A lunar soil brick virtual nanoindentation test method based on microscopic images and deep learning, a data set collection and production method, a model training method and application

Abstract

The application discloses a lunar soil brick virtual nanoindentation test method based on microscopic images and deep learning, a data set collection and production method, a model training method and application, and belongs to the technical field of lunar in-situ resource utilization and material performance prediction. The system obtains the microscopic structure images of the lunar soil brick before and after nanoindentation by a scanning electron microscope, combines the hardness and elastic modulus data measured by the nanoindentation experiment, generates a phase label by using a clustering algorithm, and thus constructs a multi-dimensional training data set. A multi-task deep learning model is constructed based on a Vision Transformer, and joint training is carried out in combination with a regression algorithm, so that phase prediction, virtual nanoindentation simulation and pixel-level mechanical performance prediction are realized. The method significantly improves the efficiency and accuracy of material performance evaluation, and can provide a scientific basis for lunar building material selection, performance monitoring and process optimization.

Description

Technical Field

[0001] This invention belongs to the field of in-situ utilization of lunar resources and prediction of advanced material properties. More specifically, it relates to a virtual nanoindentation testing method and system for lunar soil bricks based on microscopic images and deep learning, which enables rapid and non-destructive prediction of the phase distribution and mechanical properties of lunar soil bricks. Background Technology

[0002] As lunar exploration enters the phase of long-term stay and sustainable utilization, in-situ lunar resource utilization (ISRU) has become crucial for building research stations and achieving extraterrestrial survival. Lunar regolith is the most abundant raw material on the lunar surface. It can be sintered or bonded to form lunar regolith bricks for constructing habitation modules, protective walls, and other structures, significantly reducing the cost and dependence on transporting building materials from Earth and achieving self-sufficiency in lunar construction. The macroscopic mechanical properties of lunar regolith bricks (such as compressive strength and modulus of elasticity) directly affect building safety. However, traditional destructive testing methods in the lunar environment are not only material-intensive, time-consuming, and require bulky equipment, but also struggle to meet the needs for rapid assessment and feedback of material quality during lunar construction. Therefore, there is an urgent need to develop a rapid, non-destructive, and highly accurate method for predicting mechanical properties to provide real-time support for lunar regolith construction.

[0003] In recent years, deep learning has demonstrated great potential in the prediction of material structure-property relationships. Scanning electron microscopy (SEM) can reveal the microscopic features of lunar regolith bricks, such as particle arrangement, pore size, and cementation morphology, which determine their macroscopic mechanical behavior. Deep learning analysis based on SEM images can overcome the limitations of traditional manual feature extraction, automatically uncovering the intrinsic correlation between complex microstructures and mechanical properties, and providing a new approach for non-destructive prediction of lunar regolith brick properties.

[0004] Therefore, there is an urgent need for a method that can automatically identify the phase characteristics of microscopic images, output virtual nanoindentation images through intelligent algorithms, and predict the mechanical properties of materials, so as to achieve efficient and non-destructive evaluation of material properties. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a virtual nanoindentation testing system and method for lunar regolith bricks based on microscopic images and deep learning, as well as a dataset acquisition and fabrication method and a model construction and training method. The aim is to establish an intelligent mapping relationship between the microstructure and mechanical properties of lunar regolith bricks, enabling rapid and non-destructive prediction of the material's mechanical properties. This solves the technical problems of traditional mechanical testing relying on destructive experiments, having low testing efficiency, and being unable to meet the real-time evaluation requirements of in-situ lunar construction.

[0006] According to one aspect of the present invention, a method for acquiring and creating a virtual nanoindentation test dataset for lunar soil bricks based on microscopic images and deep learning is provided, comprising the following steps: Obtain SEM images of lunar soil bricks before nanoindentation experiments; extract pre-indentation SEM images of indentation points from the pre-indentation SEM images of lunar soil bricks before nanoindentation experiments, perform pixel-level clustering on the pre-indentation SEM images, obtain semantic segmentation results of different phases, and generate phase labels for the pre-indentation SEM images. Obtain SEM images of lunar soil bricks after nanoindentation experiments and mechanical property data of lunar soil bricks in the corresponding nanoindentation areas; extract SEM images of nanoindentation points from the SEM images of lunar soil bricks after nanoindentation experiments. The datasets of the pre-indentation SEM images and the SEM images of the nano-indentation points were augmented to generate the following dataset: Dataset 1: SEM images before indentation and phase labels; Dataset 2: SEM images before and after indentation; Dataset 3: SEM images and mechanical property data before indentation.

