A method for predicting macroscopic mechanical properties of soil based on SEM images

Abstract

The present application relates to the field of artificial intelligence, in particular to a soil macroscopic mechanical property prediction method based on SEM images, comprising the following steps: acquiring a low-magnification field sequence and correcting the objective table coordinates to generate a sample surface coordinate map, identifying particle contours and pore contours, determining high-magnification re-scanning coordinates according to particle contact arcs, pore neck widths and deflection degrees, generating high-magnification re-scanning images, extracting bearing nodes, constructing a transfer path and a bearing chain representation sequence, training a prediction network combined with mechanical target data and outputting results. In the present application, low-magnification positioning, high-magnification re-scanning and bearing chain representation are included in the same processing chain, so that the SEM image is converted from appearance texture input into structured information reflecting particle contact, pore shrinkage and loading direction transmission relationship, the mapping reliability between microstructure and macroscopic mechanical properties such as strength, deformation and modulus is enhanced, and the interference of field selection and texture noise on the prediction results is reduced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for predicting the macroscopic mechanical properties of soil based on SEM images. Background Technology

[0002] The field of artificial intelligence technology involves computer processing methods that utilize training samples, neural network parameters, and inference programs to establish a mapping relationship between input data and target parameters. Among these, traditional methods for predicting the macroscopic mechanical properties of soil involve a computer device reading SEM images of soil samples and corresponding strength, deformation, or modulus test data. The image pixels are then subjected to pixel size standardization, grayscale normalization, and sample labeling. The processed image data is input into a neural network, and the network weights are updated through forward and backward propagation. Finally, the trained network outputs the macroscopic mechanical property parameters of the soil under test.

[0003] Traditional processing workflows take the entire SEM image and mechanical test labels as input, focusing on size uniformity, grayscale normalization, and pixel-level sample annotation. In actual operation, the correlation between the position of the microscopic field of view and the spatial coordinates of the sample surface is weak, and it is difficult to continuously correspond between the low-magnification overall view and the high-magnification details. Bearing factors such as particle contact, pore necks, and particle orientation are easily averaged into texture differences. Neural networks mainly learn the statistical mapping between the image appearance and the test results. The prediction results are sensitive to the shooting area, magnification, and sample differences, and it is difficult to stably reflect the macroscopic strength, deformation, and modulus formation mechanism of the soil. Summary of the Invention

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a method for predicting the macroscopic mechanical properties of soil based on SEM images, comprising the following steps: The low-magnification field sequence of the soil sample formed by the scanning electron microscope is obtained, and the stage coordinates are corrected based on the overlapping edge position of adjacent low-magnification fields to generate a coordinate map of the sample surface. Based on the coordinate diagram of the sample surface, the particle outline and pore outline are identified, and the particle contact arc, pore neck width and particle long axis deflection relative to the loading direction are verified. Based on the sorting positions of the particle contact arc length, the pore neck width, and the deflection degree, the high-magnification rescan coordinate set is verified, and a high-magnification rescan image is generated based on the high-magnification rescan coordinate set. For the particle contact arc in the high-magnification rescan image, the bearing nodes are determined, and based on the bearing nodes located on the same particle contour and adjacent along the loading direction, the transfer path and bearing chain characterization sequence are generated. Obtain the mechanical target data corresponding to the soil sample, train a prediction network based on the bearing chain characterization sequence and the mechanical target data, input the bearing chain characterization sequence of the soil sample to be tested into the prediction network, and output the prediction results of the macroscopic mechanical properties of the soil.

[0005] As a further aspect of the present invention, the process of determining the offset of the edge texture feature points includes: extracting the first edge patch and the second edge patch from the overlapping edge positions of adjacent low-magnification fields of view, determining the particle edge intersection point and the pore boundary turning point based on the gray-scale gradient change, and determining the particle edge intersection point and the pore boundary turning point as candidate texture feature points; Compare the local grayscale distribution, contour direction and spacing of candidate texture feature points in the first edge patch and the second edge patch, and determine the candidate texture feature points that meet the fitting degree threshold and whose spacing deviation is less than the spacing deviation threshold as the matching texture feature points; The edge displacement is determined based on the pixel position difference of the matching texture feature points in adjacent low-magnification fields of view, and the edge displacement is converted into a stage coordinate correction amount to obtain the edge texture feature point offset.

[0006] As a further aspect of the present invention, the process of determining the particle contact arc, the pore neck width, and the degree of deflection includes: obtaining the gray-scale continuous region and gray-scale abrupt boundary in the coordinate diagram of the sample surface; dividing the soil sample surface region into particle candidate region and pore candidate region according to the degree of closure of the gray-scale abrupt boundary; and performing neighborhood connectivity correction on the boundary fracture position to generate particle profile and pore profile. The particle contact area is determined based on the common edge length and boundary spacing between adjacent particle contours, and the particle contact area is mapped to the boundary arc segment of the corresponding particle contour to obtain the particle contact arc. The pore neck position is determined based on the narrowest connection position between adjacent pore profiles. The shortest distance between the particle profiles on both sides at the pore neck position is calculated to obtain the pore neck width. The degree of deflection is determined based on the angle between the major axis of the particle profile and the loading direction.

