A Deep Learning-Based Method and System for Provenance Quantification of Detrital Zircon

Abstract

This invention discloses a method and system for probable quantification of detrital zircon based on deep learning, relating to the field of biogeochemical cycle system technology. The method includes: step S1, constructing a detrital zircon U-Pb geochronology database; step S2, using a deep temporal clustering-based probable clustering model to delineate direct probable regions within the database constructed in step S1; step S3, using a temporal convolutional network-based probable quantification model to learn the data features of the direct probable regions based on the results of step S2, and performing probable quantification on the sink regions to obtain the quantification results. This invention effectively solves the problem that mainstream temporal clustering and classification models do not pay enough attention to the absolute location information of age distribution data, making them well-suited for detrital zircon U-Pb geochronology data.

Description

Technical Field

[0001] This invention relates to the field of biogeochemical cycle system technology, and in particular to a method and system for quantifying the provenance of detrital zircon based on deep learning. Background Technology

[0002] The aeolian deposits of the Loess Plateau in China contain rich information on continental paleoclimate and atmospheric circulation changes. Research on the provenance of aeolian deposits on the Loess Plateau is crucial for identifying the formation mechanisms of silt and their relationship with the regional climatological environment, thus playing a vital role in interpreting the paleoenvironmental record of aeolian deposits. It also provides important evidence for reconstructing the regional tectonic evolution history, paleoenvironment, and paleoclimate, and is significant for accurately understanding the global biogeochemical cycle system. Current research methods employ provenance tracing based on detrital zircon U-Pb geochronology, determining the material origin by analyzing the similarity of age distribution maps between potential provenance areas and sediment samples. Deep learning, through multi-layered nonlinear networks, can abstract and interpret input data, possessing powerful feature extraction and learning capabilities. The deep learning-based method for quantifying the provenance of materials first delineates potential provenance regions using a deep time-series clustering model. Then, it learns the age distribution map data characteristics of potential provenance regions using a deep time-series classification model. The contribution of the sedimentary material is quantified based on the similarity between the age distribution map of the potential provenance region and the sedimentary material. This provides a new approach and method for provenance tracing based on detrital zircon U-Pb geochronology.

[0003] However, due to the large amount of detrital zircon U-Pb geochronology data from the Loess Plateau being scattered across different literatures, no one has yet conducted a comprehensive and systematic collection and organization of this data, which hinders research on the provenance of the Loess Plateau. Furthermore, age distribution mapping analysis requires absolute location information hidden within the data. Current deep learning-based time series clustering and classification models typically do not pay much attention to absolute location information and are not well-suited for detrital zircon U-Pb geochronology data. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for quantifying the provenance of detrital zircon based on deep learning, which effectively solves the problem that mainstream time-series clustering and classification models do not pay enough attention to the absolute location information of age distribution data, making it well applicable to detrital zircon U-Pb geochronological data.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a deep learning-based method for quantifying the provenance of detrital zircon, the method comprising the following steps:

[0007] Step S1: Construct a detrital zircon U-Pb geochronology database;

[0008] Step S2: Using the database constructed in step S1, the direct source region is divided using a source clustering model based on deep temporal clustering.

[0009] Step S3: Based on the results of step S2, the source quantization model based on temporal convolutional network is used to learn the data characteristics of the direct source region and perform source quantization on the sink region to obtain the quantization result.

[0010] Furthermore, step S1, constructing a detrital zircon U-Pb geochronology database, specifically includes:

[0011] By manually collecting literature on the provenance of detrital zircon from typical regions, and by collecting and organizing detrital zircon U-Pb geochronological data, a typical regional detrital zircon U-Pb geochronological database containing spatiotemporal dimensions was constructed.

[0012] Furthermore, the collection and organization of the detrital zircon U-Pb geochronological data includes: literature search, literature screening, data download, data collection, data organization, and data entry; among which,

[0013] The literature search involves using a literature indexing tool to find relevant papers using keywords, and collecting the titles, abstracts, keywords, and DOI information of the relevant papers for subsequent literature screening.

[0014] The literature screening refers to selecting literature that contains detrital zircon U-Pb geochronological data.

[0015] The data download refers to downloading the original text of the target document and the corresponding data file;

[0016] The data collection process involves copying data from the original target document and data files into a transition file.

