Amorphous alloy quasi-static damage and atomic structure correlation analysis method based on improved hybrid model

Abstract

The application belongs to the technical field of amorphous alloy material performance test, and discloses an amorphous alloy quasi-static damage and atomic structure correlation analysis method based on an improved hybrid model. First, a standard amorphous alloy sample is prepared and quasi-static tensile loading is performed. Time-series atomic structure parameters and multi-modal damage data in the loading process are synchronously collected. After scene-based preprocessing of the data, the improved hybrid model composed of a double-branch feature extraction module and a cross-modal attention fusion module is input. The model is converged by adopting a segmented loss function for stage-by-stage training. The method adaptively matches multi-modal data through a double-branch architecture, matches loading stage characteristics by using dynamic fusion weights, realizes high-precision and stage-sensitive quantitative correlation analysis between atomic structure and macroscopic damage, and effectively overcomes the problems of insufficient modal fusion and poor stage adaptability of traditional methods.

Description

Technical Field

[0001] This application belongs to the field of performance testing technology for amorphous alloy materials, specifically involving a method for quasi-static damage and atomic structure correlation analysis of amorphous alloys based on an improved hybrid model. Background Technology

[0002] Amorphous alloys, due to their unique disordered atomic packing structure, exhibit high strength, high hardness, and excellent corrosion resistance, showing broad application prospects in cutting-edge fields such as precision instruments and aerospace. Their macroscopic mechanical properties and damage behavior are fundamentally controlled by the dynamic evolution of their microscopic atomic structure. Under quasi-static loading conditions, the damage process of amorphous alloys, such as the initiation and propagation of microcracks, is intricately coupled with atomic-scale structural transformations, including the evolution of interatomic spacing, coordination number, and short-range ordered domains. Therefore, accurately establishing the quantitative correlation between microscopic atomic structure parameters and macroscopic damage evolution is a key scientific issue for deeply understanding the deformation mechanism of amorphous alloys and guiding their composition design and process optimization.

[0003] Existing technologies offer various analytical methods for exploring the aforementioned correlations. In experimental characterization, X-ray diffraction, high-resolution transmission electron microscopy, and extended X-ray absorption fine structure spectroscopy are typically used to obtain structural information, combined with optical microscopy and acoustic emission techniques to monitor damage. In data analysis, traditional methods often rely on phenomenological models based on physical experience or single machine learning models for correlation analysis. However, when facing the complex multiphysics problem of "structure-damage" correlation analysis in amorphous alloys, these methods struggle to meet the demands for high-precision and intelligent analysis. This is mainly reflected in the following aspects: Insufficient data modality fusion. The atomic structure data of amorphous alloys is a multi-parameter time series, while damage information includes multi-source heterogeneous data such as images and acoustic emission signals. Traditional analysis models, such as single convolutional neural networks or random forests, are usually designed for specific data types and lack the ability to collaboratively extract and deeply fuse multi-modal features, resulting in incomplete information utilization and one-sided feature representation.

[0004] The dynamic characteristics of the loading process are neglected. The quasi-static damage evolution of amorphous alloys exhibits distinct stages, typically divided into elastic deformation, damage initiation, and damage propagation. The dominant structure-damage correlation mechanism differs significantly within each stage. Existing models mostly employ fixed network architectures and uniform loss functions to fit the entire loading process, failing to dynamically adjust feature fusion strategies and optimization objectives based on the physical nature of different stages. This results in models being insensitive to stage transitions, and analysis results are prone to bias near critical points.

[0005] The quality of experimental data constrains model accuracy. In actual testing, damage optical images are easily affected by surface reflections, masking early microcrack features; during high strain stages, specimens may fracture, leading to sparse or interrupted data sequences. Traditional data preprocessing methods have limited ability to handle such noise and missing values. Directly inputting poor-quality raw data into the model introduces significant errors, affecting the reliability of correlation analysis.

[0006] In summary, existing technologies lack sufficient analytical precision for the intrinsic relationship between structure and damage in amorphous alloys, making it difficult to extract reliable and universal quantitative laws from massive, multi-source, and noisy experimental data, thus hindering the process of data-driven material performance optimization. Summary of the Invention

[0007] To address the core problems of existing technologies in analyzing the correlation between quasi-static damage and atomic structure in amorphous alloys, such as difficulties in multimodal data fusion, inability to adapt to differences in damage evolution stages, and weak resistance to experimental interference, this application provides a correlation analysis method for quasi-static damage and atomic structure in amorphous alloys based on an improved hybrid model. By providing a correlation analysis method based on an improved hybrid model, high-precision, stage-sensitive quantitative correlation analysis between atomic structure parameters and macroscopic damage behavior can be achieved, thereby providing a reliable basis for material performance optimization.

