A machine learning-based method for mineral trace element analysis
By employing a machine learning-based approach and leveraging data coupling between multi-source features and progressive transfer learning, a deep neural network model was constructed. This solved the signal aliasing problem in the analysis of trace element distribution in garnet zonation, enabling high-confidence reconstruction and causal inference, and improving the accuracy and consistency of the analysis results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF GEOSCIENCES (BEIJING)
- Filing Date
- 2026-04-02
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies struggle to accurately interpret true geochemical signals when analyzing trace element distribution in garnet zonation due to inherent limitations in data acquisition. This is particularly true at abrupt compositional transitions where signal aliasing is severe, leading to increased uncertainty and noise in the measurement data.
A machine learning-based approach is adopted, which integrates multi-source features with progressive transfer learning to construct a deep neural network model. By utilizing the spatial correlation matrix between the source domain and the target domain of the mineral core and the fusion of multi-source features, a high-confidence reconstruction and genetic inference of the trace element distribution in the complex zonal structure inside the mineral is achieved.
It effectively solves the problem of data collection limitations, improves the accuracy and consistency of trace element analysis, provides a complete spatial distribution map of trace elements and a zoning map of causal control factors, and lowers the threshold and ambiguity of subsequent interpretation.
Smart Images
Figure CN122290795A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of mineral element analysis technology, and in particular relates to a mineral trace element analysis method based on machine learning. Background Technology
[0002] Machine learning, a branch of artificial intelligence, automatically identifies hidden patterns, correlations, or regularities from existing data by building computational models with specific structures. The core of this process lies in the model's ability to self-adjust its internal parameters based on the data, without requiring explicit, fixed instruction rules for each problem. Once trained, these models possess the ability to classify, predict, or discover similar patterns in new, unseen data, and are therefore widely used in many scientific and engineering fields that require extracting information from complex data.
[0003] Mineral trace elements are elements that exist in the crystal structure of minerals but are usually present in very low amounts (often less than one-thousandth). Although they do not constitute the main chemical components of minerals, their existence, content, and distribution characteristics are extremely sensitive to the physicochemical environment during formation. Therefore, these elements act like natural "geological fingerprints" or "record recorders," used by Earth scientists to precisely interpret information such as the genesis of rocks, the properties of hydrothermal fluids, magma evolution processes, and the temperature and pressure conditions of various geological events. They are key tracers for retrieving deep processes in Earth's history.
[0004] In metamorphic petrology, the compositional zoning of garnet is a crucial indicator for revealing the metamorphic history and dynamics of rocks. The spatial distribution of trace elements (such as rare earth elements, yttrium, and scandium) contains detailed information about mineral growth environments, fluid activity, and elemental differentiation behavior, which is essential for constructing a complete metamorphic Ptt trajectory. However, obtaining high spatial resolution trace element zoning data faces inherent bottlenecks in practice. When using high-precision micro-area analysis techniques (such as laser ablation inductively coupled plasma mass spectrometry) for direct measurement, the acquired trace element signals are easily affected by spatial aliasing when the analysis beam covers narrow interface regions with abrupt compositional changes in garnet zoning or encounters tiny mineral inclusions. This aliasing results in signals near the interface that are essentially a mixture of multiple micro-area components or substances from different growth stages, significantly increasing the uncertainty and noise of the measurement data and making it difficult to accurately reflect the true in-situ chemical composition of the micro-area. Therefore, the following solution is proposed to address these issues. Summary of the Invention
[0005] The purpose of this invention is to provide a mineral trace element analysis method based on machine learning. By integrating multi-source features with a data coupling method of progressive transfer learning, it is possible to reconstruct and infer the genesis of trace element distribution in complex zonal structures within minerals with high confidence. This solves the problem that existing technologies are unable to accurately interpret real geochemical signals when analyzing microscopic regions such as abrupt compositional interfaces due to the inherent limitations of data acquisition.