[0007] As a further aspect of the present invention, the generation process of the particle contour and the pore contour includes: obtaining the pixel grayscale value, local grayscale gradient and neighborhood connectivity in the coordinate map of the sample surface; determining the first segmentation boundary based on the grayscale difference threshold between the particle region and the pore region; and performing edge adsorption adjustment on the first segmentation boundary based on the local grayscale gradient to generate the second segmentation boundary. When the second segmentation boundary has a fracture endpoint, the boundary segment to be connected is determined based on the directional consistency and distance threshold between the fracture endpoints. The boundary segment to be connected is incorporated into the second segmentation boundary, and the area with higher gray level after closure is determined as the particle profile, and the area with lower gray level after closure is determined as the pore profile.

[0008] As a further aspect of the present invention, the process of determining the rescan priority includes: generating particle contact arc length sorting positions according to the arc length from large to small, generating pore neck width sorting positions according to the width from small to large, and generating deflection degree sorting positions according to the degree of deviation from the loading direction from large to small. The sorting weights corresponding to the sorting positions of the particle contact arc length, the pore neck width, and the deflection degree in the rescan priority mapping table are read. The sorting weights are weighted and summed and summarized according to the candidate rescan coordinate attribution relationship to obtain the rescan priority.

[0009] As a further aspect of the present invention, the process of generating candidate rescan coordinates includes: obtaining particle contact arcs, pore neck positions and particle contour center positions with rescan priority greater than the priority threshold, and reading the corresponding sample surface coordinates, low magnification and high magnification rescan magnification, and determining the rescan coverage window based on the high magnification field size corresponding to the high magnification rescan magnification. When the overlap ratio between two rescan coverage windows is greater than the window merging threshold, the center positions of the two rescan coverage windows are merged into the same candidate rescan coordinates, and the higher rescan priority is retained to generate the candidate rescan coordinates.

[0010] As a further aspect of the present invention, the process of determining the bearing node includes: obtaining the grayscale distribution of the particle contact arc, the coordinates of the contact arc endpoints, and the particle outline boundaries on both sides corresponding to the local contact area; determining the midpoint of the contact arc based on the contact arc endpoint coordinates; and determining the contact normal direction based on the particle outline boundaries on both sides. When the grayscale distribution of the particle contact arc satisfies the contact density threshold, the contact arc length is greater than the node arc length threshold, and the angle between the contact normal direction and the loading direction is within the bearing angle range, the midpoint of the contact arc is mapped to the coordinate diagram of the sample surface to obtain the bearing node.

[0011] As a further aspect of the present invention, the process of determining the contact density threshold includes: obtaining the average gray value of the non-contact pore region and the average gray value of the particle interior region within the same high-magnification rescan image, and determining the contact gray value reference range based on the gray value interval between the average gray value of the non-contact pore region and the average gray value of the particle interior region; When the average gray value of a local contact area falls within the contact gray value reference range and the gray value dispersion is less than the gray value fluctuation threshold, the local contact area is determined as a dense contact area, and the contact density threshold is updated based on the dense contact area.

[0012] As a further aspect of the present invention, the training sample generation process includes: acquiring multiple bearing chain characterization sequences under the same soil sample identifier, reading the number of bearing nodes, transmission path length level, path direction level and contact strength level in each bearing chain characterization sequence, and generating a sample feature sequence according to the path position order in the coordinate diagram of the same soil sample surface. When the soil sample identifier corresponding to the sample feature sequence is consistent with the soil sample identifier in the mechanical target data, the sample feature sequence is written to the sample input terminal, and the mechanical target data is written to the sample output terminal to generate the training sample.

[0013] As a further aspect of the present invention, the process of updating the sample feature sequence includes: obtaining the collection area location and high-magnification rescan image clarity evaluation value of the bearing chain characterization sequence under the same soil sample identification, and determining the coverage distribution of the bearing chain characterization sequence in the sample surface coordinate diagram based on the collection area location. When the coverage distribution does not reach the coverage ratio threshold, or when the high-magnification rescan image sharpness evaluation value is lower than the sharpness threshold, the corresponding carrier chain representation sequence is removed from the sample feature sequence, and the sample feature sequence is regenerated based on the remaining carrier chain representation sequence.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the stage coordinates are corrected based on the overlapping edges of adjacent low-magnification fields of view to form a coordinate map of the sample surface. The microscopic image acquisition process is transformed from discrete framing to traceable spatial sampling. Particle contours, pore contours, and particle orientations are sorted under a unified coordinate system. High-magnification rescans are concentrated in the contact arc, narrow neck, and deflection regions. Bearing nodes are connected in series along the loading direction to form a transmission path and bearing chain characterization sequence. Image appearance features are transformed into structured inputs that conform to the force transmission relationship. During the training of the prediction network, microscopic bearing chain and mechanical target data are absorbed simultaneously, reducing the interference of shooting area and texture noise, and enhancing the consistency between strength, deformation, and modulus predictions and the macroscopic mechanical formation mechanism of soil. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method for predicting the macroscopic mechanical properties of soil based on scanning electron microscope images according to the present invention. Figure 2 This is a schematic diagram illustrating the low-magnification field-of-view coordinate correction effect of the present invention; Figure 3This is a schematic diagram illustrating the high-magnification rescan coordinate set generation effect of the present invention; Figure 4 This is a schematic diagram illustrating the generation effect of the carrier node and transmission path in this invention; Figure 5 This is a schematic diagram illustrating the correspondence between the characterization sequence of the bearing chain and the mechanical target data of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be described with reference to the accompanying drawings. Example

[0017] Please see Figures 1 to 5 This embodiment provides a method for predicting the macroscopic mechanical properties of soil based on SEM images. During the preparation and continuous observation of soil samples using a scanning electron microscope, a low-magnification field of view is used to cover the sample surface, while a high-magnification rescan is used to supplement details of particle contact and pore necks. The microscopic bearing chain is then correlated with the mechanical target data of the same sample. Existing local image analysis is easily affected by field-of-view position deviations, inconsistent rescan positions, and discontinuous correspondences between microscopic features and macroscopic mechanical targets. This method forms a continuous processing chain through low-magnification field-of-view coordinate correction, rescan priority filtering, bearing node connection, and prediction network training. Example

[0018] S1: Obtain the low-magnification field sequence of the soil sample formed by the scanning electron microscope, correct the stage coordinates based on the overlapping edge position of adjacent low-magnification fields, and generate a coordinate map of the sample surface.

