Volcanic rock source inversion method and system based on machine learning and evolutionary manifold

Abstract

The application discloses a volcanic rock source area inversion method and system based on machine learning and evolutionary manifold, relates to the technical field of geochemical data inversion, and provides the following scheme, which comprises the following steps: acquiring the incompatible element content, compatible element content and strontium-neodymium isotope ratio of a volcanic rock sample, aligning the measured geochemical vectors, determining candidate source area end members and candidate assimilation end members from the measured geochemical vectors, and setting a crystallization fraction and an assimilation-crystallization ratio. The application reduces the participation weight of samples deviating from the evolutionary manifold by using vertical residuals, and realizes the anti-crosstalk constraint from the coordinate layer to the parameter update layer by orthogonalizing the direction angle between the crystallization fraction residual and the assimilation-crystallization ratio residual in the local tangent space of the manifold and then returning to the two branches, so that the source area end members covered by assimilation and mixing are no longer systematically deviated from the two processes.

Description

Technical Field

[0001] This invention relates to the field of geochemical data inversion technology, and more specifically, this application relates to a method and system for inverting volcanic source regions based on machine learning and evolutionary manifolds. Background Technology

[0002] In tectonic settings such as orogenic belts, primary magma often undergoes both fractional crystallization and assimilation and contamination of crustal materials during its ascent, stagnation, and pre-eruption evolution, existing in an open system. Geological research requires inversely elucidating the source region endmember composition of magma based on the incompatible element content, compatible element content, and strontium-neodymium isotope ratio of volcanic rock samples, and distinguishing the strength of each of the two processes—fractional crystallization and assimilation and contamination—to support the determination of magma genesis and deep dynamics. Such inversion requires providing a stable estimate of the intensity of processes that do not mix with each other, even when both processes coexist and the degree of evolution varies among different samples.

[0003] Existing techniques often treat samples as linear mixing points of several endmembers, obtaining source region interpretations through graphical point projection, endmember decomposition, or forward modeling with pre-defined endmembers. Some methods directly train models using observations to regress causal types or endmember parameters. These methods can provide basic results under closed systems, single master-controlled processes, or approximately linear mixing conditions, provided that the two types of processes can be processed independently or combined into a single fit. However, separation crystallization mainly changes the relative content of elements without directly altering isotope ratios, while assimilation and contamination introduce foreign elements while changing isotope ratios. Therefore, the response strength of different components to the two types of processes is inherently inconsistent, and existing treatments typically treat each component as a similar input.

[0004] Therefore, the same observed change may be superimposed with contributions from two types of processes. Existing methods either attribute the entire change to a single process or combine them into a single fitting, but neither method can separate the respective proportions of the two types of processes in the same observation. The contribution of one type of process is easily miscounted as that of the other, causing the intensity of the two types of processes to crosstalk each other and continuously propagate during fitting or model training, resulting in systematic distortion of the source region endmembers that are obscured by the contamination. In addition, the degree to which different samples are affected by the two types of processes varies greatly. A few samples that deviate from the evolutionary system will pull the overall estimate off during unified processing, ultimately leading to inaccurate decomposition of the genetic process of volcanic rock samples, distortion of source region endmember judgment, and deviations in subsequent interpretation of magmatic genesis. Summary of the Invention

[0005] To address the aforementioned technical problems, this paper provides a method and system for inverting volcanic source regions based on machine learning and evolutionary manifolds. This technical solution solves the problems mentioned in the background section.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, this application provides a method for inverting volcanic source regions based on machine learning and evolutionary manifolds, the method comprising:

[0008] Obtain the incompatible element content, compatible element content, and strontium-neodymium isotope ratio of volcanic rock samples, and align them to the measured geochemical vector;

[0009] Candidate source region endmembers and candidate assimilation endmembers are determined by measured geochemical vectors. Crystallization fraction and assimilation-crystallization ratio are set. The combination of these two values ​​is substituted into the assimilation-separation crystallization mass balance relationship to obtain theoretical vectors, which constitute an evolutionary manifold with these two as coordinates.

[0010] The sensitivity of each component to changes in crystallization fraction and assimilation crystallization ratio is extracted on the evolutionary manifold in two directions, and the two are normalized into two attribution weights.

[0011] A mutual adjudication inversion network containing attribution units, crystallization fraction branches, and assimilation crystallization ratio branches is constructed. The attribution units decompose the offset of the measured geochemical vector relative to the candidate source region endmembers into crystallization fraction offset, assimilation crystallization ratio offset, and vertical offset according to the two attribution weights.

[0012] The two branches learn and output two coordinates based on the crystallization fraction offset and the assimilation crystallization ratio offset, respectively. The assimilation crystallization ratio coordinate is substituted into the mass balance relationship to obtain the assimilation contribution of incompatible elements. The crystallization fraction coordinate is then corrected after subtracting the contribution from the incompatible element offset.

[0013] Substituting the corrected two coordinates into the mass balance relationship yields the predicted vector. The difference between the predicted vector and the measured geochemical vector is decomposed into crystallization fraction residual, assimilation crystallization ratio residual, and vertical residual according to the sensitivity of the two directions and their remaining components.

[0014] The two branches are called back respectively using the crystallization fraction residual and the assimilation crystallization ratio residual. The participation weight of samples exceeding the preset vertical residual threshold is reduced by the vertical residual. The iteration continues until the residual change satisfies the preset convergence condition.

[0015] The crystallization fraction shift and assimilation crystallization ratio shift after convergence are used as the separation crystallization component and the contamination component, respectively. The candidate source region endmembers are inverted using the two coordinates after convergence, and the source region endmember composition and isotopic coordinates are output.

[0016] Secondly, this application provides a volcanic source region inversion system based on machine learning and evolutionary manifolds, for implementing the aforementioned volcanic source region inversion method based on machine learning and evolutionary manifolds, including:

[0017] The vector generation module is used to obtain the incompatible element content, compatible element content and strontium-neodymium isotope ratio of volcanic rock samples, and align them with the measured geochemical vectors.

[0018] The manifold construction module is used to determine candidate source region endmembers and candidate assimilation endmembers from measured geochemical vectors, set the crystallization fraction and assimilation-crystallization ratio, and substitute the values ​​of the two into the assimilation-separation crystallization mass balance relationship to obtain the theoretical vector, thus constructing an evolutionary manifold with the two as coordinates.

[0019] The weight generation module is used to extract the two-directional sensitivity of each component to changes in crystallization fraction and assimilation crystallization ratio on the evolutionary manifold, and normalize the two into two attribution weights.

[0020] The offset splitting module is used to construct a mutual adjudication inversion network containing attribution units, crystallization fraction branches, and assimilation crystallization ratio branches. The attribution units split the offset of the measured geochemical vector relative to the candidate source region endmembers into crystallization fraction offset, assimilation crystallization ratio offset, and vertical offset according to the two attribution weights.

[0021] The coordinate correction module is used to learn and output two coordinates from the two branches based on the crystallization fraction offset and the assimilation crystallization ratio offset, respectively. The assimilation crystallization ratio coordinate is substituted into the mass balance relationship to obtain the assimilation contribution of incompatible elements, and the crystallization fraction coordinate is corrected after subtracting the contribution from the incompatible element offset.

[0022] The residual decomposition module is used to substitute the corrected two coordinates into the mass balance relationship to obtain the predicted vector, and decompose the difference between the predicted vector and the measured geochemical vector into crystallization fraction residual, assimilation crystallization ratio residual and vertical residual according to the sensitivity of the two directions and their remaining components.