[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: This invention relates to a machine learning-based method for mineral trace element analysis, comprising the following steps: S1. Simultaneously acquire major element surface distribution data, backscattered electron images, and trace element line scan data passing through the target garnet particles and their associated minerals, and spatially register all data. S2. Based on the data quality factor and spatial location information of the line scan data, it is divided into a high-quality core source domain data point set and a low-quality or ring target domain data point set located at the compositional abrupt change interface; and a spatial correlation matrix is constructed between the ring target domain data points and the core source domain data points. The spatial correlation matrix is calculated by weighting based on the major element chemical distance between the two points and the estimated relative differences in physical conditions. S3. For each data point, calculate the crystal chemical characteristics based on its major element data, extract local texture features based on the backscattered electron image, and extract the surrounding symbiotic mineral assemblage features based on the mineral identification results. The three types of features are then fused to form a comprehensive feature vector. S4. Construct a deep neural network model with the comprehensive feature vector as input and trace element content as output; design a total loss function including a main prediction loss term, a dynamic domain adaptation loss term, and a spatial smoothing constraint loss term; train the deep neural network model using the data point set of the ore core source domain and its true trace element values, with the goal of minimizing the total loss function; wherein, the dynamic domain adaptation loss term uses the spatial correlation matrix to guide the model to conditionally align the feature distribution of the target domain data points to the feature distribution of the associated ore core source domain data points; S5. Input the comprehensive feature vector of the target domain data points of the ring into the trained deep neural network model to obtain the predicted value of its trace element content; at the same time, output the probability distribution of the cause-controlling factors of trace element content at each point through the auxiliary classification module of the model. S6. Integrate the measured trace element data of the source domain data points of the ore core with the predicted trace element data of the target domain data points of the ring zone to generate a complete spatial distribution map of trace elements inside the target garnet particles and a zonal map of genetic control factors.
[0007] Furthermore, in step S2, the data quality factor comprehensively considers signal strength, internal standard stability, isotope measurement error, and spatial location information of erosion pits; in the weighted calculation of the spatial correlation matrix, the relative differences in physical conditions include relative temperature difference and relative pressure difference estimated based on symbiotic mineral assemblage and major element data.
[0008] Furthermore, in step S2, the spatial correlation matrix is specifically used to represent the correlation strength between each ring target domain data point and all core source domain data points. The correlation strength is negatively correlated with the principal element Euclidean distance between the two points and negatively correlated with the absolute value of the relative difference of the physical conditions.
[0009] Further, in step S3, the crystal chemical features include the mole fraction of garnet endmember components and crystal chemical parameters; the local texture features are obtained by extracting backscattered electron image blocks centered on the point through a pre-trained convolutional neural network; and the symbiotic mineral assemblage features are represented by the types of symbiotic minerals appearing within a preset range around the point and their relative abundance.
[0010] Further, in step S4, the total loss function consists of the following parts: an error loss term between the predicted value and the true value calculated on the data points in the core source domain; a dynamic domain adaptation loss term calculated based on the spatial correlation matrix to reduce the difference in feature distribution between the core source domain and the ring target domain; and a smoothing constraint loss term to encourage the predicted values of spatially adjacent and similar data points to change smoothly.
[0011] Furthermore, the dynamic domain adaptation loss term is implemented by calculating the difference measure between the target domain data points of the ring zone and the source domain data points of each ore core in the hidden layer feature distribution of the model, and using the corresponding correlation strength value in the spatial correlation matrix to perform a weighted summation of the difference measure, which serves as a constraint in model training.
[0012] Furthermore, in step S5, the auxiliary classification module shares a feature extraction layer with the deep neural network model, and during the training process, it uses some causal-labeled mineral core source domain data for supervised training to synchronously output the causal control probability of the predicted trace element values.
[0013] Furthermore, in step S6, the predicted values of trace element content and dominant causal categories of discrete data points are converted into continuous spatial distribution maps using a spatial interpolation algorithm, and then overlaid and displayed with the surface distribution map of major elements and the backscattered electron image.