[0008] To achieve the above technical objectives, this application specifically adopts the following technical solution: In one aspect of this application, a method for quasi-static damage and atomic structure correlation analysis of amorphous alloys based on an improved hybrid model is provided, comprising the following steps: S1. Prepare standard amorphous alloy specimens and subject them to quasi-static tensile loading; S2. During the quasi-static loading process, atomic structure data and damage data of the sample are acquired simultaneously; the atomic structure data includes structural parameters of multiple time nodes; the atomic structure data includes at least interplanar spacing, short-range ordered domain size, coordination number and interatomic spacing, and the damage data includes damage images and acoustic emission signals of the corresponding time nodes. S3. Normalize the atomic structure data and enhance and complete the damage data to obtain preprocessed multimodal time series data. S4. Construct and train an improved hybrid model, which includes a dual-branch feature extraction module and a cross-modal attention fusion module. The dual-branch feature extraction module is used to extract feature vectors from the preprocessed atomic structure data and damage data, respectively. The cross-modal attention fusion module dynamically allocates weights according to the quasi-static loading stage and performs weighted fusion of the feature vectors of the two branches. S5. The improved hybrid model is trained in stages using a piecewise loss function until convergence, and the macroscopic damage assessment results and microstructure-damage correlation patterns are finally output. The staged training includes a pre-training stage and a fusion training stage.

[0009] In one embodiment, in step S1, the standard amorphous alloy sample is Zr. 55 Cu 30 Al 10 Ni5 or Ti 60 Cu 20 Ni 15 Sn5 amorphous alloy was produced by vacuum melting and copper mold spraying, and XRD verification confirmed it to be amorphous; the quasi-static loading rate was 1×10⁻⁶. -3 s -1 .

[0010] In one implementation, in step S2, the strain interval for acquiring the atomic structure data is 0.1%.

[0011] In one implementation, in step S3, the normalization process is to standardize the atomic structure data to the [0,1] interval based on the parameters of the amorphous state; the enhancement process includes using a multi-scale residual denoising unit to denoise the damage image, and then converting the acoustic emission signal parameters into a thermogram and stitching it with the denoised damage image to form multi-channel damage data; the completion process is to perform linear interpolation based on the characteristic correlation of adjacent strain nodes for the missing time-series nodes in the high strain stage.

[0012] In one implementation, in step S4, the dual-branch feature extraction module includes an atomic structure branch and a damage feature branch; The atomic structure branch includes sequentially connected 1D convolutional layers and LSTM layers, used to extract local correlations and temporal evolution features of structural parameters; The damage feature branch includes a 3D convolutional layer, a global average pooling layer, and a fully connected layer, which are used to extract spatial-temporal features of the damage image and fuse acoustic emission features.

[0013] In one implementation, the cross-modal attention fusion module dynamically allocates fusion weights according to three stages of quasi-static loading: in the elasticity stage, the weight of atomic structure features is 0.7 and the weight of damage features is 0.3; in the damage initiation stage, the weights of atomic structure features and damage features are each 0.5; and in the damage expansion stage, the weight of atomic structure features is 0.3 and the weight of damage features is 0.7.

[0014] In one implementation, in step S5, the segmented loss function is a mean squared error loss function with weights adjusted according to the stage, and the total loss function is: structural loss weight × mean squared error of atomic structure prediction value + damage loss weight × mean squared error of damage prediction value; the pre-training stage independently trains the dual-branch feature extraction module, and the fusion training stage connects the cross-modal attention fusion module for joint training.

[0015] In one implementation, the convergence criteria for the improved hybrid model are: the damage area prediction error of the validation set is less than 4% for 10 consecutive rounds, and / or the microstructure-damage correlation bias is less than 3%.

[0016] In one implementation, the macroscopic damage assessment results output in step S5 include the predicted damage area and damage level; the microstructure-damage correlation includes the contribution of coordination number and / or interatomic spacing to the damage and the damage-sensitive structure threshold.

