion, and cross-attention fusion. It can more accurately represent the intrinsic coupling law between heat treatment regime, microstructure morphology, and fracture toughness, effectively mitigating error fluctuations caused by small-sample training.

[0061] In this embodiment of the invention, a device for constructing a titanium alloy performance prediction model is also provided, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the above-described method for constructing a titanium alloy performance prediction model.

[0062] In this embodiment of the invention, a computer-readable storage medium is also provided, which includes a stored computer program, wherein the computer program controls the device where the computer-readable storage medium is located to execute the above-described method for constructing a titanium alloy performance prediction model when it is running.

[0063] For example, a computer program can be divided into one or more modules, one or more of which are stored in memory and executed by a processor to carry out the present invention. The one or more modules can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in a device for constructing a titanium alloy performance prediction model.

[0064] The device for building a titanium alloy performance prediction model can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. This device may include, but is not limited to, processors, memory, and displays. Those skilled in the art will understand that the above-mentioned components are merely examples of devices for building a titanium alloy performance prediction model and do not constitute a limitation on the device. The device may include more or fewer components than those listed above, or a combination of certain components, or different components. For example, the device for building a titanium alloy performance prediction model may also include input / output devices, network access devices, buses, etc.

[0065] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the titanium alloy performance prediction model construction device, connecting all parts of the device via various interfaces and lines.

[0066] The memory can be used to store computer programs and / or modules. The processor, by running or executing the computer programs and / or modules stored in the memory, and by calling the data stored in the memory, realizes various functions of the device for constructing the performance prediction model of titanium alloys. The memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function (such as sound playback function, text conversion function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, text message data, etc.). In addition, the memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0067] The module based on the construction of the titanium alloy performance prediction model, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. Those skilled in the art can understand and implement this without any inventive effort.

[0068] This invention provides a method for constructing a titanium alloy performance prediction model. It constructs a multimodal dataset binding heat treatment processes, mechanical labels, and microscopic images to address the problem of incomplete representation by single-class data. Randomly cropping image sub-blocks expands the sample base. Simultaneously, auxiliary image data with microstructure category labels is built. Based on this auxiliary data, the visual backbone classification is pre-trained, allowing the visual backbone to learn the unique microstructure texture and morphological features of titanium alloys in advance, forming a transferable feature extractor. This eliminates the need to train deep networks from scratch, mitigating overfitting issues with small sample sizes. Furthermore, the initial model adopts a hierarchical module design, unifying the feature dimensions of visual feature mapping to ensure matching in subsequent modulation operations. Unlike simple concatenation, the FiLM module recalibrates the image modal features based on conditional modulation parameters generated from heat treatment process parameters, dynamically correcting morphological features as constraints, fully exploring the intrinsic relationship between process, microstructure, and performance. Finally, a regression network completes the mechanical performance fitting output. By leveraging pre-trained transfer learning to reduce the difficulty of training with small samples, and by utilizing multimodal complementary information to improve representation capabilities, deep fusion of two types of data is achieved through FiLM conditional modulation. Compared with single-modal and simple fusion schemes, this can effectively improve the accuracy and generalization stability of predicting the mechanical properties of titanium alloys, and reduce experimental testing costs.

[0069] Example 2 See Figure 3 , Figure 3 This is a flowchart illustrating the application method of the titanium alloy performance prediction model provided in one embodiment of the present invention. The application method of the titanium alloy performance prediction model provided in this embodiment includes steps 201 to 202, each step of which is detailed below: Step 201: Collect the titanium alloy sample to be evaluated; wherein the titanium alloy sample includes associated heat treatment process parameters, microstructure images and / or microstructure image sub-blocks obtained by random cropping based on the microstructure images.

[0070] In this embodiment of the invention, for a novel titanium alloy specimen requiring evaluation of its mechanical properties, its supporting data are collected simultaneously. The first type is the complete heat treatment process parameters of the specimen, covering information such as solution treatment, aging-related temperatures, durations, and cooling methods. The second type is the original microstructure image of the specimen. Simultaneously, multiple image sub-blocks can be generated by randomly cropping the effective microstructure area of ​​the image to enrich the observation field of view. All data are interconnected using the specimen's unique number as an index, forming a standard multimodal sample format that matches the data input specifications during model training, avoiding prediction anomalies caused by inconsistent data formats.