[0014] The present invention has the following beneficial effects: 1. This invention constructs a dynamic correlation between core data and zonal data, and uses a transfer learning framework to constrain the model, which can effectively handle the problem of data distribution differences caused by changes in growth conditions within the same mineral grain. This enables the model to make a more reasonable estimate of trace element content based on reliable core data when facing complex zonal structures such as abrupt compositional interfaces, thereby improving the consistency and coordination of the overall analysis results across different mineral growth domains.
[0015] 2. This invention integrates crystal chemical features, microstructure image features and information on associated mineral assemblages from multiple sources to construct a comprehensive feature space. This approach enables the model to comprehensively utilize multiple geological clues related to the occurrence state of trace elements, rather than relying solely on the content of major elements, thereby incorporating richer geological constraints during the inversion process and enhancing the geochemical rationality of the model's inferences.
[0016] 3. This invention, while predicting the content of trace elements, simultaneously outputs the probability distribution of their potential causal control. This design extends traditional component analysis to the level of causal analysis, providing geologists with coherent information from data to interpretation. The results not only include quantitative concentration values, but also preliminary judgments on the possible geological meanings of these values, which helps to reduce the threshold and ambiguity of subsequent professional interpretation.
[0017] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description
[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This is a schematic flowchart of a mineral trace element analysis method based on machine learning according to the present invention. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] Please see Figure 1As shown, this invention is a mineral trace element analysis method based on machine learning, comprising the following steps: S1. Simultaneously acquire major element surface distribution data, backscattered electron images, and trace element line scan data passing through the target garnet particles and their associated minerals, and spatially register all data. S2. Based on the data quality factor and spatial location information of the line scan data, it is divided into a high-quality core source domain data point set and a low-quality or ring target domain data point set located at the compositional abrupt change interface; and a spatial correlation matrix is constructed between the ring target domain data points and the core source domain data points. The spatial correlation matrix is calculated by weighting based on the major element chemical distance between the two points and the estimated relative differences in physical conditions. S3. For each data point, calculate the crystal chemical characteristics based on its major element data, extract local texture features based on backscattered electron images, and extract the surrounding associated mineral assemblage features based on mineral identification results. Then, fuse the three types of features to form a comprehensive feature vector. S4. Construct a deep neural network model with comprehensive feature vectors as input and trace element content as output; design a total loss function including main prediction loss term, dynamic domain adaptation loss term, and spatial smoothing constraint loss term; train the deep neural network model using the core source domain data point set and its true trace element values, with the goal of minimizing the total loss function; wherein, the dynamic domain adaptation loss term uses the spatial correlation matrix to guide the model to conditionally align the feature distribution of the target domain data points to the feature distribution of the associated core source domain data points; S5. Input the comprehensive feature vector of the target domain data points of the ring zone into the trained deep neural network model to obtain the predicted value of its trace element content; at the same time, output the probability distribution of the causal control factors of trace element content at each point through the auxiliary classification module of the model. S6. Integrate the measured trace element data of the source domain data points of the ore core with the predicted trace element data of the target domain data points of the ring zone to generate a complete spatial distribution map of trace elements inside the target garnet grains and a zonal map of the genetic control factors.
[0022] In step S2, the data quality factor comprehensively considers signal strength, internal standard stability, isotope measurement error, and spatial location information of erosion pits; in the weighted calculation of the spatial correlation matrix, the relative differences in physical conditions include relative temperature difference and relative pressure difference estimated based on the symbiotic mineral assemblage and major element data.
[0023] In step S2, the spatial correlation matrix is specifically used to represent the correlation strength between each ring target domain data point and all core source domain data points. The correlation strength is negatively correlated with the principal element Euclidean distance between the two points and negatively correlated with the absolute value of the relative difference in physical conditions.
[0024] In step S3, the crystal chemical features include the mole fraction of garnet endmember components and crystal chemical parameters; the local texture features are obtained by extracting backscattered electron image blocks centered on the point through a pre-trained convolutional neural network; and the symbiotic mineral assemblage features are represented by the types and relative abundance of symbiotic minerals appearing within a preset range around the point.