[0017] In another aspect of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the method for correlation analysis of quasi-static damage and atomic structure of amorphous alloys based on an improved hybrid model.

[0018] The beneficial effects of this application are as follows: The dual-branch architecture employed in this application can fully adapt to the different characteristics of atomic structure time-series data and damage image-signal fusion data, achieving comprehensive extraction and deep fusion of multimodal features, thereby significantly improving the completeness and accuracy of correlation analysis. By introducing a stage-based dynamic weight allocation mechanism and a piecewise loss function, the model can accurately capture and adapt to the dominant damage mechanisms of amorphous alloys at different loading stages, effectively overcoming the analytical bias caused by the stage insensitivity of traditional models. Addressing the inherent noise and missing data in experimental data, the designed scenario-based preprocessing module significantly enhances the model's anti-interference ability and data robustness. Ultimately, the method in this application can not only predict macroscopic damage evolution with high accuracy, but also quantitatively reveal the contribution and threshold law of key microstructural parameters to damage behavior, providing guidance with both theoretical depth and practical value for material composition design and process optimization. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the method for quasi-static damage and atomic structure correlation analysis of amorphous alloys based on an improved hybrid model, as described in this application. Detailed Implementation

[0020] The technical solution of this application will be clearly and completely described below with reference to specific embodiments. However, those skilled in the art will understand that the embodiments described below are only some embodiments of this application, not all embodiments, and are only used to illustrate this application, and should not be regarded as limiting the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To address the shortcomings of existing analytical methods in modal fusion, stage adaptation, and data robustness, this application constructs a complete multimodal dataset by simultaneously collecting temporally sequenced atomic structure parameters and image-based and signal-based damage data during the loading process. Targeted preprocessing, including normalization, denoising, and completion, is applied to suppress experimental interference. A neural network model with independent structural and damage branches is designed to extract deep features from different data types. A fusion mechanism that dynamically adjusts weights according to the loading stage (elasticity, damage initiation, and expansion) is introduced to match the feature fusion strategy with the dominant physical mechanisms at each stage. A staged training strategy and a segmented weighted loss function are employed to guide the model to accurately learn the correlation patterns at different stages. Finally, the model simultaneously outputs quantitative predictions of macroscopic damage and contribution analyses of microscopic structural parameters, thus establishing a reliable bridge from the atomic scale to macroscopic performance.

[0022] In one specific embodiment, this application provides a method for quasi-static damage and atomic structure correlation analysis of amorphous alloys based on an improved hybrid model, referring to... Figure 1 As shown, it includes the following steps: S1. Sample preparation and quasi-static loading A standard amorphous alloy specimen was prepared and subjected to quasi-static tensile loading.

[0023] In some embodiments, the standard amorphous alloy sample is Zr. 55 Cu 30 Al 10 Ni5 or Ti 60 Cu 20 Ni 15 Sn5 amorphous alloy is produced by vacuum melting and copper mold casting. Specifically, the proportioned raw materials are placed in an electric arc melting furnace, and the vacuum level inside the furnace is evacuated to 1×10⁻⁶. -4 After Pa, high-purity argon gas is introduced as a protective atmosphere. To ensure the uniformity of the alloy composition, at least three repeated melting processes are performed. Subsequently, the fully molten alloy is poured into a copper mold preheated to 200°C for spray casting, which is then rapidly cooled and shaped to obtain a plate-shaped standard sample with dimensions of 10 mm (length) × 5 mm (width) × 2 mm (thickness).

[0024] Furthermore, XRD analysis confirmed that the sample was amorphous. The prepared sample was analyzed using an X-ray diffractometer, with a scanning range typically from 20° to 80° and a step size of 0.02°. When the obtained XRD spectrum exhibits typical amorphous diffuse scattering peaks and no obvious sharp crystalline diffraction peaks, the sample can be verified as completely amorphous and meets the testing requirements.

[0025] In some embodiments, the quasi-static loading employs a unidirectional stretching mode with a loading rate of 1×10⁻⁶. -3 s -1 To simulate quasi-static deformation conditions, a quasi-static loading is implemented using an electronic universal testing machine, such as the Instron 5969. During installation, careful adjustment using an optical alignment device ensures that the longitudinal axis of the specimen is precisely aligned with the loading direction to eliminate additional bending moments caused by eccentric loading.