[0023] The iterative callback module is used to call back the two branches with the crystallization fraction residual and the assimilation crystallization ratio residual respectively, and reduce the participation weight of samples exceeding the preset vertical residual threshold with the vertical residual, and iterate until the residual change meets the preset convergence condition.

[0024] The endmember output module is used to use the crystallization fraction offset and assimilation crystallization ratio offset after convergence as the separation crystallization component and the contamination component, respectively, to invert the candidate source region endmembers using the two coordinates after convergence, and output the source region endmember composition and isotopic coordinates.

[0025] Thirdly, this application provides a computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described method for inverting volcanic source regions based on machine learning and evolutionary manifolds.

[0026] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for inverting volcanic source regions based on machine learning and evolutionary manifolds.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] This application uses the evolved manifold obtained by forward modeling the assimilation-separation crystallization quality balance relationship as the inversion coordinate basis, and normalizes the crystallization fraction sensitivity and assimilation-crystallization ratio sensitivity of each component at the manifold position into two assignment weights to split the offset. This solves the problem in the prior art that treats each component as the same type of input, resulting in components with different response strengths being unable to indicate their respective processes. It realizes that the component assignment criterion is reconstructed from a uniform preset of the whole sample to a dynamic boundary generated point by point with the manifold position, so that the crystallization fraction offset, assimilation-crystallization ratio offset and vertical offset do not overlap with each other at the beginning of the split.

[0029] This application solves the problem of crosstalk in fitting by setting a gated correction channel, using the ratio of assimilation crystallization ratio sensitivity to crystallization fraction sensitivity at the manifold position as the gate coefficient, scaling the assimilation contribution according to the element, deducting it from the incompatible element offset, recalculating the crystallization fraction coordinate, and writing back to adjust the isotope mixing weight. This solves the problem of the contribution of two processes superimposed on the same observation, and the fact that one class is easily mistakenly included in the other class, resulting in continuous crosstalk in fitting. It realizes the reconstruction of the crystallization fraction coordinate from the directly output result quantity to the verified and reduced intermediate quantity, so that the crystallization degree on the element side and the contamination degree on the isotope side can be mutually determined in one forward calculation.

[0030] This application reduces the participation weight of samples deviating from the evolutionary manifold by using vertical residuals, and orthogonalsizes the two branches in the local tangent space of the manifold according to the directional angle between the crystallization fraction residual and the assimilation crystallization ratio residual before reverting to the two branches. This solves the problem of samples deviating from the system being biased in the unified processing of the overall estimate and the two processes being crosstalked again at the gradient level. It realizes that the anti-crosstalk constraint extends from the coordinate layer to the parameter update layer, so that the source region endmembers that are masked by assimilation and contamination no longer systematically shift with the coupling of the two processes. Attached Figure Description

[0031] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein:

[0032] Figure 1 This is a flowchart of the volcanic source region inversion method based on machine learning and evolutionary manifold proposed in this invention;

[0033] Figure 2 This is a structural block diagram of the volcanic source region inversion system based on machine learning and evolutionary manifold proposed in this invention. Detailed Implementation

[0034] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0035] For ease of understanding, the following explanation uses a Mesozoic intermediate-acidic volcanic rock sample set from an orogenic belt as a consistent example. This sample set comprises 40 whole-rock samples, all from the same magma evolution sequence. After rock crushing and acid digestion, trace element content was measured by inductively coupled plasma mass spectrometry (ICP-MS), and isotope ratios were measured by thermal ionization mass spectrometry (TIMS). One representative sample, designated as sample S, was selected. Its incompatible element content was Th = 12.0 ppm, Nb = 15.0 ppm, La = 40.0 ppm, and Zr = 180.0 ppm, while its compatible element content was Cr = 50.0 ppm and Ni = 30.0 ppm. The Sr-Nd isotope ratio was... =0.7060, εNd=−4.0, and the intermediate results and final outputs in each step below are based on these data.

[0036] Reference Figure 1 As shown, this application proposes a method for inverting volcanic source regions based on machine learning and evolutionary manifolds;

[0037] In step S1: Obtain the incompatible element content, compatible element content and strontium-neodymium isotope ratio of the volcanic rock sample, and align them to the measured geochemical vector;

[0038] It should be noted that the contents of incompatible and compatible elements are derived from the mass fraction fields (unit: ppm) of each element in the ICP-MS analysis report, and the strontium-neodymium isotope ratios are derived from the TIMS analysis report. Fields and by The converted εNd field; the alignment process is as follows: using the sample number as the primary key, concatenate the above three types of fields of the same sample into a fixed-length numerical vector. Each component The components are arranged in a fixed order of "incompatible element segment - compatible element segment - isotope segment" to ensure that the physical meaning of each component in the full sample set corresponds one-to-one. Since the element content (ppm) and the isotope ratio (dimensionless) have different dimensions, the element content segment is standardized with zero mean and unit variance by taking the logarithm to the base 10, and the isotope segment is standardized with zero mean and unit variance separately to make each component fall into a comparable scale. The mean and standard deviation used for standardization are statistically based on the full sample set and are fixed with the sample set.

[0039] If a sample contains individual elements below the detection limit, fill the missing element with half the detection limit and mark it as missing. Subsequent sensitivity calculations will then deweight the marked element. Using sample S, after alignment... The incompatible element range was obtained by normalizing (12.0, 15.0, 40.0, 180.0) ppm as described above, and the isotope range was obtained by normalizing (0.7060, −4.0).

[0040] In step S2: candidate source region endmembers and candidate assimilation endmembers are determined by measured geochemical vectors, crystallization fraction and assimilation-crystallization ratio are set, and the combination of the two values ​​is substituted into the assimilation-separation crystallization mass balance relationship to obtain theoretical vectors, forming an evolutionary manifold with the two as coordinates;

[0041] It should be noted that candidate source region endmembers Taken from the most depleted and most primitive end of the sample set in the isotope-incompatible element space ( (Lowest, εNd highest), candidate assimilation endmembers are taken from the most enriched end of the published surrounding rock crustal components or sample sets in the region; in this example, the candidate source region endmember is taken as... =0.7035, εNd=+6.0, Th=2.0ppm, and candidate assimilation endmembers are selected as follows. =0.7200, εNd=−12.0, Th=15.0ppm; the crystallization fraction F is the mass fraction of residual magma to initial magma, with a value range of (0,1]; the assimilation-crystallization ratio r is the ratio of the mass of assimilated surrounding rock to the mass of crystallized precipitation per unit time, with a value range of 0 to 0.9.

[0042] The assimilation-separation crystallization mass equilibrium relationship adopts the DePaolo-type open system evolution relationship: for the i-th component, the theoretical content With F, r and the total distribution coefficient of the component Evolution, incompatible elements take (like ), compatible elements (like Isotope ratios evolved by mass-weighted evolution of assimilation endmembers and residual magma;

[0043] Among them, the total allocation coefficient This refers to the ratio of the concentration of the i-th component in the mineral aggregate to its concentration in the melt when equilibrium is reached between the crystallized mineral aggregate and the residual magma melt. It is used to characterize the tendency of this component to be carried away by the solid phase during crystallization. This indicates that the component tends to accumulate in the residual melt, meaning it is incompatible; This indicates that the component tends to enter crystalline minerals, i.e., it is compatible; denoted as the total partition coefficient of thorium. Thorium is strongly incompatible in common silicate minerals, and its value is much less than 1; in this example, it is taken as 0.05. The total distribution coefficient for chromium is given. Chromium readily enters early-crystallized minerals such as spinel and pyroxene, and its value is greater than 1. In this example, it is taken as 3.0. The above values ​​are representative values ​​set in this example. The actual values ​​depend on the assemblage of crystallized minerals and the composition of magma. They are determined by the mineral-melt distribution coefficient library described later, weighted according to the estimated crystallized mineral assemblage.