[0007] Furthermore, the mechanical property data of the lunar soil bricks include the hardness and elastic modulus of the lunar soil brick material, which are obtained directly through nanoindentation experiments.

[0008] Furthermore, the clustering algorithm for pixel-level clustering of the SEM image before indentation is at least one of Gaussian mixture model, K-means clustering, or spectral clustering.

[0009] According to another aspect of the present invention, a method for training a virtual nanoindentation test model for lunar soil bricks based on microscopic images and deep learning is provided, comprising the following steps: Construct a multi-task deep learning model based on Vision Transformer, input the dataset obtained according to any one of claims 1 to 3 into the multi-task deep learning model, extract global structural features from the image through an attention mechanism, and jointly train the following three prediction tasks through a regression head: ① Phase prediction task: Output phase distribution diagram; ② Virtual nanoindentation task: Output simulated nanoindentation location and morphology prediction maps; ③ Mechanical property prediction task: Output the mechanical property data of lunar soil bricks in the corresponding region; The regression losses from the three tasks are aggregated, and the network parameters of the shared ViT encoder and each regression head are updated synchronously through backpropagation to complete the end-to-end joint training of the multi-task model.

[0010] Furthermore, the training process of the Vision Transformer-based multi-task deep learning model includes: The SEM image is segmented into several image blocks and flattened into a sequence; The sequence is embedded into a high-dimensional feature space via linear projection. Global dependencies between image patches in a high-dimensional feature space are captured using a multi-head self-attention mechanism. Multiple prediction tasks are accomplished through joint training of regression models.

[0011] Furthermore, the Vision Transformer-based multi-task deep learning model is a multi-task learning framework that shares the Vision Transformer encoder and outputs phase prediction maps, virtual nanoindentation maps, and lunar soil brick mechanical property data through independent task heads. The training inputs for the virtual nanoindentation map prediction task and the mechanical property prediction task are derived from the phase features obtained from the phase prediction task, or from the global features extracted by the Vision Transformer-based multi-task deep learning model.

[0012] According to another aspect of the present invention, a virtual nanoindentation test model for lunar soil bricks based on microscopic images and deep learning is provided, which is trained according to the virtual nanoindentation test model training method for lunar soil bricks as described in any of the preceding claims.

[0013] According to another aspect of the present invention, a virtual nanoindentation test method for lunar soil bricks based on microscopic images and deep learning is provided. The method inputs SEM images that have not undergone nanoindentation experiments into the virtual nanoindentation test model for lunar soil bricks as described above, and automatically outputs the corresponding phase prediction map, virtual nanoindentation prediction map, and lunar soil brick mechanical property data prediction cloud map, thereby realizing non-destructive prediction of the microscopic mechanical properties of the material.

[0014] According to another aspect of the present invention, the application of the virtual nanoindentation test method for lunar soil bricks as described above in material selection, performance evaluation or process optimization is provided.

[0015] In summary, compared with existing technologies, the technical solutions conceived in this invention, by employing a multi-task joint learning architecture that integrates phase distribution prediction, virtual nanoindentation simulation, and full-domain characterization of mechanical properties, establish a complete intelligent mapping link between the microstructure, phase characteristics, indentation response, and mechanical properties of lunar regolith bricks. This overcomes the inherent limitations of traditional testing methods and single-task prediction models, and achieves the following outstanding beneficial effects suitable for in-situ lunar resource utilization scenarios: ① This invention can achieve non-destructive prediction of the microstructure and mechanical properties of lunar soil bricks, obtain virtual nanoindentation response and local mechanical properties without damaging the material, and can jointly predict phase distribution, virtual nanoindentation morphology, elastic modulus and hardness, thus realizing coupled analysis of microstructure and mechanical properties. ② This invention employs a ViT-based attention mechanism to effectively capture global image structure information and improve prediction accuracy; through the combination of data augmentation and multi-task learning, it significantly enhances the robustness and generalization ability of the model under different SEM image conditions. ③ The prediction results of this invention can be used for material selection, performance evaluation and process optimization, providing reliable data support for the in-situ utilization of lunar resources.