[0026] S2, Multimodal Data Synchronous Acquisition During the quasi-static loading process, atomic structure data and damage data of the sample are acquired simultaneously. The atomic structure data includes structural parameters at multiple time nodes. The atomic structure data includes at least interplanar spacing, short-range ordered domain size, coordination number, and interatomic spacing. The damage data includes damage images and acoustic emission signals at corresponding time nodes.

[0027] The atomic structure data includes one or more sets of structural parameters acquired at each time node. These structural parameters are obtained through the combined use of multiple complementary microstructure characterization techniques, and at least include: Interplanar spacing: Reflects the statistical average distance of atomic stacking, obtained by analyzing diffraction patterns using X-ray diffraction (XRD); Short-range ordered domain size: describes the size of localized atomic ordered clusters, obtained by statistical analysis of atomic images using high-resolution transmission electron microscopy (HRTEM) or by fitting XRD patterns; Coordination number: The average number of nearest-neighbor atoms around a specified atom, reflecting the local atomic packing density. It is obtained by fitting the absorption edge data of a specific element through extended X-ray absorption fine structure spectroscopy (EXAFS). Interatomic spacing: the average distance between a specific pair of atoms, obtained through EXAFS analysis; Understandably, in practice, XRD, HRTEM, EXAFS, and other equipment can be used in combination to obtain one or more of the above parameters from parallel samples prepared in the same batch and subjected to the same loading conditions to reach the target strain. The set of parameters obtained at each time point constitutes a quantitative description of the atomic structure at that moment.

[0028] In some embodiments, the strain interval for acquiring the atomic structure data is 0.1%.

[0029] The damage data aims to record the physical damage and acoustic characterization of the specimen during loading. At each time point, two types of data are acquired simultaneously: damage images and acoustic emission signals. Damage images: Digital images acquired using optical microscopes, scanning electron microscopes, or other surface imaging devices, reflecting the damage morphology (such as microcrack initiation and propagation) of a sample surface or specific cross-section under specific strain. Understandably, the images should have sufficient spatial resolution to identify target damage features, such as micrometer- or submicrometer-sized cracks.

[0030] Acoustic emission signal: An elastic wave signal generated by the instantaneous release of energy during material deformation and damage, acquired by an acoustic emission sensor coupled to the sample. Recorded acoustic emission signal parameters include at least the event amplitude, energy, count, or event rate.

[0031] In some embodiments, at each predetermined strain time point, surface morphology images of the specimen and their corresponding acoustic emission activity signals are acquired synchronously. The surface morphology images are captured using an optical microscope system with a digitized format of 32×32 pixels and a theoretical optical resolution of at least 0.1 μm to ensure the identification of micrometer-scale damage features. The synchronously acquired acoustic emission signals are recorded with the following parameters: a dynamic range of amplitude for a single event between 0 and 80 dB, and a count of 0 to 50 events detected within each strain interval.

[0032] Through synchronous acquisition, for the i-th strain time series node, a corresponding atomic structure data sample (containing a set of structural parameter values) and a corresponding damage data sample (containing a damage image and a set of acoustic emission signal parameters) will be obtained. The data samples of all time series nodes are arranged in strain order, and together they constitute a multimodal, time-series dataset for subsequent correlation analysis.

[0033] S3, Data Contextualized Preprocessing The atomic structure data is normalized, and the damage data is enhanced and completed to obtain preprocessed multimodal time series data.

[0034] To eliminate the differences in absolute parameter values ​​caused by different amorphous alloy compositions, in some embodiments, the atomic structure data is normalized. The initial state of the material in a completely amorphous state without loading and deformation is selected as the reference state, and the parameter values ​​measured by characterization methods such as XRD in this state are used as the normalization reference values. For any structural parameter, the measured values ​​at all time nodes during the entire loading process are mapped to the numerical range [0,1] through a linear transformation based on the maximum and minimum values ​​of the parameter in the entire dataset.

[0035] In some embodiments, the enhancement process includes denoising the damage image using a multi-scale residual denoising unit, and then stitching the acoustic emission signal parameters into a heatmap and the denoised damage image together to form multi-channel damage data.

[0036] Specifically, a multi-scale residual denoising network unit is constructed to process the original damage image. This multi-scale residual denoising network unit uses convolutional layers with different kernel sizes (e.g., 5×5 and 3×3) in parallel to extract features. Larger kernels focus on capturing a large area of ​​background and suppressing low-frequency noise such as reflections; smaller kernels are used to extract fine edge features related to protecting microcracks (e.g., those less than 1 μm wide). Through network learning, the final output is a clear image that suppresses noise while preserving key damage morphology details to the greatest extent possible.