[0044] The process of constructing the evolutionary manifold is as follows: Let F take values ​​in groups of 0.02 steps within (0,1] and r take values ​​in groups of 0.05 steps within [0,0.9]. Substitute each group (F, r) into the above relationship to calculate a theoretical vector. All theoretical vectors form a two-dimensional surface in the geochemical vector space with F and r as coordinates, i.e., the evolutionary manifold;

[0045] Allocation coefficient The results were obtained by weighting the mineral-melt partition coefficient library according to the estimated crystallization mineral assemblage. Each mineral-element pair in the library contains at least three sets of experimental values. Missing values ​​are approximated by elements in the same group and marked with low confidence.

[0046] To further clarify, the mineral-melt partition coefficient library refers to a pre-established dataset used to retrieve single mineral-melt partition coefficients by mineral type and element type. The data sources are publicly published experimental petrological measurements and natural mineral-melt pair analyses. The library is organized using "mineral type-element type" as the primary key. Each record contains at least the following fields: mineral type (e.g., olivine, clinopyroxene, plagioclase, spinel, etc.), element type, single mineral-melt partition coefficient value, and the applicable temperature, pressure, and melt composition range for that value. For the same mineral-element pair, multiple sets of values ​​from different literature or experimental conditions are included, with at least three sets. The representative value (e.g., median) within the applicable temperature and pressure range is taken as the single mineral-melt partition coefficient for that mineral-element pair, and its dispersion is recorded as the confidence level.

[0047] The total distribution coefficient is determined from the mineral-melt distribution coefficient library. The process is as follows: First, based on the magma evolution stage and lithofacies observation, estimate the crystalline mineral assemblage and the mass crystallization ratio of each mineral. Then, for the i-th component, retrieve its single-mineral-melt partition coefficient in each participating crystallizing mineral from the database, and sum them according to the mass crystallization ratio of each mineral to obtain the total partition coefficient of that component. ;

[0048] The weighted calculation expression for the total allocation coefficient is:

[0049] ;

[0050] in, Let i be the total allocation coefficient of the i-th component. Let be the mass crystallization proportion of the j-th crystalline mineral in the crystalline mineral assemblage, and the sum of all proportions is 1. Let be the single-mineral-melt partition coefficient of the i-th component between the j-th crystalline mineral and the melt. Since the total partition coefficient is determined by the proportions of the minerals involved in crystallization, the total partition coefficient of the same component is different under different combinations of crystalline minerals. It needs to be determined separately according to the estimated mineral combination of the corresponding evolution stage of each sample.

[0051] If a component lacks a corresponding mineral-element pair value in the database, it is approximated by the single mineral-melt partition coefficient of elements in the same group or elements with similar geochemical behavior, and a low confidence marker is assigned to that component. For components with low confidence markers, the weights are reduced accordingly in the subsequent two-way sensitivity extraction and offset splitting to reduce the impact of partition coefficient uncertainty on the inversion results. The corresponding records in the database are updated when new experimental measurements or literature values ​​with a more suitable scope are obtained.

[0052] In step S3: Extract the two-directional sensitivity of each component to the changes in crystallization fraction and assimilation crystallization ratio on the evolving manifold, and normalize the two into two attribution weights.

[0053] In step S4: a mutual decision inversion network containing attribution units, crystallization fraction branches and assimilation crystallization ratio branches is constructed. The attribution units decompose the offset of the measured geochemical vector relative to the candidate source region endmembers into crystallization fraction offset, assimilation crystallization ratio offset and vertical offset according to the two attribution weights.

[0054] It should be noted that the attribution unit, crystallization fraction branch, and assimilation crystallization ratio branch are the three components of the mutual adjudication inversion network: the attribution unit receives the measured geochemical vectors and outputs three types of migrations; the crystallization fraction branch receives the crystallization fraction migrations and outputs the crystallization fraction coordinates; and the assimilation crystallization ratio branch receives the assimilation crystallization ratio migrations and outputs the assimilation crystallization ratio coordinates. A simple implementation example that can support complete operation is as follows: the attribution unit uses a projection operator obtained by looking up a table based on the manifold position; each of the two branches uses a three-layer fully connected network with 32 neurons in each hidden layer, the activation function is a modified linear unit, and the output layer is a single-node linear output.

[0055] In step S5: the two branches learn and output the crystallization fraction coordinates and the assimilation crystallization ratio coordinates according to the crystallization fraction offset and the assimilation crystallization ratio offset respectively. The assimilation crystallization ratio coordinates are substituted into the mass balance relationship to obtain the assimilation contribution of incompatible elements. The crystallization fraction coordinates are then corrected after the contribution is subtracted from the incompatible element offset.

[0056] In step S6: Substitute the corrected crystallization fraction coordinates and assimilation crystallization ratio coordinates into the mass balance relationship to obtain the predicted vector. Decompose the difference between the predicted vector and the measured geochemical vector into crystallization fraction residual, assimilation crystallization ratio residual and vertical residual according to the crystallization fraction sensitivity direction, the assimilation crystallization ratio sensitivity direction and their remaining components.

[0057] It should be noted that the process of obtaining the prediction vector is as follows: The corrected crystallization fraction coordinates are... Coordinates of assimilation crystallization ratio Substituting the mass balance relationship from step S2, the theoretical content of each component is calculated and standardized using the same standardized parameters as in step S1, then combined into a prediction vector. The difference between the predicted vector and the measured geochemical vector is denoted as the inversion bias. The inversion bias is decomposed into three scalar residuals by the crystallization fraction sensitivity direction, the assimilation crystallization ratio sensitivity direction, and the orthogonal residual direction of the space spanned by these two directions. The component along the crystallization fraction sensitivity direction is the crystallization fraction residual, the component along the assimilation crystallization ratio sensitivity direction is the assimilation crystallization ratio residual, and the magnitude in the residual direction is the vertical residual. The vertical residual characterizes the degree to which the sample deviates from the evolutionary manifold, that is, the part that cannot be explained by the mass balance relationship from any value of crystallization fraction and assimilation crystallization ratio.

[0058] In step S7: the crystallization fraction residual and the assimilation crystallization ratio residual are used to call back the two branches respectively, and the participation weight of samples exceeding the preset vertical residual threshold is reduced by the vertical residual, and the iteration continues until the residual change satisfies the preset convergence condition.

[0059] It should be noted that the participation weight refers to the weight of the residual term of each sample when updating the parameters of the two branches, with a value range of [0, 1] and an initial value of 1; the reduction process is as follows: when the vertical residual of a sample is greater than the preset vertical residual threshold. When the vertical residual exceeds the threshold, the participation weight of the sample is reduced in a monotonically decreasing manner. The larger the vertical residual, the lower the participation weight, until it approaches 0, so that samples that deviate significantly from the evolutionary manifold do not dominate the update direction of the two branches.

[0060] Since the weight reduction is based on the impact of the deviation samples on parameter updates, rather than the vertical residuals themselves, the convergence direction is consistent with the goal of suppressing the interference of deviation samples. The preset convergence condition is: the relative decrease of the sum of squares of the residuals of the whole sample is less than 0.5% in 5 consecutive iterations, which is considered as convergence, or the maximum number of iterations (typically 200) is reached and then stopped.