[0071] Step 202: Input the titanium alloy sample into the titanium alloy performance prediction model so that the titanium alloy performance prediction model outputs the prediction results of the titanium alloy sample in terms of the target mechanical properties; wherein, the titanium alloy performance prediction model is constructed using the construction method of the titanium alloy performance prediction model described in Example 1.

[0072] In this embodiment of the invention, standardized multimodal samples are fed into a titanium alloy performance prediction model trained using the method described above. The model sequentially performs a complete process including image preprocessing, process parameter standardization, image feature extraction, feature dimension mapping, FiLM process condition modulation, regression prediction, and label inverse normalization, ultimately outputting the target mechanical properties of the sample (such as fracture toughness K). IC The model is built upon multimodal fusion and pre-trained transfer learning mechanisms, enabling it to simultaneously utilize historical information from heat treatment processes and microstructure morphology information. It can quickly obtain reliable mechanical property prediction results without the need for complex and costly mechanical destructive tests.

[0073] For example, a brand-new heat-treated TB18 sample was taken, and all seven heat treatment process parameters (solution treatment and aging) were recorded. Microscopic images of the sample were captured using SEM, and four 224×224 pixel image sub-blocks were randomly cropped from the effective microstructure area. All data were bound together with the sample number to form a multimodal sample to be evaluated. The compiled process parameters and SEM image sub-blocks were input into the TinyEfficientNet-FiLM titanium alloy performance prediction model trained using the construction method of Example 1. The model automatically completed image normalization, process feature standardization, visual feature extraction, FiLM condition modulation, regression prediction, and inverse normalization calculation, finally outputting the plane strain fracture toughness K of the sample. IC Predicted values ​​enable rapid, non-destructive performance evaluation.

[0074] This invention provides a method for applying a titanium alloy performance prediction model. First, the heat treatment process parameters, microstructure images, and cut sub-blocks of the titanium alloy sample to be evaluated are collected as model inputs. These are then fed into the titanium alloy performance prediction model to complete the mechanical property prediction. This model is built upon a pre-trained visual feature extractor and a FiLM conditional modulation fusion structure, simultaneously utilizing process evolution information and microstructure features. Compared to evaluation methods relying solely on a single process or image, it provides more complete information dimensions and higher prediction accuracy. The entire application process only requires sample characterization and process information recording, eliminating the need for destructive mechanical tests such as fracture toughness tests. This significantly shortens the sample testing cycle and reduces the cost of test consumables and equipment wear. Simultaneously, the microstructure image cut sub-blocks provide multiple perspectives, reducing prediction bias caused by local microstructure. Combined with a model optimized through multi-fold cross-validation, prediction stability is ensured, enabling rapid non-destructive prediction of the mechanical properties of new titanium alloy samples. This provides a convenient and feasible technical means for rapid screening of heat treatment processes and early evaluation of material properties.

[0075] Example 3 See Figure 4 , Figure 4 This is a schematic diagram of the modules of a titanium alloy performance prediction model construction device provided in one embodiment of the present invention. The embodiment of the present invention provides a titanium alloy performance prediction model construction device, which includes: a first data module 301, a second data module 302, a network training module 303, and a model construction module 304; The first data module 301 constructs a multimodal performance dataset based on titanium alloy samples from different heat treatment batches; wherein, each titanium alloy sample includes associated heat treatment process parameters, mechanical property test labels, microstructure images and / or microstructure image sub-blocks obtained by randomly cropping microstructure images; The second data module 302 constructs auxiliary image data based on the microstructure images of titanium alloys acquired under different heat treatment process conditions and their corresponding microstructure category labels; wherein, the microstructure images of titanium alloys include a variety of typical microstructure morphologies; The network training module 303 trains the visual backbone network on the auxiliary image data to perform a category classification task, thereby obtaining a visual feature extractor for extracting the morphological features of titanium alloys. The model building module 304 is used to transfer the visual feature extractor to the pre-built initial neural network model, and iteratively train the initial neural network model based on the multimodal performance dataset until the loss function converges or the preset number of iterations is reached. The iteration training is then determined to be complete, and a titanium alloy performance prediction model for outputting mechanical performance prediction results is obtained. The initial neural network model includes a visual feature mapping module, a FiLM conditional modulation fusion module, and a regression prediction network connected in sequence. The visual feature mapping module maps the microstructure image feature vectors output by the visual feature extractor to a preset dimension to obtain image modal features; the FiLM conditional modulation fusion module recalibrates the image modal features based on the conditional modulation parameters generated by the heat treatment process parameters; and the regression prediction network makes predictions based on the recalibrated image modal features and outputs normalized mechanical performance prediction results.