[0019] The low-magnification field-of-view sequence refers to a set of low-magnification images acquired in the order of stage movement, each image containing field-of-view identifiers, acquisition order, nominal stage coordinates, image orientation, and magnification information. Adjacent low-magnification fields of view refer to consecutive fields of view with overlapping edges within their nominal coverage areas. The sample surface coordinate map refers to the image and its positional correspondence, stitched together using the corrected stage coordinates as a reference, converting the pixel positions of each field of view to the sample surface positions. Before matching, fields of view with missing content, unconfirmed orientations, or conflicting identifiers are excluded, and their ranges are recorded as areas to be acquired.

[0020] S101: Extract the first edge patch and the second edge patch from the overlapping edge positions of adjacent low-magnification fields of view respectively, determine the particle edge intersection point and the pore boundary inflection point based on the gray-scale gradient change, and determine the particle edge intersection point and the pore boundary inflection point as candidate texture feature points.

[0021] The first and second edge patches are taken from the expected overlapping edge ranges and their respective fields of view, original image pixel ranges, and nominal positions are recorded. Particle edge intersections are locations where multiple particle boundaries form stable intersections, and pore boundary inflection points are locations where the pore contour direction undergoes continuous inflection. Candidate texture feature points record pixel positions, local grayscale arrangement, adjacent boundary directions, and neighboring point order. Pollution spots, charging bright spots, edge gaps, and locations where contours are not discontinuously discernible are not considered as candidate texture feature points.

[0022] S102: Compare the local grayscale distribution, contour direction and spacing of candidate texture feature points in the first edge patch and the second edge patch, and determine the candidate texture feature points that meet the fitting degree threshold and whose spacing deviation is less than the spacing deviation threshold as matching texture feature points.

[0023] The fit threshold is derived from the matching records of confirmed overlapping fields of view under the same sample preparation method, and is used to limit the minimum consistency state that local gray-level arrangement, contour direction, and neighboring point order should achieve. The spacing deviation threshold is derived from the calibration records of stage repetitive movement deviation and image edge deformation. During comparison, candidate points with conflicting contour directions are first excluded, and then the gray-level change order and neighboring point spacing are compared. When multiple combinations simultaneously meet the conditions, the combination with continuous contours and no conflicting neighboring point order is retained. If a stable combination cannot be formed, the corresponding edge range is transferred to re-sampling or the overlapping range is expanded before re-matching.

[0024] S103: Determine the edge displacement based on the pixel position difference of the matching texture feature points in adjacent low magnification fields of view, and convert the edge displacement into the stage coordinate correction amount to obtain the edge texture feature point offset.

[0025] Among them, the edge displacement is the change in pixel position and direction of the same matched texture feature point in two fields of view, and the stage coordinate correction is the sample surface position correction state obtained by converting the pixel size calibration relationship based on the current magnification. Records that are opposite in direction to most matching records or that may cause contour misalignment are first eliminated, and then the edge texture feature point offset is determined from the remaining records. The confirmed position of the previous field of view is corrected for the next field of view along the acquisition sequence, and the cumulative offset is verified by using the closed overlap relationship.

[0026] S104: Generate a coordinate map of the sample surface based on the corrected stage coordinates, the pixel position relationship of each low magnification field of view, and the consistency of overlapping edges.

[0027] Each field of view pixel is mapped to a unified position on the sample surface. When multiple grayscale records exist at overlapping locations, records with continuous particle edges, clear pore boundaries, and no contamination are retained. If both records are valid, they are smoothly connected from the inside of the overlapping area towards the edge. The sample surface coordinate map simultaneously saves the field of view source identifier, allowing subsequent particle contact arcs, pore necks, and rescan coordinates to be traced back to the original low-magnification field of view. Through continuous coordinate correction, particles and pore boundaries in different fields of view are connected under the same positional reference.

[0028] S2: Identify particle and pore contours based on the coordinate diagram of the sample surface, and verify the particle contact arc, pore neck width, and the degree of deflection of the particle long axis relative to the loading direction.

[0029] Among them, particle profile is the closed boundary formed by surrounding the continuous gray-scale region of the particle and undergoing fracture correction, and pore profile is the closed boundary surrounding the continuous gray-scale region of the pore. Particle contact arc is the boundary arc segment corresponding to adjacent particles within the contact area. Pore neck width is the shortest interval between the particle profiles on both sides of the connection point between adjacent pores. Deflection degree is the level of deviation of the particle's long axis relative to the loading direction. The loading direction is taken from the mechanical test record of the same specimen; if the record is missing, deflection sorting and training are not performed.

[0030] S201: Obtain the pixel grayscale value, local grayscale gradient and neighborhood connectivity in the coordinate map of the sample surface, determine the first segmentation boundary based on the grayscale difference threshold between the particle region and the pore region, and perform edge adsorption adjustment on the first segmentation boundary based on the local grayscale gradient to generate the second segmentation boundary.