[0061] In step S8: the crystallization fraction shift and assimilation crystallization ratio shift after convergence are used as the separation crystallization component and the contamination component, respectively. The candidate source region endmembers are inverted using the crystallization fraction coordinates and assimilation crystallization ratio coordinates after convergence, and the source region endmember composition and isotope coordinates are output.

[0062] It should be noted that the process of inverting candidate source region endmembers is as follows: using the crystallization fraction coordinates and assimilation-crystallization ratio coordinates of each sample after convergence as known quantities, the mass balance relationship is reversed to solve for the initial magma endmember composition that minimizes the overall deviation between the predicted vector and the measured geochemical vector of the entire sample, which is then used as the source region endmember. The output results are structured as follows: the elemental content (ppm) of each source region endmember, the source region endmember's... And εNd, as well as the separated crystallization component and contamination component for each sample. For example, one output of sample S is: source region endmembers. =0.7038, εNd=+5.6, Th=2.1ppm, the separated crystallization components of sample S correspond to =0.62, corresponding to the mixed-color component =0.28.

[0063] Through the above technical solution, this embodiment uses the evolutionary manifold obtained by forward modeling the assimilation-separation crystallization quality balance relationship as the coordinate basis for inversion, and uses the vertical residual to reduce the weight of samples that deviate from the manifold. This transfers the decoupling of the two types of processes from the overall fitting of Euclidean space to the physical evolution surface. The determination of the proportion of each type of process is organized into two non-confusing quantities: the coordinates of the sample on the evolutionary manifold and the vertical residual of the deviation from the manifold. This ensures that the source region endmembers that are masked by assimilation and contamination no longer systematically shift due to crosstalk between the two processes.

[0064] In an optional embodiment, the bidirectional sensitivity of each component to changes in crystallization fraction and assimilation crystallization ratio is extracted on the evolving manifold, specifically including:

[0065] The measured geochemical vector of each volcanic rock sample is projected onto the evolutionary manifold, and the closest corresponding point on the evolutionary manifold for that sample is taken as the manifold position of that sample.

[0066] It should be noted that the projection process is as follows: In the set of discrete grid points of the evolutionary manifold generated in step S2, the standardized Euclidean distance between the measured geochemical vector of the sample and the theoretical vector of the grid point is calculated point by point. The grid point with the smallest distance is selected as the initial corresponding point. Then, in the local interval of F and r near the initial point, the coordinates are refined by quadratic interpolation to obtain the continuous coordinates that minimize the distance. The theoretical vector corresponding to this coordinate is the manifold position;

[0067] If the minimum distance is still greater than the preset upper manifold distance limit (indicating that the sample may not belong to the current evolutionary system), then the sample is marked as an out-of-system candidate sample, and its manifold position is approximated by the nearest grid point and subsequently weighted; using sample S, its nearest corresponding point falls on =0.62、 =0.28.

[0068] Within the neighborhood of the manifold location, preset coordinate increments are applied along the crystallization fraction direction and the assimilation crystallization ratio direction, while keeping another coordinate constant. The changes in each component before and after the increments are calculated using the mass balance relationship, thus obtaining the crystallization fraction sensitivity and assimilation crystallization ratio sensitivity of each component.

[0069] To further clarify, the preset coordinate increment refers to the small change applied to the crystallization fraction or assimilation crystallization ratio to estimate the local orientation of the manifold. It is used to differentially calculate the local response of each component, and its value ranges from 0.5% to 2% of the corresponding coordinate range. Typically, the crystallization fraction increment is used. =0.01, Increment of assimilation crystallization ratio r=0.01; Since a value that is too large will cause the difference to cross sections with significant manifold curvature and become distorted, while a value that is too small will amplify numerical noise, the value is set within the above range. The calculation process is as follows: Maintain... Unchanged, crystallization fraction from Increase to Substituting the mass balance relation yields the new theoretical vector. The difference between the components of the old and new theoretical vectors is divided by... That is, the sensitivity of the crystallization fraction of the component. Similarly, maintain Unchanged, applying to the assimilation crystallization ratio Obtain the assimilation crystallization specific sensitivity Using sample S, the two sensitivities of the incompatible element Th at its manifold location are approximately , (Standardized scale), the two sensitivities of the isotope εNd are approximately , This demonstrates that separation crystallization mainly alters incompatible elements while hardly changing isotopes, while assimilation and mixing significantly alter both.

[0070] Normalize the crystallization fraction sensitivity and assimilation crystallization ratio sensitivity for the same component to obtain the two assignment weights of the component at the manifold position.

[0071] It should be noted that the normalization process is as follows: for the same component, take the absolute values ​​of its crystallization fraction sensitivity and assimilation crystal ratio sensitivity, divide each by the sum of their absolute values, and obtain the weight of the component belonging to the crystallization fraction direction. Weights associated with the direction of assimilation crystallization ratio The sum of the two is 1; when both sensitivities of a component are close to 0 (the component is insensitive to both processes at this manifold location), its two assignment weights are reset to be equal and the overall weight of the component is reduced to avoid amplifying noise by dividing by the minimum value; using the sample S, the two assignment weights of Th are approximately , The two attribution weights of εNd are approximately , .

[0072] Through the above technical solution, this embodiment obtains the local sensitivity of each component to two coordinate directions by difference at the manifold position of each sample, and generates two assignment weights per sample and per component, so that which component can better indicate which type of process is quantitatively given by the local structure of the evolutionary manifold rather than uniformly preset for the whole sample.

[0073] In an optional embodiment, the attribution unit decomposes the offset of the measured geochemical vector relative to the candidate source region endmember into crystallization fraction offset, assimilation-crystallization ratio offset, and vertical offset according to two attribution weights, specifically including:

[0074] At the manifold location, the direction of the crystallization fraction sensitivity is taken as the first coordinate axis, and the direction of the assimilation crystallization ratio sensitivity is taken as the second coordinate axis. The local tangent space is determined by the first coordinate axis and the second coordinate axis, forming a local frame that varies with the manifold location.

[0075] It should be noted that the first coordinate axis The second coordinate axis is obtained after normalizing the direction vector composed of the crystallization fraction sensitivity of each component. The direction vector, composed of the assimilation crystallization ratio sensitivity of each component, is obtained after normalization; the process of determining the local tangent space using the first and second coordinate axes is as follows: , The local frame is a two-dimensional subspace in the geochemical vector space of the basis, that is, the tangent plane of the evolutionary manifold at the location of the manifold. Since the two coordinate axes are determined by the sensitivity and the sensitivity varies with the location of the manifold, the local frame is different with different locations of the sample manifold. When the two coordinate axes are nearly parallel (separation crystallization and assimilation contamination are difficult to distinguish in this direction), a low separability label is set for the sample and it is given priority in the tangent space residual decoupling in the following text.

[0076] The component of the measured geochemical vector offset relative to the candidate source region endmember along the first coordinate axis is taken as the crystallization fraction offset, the component along the second coordinate axis is taken as the assimilation-crystallization ratio offset, and the remaining component relative to the local tangent space is taken as the vertical offset.

[0077] To further clarify, the offset of the measured geochemical vector relative to the candidate source region endmembers, i.e. The decomposition process is as follows: project the offset onto the local frame, and its position... The coordinates on the graph represent the crystallization fraction offset. ,exist The coordinates on the graph represent the assimilation crystallization ratio offset. The magnitude of the orthogonal portion remaining after subtracting the offset from its projection in the local tangent space is the vertical offset. When the first coordinate axis and the second coordinate axis are not orthogonal, the projection adopts the oblique projection spanned by the frame basis (the coordinates are obtained according to the dual basis of the basis vectors), rather than the simple dot product, so as to ensure that the crystallization fraction shift and the assimilation crystal ratio shift correspond to the pure crystallization fraction change and the pure assimilation and contamination change, respectively.