[0037] The acoustic emission signal parameters corresponding to each strain time-series node, namely the event amplitude and event count, are transformed into two-dimensional matrices (heatmaps) with the same spatial dimensions as the damage image. Each pixel in the matrix represents the parameter intensity at its corresponding spatial location. Then, the denoised damage image is used as the first channel, the amplitude heatmap as the second channel, and the event count heatmap as the third channel. These are stitched together along the channel dimension to generate a three-channel composite data sample. This sample simultaneously contains visual information about the damage's morphology and its acoustic activity.

[0038] In some embodiments, the completion process is performed by linear interpolation based on the characteristic correlation of adjacent strain nodes for missing time-series nodes in the high-strain stage.

[0039] To address the issue of missing data at subsequent time-series nodes due to potential fracture of specimens at high strain stages (strain exceeding 1.2%), data completion is implemented. Based on the continuity of material deformation and damage evolution, complete data (including structural parameters and three-channel damage data) from the nearest effective time-series nodes before and after the missing node (e.g., the preceding 1.2% strain node and the following 1.4% strain node) are used to analyze the changing trends of various characteristics between nodes. A linear interpolation algorithm is then employed to estimate the characteristic values ​​corresponding to the missing node (e.g., the 1.3% strain node) based on these trends, thereby generating a complete data sample for that node.

[0040] S4. Construct and train the improved hybrid model. The improved hybrid model includes a dual-branch feature extraction module and a cross-modal attention fusion module. The dual-branch feature extraction module is used to extract feature vectors from the preprocessed atomic structure data and damage data, respectively. The cross-modal attention fusion module dynamically allocates weights according to the quasi-static loading stage and performs weighted fusion of the feature vectors of the two branches.

[0041] In some embodiments, the dual-branch feature extraction module includes two parallel and structurally independent sub-networks: an atomic structure branch and a damage feature branch.

[0042] The atomic structure branch receives normalized atomic structure data as input. First, a 1D convolutional layer operates on the input parameter sequence. This convolutional layer is configured with multiple kernels, which slide across the sequence to extract local correlations and combination patterns between different structural parameters. Subsequently, the feature sequence output from the convolutional layer is fed into a Long Short-Term Memory (LSTM) network layer. The LSTM layer, through its internal gating mechanism, learns and memorizes the evolution and long-term dependencies of atomic structure features over the entire loading time series. Finally, it outputs a high-dimensional feature vector to characterize the temporal evolution of the atomic structure.

[0043] In some embodiments, the 1D convolutional layer is configured with 32 convolutional kernels with a width of 3. The LSTM layer contains 64 hidden units and is equipped with a dropout mechanism with a dropout rate of 0.3 to prevent overfitting. The final output is a structural feature vector with dimensions of 128 (temporal length) × 64 (feature dimension), which serves as a deep representation of the evolution of atomic structure.

[0044] The damage feature branch receives a multi-channel damage data cube as input after enhancement and fusion processing. First, a 3D convolutional layer operates on this data cube. The 3D convolutional kernel performs sliding convolution simultaneously in the 2D image space (length and width) and the temporal dimension, thereby extracting joint features of the damage morphology and its acoustic emission representation in spatial distribution and temporal evolution. Then, a global average pooling layer aggregates the spatial dimensions of the feature map, compressing it into a series of feature values. Finally, a fully connected layer performs further nonlinear transformation and fusion on these features, deeply combining the abstract features of acoustic emission with the image visual features. The final output is a high-dimensional feature vector used to characterize the spatiotemporal evolution of the damage.

[0045] In some embodiments, the preprocessed multi-channel damage data cube (dimensions: temporal length × image height × image width × number of channels) is first fed into a 3D convolutional layer. This layer is configured with 64 convolutional kernels, each with a size of 3 in the spatial (height, width) and temporal dimensions. By performing a 3D convolution operation on the input data cube, the spatial distribution features of the damage morphology on the 2D image plane and its evolution features along loading time are extracted simultaneously. Subsequently, a global average pooling layer is applied to aggregate the feature maps output by the convolutional layer in the spatial dimensions (height and width), compressing them into a set of one-dimensional feature vectors arranged temporally, thereby achieving significant dimensionality reduction. This feature vector is then fed into a fully connected layer containing 128 neurons. The fully connected layer performs deep fusion and abstraction of the aforementioned extracted visual spatiotemporal features and features derived from acoustic emission signals and encoded in the multi-channel data through nonlinear transformations. After the above processing, a damage feature vector with dimensions of 128 (time series length) × 64 (feature dimension) is finally generated as a unified deep representation of the damage evolution process.