[0031] The grayscale difference threshold is derived from the confirmed internal regions of particles and non-contact pore regions in the current image, used to distinguish between particle candidate regions and pore candidate regions. When illumination or charging state changes, it is redefined within the corresponding field of view. The first segmentation boundary is located adjacent to both types of candidate regions. Edge snapping adjustment searches for positions with continuous and stable grayscale changes along both sides of the first segmentation boundary, moving the boundary to the location of the true grayscale abrupt change, while suppressing isolated sharp corners and repeated folds caused by local noise, thus obtaining the second segmentation boundary.

[0032] S202: When there are broken endpoints in the second segmentation boundary, the boundary segment to be connected is determined based on the directional consistency and distance threshold between the broken endpoints, and the boundary segment to be connected is incorporated into the second segmentation boundary.

[0033] Distance thresholds are derived from local interruption calibration records of clear particle boundaries at the same magnification. Orientation consistency is determined based on whether adjacent endpoint boundaries extend towards each other, whether the original directional trend is maintained after connection, and whether the connecting line passes through the internal region of a confirmed particle. When multiple endpoints compete for connection, connections that can form closed boundaries and do not engulf adjacent pores are prioritized. Endpoints that do not meet the conditions remain broken, and their areas are marked as contour-unconfirmed areas, not included in contact arc and neck identification.

[0034] S203: The region whose grayscale value meets the criteria for determining a particle region after closure is defined as the particle outline, and the region whose grayscale value meets the criteria for determining a pore region after closure is defined as the pore outline. The candidate regions for particles and pores are verified based on the degree of closure of the grayscale abrupt change boundary.

[0035] The criteria for determining particle regions include continuous internal grayscale, grayscale transitions between pores on the outer edge of the boundary, and the absence of through pores within the particle. The criteria for determining pore regions include internal grayscale consistent with non-contact pores, boundaries formed by abrupt grayscale changes on the particle side, and connectivity with adjacent pores. Closed dark areas formed by internal particle textures are incorporated into the particles, while the actual intervals between particles are preserved as pores. Each contour records the boundary point sequence, center position, orientation, and adjacent contour identifiers.

[0036] S204: Determine the particle contact area based on the common edge length and boundary spacing between adjacent particle contours, and map the particle contact area to the boundary arc segment of the corresponding particle contour to obtain the particle contact arc.

[0037] The particle contact zone is a region in which the contours of two particles extend towards each other within a local area, and the boundary spacing meets the contact determination criteria. The contact determination criteria are derived from the boundary spacing and continuous state of shared edges of the confirmed contact areas at the same magnification. Boundaries that only briefly approach each other and then quickly separate are not considered contact zones. After confirmation, the two sides of the contact zone are mapped to particle contact arcs, and the endpoints, adjacent particle identifiers, and sample surface coordinates are recorded.

[0038] S205: Determine the pore neck position based on the narrowest connection position between adjacent pore profiles, calculate the shortest distance between the particle profiles on both sides at the pore neck position, and obtain the pore neck width.

[0039] Examine the particle profiles on both sides of the channel along the pore connectivity direction. Determine the pore neck position according to the sequence of boundary interval transitioning from contraction to expansion, and record the shortest interval between the particle profiles on both sides of that position as the pore neck width. Record each contraction position separately if multiple contraction positions occur in a channel. Internal particle cracks, image gaps, and contraction positions falling into the area to be confirmed are not considered valid pore necks.

[0040] S206: Determine the degree of deflection based on the angle between the major axis of the particle profile and the loading direction.

[0041] The major axis of the particles is determined by the main extension direction of the profile and unified to the coordinate reference of the sample surface. The degree of deflection is expressed using a dimensionless grade field, with the grade boundaries derived from historical calibration records of the relationship between the particle orientation and the load path under the same loading method. The greater the deviation of the particle's major axis from the loading direction, the higher its sorting position. When particles are nearly equiaxed but cannot form a stable major axis, they are marked as having an uncertain orientation, and this field is not used to increase the rescan priority.

[0042] S3: Based on the sorting positions of the particle contact arc length, the pore neck width, and the deflection degree, verify the high-magnification rescan coordinate set, and generate a high-magnification rescan image according to the high-magnification rescan coordinate set.

[0043] The high-magnification rescan coordinate set is a collection of coordinate records formed after priority determination, coverage window merging, and coordinate validity verification. Each record includes the candidate source, sample surface coordinates, corresponding low-magnification field of view, rescan priority, rescan coverage window, and type of feature to be observed. The rescan priority is determined by relative sorting within the same sample, without directly comparing arc length, width, and directional angle.

[0044] S301: Sort the particle contact arcs by length from largest to smallest to generate the particle contact arc length sorting position, sort the pore neck width by width from smallest to largest to generate the pore neck width sorting position, and sort the deflection degree by deviation from the loading direction from largest to smallest to generate the deflection degree sorting position.

[0045] Records with invalid source contours, coordinates falling into areas to be sampled again, or incomplete feature boundaries are excluded before sorting. The three sorting categories correspond to continuous contact range, pore channel contraction state, and particle orientation deviation state, respectively. When the same sorting field exhibits the same level, a fixed order is maintained according to the path position in the sample surface coordinate diagram, ensuring a consistent rescanning sequence for repeated processing.

[0046] S302: Read the sorting weights corresponding to the sorting positions of the particle contact arc length, the pore neck width, and the deflection degree in the rescan priority mapping table, sum the sorting weights by weight and merge them according to the candidate rescan coordinate attribution relationship to obtain the rescan priority.