[0076] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0077] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and substitutions without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing a performance prediction model for titanium alloys, characterized in that, include: A multimodal performance dataset is constructed based on titanium alloy samples from different heat treatment batches; wherein each titanium alloy sample includes associated heat treatment process parameters, mechanical property test labels, microstructure images and / or microstructure image sub-blocks obtained by randomly cropping the microstructure images; Auxiliary image data is constructed based on the microstructure images of titanium alloys acquired under different heat treatment conditions and their corresponding microstructure category labels; wherein, the microstructure images of titanium alloys include a variety of typical microstructure morphologies; The visual backbone network is trained on the auxiliary image data to perform an organizational category classification task, resulting in a visual feature extractor for extracting the morphological features of titanium alloys. The visual feature extractor is transferred to a pre-constructed initial neural network model. The initial neural network model is iteratively trained based on the multimodal performance dataset until the loss function converges or the preset number of iterations is reached. The iterative training is then considered complete, resulting in a titanium alloy performance prediction model for outputting mechanical performance prediction results. The initial neural network model includes a visual feature mapping module, a FiLM conditional modulation fusion module, and a regression prediction network connected in sequence. The visual feature mapping module is used to map the microstructure image feature vector output by the visual feature extractor to a preset dimension to obtain image modal features; the FiLM conditional modulation fusion module is used to recalibrate the image modal features according to the conditional modulation parameters generated by the heat treatment process parameters; the regression prediction network is used to make predictions based on the recalibrated image modal features and output normalized mechanical performance prediction results.

2. The method for constructing the titanium alloy performance prediction model as described in claim 1, characterized in that, The multimodal performance dataset constructed based on titanium alloy samples from different heat treatment batches specifically includes: Obtain titanium alloy samples from different heat-treated batches; A microstructure image of each titanium alloy sample is obtained, and the effective tissue area in the microstructure image is randomly cropped to obtain several microstructure image sub-blocks; Obtain the heat treatment process parameters for each of the titanium alloy samples; wherein, the heat treatment process parameters include at least one of the following: solution type, solution temperature, solution time, solution cooling method, aging temperature, aging time, and aging cooling method; Obtain mechanical property test labels for each of the titanium alloy samples; wherein the mechanical property test labels include at least one of fracture toughness, tensile strength, yield strength, elongation after fracture, reduction of area, impact toughness, fatigue life, and fatigue crack propagation performance; Based on the unique sample number of each titanium alloy sample, the heat treatment process parameters, mechanical property test labels, microstructure images, and / or microstructure image sub-blocks of the same titanium alloy sample are associated as a multimodal sample; wherein, the data expression of the multimodal sample is: ； In the formula, For the first Data records for one titanium alloy sample; For the first The first titanium alloy sample corresponding to the first Zhang microscopic tissue images or image sub-blocks; For the first Number of microstructure images or image sub-blocks corresponding to each titanium alloy sample; For the first Heat treatment process parameters for one titanium alloy sample; For the first Mechanical property test labels for individual titanium alloy samples; The multimodal performance dataset is constructed based on the multimodal samples; wherein the multimodal samples undergo integrity checks and consistency verification.

3. The method for constructing the titanium alloy performance prediction model as described in claim 1, characterized in that, The auxiliary image data constructed based on the microstructure images of titanium alloys acquired under different heat treatment conditions and their corresponding microstructure category labels specifically includes: Collect images of the microstructure of titanium alloys under different heat treatment conditions and their corresponding microstructure category labels; The cracked areas in the microstructure image of the titanium alloy are marked, and the microstructure image of the titanium alloy is cropped by a pixel sliding window of a preset size to obtain several image sub-blocks that do not contain cracked areas. Image sub-blocks derived from the same titanium alloy microstructure image are divided into the same data subset, and the auxiliary image data is constructed based on the data subset.