[0046] In some embodiments, the cross-modal attention fusion module receives feature vectors output from the atomic structure branch and the damage feature branch. Based on the quasi-static loading physical stage corresponding to the input data sample, it presets and applies different fusion weight coefficients to the two types of features. If the sample is in the elastic deformation stage, the atomic structure feature vector is assigned a weight of 0.7, and the damage feature vector is assigned a weight of 0.3; if the sample is in the damage initiation stage, both feature vectors are assigned a weight of 0.5; if the sample is in the damage stable expansion stage, the atomic structure feature vector is assigned a weight of 0.3, and the damage feature vector is assigned a weight of 0.7. Finally, the weighted atomic structure feature vector and the damage feature vector are summed to generate a unified fusion feature vector.

[0047] S5. Model Convergence and Result Output The improved hybrid model is trained in stages using a piecewise loss function until convergence, and finally outputs macroscopic damage assessment results and microstructure-damage correlation patterns; the staged training includes a pre-training stage and a fusion training stage.

[0048] In some embodiments, the atomic structure branch and the damage feature branch in the dual-branch feature extraction module are disconnected and trained independently. Training continues for a preset number of rounds (e.g., 50 rounds). The Adam optimizer is used, and a relatively high initial learning rate is set (e.g., 1×10⁻⁶). -4 This enables both branches to initially learn to extract effective basic features from the input data of their respective modalities.

[0049] After pre-training, the two branches are connected via a cross-modal attention fusion module to form a complete hybrid model for joint training. Training continues for more epochs (e.g., 200 epochs). During this stage, the learner's learning rate is lowered (e.g., reduced to 5×10⁻⁶). -5 The model is trained using a periodic decay strategy, such as a 10% decay of the learning rate every 20 training rounds. This allows the model to learn how to dynamically fuse multimodal features and accurately map the complex relationship between structure and damage.

[0050] In some embodiments, the piecewise loss function is a mean squared error loss function with weights adjusted according to the stage, and the total loss function is: structural loss weight × mean squared error of atomic structure prediction value + damage loss weight × mean squared error of damage prediction value. Here, "structural loss weight" and "damage loss weight" are not fixed values, but are dynamically set according to the quasi-static loading stage (elasticity, damage initiation, damage propagation) to which the current training sample data belongs.

[0051] In some embodiments, during training, model performance is monitored using a reserved validation set (e.g., 30% of the total samples). Model training terminates when any of the following preset quantization convergence conditions are met: The prediction error for damage area (or equivalent damage parameter) on the validation set remained below 4% for 10 consecutive training rounds. The correlation curves between the microstructural parameters (such as coordination number) and damage parameters revealed by the model have an average deviation of less than 3% from independent experimental observations or curves of known physical laws.

[0052] In some embodiments, the trained and converged model can be used to analyze new data, and its output includes macroscopic damage assessment results and microstructure-damage correlation patterns.

[0053] The macroscopic damage assessment results include predicted damage area and damage level. Specifically, the predicted damage area is the predicted damage area of ​​the specimen under a specific loading state (or when loaded to a specific strain), directly output by the model, and the unit is typically square millimeters (mm). 2 The damage level is determined based on the predicted damage area and is classified according to a preset threshold. For example, it is defined as: Level I damage (microcracks, area less than 0.1 mm). 2 Grade II damage (area between 0.1-0.5 mm) 2 Grade II (between) and Grade III (area greater than 0.5 mm) injuries. 2 ).

[0054] The microstructure-damage correlation includes the contribution of key structural parameters and the damage-sensitive structural threshold. The contribution of key structural parameters is the intensity of the influence of specific atomic structural parameters on damage evolution, which can be quantified through model analysis and expressed as a mathematical relationship. For example, a 1-unit decrease in coordination number results in an average increase of 0.2 mm in the predicted damage area. 2 .