[0047] The rescan priority mapping table is a lookup record that maps the three types of sorting positions to priority levels and dimensionless sorting weights. The weights are derived from historical region records where high-magnification rescans have been completed and the carrying nodes can be identified. During processing, the three types of weights are read sequentially and merged according to the primary and secondary relationships in the table. When the contact arc and the neck both point to the same region, the priority is adjusted forward. When only the degree of deflection provides support, it is retained as a secondary candidate. In case of conflict, the effective contact arc, effective pore neck, and effective degree of deflection are judged in that order.

[0048] S303: Obtain the particle contact arc, pore neck position and particle outline center position that meet the priority threshold for rescanning priority, and read the corresponding sample surface coordinates, low magnification and high magnification rescanning magnification, and determine the rescanning coverage window based on the high magnification field size corresponding to the high magnification rescanning magnification.

[0049] The priority threshold is the lowest priority level for a candidate location to enter a high-magnification rescan, derived from the calibration record of the current allowable rescan range of the sample and the effective state of historical high-magnification images. The rescan coverage window is the coverage area centered on the candidate location and mapped onto the sample surface coordinate map according to the high-magnification field of view. The contact arc candidate location covers the arc segment and its two side boundaries, the narrow neck candidate location covers the narrowest connection point, and the deflection candidate location covers the particle center and the long axis direction.

[0050] S304: When the overlap area ratio between two rescan coverage windows meets the window merging threshold, the center positions of the two rescan coverage windows are merged into the same candidate rescan coordinates, and the higher rescan priority is retained to generate candidate rescan coordinates.

[0051] The window merging threshold is derived from the high-magnification field of view and repeated coverage calibration records. Before merging, it is confirmed that the two windows belong to the same continuous region, and it is checked whether the adjusted center position can simultaneously cover the main features of the two candidate records. If they can both cover each other, the center position and feature source identifier are merged, and the earlier rescan priority is retained. If they cannot both cover each other, they are retained separately, and merging is not forced due to local edge overlap.

[0052] S305: Verify whether the candidate rescan coordinates are within the effective coverage area of ​​the sample surface coordinate map, and generate a set of high-magnification rescan coordinates according to the rescan priority and path position, and then obtain the corresponding high-magnification rescan images.

[0053] The effective coverage area is the region where low-magnification stitching is complete, the outline is traceable, and the high-magnification field of view does not extend beyond the observable surface. When candidate coordinates fall into the area to be re-sampled, the outline to be confirmed, or the edge of the sample, the center position is adjusted along the target feature. If complete coverage is still not achieved after adjustment, it is marked as invalid for rescanning. After the high-magnification image is formed, the target feature is checked to see if it is within the image, if the boundary is identifiable, and if there is contamination or defocusing. If the sharpness threshold is not met, the original coordinates are retained and marked as rescanning to be updated. The sharpness threshold is derived from the lowest image state that can identify the contact arc endpoints, the outlines on both sides, and the local grayscale.

[0054] S4: For the particle contact arc in the high-magnification rescan image, determine the bearing node, and based on the bearing nodes located on the same particle contour and adjacent along the loading direction, generate the transfer path and bearing chain characterization sequence.

[0055] Among them, the bearing node is the record of the midpoint position of the contact arc that satisfies the contact compactness state, contact arc length state, and contact normal direction state. The transfer path is the position sequence formed by connecting adjacent bearing nodes along the loading direction according to the particle profile and contact sequence. The bearing chain characterization sequence records the bearing node quantity level, transfer path length level, path direction level, and contact strength level according to the path position. The boundaries of each level are derived from historical calibration records under the same imaging and loading conditions.

[0056] S401: Obtain the grayscale distribution of the particle contact arc, the coordinates of the contact arc endpoints, and the particle outline boundaries on both sides corresponding to the local contact area. Determine the midpoint of the contact arc based on the contact arc endpoint coordinates, and determine the contact normal direction based on the particle outline boundaries on both sides.

[0057] The local contact area is enclosed by a pair of adjacent particle profiles and their corresponding contact arcs. The midpoint of the contact arc is determined according to the continuous sequence of the arc segment boundary points, and the contact normal direction is determined according to the orientation relationship of the particle profiles on both sides, and is unified to the coordinate reference of the sample surface. When the endpoint falls into the image edge, the profile is broken, or the normal direction continues to conflict, the area is marked as a node pending confirmation and is not included in the current set of bearing nodes.

[0058] S402: Obtain the average grayscale value of the non-contact pore area and the average grayscale value of the internal area of ​​the particle within the same high-magnification rescan image, and determine the contact grayscale reference range based on the grayscale interval between the two grayscale values.

[0059] The contact grayscale reference range is the relative grayscale boundary within the current high-magnification image used to determine dense contact. For non-contact pore regions, locations with clear outlines and far from the contact arc are selected; for internal particle regions, locations far from boundaries and internal cracks are selected. Both types of regions must have continuous grayscale and be free of contaminating bright spots. A separate reference range is established for each high-magnification image, without directly using ranges across images. If the reference range is insufficient, the dense contact determination state is marked as unavailable and the image is rescanned for updates.

[0060] S403: When the average gray value of a local contact area falls within the contact gray value reference range and the gray value dispersion meets the gray value fluctuation threshold, the local contact area is determined as a dense contact area, and the contact density threshold is updated based on the dense contact area.

[0061] The grayscale fluctuation threshold is derived from the continuous grayscale records of the internal regions of particles and the confirmed dense contact regions within the current high-magnification image. The contact density threshold uses the contact grayscale reference range as its initial boundary and is updated as the confirmed dense contact regions are identified. Local grayscale states falling within the reference range but exhibiting porosity-like abrupt changes are not considered dense contact. Records whose continuous state cannot be restored after contaminated pixels are excluded are not included in the threshold update.