[0025] In step S4, the total loss function consists of the following parts: the error loss term between the predicted value and the true value calculated on the data points in the core source domain; the dynamic domain adaptation loss term, which is calculated based on the spatial correlation matrix and used to reduce the difference in feature distribution between the core source domain and the ring target domain; and the smoothing constraint loss term, which encourages the predicted values of spatially adjacent and similar data points to change smoothly.
[0026] The dynamic domain adaptation loss term is implemented by calculating the difference measure between the target domain data points of the ring zone and the source domain data points of each ore core in the hidden layer feature distribution of the model, and using the corresponding correlation strength value in the spatial correlation matrix to perform a weighted sum of the difference measure, which is used as a constraint in model training.
[0027] In step S5, the auxiliary classification module shares the feature extraction layer with the deep neural network model, and during the training process, it uses some causal source domain data of the ore core that has been labeled to conduct supervised training, so as to synchronously output the causal control probability of the predicted value of trace elements.
[0028] In step S6, the predicted values of trace element content and dominant causal categories of discrete data points are converted into continuous spatial distribution maps using a spatial interpolation algorithm, and then overlaid and displayed with the surface distribution map of major elements and the backscattered electron image.
[0029] One specific analysis method in this embodiment is as follows: Step S100: Synchronous Acquisition and Preprocessing of Multimodal Mineralogical Data An electron probe microanalysis system equipped with a backscattered electron detector was used to perform surface scanning analysis on target garnet grains in polished rock thin sections. S110: Set the analysis step size (e.g., 2 micrometers) and collect a quantitative matrix of major element (SiO2, TiO2, Al2O3, FeO, MnO, MgO, CaO, Na2O, K2O, etc.) concentration surface distribution data of the target garnet grains and their adjacent associated minerals (e.g., biotite, amphibole, clinopyroxene, etc.). ,in The total number of pixels scanned on the surface. The number of major element types.
[0030] S120: Simultaneously acquires grayscale images of backscattered electrons (BSE) in the same region. By utilizing image segmentation algorithms (such as the threshold-based watershed algorithm) and combining principal element data, the contour mask of the target garnet grains is accurately extracted. The process involves identifying relatively homogeneous regions within the garnet grain (preliminarily divided into core, mantle, and edge regions). S130: Trace element analysis is performed on adjacent regions of the same sample or on the same particle in the same thin section that has undergone carbon coating treatment, using laser ablation inductively coupled plasma mass spectrometry (LA-ICP-MS). First, line scan ablation is performed along one or more pre-designed trajectories traversing the core to the edge of the garnet grain. Time-series data containing the intensities of dozens of trace elements, including Y, HREE, Sc, Zr, Sr, and Ba, are collected for each trajectory. By comparing the data with standard materials, the intensities are converted into elemental concentrations to obtain the line scan data. ,in This represents the total number of data points in the line scan. This represents the number of trace element types. Simultaneously, the spatial coordinates corresponding to each data point are recorded and registered onto the principal element surface distribution map in step S110.
[0031] Step S200: Data quality assessment and construction of core ring-coupled data pairs This step is a key data preparation step in this invention, designed to identify high-quality source domain data and target domain data to be predicted. S210: For LA-ICP-MS line scan data Each data point in Calculate its data quality factor This factor integrates signal intensity, internal standard element stability, isotope ratio error, and the presence of erosion pits on the garnet mask. The location information is determined by judging whether the center coordinates of the erosion pit are located within the compositionally homogeneous region identified in step S120. For areas located at compositional abrupt transition interfaces... Within the micrometer range ( The position score is calculated for data points that are 1.5 times the radius of the laser beam spot (typically 10-15 micrometers). Significantly reduced S220: Set quality threshold .Will. and High-value line scan data points are defined as the "core" high-quality data point set. .in The distribution map of the registered EPMA surface The extracted principal element concentration vector corresponding to the spatial location. for The corresponding trace element concentration vector.
[0032] S230: Define the data points that satisfy any of the following conditions as the ring zone target data point set. :a)
[0033] b) It is a low value (i.e., located at or adjacent to the component mutation interface).
[0034] At the same time, for Each point in Similarly, extract the principal element concentration vector at its corresponding spatial location. However, its trace element concentration vector It is considered incomplete or unreliable as the target to be predicted by the model.