4. The method for constructing the titanium alloy performance prediction model as described in claim 1, characterized in that, The step of training the visual backbone network on the auxiliary image data to perform an organizational category classification task, resulting in a visual feature extractor for extracting the morphological features of titanium alloys, specifically includes: A visual backbone network is constructed based on a pre-defined neural network model; wherein, the neural network includes at least one of ResNet, MobileNet, EfficientNet, ConvNeXt, VisionTransformer, or improved networks thereof; The visual backbone network is iteratively trained based on the auxiliary image data until the performance index of the visual backbone network reaches the preset requirements. The training of the visual backbone network is then determined to be complete, resulting in a visual feature extractor capable of extracting texture, boundary, phase distribution, tissue scale differences, and morphological differences in microscopic tissue images. Among these features, when training the visual backbone network for tissue category classification, a category-weighted cross-entropy loss is used.

5. The method for constructing the titanium alloy performance prediction model as described in claim 1, characterized in that, Before iteratively training the initial neural network model based on the multimodal performance dataset, the method further includes preprocessing the multimodal performance dataset for model input, including: The microstructure image is preprocessed and pixel normalized to obtain a standardized microstructure image; wherein the preprocessing operations include at least one of grayscale conversion, noise reduction, brightness correction, contrast enhancement, histogram equalization, adaptive histogram equalization, scale unification, and target region cropping; the expression for the standardized microstructure image is: ； In the formula, To standardize microscopic tissue images; Pixel normalization processing; For optional data augmentation operations; For image preprocessing operations; For the first The first titanium alloy sample corresponding to the first Zhang microscopic tissue images or image sub-blocks; The heat treatment process parameters are standardized according to their type to generate standardized process features, which are then spliced ​​together to obtain a structured heat treatment process parameter vector; wherein, the expression of the structured heat treatment process parameter vector is: ； In the formula, For the first A structured heat treatment process parameter vector for a titanium alloy sample; It is a numerical process feature; Indicates categorical process characteristics; This indicates the number of numerical heat treatment process parameters; Indicates the number of category-type heat treatment process parameters; Indicates feature concatenation operation; The mechanical property test labels are normalized to obtain normalized performance labels; wherein the expression for the normalized performance labels is: ； In the formula, For the first Normalized mechanical property test labels for individual titanium alloy samples; For the first Mechanical property test labels for individual titanium alloy samples; and These are the minimum and maximum values ​​of the corresponding mechanical performance test labels in the current training subset, respectively; To prevent extremely small constants with a denominator of zero.

6. The method for constructing the titanium alloy performance prediction model as described in claim 5, characterized in that, The visual feature mapping module is used to map the microstructure image feature vector output by the visual feature extractor to a preset dimension to obtain image modal features; the FiLM conditional modulation fusion module is used to recalibrate the image modal features according to the conditional modulation parameters generated by the heat treatment process parameters; the regression prediction network is used to make predictions based on the recalibrated image modal features and output normalized mechanical performance prediction results, specifically including: The standardized microstructure image is input to the visual feature extractor, which extracts texture, boundary, phase distribution, tissue scale differences, or morphological differences from the standardized microstructure image to obtain an image feature vector, which is then input to the visual feature mapping module; wherein, the expression of the image feature vector is: ； In the formula, The feature vector of the microscopic tissue image; The parameter is Visual feature extractor; To standardize microscopic tissue images; The visual feature mapping space maps the received image feature vectors to a fusion feature space that matches the conditional modulation parameters, obtaining image modal features, which are then input to the FiLM conditional modulation fusion module; wherein, the expression of the image modal features is: ； In the formula, The mapped image modal features; For visual feature mapping module; The FiLM conditional modulation fusion module generates FiLM conditional modulation parameters of a preset dimension based on the structured heat treatment process parameter vector, and recalibrates the image modal features based on the FiLM conditional modulation parameters to obtain multimodal fusion features constrained by heat treatment process conditions, which are then input into the regression prediction network. The regression prediction network performs nonlinear mapping on the received multimodal fusion features to obtain normalized prediction results of the mechanical properties of titanium alloys.