[0055] The damage-sensitive structure threshold is a critical value of a key structural parameter that the model can identify that causes a significant change in damage behavior (such as a jump in damage level). It is expressed as a threshold condition; for example, when the interatomic spacing expands to more than 0.30 nm, the damage level easily increases from level I to level II.

[0056] Example (I) Sample preparation and loading implementation Smelting process: Zr 55 Cu 30 Al 10 Ni5 raw materials are mixed in a certain proportion and placed in an electric arc melting furnace, with the vacuum level reduced to 1×10⁻⁶. -4 After Pa, argon gas is introduced for protection, and the melting is repeated 3 times to ensure uniform composition. Spray casting: Pour the alloy melt into a copper mold preheated to 200℃, spray cast under a pressure of 0.5MPa, and remove it after natural cooling to room temperature. XRD scanning (20-80°, step size 0.02°) confirms the absence of crystalline peaks. Loading settings: Force sensor calibration (error ±0.5%) on the electronic universal testing machine (model: Instron 5969). During specimen installation, optical alignment is used to ensure the axis is consistent with the loading direction. Loading rate: 1×10⁻⁶. -3 s -1 .

[0057] (II) Data Collection Implementation Atomic structure data: XRD (model: Bruker D8 Advance) measured interplanar spacing (accuracy ±0.001 nm); HRTEM (model: Tecnai G2 F20) accelerated at 200 kV to capture atomic arrangement images (resolution 0.14 nm); EXAFS (Beijing Synchrotron Radiation Facility 1W1B beamline) measured the Zr / Ti atomic coordination environment; Damage data: An optical microscope (model: Olympus MX63) connected to a CCD camera automatically captures a 32×32 pixel image every 0.1% strain; an acoustic emission detector (model: PAC Micro-II) sensor is attached to both sides of the sample, with a threshold of 40dB and a sampling frequency of 1MHz.

[0058] (III) Data Preprocessing Implementation Structural parameter normalization: A normalization script was written in Python, with the interatomic spacing (0.27 nm) and coordination number (12) of the Zr-based amorphous fully amorphous state as the reference. The formula is: Normalized value = (measured value - reference value) / (maximum value - minimum value), and normalized to [0,1]. Damage image denoising: A multi-scale residual network was constructed based on TensorFlow. The original damage image was taken as input, and the denoised image was output. The microcrack edge retention rate was >95%. Data completion: Missing strain data for 1.3% in Ti-based samples, based on 1.2% (atomic spacing 0.29 nm, damage area 0.25 mm). 2 ) and 1.4% (interatomic spacing 0.31nm, damage area 0.32mm) 2 Linear interpolation of the data was used to complete the 1.3% strain data (atomic spacing 0.30 nm, damage area 0.28 mm). 2 ).

[0059] (iv) Model training and convergence implementation Training environment: GPU (NVIDIA RTX 3090), PyTorch 1.12 framework, Ubuntu 20.04 system; Training process: 70% of the samples are used as the training set and 30% as the validation set, with a batch size of 32. After 50 rounds of pre-training, the structural branch loss is reduced to 0.08 and the damage branch loss is reduced to 0.1. When the fusion training reaches 180 rounds, the damage area prediction error of the validation set is 3.5% for 10 consecutive rounds, which meets the convergence condition, and training is stopped.

[0060] (v) Implementation of Results Analysis results: At 1.0% strain, the model predicts a damage area of ​​0.32 mm for the Zr-based specimen. 2 (Actual measurement: 0.33mm) 2 (Error 2.9%), damage grade II; microscopic results show that for every 1 decrease in coordination number, the damage area increases by 0.21 mm. 2 When the interatomic spacing is >0.30nm, the damage level is easily escalated; Performance optimization application: Based on the microscopic results, the proportion of Zr-based amorphous Cu element was increased from 30% to 35%, and the coordination number was improved to 12.5.

[0061] Although the embodiments of this application have been described above in conjunction with the accompanying drawings, this application is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of this application, and all of these are within the scope of protection of this application.