[0078] The calculation expressions for crystallization fraction shift and assimilation crystallization ratio shift are as follows:

[0079] ;

[0080] Where E is a matrix with the first and second coordinate axes as columns. To measure the shift of geochemical vectors relative to candidate source region endmembers, For crystallization fraction shift, The assimilation crystallization ratio is shifted; since the first and second coordinate axes may not be orthogonal, the least squares projection of the above formula is used to obtain a unique coordinate under the sloping basis, and the vertical shift is calculated. Pick The model.

[0081] Through the above technical solution, this embodiment decomposes the sample offset into two tangential components and one vertical component corresponding to the two types of processes within a local frame that changes with the manifold position, so that the changes caused by separation crystallization and the changes caused by assimilation and contamination fall in non-overlapping directions at the beginning of the separation.

[0082] In an optional embodiment, the assimilation crystallization ratio coordinates are substituted into the mass balance relationship to obtain the assimilation contribution of incompatible elements, and the crystallization fraction coordinates are corrected after subtracting this contribution from the incompatible element offset, specifically including:

[0083] A gated correction channel is set between the crystallization fraction branch and the assimilation crystallization ratio branch;

[0084] It should be noted that the gated correction channel refers to the data path that connects the crystallization fraction branch and the assimilation crystallization ratio branch and is used to transfer the correction amount between the two branches according to the physical relationship. It is different from the forward calculation path of each branch and only carries the offset amount and write-back amount across branches.

[0085] Substitute the assimilation crystallization ratio coordinates output from the assimilation crystallization ratio branch into the mass balance relationship to calculate the assimilation contribution of the coordinates to each incompatible element; take the ratio of the assimilation crystallization ratio sensitivity to the crystallization fraction sensitivity of each incompatible element at the manifold position as the gate coefficient.

[0086] It should be noted that the assimilation contribution refers to the increase in the content of each incompatible element caused solely by assimilation contamination. The calculation process is as follows: Substitute the assimilation crystallization ratio coordinates output by the assimilation crystallization ratio branch and the current crystallization fraction coordinates into the mass balance relationship, and calculate the theoretical content of each incompatible element in the two cases of pure separation crystallization with assimilation term and with assimilation crystallization ratio set to zero. The difference between the two is the assimilation contribution of the incompatible element.

[0087] Gating coefficient The calculation expression is:

[0088] ;

[0089] in, Let be the gating coefficient for the i-th incompatible element. The assimilation crystallization ratio sensitivity of this element. Sensitivity to the crystallization fraction of this element. To prevent small positive numbers from being divided by zero (typically taking...) Since the gating coefficient is taken from the ratio of the two sensitivities at the manifold position and changes point by point with the sample manifold position, the element whose assimilation and mixing response is stronger than its separation and crystallization response has a higher proportion of its assimilation contribution written into the crystallization fractional branch.

[0090] After weighting the assimilation contribution with this gating coefficient, the assimilation crystallization ratio branch is written into the crystallization fraction branch through the gating correction channel, and the weighted assimilation contribution is deducted from the incompatible element offset of the input crystallization fraction branch.

[0091] It should be noted that the deduction process is as follows: for each incompatible element, its gating coefficient is multiplied by the assimilation contribution of that element to obtain the weighted assimilation contribution, which is then sent to the crystallization fraction branch through the gating correction channel; when the crystallization fraction branch reads its input incompatible element offset, it subtracts the corresponding weighted assimilation contribution from each element to obtain the incompatible element offset that has been removed from the assimilation contamination effect, and then outputs the crystallization fraction coordinates accordingly.

[0092] If all assimilation contributions are indiscriminately deducted from the crystallization fraction offset, elements that are mainly controlled by separation crystallization will be over-deducted, which will introduce new biases. Therefore, a mechanism is set up to weight each element separately with a gating coefficient, so that the deduction amount is adaptively scaled according to the relative sensitivity of the element to the two processes.

[0093] The crystallization fraction branch is written back to the assimilation crystallization ratio branch through a gated correction channel based on the crystallization fraction coordinates obtained after deduction and offset, and the weight of controlling isotope mixing in the assimilation crystallization ratio branch is adjusted.

[0094] It should be noted that the write-back process is as follows: the crystallization fraction coordinates output by the crystallization fraction branch are sent back to the assimilation crystallization ratio branch through the gated correction channel. The assimilation crystallization ratio branch updates its internal weight parameters used to correct the isotopic mixing ratio according to the crystallization fraction, so that the estimated assimilation crystallization ratio coordinates take into account the influence of the determined degree of crystallization on the residual magma isotopic benchmark. This write-back and the aforementioned write constitute a two-way mutual adjudication: one direction corrects the input of the other direction, and the two directions are mutually constrained in the same forward calculation.

[0095] For example, the following provides a complete numerical implementation of the gating mutual adjudication process: Taking sample S as an example, we take its incompatible elements Th and La, and the two sensitivities at the manifold position are... , , , Therefore, the gating coefficient , Assuming the initial assimilation crystallization ratio coordinate of the assimilation crystallization ratio branch is 0.30, substituting this into the mass balance relationship, the assimilation contribution of Th is calculated to be +3.0 ppm and the assimilation contribution of La is +6.0 ppm. Weighted assimilation contribution: Th is... (Standardized scale corresponding value), La is After subtracting the Th and La incompatible element offsets from the input of the crystallization fraction branch, the corrected crystallization fraction coordinates output by the crystallization fraction branch are 0.62 (initial value 0.55). After writing this coordinate back to the assimilation crystallization ratio branch and adjusting its isotopic mixing weights, the secondary output of the assimilation crystallization ratio branch converges to 0.28. At this point, the crystallization fraction coordinates and the assimilation crystallization ratio coordinates... They are mutually consistent.

[0096] Through the above technical solution, this embodiment sets up a gated correction channel that carries the cross-branch correction amount, and uses the ratio of the two sensitivities at the manifold position as the gate coefficient to adaptively scale the assimilation contribution of the element for deduction and write-back. This realizes the action of deducting the assimilation contribution and correcting the crystallization fraction coordinate in the above-mentioned overall process as a two-way structured mutual adjudication between the two branches in the same forward calculation. The cross-branch deduction is performed according to the relative sensitivity of each element to the two processes at the current manifold position, so that the crystallization degree judgment on the element side and the contamination degree judgment on the isotope side mutually correct each other rather than being independent of each other.

[0097] In an optional embodiment, before constructing the mutual adjudication inversion network containing the attribution unit, crystallization fraction branch, and assimilation crystallization ratio branch, a process is also included in which the two branches are pre-trained on forward-modeled samples on the evolutionary manifold, specifically including:

[0098] Based on the combination of crystallization fraction and assimilation crystallization ratio, theoretical vectors are calculated group by group according to the mass balance relationship, and the theoretical vectors are paired with the corresponding crystallization fraction and assimilation crystallization ratio values ​​as pre-training samples.

[0099] It should be noted that the process of pairing samples for pre-training is as follows: for each group (F, r) traversed during the construction of the evolving manifold in step S2, its theoretical vector is used as input, and the crystallization fraction and assimilation crystallization ratio used to generate the theoretical vector are used as labels to form input-label pairs; since these labels are known quantities rather than estimated quantities during forward generation, the pre-training samples inherently carry error-free supervision signals.