[0062] S404: When the grayscale distribution of the particle contact arc meets the contact density threshold, the contact arc length meets the node arc length threshold, and the angle between the contact normal direction and the loading direction is within the bearing angle range, the midpoint of the contact arc is mapped to the coordinate diagram of the sample surface to obtain the bearing node.

[0063] The node arc length threshold is derived from the boundary calibration records of effective load-bearing contact and brief near-contact contact. The load-bearing angle range is derived from the loading direction of the mechanical test and the confirmed load-bearing path direction records. The judgment order is: image and contour validity, contact compactness, arc length status, and normal direction status. If any condition is not met, no node is generated, and the failed field is recorded. When all conditions are met, the midpoint of the contact arc is converted to the coordinates of the sample surface, and the adjacent particles, contact arc, normal direction, contact strength level, and image sharpness evaluation value are recorded.

[0064] S405: Generate a transfer path based on the load-bearing nodes that are located on the same particle profile and are adjacent along the loading direction.

[0065] First, establish the node sequence according to the sample surface coordinates and loading direction, then check whether adjacent nodes share the same particle profile. If they share particles and the transfer direction is continuous, connect them as path segments. When one node corresponds to multiple subsequent nodes, determine the main path based on the consistency level of the normal direction and loading direction, the continuity of the shared profile, and the contact strength level; the rest are retained as branch paths. Paths must not cross unidentified particles, nor should they be connected solely due to coordinate proximity. Interruptions caused by areas awaiting re-sampling are recorded as path gaps.

[0066] S406: Generate a load-bearing chain characterization sequence based on the number of load-bearing nodes in the transmission path, the continuous length along the loading direction, the consistency of the path direction, and the contact strength level of each load-bearing node.

[0067] The number of load-bearing nodes is determined based on the relative number of effective nodes within the path segment; the path length is determined based on the relative segments extending continuously along the loading direction; the path direction is determined based on the consistency and abrupt changes in direction of each path segment with the loading direction; and the contact strength is determined by looking up tables based on the compaction state, arc length state, and normal direction state. All levels are uniformly represented as dimensionless fields and written into the load-bearing chain characterization sequence according to the path location, ensuring a stable input for the microscopic contact state, spatial continuity, and loading direction relationship.

[0068] S5: Obtain the mechanical target data corresponding to the soil sample, train the prediction network based on the bearing chain characterization sequence and the mechanical target data, input the bearing chain characterization sequence of the soil sample to be tested into the prediction network, and output the prediction results of the macroscopic mechanical properties of the soil.

[0069] The mechanical target data consists of supervised target fields corresponding to the specimen identification and formed from macroscopic mechanical test records, including target type, specimen identification, loading direction, effective test state, and target result. The prediction network is a hierarchical transformation structure that receives the load-bearing chain characterization sequence and outputs the corresponding macroscopic mechanical target result, comprising a sequence input processing layer, a local feature extraction layer, a sequential relationship encoding layer, a feature aggregation layer, and a regression output layer. The input processing layer checks field range, missing states, and path order. The local feature extraction layer extracts the local combination relationships of node number, path length, direction, and contact strength, and uses rectified activation to retain positive responses. The sequential relationship encoding layer uses a gating update mechanism to pass the preceding and following path states. The feature aggregation layer eliminates invalid positions and aggregates multiple load-bearing chains. The regression output layer uses a linear output mechanism to generate the prediction result.

[0070] S501: Obtain multiple load-bearing chain characterization sequences under the same soil sample identifier, read the load-bearing node quantity level, transmission path length level, path direction level and contact strength level in each load-bearing chain characterization sequence, and generate sample feature sequences according to the path position order in the coordinate diagram of the same soil sample surface.

[0071] The sample feature sequence is a training input record formed by connecting multiple load-bearing chains of the same sample in a stable spatial order. It is preferentially arranged according to their front-to-back positions in the loading direction; when multiple paths exist at the same location, they are maintained in a fixed order according to path direction level and contact strength level. The sequence simultaneously records the acquisition area location, image sharpness evaluation value, path gap status, and valid path identifier. When the number of load-bearing chains changes, the scope entering the convergence processing is limited by the valid path identifier, and fictitious data is not used to fill in the gaps.

[0072] S502: Obtain the location of the sampling area and the image clarity evaluation value of the high-magnification rescan image of the bearing chain characterization sequence under the same soil sample identification, and determine the coverage distribution of the bearing chain characterization sequence on the sample surface coordinate map based on the sampling area location.

[0073] Coverage distribution refers to the spatial distribution of the effective load-bearing chain's sampling area relative to the effective range of the sample surface. The load-bearing chain is checked against the loading direction and its perpendicular direction to determine if it is concentrated in a localized area and whether there are continuous gaps caused by areas to be resampled. The coverage ratio threshold is derived from historical sample records that stably correspond to the mechanical target data under the same sample preparation and rescanning process. The sharpness evaluation value is formed based on the identifiable state of the contact arc endpoints, side contours, and local grayscale; the sharpness threshold is the lowest state required to complete the confirmation of the load-bearing node.

[0074] S503: When the coverage distribution does not reach the coverage ratio threshold, or the high magnification rescan image sharpness evaluation value is lower than the sharpness threshold, the corresponding carrier chain characterization sequence is removed from the sample feature sequence, and the sample feature sequence is regenerated based on the remaining carrier chain characterization sequence.