Claims

1. A method for quasi-static damage and atomic structure correlation analysis of amorphous alloys based on an improved hybrid model, characterized in that, Includes the following steps: S1. Prepare standard amorphous alloy specimens and subject them to quasi-static tensile loading; S2. During the quasi-static loading process, atomic structure data and damage data of the sample are collected simultaneously. The atomic structure data includes structural parameters of multiple time nodes; the atomic structure data includes at least interplanar spacing, short-range ordered domain size, coordination number and atomic spacing; the damage data includes damage images and acoustic emission signals of the corresponding time nodes. S3. Normalize the atomic structure data and enhance and complete the damage data to obtain preprocessed multimodal time series data. S4. Construct and train an improved hybrid model, which includes a dual-branch feature extraction module and a cross-modal attention fusion module. The dual-branch feature extraction module is used to extract feature vectors from the preprocessed atomic structure data and damage data, respectively. The cross-modal attention fusion module dynamically allocates weights according to the quasi-static loading stage and performs weighted fusion of the feature vectors of the two branches. S5. The improved hybrid model is trained in stages using a piecewise loss function until convergence, and the macroscopic damage assessment results and microstructure-damage correlation patterns are finally output. The staged training includes a pre-training stage and a fusion training stage.

2. The method for correlation analysis of quasi-static damage and atomic structure of amorphous alloys based on an improved hybrid model according to claim 1, characterized in that, The standard amorphous alloy sample was Zr. 55 Cu 30 Al 10 Ni5 or Ti 60 Cu 20 Ni 15 Sn5 amorphous alloy was produced by vacuum melting and copper mold spraying, and XRD verification confirmed it to be amorphous; the quasi-static loading rate was 1×10⁻⁶. -3 s -1 .

3. The method for correlation analysis of quasi-static damage and atomic structure of amorphous alloys based on an improved hybrid model according to claim 1, characterized in that, In step S2, the strain interval for acquiring the atomic structure data is 0.1%.

4. The method for correlation analysis of quasi-static damage and atomic structure of amorphous alloys based on an improved hybrid model according to claim 1, characterized in that, In step S3, the normalization process is to standardize the atomic structure data to the [0,1] interval based on the parameters of the amorphous state; the enhancement process includes using a multi-scale residual denoising unit to denoise the damage image, and then converting the acoustic emission signal parameters into a thermogram and stitching it with the denoised damage image to form multi-channel damage data; the completion process is to perform linear interpolation based on the characteristic correlation of adjacent strain nodes for the missing time-series nodes in the high strain stage.

5. The method for correlation analysis of quasi-static damage and atomic structure of amorphous alloys based on an improved hybrid model according to claim 1, characterized in that, In step S4, the dual-branch feature extraction module includes an atomic structure branch and a damage feature branch; The atomic structure branch includes sequentially connected 1D convolutional layers and LSTM layers, used to extract local correlations and temporal evolution features of structural parameters; The damage feature branch includes a 3D convolutional layer, a global average pooling layer, and a fully connected layer, which are used to extract spatial-temporal features of the damage image and fuse acoustic emission features.

6. The method for correlation analysis of quasi-static damage and atomic structure of amorphous alloys based on an improved hybrid model according to claim 5, characterized in that, The cross-modal attention fusion module dynamically allocates fusion weights according to the three stages of quasi-static loading: in the elastic stage, the weight of atomic structure features is 0.7, and the weight of damage features is 0.

3. In the initial stage of damage, the weights of atomic structure features and damage features are each 0.5; During the damage propagation stage, the atomic structure feature weight is 0.

3. The damage feature weight is 0.

7.

7. The method for correlation analysis of quasi-static damage and atomic structure of amorphous alloys based on an improved hybrid model according to claim 1, characterized in that, In step S5, the segmented loss function is a mean squared error loss function with weights adjusted according to the stage, and the total loss function is: structural loss weight × mean squared error of atomic structure prediction value + damage loss weight × mean squared error of damage prediction value; the pre-training stage independently trains the dual-branch feature extraction module, and the fusion training stage connects the cross-modal attention fusion module for joint training.

8. The method for correlation analysis of quasi-static damage and atomic structure of amorphous alloys based on an improved hybrid model according to claim 7, characterized in that, The convergence criteria for the improved hybrid model are: the damage area prediction error of the validation set is less than 4% for 10 consecutive rounds, and / or the microstructure-damage correlation deviation is less than 3%.

9. The method for quasi-static damage and atomic structure correlation analysis of amorphous alloys based on an improved hybrid model according to claim 1, characterized in that, The macroscopic damage assessment results output in step S5 include the predicted damage area and damage level; the microstructure-damage correlation includes the contribution of coordination number and / or interatomic spacing to damage and the damage-sensitive structure threshold.

Images