[0100] However, real volcanic rock samples are often unevenly distributed across the evolutionary range, with sparse samples in some sections. Training directly with real samples will cause the two branches to underfit in sparse sections. To address this, a mechanism is set up to use forward modeling vectors covering the entire evolutionary range as pre-training samples, so that the two branches can first establish responses in two coordinate directions across the complete range.

[0101] The theoretical vector is taken as the input, and its corresponding crystallization score is taken as the target pre-training crystallization score branch. The theoretical vector is taken as the input, and its corresponding assimilation crystallization ratio is taken as the target pre-training assimilation crystallization ratio branch. The initial regression parameters of the two branches are obtained, and the initial regression parameters are set as the starting parameters of the two branches in the training of volcanic rock samples.

[0102] To further clarify, the pre-training consists of two parallel regression tasks: the crystallization score branch regresses the crystallization score using the theoretical vector, and the assimilation crystallization ratio branch regresses the assimilation crystallization ratio using the theoretical vector. The loss is the mean square error between the predicted coordinates and the label coordinates. The training stops when the relative decrease in loss on the validation set is less than 0.5% for 5 consecutive rounds or reaches 100 rounds. The resulting parameters of the two branches are the initial regression parameters, which are used as the starting parameters for the two branches when training with real volcanic rock samples, replacing random initialization.

[0103] Through the above technical solution, this embodiment uses forward modeling vectors covering the entire evolutionary range to perform supervised pre-training on the two branches and uses the obtained initial regression parameters as the starting point for training on real samples, so that the two branches still have the ability to regress along the corresponding coordinate direction in the evolutionary segment where real samples are sparse.

[0104] In an optional embodiment, the process of reversing the two branches using the crystallization fraction residual and the assimilation crystallization ratio residual respectively further includes decoupling the two-directional residuals within the manifold tangent space, specifically including:

[0105] Extract the local tangent space at the manifold location, and project the crystallization fraction residual and the assimilation crystallization ratio residual onto the local tangent space to obtain two tangential residual components;

[0106] It should be noted that extracting the local tangent space means calling the tangent plane spanned by the first and second coordinate axes at the location of the sample manifold; the projection process is as follows: after restoring the crystallization fraction residual and the assimilation crystallization ratio residual to vectors along the corresponding coordinate axes in the geochemical vector space, take their components in the local tangent space to obtain two tangential residual components, denoted as the component along the first coordinate axis and the component along the second coordinate axis; the vertical residual does not participate in this decoupling step because it falls in the orthogonal direction outside the tangent space.

[0107] Extract the angle between the directions of two tangential residual components. When the angle between the directions meets the preset conflict condition, orthogonalize at least one tangential residual component to obtain a component that does not conflict with the other tangential residual component.

[0108] It should be noted that the direction angle is the angle between the two tangential residual component vectors, measured by the cosine value. The preset conflict condition refers to the condition for determining whether the two tangential residual components will cancel each other's update direction. The value range is satisfied when the cosine of the angle is less than a certain negative threshold (i.e. the angle is obtuse), typically less than -0.2. Since when the angle between the two residual components is obtuse, the update in one direction will partially cancel the update in the other direction, causing crosstalk between the two processes during training, the obtuse angle is used as the conflict criterion.

[0109] The orthogonalization process is as follows: when the conflict condition is met, subtract the projection of one tangential residual component onto the other tangential residual component to obtain a component that is orthogonal to the latter and no longer cancels each other out; for which of the two components to orthogonalize, retain the one with the smaller modulus and orthogonalize the one with the larger modulus, so as to preserve the update direction of the dominant process as much as possible; the boundary in the execution is: when the two tangential residual components are nearly parallel to each other (the cosine of the included angle is close to −1), the residual after orthogonalization approaches zero. At this time, the two branches are updated alternately in adjacent iterations to avoid the update amount in a single round being completely canceled out.

[0110] The network parameters of the crystallization fraction branch and the assimilation crystallization ratio branch are updated with the two processed tangential residual components respectively, and only the cross-branch correction amount transmitted through the gated correction channel is retained between the two branches.

[0111] It should be noted that the update process is as follows: the parameters of the crystallization fraction branch are updated by the processed components along the first coordinate axis, and the parameters of the assimilation crystallization ratio branch are updated by the processed components along the second coordinate axis. The two updates are performed independently within their respective branches. Apart from the weighted assimilation contribution and crystallization fraction coordinate write-back transmitted by the aforementioned gated correction channel, there are no other mutually influential update quantities between the two branches, so that the two branches do not contaminate each other due to residual direction coupling during the parameter update stage.

[0112] For example, the following provides a complete numerical implementation of the decoupling process of the tangent space residual: taking a certain iteration of sample S as an example (the assimilation-crystallization ratio coordinates of the manifold position in this iteration have converged to...) The crystallization fraction residual is restored to a vector along the first coordinate axis, and the assimilation crystallization ratio residual is restored to a vector along the second coordinate axis. The magnitudes of the two tangential residual components are 0.30 and 0.22, respectively, and the cosine of the direction angle is −0.35, which is less than −0.2, satisfying the conflict condition. The smaller magnitude (the assimilation crystallization ratio direction component of 0.22) is retained, and the larger magnitude (the crystallization fraction direction component of 0.30) is orthogonalized, and its projection on the assimilation crystallization ratio direction component is subtracted (the projection amount is...). The orthogonalized crystallization fraction direction component has a modulus of approximately 0.28 and is orthogonal to the assimilation crystallization ratio direction component. The crystallization fraction branch is then updated with a modulus of 0.28 and the assimilation crystallization ratio branch is updated with a modulus of 0.22. In this round, only the aforementioned gating correction amount is additionally passed between the two branches.

[0113] Through the above technical solution, this embodiment detects the angle between the directions of the two tangential residual components in the local tangential space of the sample manifold position, and orthogonalizes those that meet the conflict conditions before updating the two branches respectively. This extends the anti-crosstalk of the two types of processes from the coordinate layer and the structure layer to the parameter update stage. The decoupling at the gradient level is not performed by making a universal orthogonal projection in the Euclidean residual space, but in the local tangential space determined by the mass balance relationship. Moreover, the conflict criterion and the projection direction are given by the two sensitivity coordinate axes. Without this evolutionary manifold, this decoupling cannot be defined.

[0114] In an optional embodiment, after outputting the source region endmember composition and isotopic coordinates, the method further includes local retraining according to the manifold region, specifically including:

[0115] Substitute the source region endmembers and candidate assimilation endmembers into the mass balance relationship to recalculate the theoretical vector and update the evolutionary manifold; project each volcanic rock sample onto the updated evolutionary manifold, and take the difference between each sample and its nearest corresponding point as the reprojection residual.

[0116] It should be noted that the process of updating the evolutionary manifold is as follows: the source region endmembers obtained by the inversion of the aforementioned overall process are used to replace the initial candidate source region endmembers, and the candidate assimilation endmembers are substituted into the mass balance relation. All theoretical vectors are recalculated by traversing F and r in the same way as in step S2, and the updated evolutionary manifold is obtained. The reprojection residual refers to the standardized Euclidean distance between the measured geochemical vector of each sample and its nearest corresponding point on the updated evolutionary manifold, which represents the degree to which the sample still cannot be explained by the evolution relation under the updated endmembers.

[0117] The evolving manifold is divided into several manifold intervals according to the range of values ​​for crystallization fraction and assimilation crystallization ratio, and the reprojection residuals of samples falling into each manifold interval are statistically analyzed.