[0035] S240: Constructing the spatial correlation matrix between ore core and circumferential zone For each ring-shaped target point Calculate its relationship with all core source points Comprehensive distance The distance is defined as: In the formula, The Euclidean distance between the principal element concentration vectors; To utilize garnet thermobarometers (such as Fe-Mg exchange thermometers) and And the absolute value of the relative temperature difference estimated from the composition of the symbiotic biotite / amphibole; This represents the absolute value of the relative pressure difference estimated using a garnet thermobarometer. These are all preset weighting coefficients used to balance the contributions of chemical distance and physical condition distance; based on distance The incidence matrix is calculated using a kernel density function (such as a Gaussian kernel). elements , The larger the value, the higher the core point. For the ring point The higher the reference value, the better. Step S300: Construction of multi-source feature space. S310: For each data point (including...) and, All points in the vector, based on their principal element vectors Calculate a set of crystal chemical characteristics Including but not limited to: mole fraction of end-member components (pyrope, almandine, spessartine, grossular), total cation charge, average ionic radius, etc.
[0036] S320: Centered on this data point, from the BSE image Extract one A neighborhood image patch of a pixel. Features are extracted from this image patch using a pre-trained shallow convolutional neural network (such as the first few layers of VGG), resulting in a fixed-dimensional texture feature vector. It is used to characterize microstructural information such as inclusions, fissures, and compositional oscillations within micro-regions.
[0037] S330: Extract the features of the symbiotic mineral assemblage within a certain radius (e.g., 100 micrometers) around the data point from the mineral identification results at the entire thin section scale. , which represents the unique thermal code or continuous probability vector of the presence and relative area of each associated mineral (such as clinopyroxene, rutile, apatite, polysilicon, etc.).
[0038] S340: Combine the above three types of features to form the comprehensive feature vector of this data point. .for The points in the vector are labeled with their corresponding infinitesimal element vectors. .
[0039] Step S400: Construction and Training of Progressive Domain Adaptive Transfer Learning Model This invention proposes a novel loss function-driven transfer learning framework. S410: Construct a framework based on the comprehensive feature vector... As input, to predict the vector of trace elements The deep neural network model for output ,in These are the network parameters. The network output layer uses the ReLU activation function to ensure that the predicted concentration is non-negative.
[0040] S410: During the training phase, design a dual-constraint progressive loss function. :
[0041] In the formula, This is the total loss function value (scalar) that needs to be minimized during model training. The value of the primary forecast loss term (scalar); The value of the dynamic domain adaptation loss term (scalar); The value of the space-component smoothing constraint loss term (scalar); These are hyperparameter scalars, used to control the relative importance of domain adaptation loss and smoothing constraint loss in the total loss, respectively; S421: Master Prediction Loss In high-quality source domain data Above, the mean squared error loss between the model's predicted values and the actual values is calculated:
[0042] In the formula, To achieve average operation; This refers to the number of core points. For the first The comprehensive feature vector of each mineral core data point; For the model For input The predicted output is the predicted vector of trace elements. For the first The actual trace element vector of each mineral core data point; S422: The square of the L2 norm of the difference vector between the predicted and true values, i.e., the mean squared error; S422: Dynamic domain adaptation loss The aim is to minimize the conditional distribution differences between the core source domain and the annular target domain in the hidden feature space of the model, while utilizing the spatial correlation matrix constructed in step S240. Provide guidance. Specifically, set up... and They are the core points and circumferential point The feature vector output by a hidden layer (such as the penultimate layer) of the model. Definition for:
[0043] In the formula, For the looped points in the correlation matrix With core point The strength of the association; It is a probability distribution; , To represent the core points respectively and ring point The feature vector output by a certain intermediate hidden layer of a neural network; Given a principal element vector Under these conditions, the hidden layer characteristics of the ore core point Conditional distribution; Given a principal element vector Under these conditions, the hidden layer features of the annular zone Conditional distribution; This represents the maximum mean difference, used to measure the difference between two conditional distributions. and The differences between them. (Here, through...) Weighted, forced model learning makes the connection with the ring zone point The ore core point that is most similar in physicochemical and spatial terms The characteristic distribution of points The feature distributions themselves are aligned as closely as possible. This achieves gradual and conditional domain adaptation, rather than a globally forced alignment. S423: Spatial-Composition Smoothing Constraint Loss To leverage prior knowledge of the spatial continuity of garnet compositional zoning, a smoothing constraint is introduced. This constraint applies to adjacent (4-neighbor or 8-neighbor) garnet masks within the EPMA surface scan space. Any two points within and (regardless of its belonging) or The model encourages that the predicted changes in trace element concentrations are proportional to the changes in major element concentrations, with the proportionality coefficient determined by the feature distance between the two: In the formula, These are the indices of two spatially adjacent arbitrary data points; Points and points The comprehensive feature vector; These are model pairs of points. and points The microelement prediction vector; The squared L2 norm of the difference vector between two predicted values; For point and points The principal element concentration vector; Let L2 norm be the square of the vector of principal element concentration differences between two points; It is a very small positive scalar (such as 1e-8) used to prevent the denominator from being zero and to enhance numerical stability; It is a natural exponential function; It is the L2 norm of the difference between the combined eigenvectors of the two points; It is the set of all adjacent pairs of points; Prevent division by zero for small positive numbers. S430: Minimize the total loss using a gradient descent algorithm (such as the Adam optimizer). Iteratively update model parameters Hyperparameters and Through a reserved verification set (from The grid search is used to determine the division of the area.
[0044] Step S500: Intelligent Inversion of Trace Elements in the Ring Zone and Output of Genetic Probability S510: The material constructed in step S300, corresponding to... The comprehensive feature vector of all target points Input into the trained model In this process, the predicted values of its trace element concentrations are obtained. S520: Model Design - A Parallel, Lightweight Auxiliary Classification Head ,and They share most of the feature extraction layers. This classification head uses the same hidden layer features. As input, output a probability distribution vector of the causal control factors. ,in The number of predefined causal categories (e.g., C=4, corresponding to equilibrium distribution dominance, inclusion dissolution dominance, diffusion reequilibrium dominance, and external fluid permeation dominance, respectively).
[0045] S530: Auxiliary Sorting Head During the training phase, utilize Supervised training is conducted on data points with clear trace element patterns that can be manually or semi-automatically labeled based on geochemical discriminant maps (such as Y vs. HREE models, Sc / Y ratios, etc.), using the cross-entropy loss function. During the inference phase, for each zonal target point... The model not only outputs predicted values of trace elements It also outputs its causal probability vector simultaneously. This provides direct evidence for geological interpretation.
[0046] Step S600: High-resolution 3D inversion map generation and integration S610: The inverted... The predicted results of the data points (trace element concentration and causal probability) and The measured results of the data points are integrated.
[0047] S620: Based on all data points ( The spatial coordinates of the target garnet grain and its corresponding complete trace element data (measured or predicted) are used to generate a complete, high spatial resolution spectrum of the concentration distribution of various trace elements inside the grain using spatial interpolation algorithms (such as Kriging interpolation).
[0048] S630: Similarly, for the causal probability vector of each data point Spatial interpolation is performed on the dominant genetic categories (the categories corresponding to the highest probability values) to generate a zoning map of the genetic control factors within garnet grains.
[0049] S640: Combine the generated trace element distribution map and genetic zoning map with the original principal element surface distribution map. BSE images Spatial overlay and visualization output are performed to form a multi-layered, interactive digital core archive for genetic interpretation, which can be used for subsequent quantitative simulation of metamorphic processes.
[0050] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0051] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.