[0016] ④ This invention introduces clustering algorithms such as Gaussian mixture models to automatically generate phase labels, achieving efficient labeling of large-scale data and significantly reducing the cost of manual labeling.

[0017] ⑤ The method of the present invention is not only applicable to lunar soil brick materials, but can also be extended to various material systems such as ceramics, bricks, geopolymers and metal composites, and has broad application prospects. Attached Figure Description

[0018] Figure 1 This describes the process of data collection and creation.

[0019] Figure 2 This is the multi-level prediction architecture established for this invention.

[0020] Figure 3 Detailed diagram of Vision Transformer. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] This invention addresses the technical problem of predicting the macroscopic mechanical properties of lunar regolith bricks through microstructure. It proposes a virtual nanoindentation testing system for lunar regolith bricks based on microscopic images and deep learning. This method uses deep learning to identify phase characteristics in microscopic images and establishes a mapping relationship between phase characteristics and mechanical properties, thereby accurately predicting the mechanical properties of lunar regolith bricks.

[0023] Preferably, the present invention provides a virtual nanoindentation testing system and method for lunar soil bricks based on microscopic images and deep learning, comprising the following steps: S1. SEM images of lunar soil bricks before and after nanoindentation experiments were obtained. Images with nanoindentation points no smaller than 128 pixels × 128 pixels were extracted, and the corresponding hardness and elastic modulus data were acquired. Pixel-level clustering was performed on the obtained SEM images to obtain semantic segmentation results for different phases, generating phase labels. Subsequently, the SEM images before and after nanoindentation, phase labels, and corresponding mechanical property data were integrated to construct a sample dataset for multi-task training.

[0024] S2. Construct a multi-task deep learning model based on Vision Transformer (ViT). Input the training data into the model and extract global structural features of the image through an attention mechanism. The model is jointly trained for three tasks using end-to-end regression or independent regression heads: a phase prediction task to output phase distribution maps of lunar regolith bricks; a virtual nanoindentation task to output simulated nanoindentation location and morphology prediction maps; and a mechanical property prediction task to output elastic modulus and hardness contour maps of the corresponding regions. During training, the inputs for the virtual nanoindentation and mechanical property prediction tasks can come from the phase features obtained from the phase prediction task or from the global structural features extracted by the ViT model. Simultaneously, data augmentation operations such as rotation, scaling, flipping, and noise perturbation are performed on the microscopic images to improve the model's robustness and generalization ability.

[0025] S3. Inputting SEM images without nanoindentation experiments into the trained model automatically generates phase prediction maps, virtual nanoindentation prediction maps, and elastic modulus and hardness prediction contour maps. Mechanical property prediction can be achieved through an end-to-end VisionTransformer multi-task regression head, or the phase and structural features extracted from the model can be input into independent regression algorithms (such as linear regression, random forest, gradient boosting, MLP regression, or deep regression networks) for training and prediction. The prediction results can be used for material selection, performance evaluation, and process optimization, enabling the reconstruction of the local mechanical distribution of materials under non-destructive conditions.

[0026] Preferably, in step S1, the hardness and elastic modulus of the lunar soil brick material are obtained directly by nanoindentation experiment.

[0027] Preferably, in step S1, the phase segmentation is achieved by a clustering algorithm, wherein the clustering algorithm is at least one of Gaussian mixture model, K-means clustering or spectral clustering.

[0028] Preferably, in step S2, the training process of the Vision Transformer model includes: ① Divide the SEM image into several image patches and flatten the image patches into a sequence; ② Embed the sequence into a high-dimensional feature space through linear projection; ③ Model the global dependencies between image patches using a multi-head self-attention mechanism; ④ Jointly train the multi-task prediction results using a regression model.

[0029] Preferably, the model adopts a multi-task learning framework, shares the Vision Transformer encoder structure, and outputs phase prediction maps, virtual nanoindentation prediction maps, and elastic modulus and hardness prediction cloud maps through independent task heads.