[0075] When the image corresponding to a single carrier chain is unclear, only that carrier chain is removed, and the image is kept for rescanning and updating. When the overall coverage is insufficient, the current sample is marked as training paused, and local carrier chains are not used to represent the complete sample. After rescanning and updating, the carrier chains are regenerated according to the original sample surface coordinates, and the coverage distribution is judged again. Repeated paths are merged according to the node order and the location of the acquisition area, and records with sufficient clarity and few path gaps are retained.

[0076] S504: When the soil sample identifier corresponding to the sample feature sequence is consistent with the soil sample identifier in the mechanical target data, the sample feature sequence is written into the sample input terminal, and the mechanical target data is written into the sample output terminal to generate a training sample.

[0077] During matching, the specimen identifier, loading direction, and target type are verified simultaneously. If the mechanical target data is marked as invalid for the test, has a conflicting identifier, or the loading direction cannot be confirmed, no training sample is generated, and the data remains in a pending verification state. The training sample saves the input field definitions, valid path identifiers, target field definitions, and source relationships, enabling the output target to be traced back to the corresponding load-bearing chain input range.

[0078] S505: Train the prediction network based on training samples.

[0079] Samples for parameter adjustment and samples for stopping the test are separated according to their identification to avoid the same sample entering both types of samples simultaneously. The sample feature sequences are sequentially processed through each layer to form training prediction results, which are then compared with the corresponding mechanical target data to obtain the output deviation direction. Parameter adjustment is fed back layer by layer in the direction of reducing deviation, sequentially adjusting output connectivity, feature convergence, preservation of preceding and following path states, and local combined feature responses. Stopping the test is triggered when the output deviation no longer continues to improve, continuous adjustment fails to achieve stable improvement, or a preset training termination state is reached. The termination state is derived from historical convergence records of similar targets, not from fixed training iterations or performance values.

[0080] S506: Obtain the bearing chain characterization sequence of the soil sample to be tested, generate the feature sequence of the sample to be tested according to the same field definitions, grade boundaries, path position order and validity judgment rules in the training stage, and input it into the trained prediction network to output the prediction results of the macroscopic mechanical properties of the soil.

[0081] Before input, checks are performed for undefined levels, continuous path gaps, and insufficient overall coverage. Undefined levels are mapped to existing adjacent levels based on the currently valid calibration records, and the mapping status is recorded. When there are continuous path gaps or insufficient coverage, the predicted supplementary status and corresponding rescan positions are output, but the formal prediction results are not output. When the input conditions are met, the prediction network processes the data according to the predetermined interlayer order and generates prediction results corresponding to the target type, while also associating the specimen identifier, loading direction, load-bearing chain input range, and validity status.

[0082] In this embodiment, the low-magnification field of view position is corrected by overlapping edge textures to maintain a unified positional reference for particle outlines, pore outlines, and high-magnification rescan coordinates. By summarizing the three types of sorted positions through table lookup and merging the coverage windows, the rescan range is concentrated on areas related to particle load-bearing capacity and pore shrinkage. The load-bearing nodes are determined jointly by in-image grayscale reference, contact arc length status, and normal direction status, ensuring that the load-bearing chain retains contact density, spatial continuity, and loading direction relationships. Through coverage distribution, image sharpness, and specimen identification verification, the predicted input and mechanical target data are kept consistent with each other, thus completing the continuous processing from SEM image to prediction of soil macroscopic mechanical properties.

Claims

1. A method for predicting the macroscopic mechanical properties of soil based on SEM images, characterized in that, Includes the following steps: The low-magnification field sequence of the soil sample formed by the scanning electron microscope is obtained, and the stage coordinates are corrected based on the overlapping edge position of adjacent low-magnification fields to generate a coordinate map of the sample surface. Based on the coordinate diagram of the sample surface, the particle outline and pore outline are identified, and the particle contact arc, pore neck width and particle long axis deflection relative to the loading direction are verified. Based on the sorting positions of the particle contact arc length, the pore neck width, and the deflection degree, the high-magnification rescan coordinate set is verified, and a high-magnification rescan image is generated based on the high-magnification rescan coordinate set. For the particle contact arc in the high-magnification rescan image, the bearing nodes are determined, and based on the bearing nodes located on the same particle contour and adjacent along the loading direction, the transfer path and bearing chain characterization sequence are generated. Obtain the mechanical target data corresponding to the soil sample, train a prediction network based on the bearing chain characterization sequence and the mechanical target data, input the bearing chain characterization sequence of the soil sample to be tested into the prediction network, and output the prediction results of the macroscopic mechanical properties of the soil.

2. The method for predicting the macroscopic mechanical properties of soil based on SEM images according to claim 1, characterized in that, The process of determining the offset of the edge texture feature points includes: extracting the first edge patch and the second edge patch from the overlapping edge positions of adjacent low-magnification fields of view, determining the particle edge intersection point and the pore boundary inflection point based on the gray-level gradient change, and determining the particle edge intersection point and the pore boundary inflection point as candidate texture feature points; Compare the local grayscale distribution, contour direction and spacing of candidate texture feature points in the first edge patch and the second edge patch, and determine the candidate texture feature points that meet the fitting degree threshold and whose spacing deviation is less than the spacing deviation threshold as the matching texture feature points; The edge displacement is determined based on the pixel position difference of the matching texture feature points in adjacent low-magnification fields of view, and the edge displacement is converted into a stage coordinate correction amount to obtain the edge texture feature point offset.