[0118] To further explain, the process of dividing the manifold intervals is as follows: the crystallization fraction interval and the assimilation crystallization ratio interval are each divided into several segments (typically 4 to 6 segments each). The two are combined to form several two-dimensional manifold intervals. Each sample is assigned to a unique interval according to its manifold position. The statistical process is as follows: for each manifold interval, the reprojection residuals of each sample falling into the interval are summarized, and its statistical value is calculated (the mean, upper quartile, or percentage of samples exceeding the threshold can be taken; in this example, the mean is taken).

[0119] When the statistical value of the reprojection residual in a certain manifold interval exceeds the preset reprojection residual threshold, the two-directional sensitivity of the interval is recalculated for the samples falling into the manifold interval, and only the part of the belonging unit and the two branches that processes the samples in the interval is retrained, and iterated until the change of the reprojection residual of each manifold interval satisfies the preset convergence condition.

[0120] It should be noted that the preset reprojection residual threshold is the boundary for determining whether the estimation of a certain manifold interval is inaccurate and whether local retraining is necessary. The value ranges from 1.5 to 3 times the median of the full sample reprojection residuals, typically set at 2 times. Because a threshold that is too low will cause intervals with normal fluctuations to be frequently retrained, while a threshold that is too high will miss truly inaccurate intervals, the above range is used. The processing for intervals that meet the conditions is as follows: for the inaccurate manifold interval, the sensitivity of both directions of the interval is recalculated using its updated manifold position (i.e., the first and second coordinate axes of the interval are re-estimated under the new endmembers), and only the projection operator serving the interval in the home unit and the parameters of the samples processing the interval in the two branches are retrained; the parameters of other intervals remain unchanged. Manifold intervals that do not meet the conditions (statistical values ​​do not exceed the threshold) are not retrained.

[0121] The retraining iteration continues until the relative change of the reprojection residual of each manifold interval is less than the preset convergence condition (typically 1%) for 3 consecutive rounds. The applicable boundary condition in the execution is: if the reprojection residual of a certain interval does not decrease after multiple rounds of local retraining, it indicates that the sample in that interval may not belong to the current assimilation-separation crystallization system. In this case, the sample in that interval is marked as an external sample and removed from the current inversion object, and it will no longer be retrained.

[0122] Through the above technical solution, this embodiment reconstructs the evolving manifold using the source region endmembers obtained from the inversion, statistically reprojects the residuals according to the manifold intervals, and only locally retrains the overthreshold intervals. This allows the bias caused by the manual initial selection of candidate source region endmembers to be locally corrected in the self-consistent iteration without disturbing the other intervals. The resulting updated source region endmembers and the two-directional sensitivity of each interval are used as the final basis for outputting the source region endmember composition and isotope coordinates in the aforementioned overall process.

[0123] See Figure 2As shown, this scheme proposes a volcanic source region inversion system based on machine learning and evolutionary manifolds to implement the aforementioned volcanic source region inversion method based on machine learning and evolutionary manifolds, including:

[0124] The vector generation module is used to obtain the incompatible element content, compatible element content and strontium-neodymium isotope ratio of volcanic rock samples, and align them with the measured geochemical vectors.

[0125] The manifold construction module is used to determine candidate source region endmembers and candidate assimilation endmembers from measured geochemical vectors, set the crystallization fraction and assimilation-crystallization ratio, and substitute the values ​​of the two into the assimilation-separation crystallization mass balance relationship to obtain the theoretical vector, thus constructing an evolutionary manifold with the two as coordinates.

[0126] The weight generation module is used to extract the two-directional sensitivity of each component to changes in crystallization fraction and assimilation crystallization ratio on the evolutionary manifold, and normalize the two into two attribution weights.

[0127] The offset splitting module is used to construct a mutual adjudication inversion network containing attribution units, crystallization fraction branches, and assimilation crystallization ratio branches. The attribution units split the offset of the measured geochemical vector relative to the candidate source region endmembers into crystallization fraction offset, assimilation crystallization ratio offset, and vertical offset according to the two attribution weights.

[0128] The coordinate correction module is used to learn and output two coordinates from the two branches based on the crystallization fraction offset and the assimilation crystallization ratio offset, respectively. The assimilation crystallization ratio coordinate is substituted into the mass balance relationship to obtain the assimilation contribution of incompatible elements, and the crystallization fraction coordinate is corrected after subtracting the contribution from the incompatible element offset.

[0129] The residual decomposition module is used to substitute the corrected two coordinates into the mass balance relationship to obtain the predicted vector, and decompose the difference between the predicted vector and the measured geochemical vector into crystallization fraction residual, assimilation crystallization ratio residual and vertical residual according to the sensitivity of the two directions and their remaining components.

[0130] The iterative callback module is used to call back the two branches with the crystallization fraction residual and the assimilation crystallization ratio residual respectively, and reduce the participation weight of samples exceeding the preset vertical residual threshold with the vertical residual, and iterate until the residual change meets the preset convergence condition.

[0131] The endmember output module is used to use the crystallization fraction offset and assimilation crystallization ratio offset after convergence as the separation crystallization component and the contamination component, respectively, to invert the candidate source region endmembers using the two coordinates after convergence, and output the source region endmember composition and isotopic coordinates.

[0132] In another embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiments.

[0133] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described above.

[0134] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps described above.

[0135] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for inverting volcanic source regions based on machine learning and evolutionary manifolds, characterized in that, The method includes: Obtain the incompatible element content, compatible element content, and strontium-neodymium isotope ratio of volcanic rock samples, and align them to the measured geochemical vector; Candidate source region endmembers and candidate assimilation endmembers are determined by measured geochemical vectors. Crystallization fraction and assimilation-crystallization ratio are set. The combination of these two values ​​is substituted into the assimilation-separation crystallization mass balance relationship to obtain theoretical vectors, which constitute an evolutionary manifold with these two as coordinates. The sensitivity of each component to changes in crystallization fraction and assimilation crystallization ratio is extracted on the evolutionary manifold in two directions, and the two are normalized into two attribution weights. A mutual adjudication inversion network containing attribution units, crystallization fraction branches, and assimilation crystallization ratio branches is constructed. The attribution units decompose the offset of the measured geochemical vector relative to the candidate source region endmembers into crystallization fraction offset, assimilation crystallization ratio offset, and vertical offset according to the two attribution weights. The two branches learn and output two coordinates based on the crystallization fraction offset and the assimilation crystallization ratio offset, respectively. The assimilation crystallization ratio coordinate is substituted into the mass balance relationship to obtain the assimilation contribution of incompatible elements. The crystallization fraction coordinate is then corrected after subtracting the contribution from the incompatible element offset. Substituting the corrected two coordinates into the mass balance relationship yields the predicted vector. The difference between the predicted vector and the measured geochemical vector is decomposed into crystallization fraction residual, assimilation crystallization ratio residual, and vertical residual according to the sensitivity of the two directions and their remaining components. The two branches are called back respectively using the crystallization fraction residual and the assimilation crystallization ratio residual. The participation weight of samples exceeding the preset vertical residual threshold is reduced by the vertical residual. The iteration continues until the residual change satisfies the preset convergence condition. The crystallization fraction shift and assimilation crystallization ratio shift after convergence are used as the separation crystallization component and the contamination component, respectively. The candidate source region endmembers are inverted using the two coordinates after convergence, and the source region endmember composition and isotopic coordinates are output.