[0030] Preferably, in step S2, the training inputs for the virtual nanoindentation task and the mechanical property prediction task can be derived from the phase characteristics output by the phase prediction task or from the global structural features extracted by the Vision Transformer model.

[0031] Preferably, the mechanical property prediction task is a pixel-level regression task, and the predicted elastic modulus and hardness results are visualized in the form of cloud maps.

[0032] Preferably, during model training, SEM microscopic images are subjected to data augmentation processing, including rotation, scaling, flipping, and noise perturbation, to improve the robustness and generalization ability of the model under different image conditions.

[0033] Preferably, the SEM image is a grayscale image, and the image resolution is not less than 128 pixels × 128 pixels.

[0034] Preferably, in the prediction output stage, the model can generate a virtual nanoindentation response based on the unindented SEM image, thereby reconstructing the local mechanical property distribution of the material under non-destructive conditions.

[0035] Preferably, the predicted material elastic modulus, hardness, and strength can be used for material selection, performance evaluation, and process optimization.

[0036] According to another aspect of the present invention, a method for acquiring and creating a dataset for implementing the above-described virtual nanoindentation testing system is provided, the method comprising the following steps: S11. Lunar soil brick samples were prepared using different lunar soil brick preparation processes; S12. Perform nanoindentation experiments on the sample and record the indentation location, indentation depth, hardness and elastic modulus data. S13. Before and after the nanoindentation experiment, the sample surface was imaged by scanning electron microscopy (SEM) to obtain microscopic images of the corresponding areas. S14. Crop the SEM images before and after indentation into image samples at a fixed size, and establish a mapping relationship with the corresponding nanoindentation experimental data; S15. Perform unsupervised clustering or image segmentation on the SEM images to generate phase labels; S16. Perform unified formatting processing on SEM images, phase labels, and corresponding mechanical property data to construct a dataset for deep learning training.

[0037] Preferably, the lunar soil brick samples are prepared by changing the molding process route and its process parameters. The process includes, but is not limited to: sintering molding, melt solidification molding, fused deposition modeling, pressing and sintering molding, and binder-assisted molding, in order to obtain sample materials with different microstructural characteristics.

[0038] Preferably, the SEM images are preprocessed before the dataset is created, including grayscale normalization, image denoising, and contrast enhancement, to improve the recognizability of microscopic structural features.

[0039] Preferably, the nanoindentation locations are marked in the SEM image to generate an indentation area annotation map, which is used for training the virtual nanoindentation prediction task.

[0040] Preferably, the dataset samples simultaneously include: SEM images before nanoindentation, SEM images after nanoindentation, phase segmentation label images, nanoindentation location annotation images, and corresponding hardness and elastic modulus data, thereby forming a unified training sample for multi-task learning.

[0041] Preferably, the resolution of the SEM image is not less than 128×128 pixels.

[0042] Preferably, a method for building and training a model based on Vision Transformer is also provided, including the following steps: S21. Construct a deep learning network structure based on Vision Transformer, and set up a phase segmentation prediction head, a virtual nanoindentation prediction head, and a mechanical property regression prediction head. S22. Input the constructed dataset into the model for training, and optimize the model parameters through the multi-task joint loss function; S23. Use cross-validation of training and validation sets to evaluate the model's performance. S24. Once the model converges, the final model for virtual nanoindentation prediction and mechanical property prediction is obtained.

[0043] Preferably, the Vision Transformer model includes an image patch embedding module, a Transformer encoder module, and a multi-task prediction head module.

[0044] Preferably, during model training, the input SEM image is divided into multiple image patches, and the image patches are embedded into a high-dimensional feature space through linear mapping.

[0045] Preferably, the Transformer encoder models the global dependencies between different image patches through a multi-head self-attention mechanism to extract global feature information of the microstructure.

[0046] Preferably, the multi-task joint loss function includes: a phase segmentation loss function, a virtual nanoindentation prediction loss function, and a mechanical property regression loss function, which are jointly optimized through a weighted approach.

[0047] Preferably, during model training, the input SEM image is subjected to data augmentation processing, including rotation, scaling, flipping, and noise perturbation, to improve the model's generalization ability.