3. The method for predicting the macroscopic mechanical properties of soil based on SEM images according to claim 1, characterized in that, The process of determining the particle contact arc, the pore neck width, and the degree of deflection includes: obtaining the gray-scale continuous region and gray-scale abrupt boundary in the coordinate diagram of the sample surface; dividing the soil sample surface region into particle candidate region and pore candidate region according to the degree of closure of the gray-scale abrupt boundary; and performing neighborhood connectivity correction on the boundary fracture position to generate particle profile and pore profile. The particle contact area is determined based on the common edge length and boundary spacing between adjacent particle contours, and the particle contact area is mapped to the boundary arc segment of the corresponding particle contour to obtain the particle contact arc. The pore neck position is determined based on the narrowest connection position between adjacent pore profiles. The shortest distance between the particle profiles on both sides at the pore neck position is calculated to obtain the pore neck width. The degree of deflection is determined based on the angle between the major axis of the particle profile and the loading direction.

4. The method for predicting the macroscopic mechanical properties of soil based on SEM images according to claim 3, characterized in that, The generation process of the particle profile and the pore profile includes: obtaining the pixel gray value, local gray gradient and neighborhood connectivity in the coordinate map of the sample surface; determining the first segmentation boundary based on the gray difference threshold between the particle region and the pore region; and adjusting the edge of the first segmentation boundary based on the local gray gradient to generate the second segmentation boundary. When the second segmentation boundary has a fracture endpoint, the boundary segment to be connected is determined based on the directional consistency and distance threshold between the fracture endpoints. The boundary segment to be connected is incorporated into the second segmentation boundary, and the area with higher gray level after closure is determined as the particle profile, and the area with lower gray level after closure is determined as the pore profile.

5. The method for predicting the macroscopic mechanical properties of soil based on SEM images according to claim 1, characterized in that, The process of determining the rescan priority includes: generating particle contact arc length sorting positions according to the arc length from large to small, generating pore neck width sorting positions according to the width from small to large, and generating deflection degree sorting positions according to the degree of deviation from the loading direction from large to small. The sorting weights corresponding to the sorting positions of the particle contact arc length, the pore neck width, and the deflection degree in the rescan priority mapping table are read. The sorting weights are weighted and summed and summarized according to the candidate rescan coordinate attribution relationship to obtain the rescan priority.

6. The method for predicting the macroscopic mechanical properties of soil based on SEM images according to claim 5, characterized in that, The process of generating candidate rescan coordinates includes: obtaining particle contact arcs, pore neck positions and particle contour center positions with rescan priority greater than the priority threshold, and reading the corresponding sample surface coordinates, low magnification and high magnification rescan magnification, and determining the rescan coverage window based on the high magnification field size corresponding to the high magnification rescan magnification. When the overlap ratio between two rescan coverage windows is greater than the window merging threshold, the center positions of the two rescan coverage windows are merged into the same candidate rescan coordinates, and the higher rescan priority is retained to generate the candidate rescan coordinates.

7. The method for predicting the macroscopic mechanical properties of soil based on SEM images according to claim 1, characterized in that, The process of determining the bearing node includes: obtaining the grayscale distribution of the particle contact arc, the coordinates of the contact arc endpoints, and the particle outline boundaries on both sides corresponding to the local contact area; determining the midpoint of the contact arc based on the contact arc endpoint coordinates; and determining the contact normal direction based on the particle outline boundaries on both sides. When the grayscale distribution of the particle contact arc satisfies the contact density threshold, the contact arc length is greater than the node arc length threshold, and the angle between the contact normal direction and the loading direction is within the bearing angle range, the midpoint of the contact arc is mapped to the coordinate diagram of the sample surface to obtain the bearing node.

8. The method for predicting the macroscopic mechanical properties of soil based on SEM images according to claim 7, characterized in that, The process of determining the contact density threshold includes: obtaining the average gray value of the non-contact pore area and the average gray value of the particle interior area in the same high-magnification rescan image, and determining the contact gray value reference range based on the gray value interval between the average gray value of the non-contact pore area and the average gray value of the particle interior area; When the average gray value of a local contact area falls within the contact gray value reference range and the gray value dispersion is less than the gray value fluctuation threshold, the local contact area is determined as a dense contact area, and the contact density threshold is updated based on the dense contact area.

9. The method for predicting the macroscopic mechanical properties of soil based on SEM images according to claim 1, characterized in that, The training sample generation process includes: obtaining multiple bearing chain characterization sequences under the same soil sample identifier, reading the number of bearing nodes, transmission path length level, path direction level and contact strength level in each bearing chain characterization sequence, and generating sample feature sequences according to the path position order in the coordinate diagram of the same soil sample surface. When the soil sample identifier corresponding to the sample feature sequence is consistent with the soil sample identifier in the mechanical target data, the sample feature sequence is written to the sample input terminal, and the mechanical target data is written to the sample output terminal to generate the training sample.

10. The method for predicting the macroscopic mechanical properties of soil based on SEM images according to claim 9, characterized in that, The process of updating the sample feature sequence includes: obtaining the collection area location and high-magnification rescan image clarity evaluation value of the bearing chain characterization sequence under the same soil sample identification, and determining the coverage distribution of the bearing chain characterization sequence in the sample surface coordinate map based on the collection area location. When the coverage distribution does not reach the coverage ratio threshold, or when the high-magnification rescan image sharpness evaluation value is lower than the sharpness threshold, the corresponding carrier chain representation sequence is removed from the sample feature sequence, and the sample feature sequence is regenerated based on the remaining carrier chain representation sequence.

Images