Claims
1. A mineral trace element analysis method based on machine learning, characterized in that, The analytical method includes the following steps: S100: Simultaneously acquire major element surface distribution data, backscattered electron images, and trace element line scan data passing through the target garnet particles and their associated minerals, and spatially register all data. S200. Based on the data quality factor and spatial location information of the line scan data, it is divided into a high-quality core source domain data point set and a low-quality or ring target domain data point set located at the compositional abrupt change interface; and a spatial correlation matrix is constructed between the ring target domain data points and the core source domain data points. The spatial correlation matrix is calculated by weighting based on the major element chemical distance between the two points and the estimated relative differences in physical conditions. S300. For each data point, calculate the crystal chemical characteristics based on its major element data, extract local texture features based on the backscattered electron image, and extract the surrounding associated mineral assemblage features based on the mineral identification results. The three types of features are then fused to form a comprehensive feature vector. S400. Construct a deep neural network model with the comprehensive feature vector as input and trace element content as output; design a total loss function including a main prediction loss term, a dynamic domain adaptation loss term, and a spatial smoothing constraint loss term; train the deep neural network model using the data point set of the ore core source domain and its true trace element values, with the goal of minimizing the total loss function; wherein, the dynamic domain adaptation loss term uses the spatial correlation matrix to guide the model to conditionally align the feature distribution of the target domain data points to the feature distribution of the associated ore core source domain data points; S500: Input the comprehensive feature vector of the target domain data points of the ring zone into the trained deep neural network model to obtain the predicted value of its trace element content; at the same time, output the probability distribution of the causal control factors of the trace element content at each point through the auxiliary classification module of the model. S600. Integrate the measured trace element data of the source domain data points of the ore core with the predicted trace element data of the target domain data points of the ring zone to generate a complete spatial distribution map of trace elements inside the target garnet particles and a zonal map of genetic control factors.
2. The mineral trace element analysis method based on machine learning according to claim 1, characterized in that, In step S2, the data quality factor comprehensively considers signal strength, internal standard stability, isotope measurement error, and spatial location information of erosion pits; in the weighted calculation of the spatial correlation matrix, the relative differences in physical conditions include relative temperature difference and relative pressure difference estimated based on symbiotic mineral assemblage and major element data.
3. The mineral trace element analysis method based on machine learning according to claim 1, characterized in that, In step S2, the spatial correlation matrix is specifically used to represent the correlation strength between each ring target domain data point and all core source domain data points. The correlation strength is negatively correlated with the principal element Euclidean distance between the two points and negatively correlated with the absolute value of the relative difference of the physical conditions.
4. The mineral trace element analysis method based on machine learning according to claim 1, characterized in that, In step S3, the crystal chemical features include the mole fraction of garnet endmember components and crystal chemical parameters; the local texture features are obtained by extracting backscattered electron image blocks centered on the point through a pre-trained convolutional neural network; and the symbiotic mineral assemblage features are represented by the types and relative abundance of symbiotic minerals appearing within a preset range around the point.
5. The mineral trace element analysis method based on machine learning according to claim 1, characterized in that, In step S4, the total loss function consists of the following parts: an error loss term between the predicted value and the true value calculated on the data points in the core source domain; a dynamic domain adaptation loss term calculated based on the spatial correlation matrix to reduce the difference in feature distribution between the core source domain and the ring target domain; and a smoothing constraint loss term to encourage the predicted values of spatially adjacent and similar data points to change smoothly.
6. The mineral trace element analysis method based on machine learning according to claim 5, characterized in that, The dynamic domain adaptation loss term is implemented by calculating the difference measure between the target domain data points of the ring zone and the source domain data points of each ore core in the hidden layer feature distribution of the model, and using the corresponding correlation strength value in the spatial correlation matrix to perform a weighted summation of the difference measure, which is used as a constraint in model training.
7. The mineral trace element analysis method based on machine learning according to claim 1, characterized in that, In step S5, the auxiliary classification module shares a feature extraction layer with the deep neural network model, and during the training process, it uses some causal source domain data of the ore core that has been labeled with origin to perform supervised training, so as to synchronously output the causal control probability of the predicted value of trace elements.
8. The mineral trace element analysis method based on machine learning according to claim 1, characterized in that, In step S6, the predicted values of trace element content and dominant causal categories of discrete data points are converted into a continuous spatial distribution map using a spatial interpolation algorithm, and then overlaid with the surface distribution map of major elements and the backscattered electron image for display.