[0048] Preferably, the mechanical property prediction is a pixel-level regression task, and the output elastic modulus and hardness results are visualized in the form of cloud maps.

[0049] The present invention will now be described with reference to a more specific embodiment.

[0050] Using lunar soil-sintered bricks in situ on the lunar surface requires real-time strength testing due to the uncertainty of material compatibility and process parameters, in order to ensure that the research base or habitat built with lunar soil bricks meets basic safety requirements. This embodiment follows the method flow of steps (1) to (3) to complete the construction and verification of the virtual nanoindentation testing system for lunar soil bricks.

[0051] In this embodiment, 200 sets of SEM images of lunar soil bricks and corresponding mechanical property data are selected as the original samples. The dataset construction and labeling are completed according to step (1). After data augmentation, the dataset is expanded to 20,000 sets of samples. The training set, validation set and test set are divided proportionally. The construction and joint training of the multi-task deep learning model based on Vision Transformer are completed according to step (2). The prediction verification of the model is completed according to step (3).

[0052] Specifically, this embodiment provides the design and training of a virtual nanoindentation testing system for lunar soil bricks based on microscopic images and deep learning, which includes the following steps: (1) Dataset construction and labeling: Scanning electron microscope was used to obtain SEM images of lunar soil bricks before and after the nanoindentation experiment and the mechanical property data (hardness and elastic modulus) of the lunar soil bricks in the corresponding nanoindentation area. The specific execution process is as follows: ① After completing the preparation of lunar soil brick samples and the pre-polishing of the test surface, SEM imaging of the target test area before nanoindentation was performed to obtain the initial SEM image, and the pre-indentation SEM image with a resolution of not less than 128 pixels × 128 pixels corresponding to the preset indentation points was accurately extracted; ② A dot matrix nanoindentation experiment was carried out on the same target test area, and the measured data of hardness and elastic modulus corresponding to each indentation point were collected simultaneously. After the experiment was completed, SEM imaging was performed again on the same field of view to obtain the SEM image after nanoindentation, and the SEM image of the indentation point corresponding to the indentation point was extracted one by one; The SEM images are clustered at the pixel level using a clustering algorithm to obtain semantic segmentation results for different phases, generating phase labels for the SEM images. Specifically, for the extracted pre-indentation SEM images, an unsupervised clustering algorithm is used to perform pixel-level feature clustering, automatically generating pixel-level phase labels for the corresponding images, replacing costly manual semantic annotation. The SEM images before and after nanoindentation, phase labels, and corresponding hardness and elastic modulus together constitute the training dataset. After completing the basic data pairing, the nanoindentation point images are augmented using rotation, mirroring, cropping, random noise addition, and brightness transformation to expand the dataset. Subsequently, the expanded dataset is split into three corresponding sub-datasets according to the multi-task training requirements: Dataset 1 consists of the pre-indentation SEM image and corresponding phase label; Dataset 2 consists of the paired pre-indentation SEM image and post-indentation SEM image; and Dataset 3 consists of the pre-indentation SEM image and the mechanical property data of the corresponding point. Finally, all sub-datasets are uniformly divided according to the ratio of 70%-80% training set, 10%-15% validation set, and 10%-15% test set to complete the construction of a complete dataset adapted for multi-task joint training.

[0053] To further standardize the dataset construction process, the following standardized procedure can be followed: Complete SEM imaging before and after the nanoindentation experiment on lunar soil bricks, extract corresponding images of nanoindentation points with a resolution of at least 128 pixels × 128 pixels before and after the experiment, and simultaneously acquire measured data of hardness and elastic modulus at the corresponding points; perform pixel-level clustering processing on the SEM images before indentation to obtain semantic segmentation results of different phases and generate corresponding phase labels; finally, pair and integrate the SEM images before and after nanoindentation, phase labels, and corresponding mechanical property data to construct a complete sample dataset suitable for multi-task joint training.