7. The method for constructing the titanium alloy performance prediction model as described in claim 6, characterized in that, The FiLM conditional modulation fusion module generates FiLM conditional modulation parameters of a preset dimension based on the structured heat treatment process parameter vector, and recalibrates the image modal features based on the FiLM conditional modulation parameters to obtain multimodal fusion features constrained by heat treatment process conditions, which are then input into the regression prediction network. Specifically, this includes: The structured heat treatment process parameter vector is input into the process condition parameter generation network to generate conditional modulation parameters for modulating the image modal features; wherein, the conditional modulation parameters include at least one of scaling parameters, offset parameters, gating coefficients, attention weights, or combinations thereof; The image modal features are modulated element-wise based on the conditional modulation parameters to generate multimodal fusion features constrained by heat treatment process conditions; wherein, the expression of the multimodal fusion features is: ； In the formula, This is a multimodal fusion feature; and These represent the scaling parameter and offset parameter generated based on the structured heat treatment process parameter vector, respectively; This is the identifier for element-wise multiplication; These are the mapped image modal features.

8. The method for constructing the titanium alloy performance prediction model as described in claim 1, characterized in that, The process of transferring the visual feature extractor to a pre-constructed initial neural network model, iteratively training the initial neural network model based on the multimodal performance dataset until the loss function converges or a preset number of iterations is reached, and determining that the iterative training is complete, yields a titanium alloy performance prediction model for outputting mechanical performance prediction results, specifically including: Leave-one-out cross-validation training is performed based on the unique sample number of each titanium alloy sample; wherein, each leave-one-out cross-validation training involves selecting one titanium alloy sample as the validation sample and the remaining titanium alloy samples as the training samples. In each training iteration, the parameters of the initial neural network model are trained based on the training sample parameters in the current training iteration, and the training loss function between the output mechanical performance prediction result and the mechanical performance prediction result of the titanium alloy sample is calculated; wherein, when training the initial neural network model, the feature extraction part of the visual feature extractor is transferred to the initial neural network model, and the parameters of the feature extraction part are frozen according to the data scale of the titanium alloy sample; The parameters of the initial neural network model in the next training session are optimized based on the training loss function output from the previous training session, until the leave-one-out cross-validation training is completed, thus obtaining the titanium alloy performance prediction model.

9. The application method of the titanium alloy performance prediction model, characterized in that, include: Collect titanium alloy samples to be evaluated; wherein, the titanium alloy samples include associated heat treatment process parameters, microstructure images and / or microstructure image sub-blocks obtained by randomly cropping the microstructure images; The titanium alloy sample is input into the titanium alloy performance prediction model so that the titanium alloy performance prediction model outputs the prediction results of the titanium alloy sample in terms of the target mechanical properties; wherein, the titanium alloy performance prediction model is constructed using the construction method of the titanium alloy performance prediction model as described in any one of claims 1 to 8.

10. A device for constructing a performance prediction model for titanium alloys, characterized in that, include: The system comprises a first data module, a second data module, a network training module, and a model building module. The first data module constructs a multimodal performance dataset based on titanium alloy samples from different heat treatment batches; wherein, each titanium alloy sample includes associated heat treatment process parameters, mechanical property test labels, microstructure images and / or microstructure image sub-blocks obtained by randomly cropping the microstructure images; The second data module constructs auxiliary image data based on the microstructure images of titanium alloys acquired under different heat treatment process conditions and their corresponding microstructure category labels; wherein, the microstructure images of titanium alloys include a variety of typical microstructure morphologies; The network training module trains the visual backbone network on the auxiliary image data to perform an organizational category classification task, thereby obtaining a visual feature extractor for extracting the morphological features of titanium alloys. The model building module is used to transfer the visual feature extractor to a pre-built initial neural network model, and iteratively train the initial neural network model based on the multimodal performance dataset until the loss function converges or the preset number of iterations is reached, at which point the iterative training is determined to be complete, and a titanium alloy performance prediction model for outputting mechanical performance prediction results is obtained; wherein, the initial neural network model includes a visual feature mapping module, a FiLM conditional modulation fusion module, and a regression prediction network connected in sequence. The visual feature mapping module is used to map the microstructure image feature vector output by the visual feature extractor to a preset dimension to obtain image modal features; the FiLM conditional modulation fusion module is used to recalibrate the image modal features according to the conditional modulation parameters generated by the heat treatment process parameters; the regression prediction network is used to make predictions based on the recalibrated image modal features and output normalized mechanical performance prediction results.

Images