2. The method according to claim 1, characterized in that, The sensitivity of each component to changes in crystallization fraction and assimilation crystallization ratio is extracted from the evolutionary manifold in two directions, specifically including: The measured geochemical vector of each volcanic rock sample is projected onto the evolutionary manifold, and the corresponding point on the evolutionary manifold that is closest to the sample is taken as the manifold position of the sample. Within the neighborhood of the manifold location, preset coordinate increments are applied along the crystallization fraction direction and the assimilation crystallization ratio direction, while keeping another coordinate constant. The changes in each component before and after the increments are calculated using the mass balance relationship, and the crystallization fraction sensitivity and assimilation crystallization ratio sensitivity of each component are obtained. Normalize the crystallization fraction sensitivity and assimilation crystallization ratio sensitivity for the same component to obtain the two assignment weights of the component at the manifold position.

3. The method according to claim 2, characterized in that, The shift of the measured geochemical vector relative to the candidate source region endmembers is decomposed into crystallization fraction shift, assimilation-crystallization ratio shift, and vertical shift based on the attribution unit and two attribution weights. Specifically, this includes: At the manifold location, the direction of the crystallization fraction sensitivity is taken as the first coordinate axis and the direction of the assimilation crystallization ratio sensitivity is taken as the second coordinate axis. The local tangent space is determined by the first coordinate axis and the second coordinate axis, forming a local frame that varies with the manifold location. The component of the measured geochemical vector offset relative to the candidate source region endmember along the first coordinate axis is taken as the crystallization fraction offset, the component along the second coordinate axis is taken as the assimilation-crystallization ratio offset, and the remaining component relative to the local tangent space is taken as the vertical offset.

4. The method according to claim 3, characterized in that, Substituting the assimilation crystallization ratio coordinates into the mass balance relationship yields the assimilation contribution of incompatible elements. This contribution is then subtracted from the incompatible element offset before correcting the crystallization fraction coordinates. Specifically, this includes: A gated correction channel is set between the crystallization fraction branch and the assimilation crystallization ratio branch; Substitute the assimilation crystallization ratio coordinate output from the assimilation crystallization ratio branch into the mass balance relationship to calculate the assimilation contribution of the coordinate to each incompatible element. The gating coefficient is the ratio of the assimilation crystallization ratio sensitivity to the crystallization fraction sensitivity of each incompatible element at the manifold position. After weighting the assimilation contribution with this gating coefficient, the assimilation crystallization ratio branch is written into the crystallization fraction branch through the gating correction channel, and the weighted assimilation contribution is deducted from the incompatible element offset of the input crystallization fraction branch. The crystallization fraction branch is written back to the assimilation crystallization ratio branch based on the crystallization fraction coordinates obtained after deduction and offset, and the weight of controlling isotope mixing in the assimilation crystallization ratio branch is adjusted.

5. The method according to claim 1, characterized in that, Before constructing the mutual adjudication inversion network containing the attribution unit, crystallization fraction branch, and assimilation crystallization ratio branch, the process also includes pre-training the two branches on the evolutionary manifold using forward modeling samples. Specifically, this includes: Based on the combination of crystallization fraction and assimilation crystallization ratio, theoretical vectors are calculated group by group according to the mass balance relationship, and the theoretical vectors are paired with the corresponding crystallization fraction and assimilation crystallization ratio values ​​as pre-training samples. The theoretical vector is taken as the input, and its corresponding crystallization score is taken as the target pre-training crystallization score branch. The theoretical vector is taken as the input, and its corresponding assimilation crystallization ratio is taken as the target pre-training assimilation crystallization ratio branch. The initial regression parameters of the two branches are obtained, and these initial regression parameters are set as the starting parameters of the two branches in the training of volcanic rock samples.

6. The method according to claim 4, characterized in that, The process of reversing the two branches using the crystallization fraction residual and the assimilation crystallization ratio residual respectively also includes decoupling the two-directional residuals within the manifold tangent space, specifically including: Extract the local tangent space at the manifold location, and project the crystallization fraction residual and the assimilation crystallization ratio residual onto the local tangent space to obtain two tangential residual components; Extract the angle between the directions of two tangential residual components. When the angle between the directions meets the preset conflict condition, orthogonalize at least one tangential residual component to obtain a component that does not conflict with the other tangential residual component. The network parameters of the crystallization fraction branch and the assimilation crystallization ratio branch are updated with the two processed tangential residual components respectively, and only the cross-branch correction amount transmitted through the gated correction channel is retained between the two branches.

7. The method according to claim 6, characterized in that, After outputting the source region endmember composition and isotope coordinates, the process also includes local retraining according to the manifold region, specifically including: Substitute the source region endmembers and candidate assimilation endmembers into the mass balance relationship to recalculate the theoretical vector and update the evolutionary manifold; Project each volcanic rock sample onto the updated evolutionary manifold, and take the difference between each sample and its nearest corresponding point as the reprojection residual; The evolutionary manifold is divided into several manifold intervals according to the range of values ​​of crystallization fraction and assimilation crystallization ratio, and the reprojection residuals of samples falling into each manifold interval are statistically analyzed. When the statistical value of the reprojection residual in a certain manifold interval exceeds the preset reprojection residual threshold, the two-directional sensitivity of the interval is recalculated for the samples falling into the manifold interval, and only the part of the belonging unit and the two branches that processes the samples in the interval is retrained, and iterated until the change of the reprojection residual of each manifold interval satisfies the preset convergence condition.

8. A volcanic source region inversion system based on machine learning and evolutionary manifold, characterized in that, A method for implementing the volcanic source region inversion method based on machine learning and evolutionary manifold as described in any one of claims 1-7, comprising: The vector generation module is used to obtain the incompatible element content, compatible element content and strontium-neodymium isotope ratio of volcanic rock samples, and align them with the measured geochemical vectors. The manifold construction module is used to determine candidate source region endmembers and candidate assimilation endmembers from measured geochemical vectors, set the crystallization fraction and assimilation-crystallization ratio, and substitute the values ​​of the two into the assimilation-separation crystallization mass balance relationship to obtain the theoretical vector, thus constructing an evolutionary manifold with the two as coordinates. The weight generation module is used to extract the two-directional sensitivity of each component to changes in crystallization fraction and assimilation crystallization ratio on the evolutionary manifold, and normalize the two into two attribution weights. The offset splitting module is used to construct a mutual adjudication inversion network containing attribution units, crystallization fraction branches, and assimilation crystallization ratio branches. The attribution units split the offset of the measured geochemical vector relative to the candidate source region endmembers into crystallization fraction offset, assimilation crystallization ratio offset, and vertical offset according to the two attribution weights. The coordinate correction module is used to learn and output two coordinates from the two branches based on the crystallization fraction offset and the assimilation crystallization ratio offset, respectively. The assimilation crystallization ratio coordinate is substituted into the mass balance relationship to obtain the assimilation contribution of incompatible elements, and the crystallization fraction coordinate is corrected after subtracting the contribution from the incompatible element offset. The residual decomposition module is used to substitute the corrected two coordinates into the mass balance relationship to obtain the predicted vector, and decompose the difference between the predicted vector and the measured geochemical vector into crystallization fraction residual, assimilation crystallization ratio residual and vertical residual according to the sensitivity of the two directions and their remaining components. The iterative callback module is used to call back the two branches with the crystallization fraction residual and the assimilation crystallization ratio residual respectively, and reduce the participation weight of samples exceeding the preset vertical residual threshold with the vertical residual, and iterate until the residual change meets the preset convergence condition. The endmember output module is used to use the crystallization fraction offset and assimilation crystallization ratio offset after convergence as the separation crystallization component and the contamination component, respectively, to invert the candidate source region endmembers using the two coordinates after convergence, and output the source region endmember composition and isotopic coordinates.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images