[0054] (2) Model Construction and Training: A multi-task deep learning model based on Vision Transformer is constructed. It is an end-to-end deep learning architecture with feature sharing and multi-task joint optimization. The core innovation lies in the design of an architecture with encoder sharing, multi-regression head collaboration, and dual-path feature input. It realizes the coupled training and inference of three tasks: phase prediction, virtual indentation simulation, and mechanical performance prediction. The specific architecture and its relationship with each module are as follows: The model is divided into two core parts: a shared backbone feature extraction module and a multi-task parallel regression head. The shared backbone adopts the Vision Transformer encoder to uniformly extract the global microstructure features and local texture features of the input image. The multi-task parallel regression head sets up three independent and feature-interconnected task branches, which correspond to three types of prediction tasks. During the training phase, an auxiliary feature extraction branch is set up only for supervision signal extraction. During the inference phase, this branch is completely cropped to achieve lossless prediction of all tasks by only inputting the SEM image before indentation. Input the data obtained in step (1) into the model, extract global structural features from the image through the attention mechanism, and jointly train the following three prediction tasks through the regression head: ① Phase prediction task: corresponding to Figure 2 In the regression head 1 branch, the supervision signal is the phase feature label generated in step (1), the input is the global structural features extracted by the shared backbone, and the output is the phase distribution map of lunar soil bricks with the same size as the input image. The phase features output by this branch can be synchronously input to the other two task branches as auxiliary feature inputs for prediction, and output phase distribution map. ② Virtual nanoindentation task: corresponding to Figure 2 The regression head has two branches. The supervision signal is the indentation morphology features extracted by the auxiliary feature extraction branch from the SEM image after nanoindentation. The input supports dual path selection: you can choose the phase features output by the phase prediction branch, you can choose the global structural features extracted by the shared backbone, or you can choose the fusion result of the two types of features. Finally, the output is the simulated nanoindentation location and morphology prediction map. ③ Mechanical property prediction task: corresponding to Figure 2 The regression head 3 branches, the supervision signal is the hardness and elastic modulus data measured in step (1), the input is consistent with the virtual nanoindentation task, supports dual path feature input, and finally outputs the pixel-level elastic modulus and hardness cloud map of the corresponding region. During model training, data augmentation operations such as rotation, scaling, flipping, and noise perturbation can be performed on the input microscopic images simultaneously to further improve the robustness and generalization ability of the model. The prediction output of each task can be achieved by the end-to-end Vision Transformer multi-task regression head, or the phase and structural features extracted from the shared backbone can be input into independent regression algorithms (such as linear regression, random forest, gradient boosting tree, MLP regression, or deep regression network) to complete the training and prediction of a single task respectively.

[0055] (3) Predictive output: Input the SEM image without nanoindentation experiment into the trained model, and automatically output the corresponding phase prediction map, virtual nanoindentation prediction map, and elastic modulus and hardness prediction cloud map to realize non-destructive prediction of the micromechanical properties of the material.

[0056] The phase prediction map, virtual nanoindentation prediction map, and elastic modulus and hardness prediction cloud map output by the model can be directly used for material selection, performance evaluation and preparation process optimization of lunar soil bricks, realizing accurate reconstruction and rapid evaluation of the local mechanical distribution of materials under non-destructive conditions.

[0057] Preferably, in step (1), the hardness and elastic modulus data of the lunar soil brick material are obtained directly by nanoindentation experiment.

[0058] Preferably, in step (1), the phase segmentation is achieved by a clustering algorithm, which is at least one of Gaussian mixture model, K-means clustering or spectral clustering.

[0059] Preferably, in step (2), the Vision Transformer model training process is similar to... Figure 2 , Figure 3 The complete data flow and task-based training process follows a one-to-one correspondence: ① Image embedding and feature encoding: For input SEM images with a resolution of at least 128 pixels × 128 pixels, first follow the... Figure 3 The Embedded Patches process completes feature embedding by dividing the image into N non-overlapping image patches. Each patch is flattened into a one-dimensional vector, linearly mapped to a D-dimensional vector, a D-dimensional classification token is inserted, and then summed with the D-dimensional position vector to complete the embedding process of the image sequence. ② Global feature extraction: Input the embedded image sequence Figure 3The L-layer TransformerEncoder shown first undergoes layer normalization and then inputs into a multi-head attention mechanism. It calculates the global dependencies between image patches through Q, K, and V vectors, then inputs them into a layer normalization and a two-layer multilayer perceptron after residual connection. Finally, it outputs the global features of the image after residual connection, thus completing the feature extraction of the shared backbone. ③ Branch path data flow and supervised matching in multi-task joint training: Based on global features extracted from the shared backbone, according to... Figure 2 The multi-level prediction architecture, corresponding to three subsets of datasets, completes the collaborative training of three tasks respectively: Phase prediction task: Corresponding to the regression head 1 branch, the SEM image before indentation of dataset 1 is used as input. After global features are extracted through the shared backbone, they are input into regression head 1. The phase labels paired with dataset 1 are used as supervision signals. The training output is a phase distribution map of lunar soil bricks with the same size as the input image. At the same time, the phase features output by this branch are synchronously transferred to the other two task branches as auxiliary feature inputs. Virtual nanoindentation task: corresponds to the first two branches of regression, enabled during the training phase. Figure 2 The auxiliary feature extraction branch within the dashed box: Taking the indented SEM image of dataset 2 as input, the global features of the indentation morphology are extracted by the ViT model consistent with the shared backbone structure, which serves as the supervision signal for this branch; the input of this branch supports dual path selection, and can select the phase features output by the phase prediction branch, the global features of the pre-indentation image extracted by the shared backbone, or the fusion result of the two types of features, and train to output the simulated nano-indentation location and morphology prediction map; Mechanical property prediction task: corresponding to the first 3 branches of regression, taking the pre-indentation SEM image of dataset 3 as input, the input source is the same as the virtual nanoindentation task, supports dual-path feature input, and uses the measured hardness and elastic modulus data paired in dataset 3 as supervision signals, and trains to output the pixel-level elastic modulus and hardness cloud map of the corresponding region. ④ Joint Optimization of Multi-Task Losses: The regression losses from the three tasks are aggregated, and the network parameters of the shared ViT encoder and each regression head are synchronously updated through backpropagation to complete the end-to-end joint training of the multi-task model. Preferably, the model is a multi-task learning framework that shares the Vision Transformer encoder and outputs phase prediction maps, virtual nanoindentation maps, and elastic modulus and hardness prediction contour maps through independent task heads.

[0060] Preferably, in step (2), the training inputs for the virtual nanoindentation task and the mechanical property prediction task can be derived from the phase characteristics obtained by the phase prediction task or from the global features extracted by the VIT model.

[0061] Preferably, the mechanical property prediction task is a pixel-level regression task, and the output elastic modulus and hardness prediction results are visualized in the form of a cloud map.

[0062] Preferably, the method can perform multi-task prediction, including phase prediction, virtual nanoindentation prediction, and prediction of elastic modulus and hardness, wherein: ① The prediction can be achieved through an end-to-end Vision Transformer model, utilizing its multi-task regression head to simultaneously output phase characteristics, virtual nanoindentation information, and elastic modulus and hardness; or ② Input the phase and structural features extracted from the model into an independent regression algorithm for training and prediction.

[0063] The regression algorithm can be any one of linear regression, random forest, gradient boosting, MLP regression or deep regression network, to predict phase, virtual nanoindentation and mechanical properties individually or in combination.

[0064] Preferably, the method further includes performing data augmentation operations on the microscopic images during model training, including rotation, scaling, flipping, and noise perturbation, to improve the robustness and generalization ability of the model.

[0065] Preferably, the SEM image is a grayscale image with a resolution of not less than 128 pixels × 128 pixels.

[0066] Preferably, the model can generate a virtual indentation response based on an unindented SEM image during the prediction output stage, thereby reconstructing the local mechanical distribution of the material under non-destructive conditions.

[0067] Preferably, the predicted material elastic modulus, hardness, and strength are used for material selection, performance evaluation, and process optimization.

[0068] Verification has shown that the system constructed in this embodiment only requires inputting SEM images of lunar soil bricks that have not undergone nanoindentation experiments, and can automatically output corresponding phase prediction maps, virtual nanoindentation prediction maps, and mechanical property cloud maps. The relative error between the prediction results and the actual experimental data is less than 5%. It can accurately reconstruct the local mechanical distribution of lunar soil bricks under non-destructive conditions, providing reliable technical support for material selection, performance evaluation, and preparation process optimization of lunar soil bricks.

[0069] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for acquiring and creating a virtual nanoindentation test dataset for lunar soil bricks based on microscopic images and deep learning, characterized in that, Includes the following steps: Obtain SEM images of lunar soil bricks before nanoindentation experiments; extract pre-indentation SEM images of indentation points from the pre-indentation SEM images of lunar soil bricks before nanoindentation experiments, perform pixel-level clustering on the pre-indentation SEM images, obtain semantic segmentation results of different phases, and generate phase labels for the pre-indentation SEM images. Obtain SEM images of lunar soil bricks after nanoindentation experiments and mechanical property data of lunar soil bricks in the corresponding nanoindentation areas; extract SEM images of nanoindentation points from the SEM images of lunar soil bricks after nanoindentation experiments. The datasets of the pre-indentation SEM images and the SEM images of the nano-indentation points were augmented to generate the following dataset: Dataset 1: SEM images before indentation and phase labels; Dataset 2: SEM images before and after indentation; Dataset 3: SEM images and mechanical property data before indentation.

2. The method for collecting and creating a virtual nanoindentation test dataset for lunar soil bricks according to claim 1, characterized in that, The mechanical property data of the lunar soil bricks include the hardness and elastic modulus of the lunar soil brick material, which are obtained directly through nanoindentation experiments.

3. The method for collecting and creating a virtual nanoindentation test dataset for lunar soil bricks according to claim 1, characterized in that, The clustering algorithm for pixel-level clustering of the SEM image before indentation is at least one of Gaussian mixture model, K-means clustering, or spectral clustering.

4. A training method for a virtual nanoindentation test model of lunar soil bricks based on microscopic images and deep learning, characterized in that, Includes the following steps: Construct a multi-task deep learning model based on Vision Transformer, input the dataset obtained according to any one of claims 1 to 3 into the multi-task deep learning model, extract global structural features from the image through an attention mechanism, and jointly train the following three prediction tasks through a regression head: ① Phase prediction task: Output phase distribution diagram; ② Virtual nanoindentation task: Output simulated nanoindentation location and morphology prediction maps; ③ Mechanical property prediction task: Output the mechanical property data of lunar soil bricks in the corresponding region; The regression losses from the three tasks are aggregated, and the network parameters of the shared ViT encoder and each regression head are updated synchronously through backpropagation to complete the end-to-end joint training of the multi-task model.

5. A training method for a virtual nanoindentation test model of lunar soil bricks based on microscopic images and deep learning, as described in claim 4, is characterized in that... The training process of the multi-task deep learning model based on Vision Transformer includes: The SEM image is segmented into several image blocks and flattened into a sequence; The sequence is embedded into a high-dimensional feature space via linear projection. Global dependencies between image patches in a high-dimensional feature space are captured using a multi-head self-attention mechanism. Multiple prediction tasks are accomplished through joint training of regression models.

6. A training method for a virtual nanoindentation test model of lunar soil bricks based on microscopic images and deep learning, as described in claim 5, is characterized in that... The Vision Transformer-based multi-task deep learning model is a multi-task learning framework that shares the Vision Transformer encoder and outputs phase prediction maps, virtual nanoindentation maps, and lunar soil brick mechanical property data through independent task heads. The training inputs for the virtual nanoindentation map prediction task and the mechanical property prediction task are derived from the phase features obtained from the phase prediction task, or from the global features extracted by the Vision Transformer-based multi-task deep learning model.

7. A virtual nanoindentation testing model for lunar soil bricks based on microscopic images and deep learning, characterized in that, The virtual nanoindentation test model for lunar soil bricks was trained according to any one of claims 4 to 6.

8. A virtual nanoindentation testing method for lunar soil bricks based on microscopic images and deep learning, characterized in that, By inputting the SEM image of the lunar soil brick without nanoindentation experiment into the virtual nanoindentation test model of claim 7, the corresponding phase prediction map, virtual nanoindentation prediction map and lunar soil brick mechanical property data prediction cloud map are automatically output, realizing non-destructive prediction of the micromechanical properties of the material.

9. The application of the virtual nanoindentation test method for lunar soil bricks as described in claim 8 in material selection, performance evaluation, or process optimization.

Images