[0017] The data processing involves organizing the collected data, supplementing the field information in the database, and reviewing the data. If the review is successful, the next step of data entry is performed; if the review fails, the process returns to the data collection step to organize and modify the parts that failed.

[0018] The data entry refers to the process of inputting the organized data into the database.

[0019] Furthermore, the material source clustering model based on deep temporal clustering consists of three parts: an encoder, a decoder, and a time-series clustering layer.

[0020] Furthermore, step S2, which involves using a deep temporal clustering-based source clustering model to delineate direct source regions within the database constructed in step S1, specifically includes:

[0021] The zircon age distribution data obtained from the database constructed in step S1 is used to extract latent feature vectors through an encoder. The latent feature vectors are then reconstructed into reconstructed vectors by a decoder to optimize the latent feature vectors. At the same time, the latent feature vectors are also fed into a time series clustering layer for clustering. Euclidean distance and KL divergence are used to optimize the encoder parameters to obtain latent feature vectors that are more suitable for clustering. Finally, the reconstruction loss and clustering loss of the model are optimized simultaneously through joint training to obtain the optimal clustering results.

[0022] Furthermore, the encoder employs three network structures: convolutional network, bidirectional GRU network, and global attention.

[0023] The decoder adopts a two-layer structure, namely an upsampling layer and a deconvolution layer;

[0024] The time series clustering layer consists of three steps: first, initializing the cluster centers; second, calculating the distance between the input data and the cluster centers and converting it into a probability distribution; and finally, minimizing the KL divergence between the calculated probability distribution and the target distribution, and updating the cluster centers.

[0025] Furthermore, the object source quantization model based on temporal convolutional networks consists of three parts: a location embedding layer, a feature extraction layer, and an output layer.

[0026] Furthermore, in step S3, based on the result of step S2, a source quantization model based on a temporal convolutional network is used to learn the data features of the direct source region, and source quantization is performed on the sink region to obtain the quantization result. Specifically, this includes:

[0027] The zircon age distribution data obtained from the partitioning results of the provenance clustering model based on deep temporal clustering in step S2 is used to embed the absolute position information into the input data through the position embedding layer; then the feature extraction layer uses a bidirectional temporal convolutional network to extract the temporal features and absolute position information of the input data from two directions; finally, the softmax function is used in the output layer to output the probability distribution and obtain the classification result.

[0028] Furthermore, the position embedding layer uses trigonometric function-based absolute position encoding to embed absolute position information into the input data, and the calculation formula is as follows:

[0029] p k,2i =sin(k / 10000) 2i / d );

[0030] p k,2i+1 =cos(k / 10000) 2i / d );

[0031] In the formula, P k,2i P represents the even-numbered row encoding information at the k-th time step.k,2i+1 This represents the encoding information of the odd-numbered rows at the k-th time step, where d represents the dimension of the input data and k represents the time step corresponding to the input data.

[0032] The feature extraction layer uses a bidirectional temporal convolutional network to extract deep features and absolute position information of age distribution data from the position embedding layer, and uses two types of causal dilated convolutions to process the input data in parallel.

[0033] The output layer employs a fully connected network and a softmax function; wherein, the fully connected network maps the extracted features to the similarity score of the input data to the features of the potential source region; the softmax function normalizes the similarity score to obtain the quantified result of the contribution of the potential source region to the input sample.

[0034] This invention also provides a deep learning-based system for quantifying the provenance of detrital zircon, the system comprising: a construction unit, a partitioning unit, and a quantization unit; wherein,

[0035] Building blocks are used to construct a detrital zircon U-Pb geochronology database;

[0036] The partitioning unit is used to partition the direct source region in the database constructed by the building unit using a source clustering model based on deep temporal clustering;

[0037] The quantization unit is used to learn the data characteristics of the direct source region by using a source quantization model based on a temporal convolutional network, based on the result of the partitioning unit, and to perform source quantization on the sink region to obtain the quantization result.

[0038] The technical effects and advantages of this invention are as follows:

[0039] First, this invention addresses the problem that mainstream temporal clustering and classification models do not pay enough attention to the absolute location information of age distribution data.

[0040] Secondly, in the task of determining the source region of a substance, this invention provides a more objective standard for the division of direct source regions and offers a novel division approach for reference in subsequent research.

[0041] Third, in the task of quantifying the proportion of material contribution sources, this invention breaks the limitation of the number of zircon age data, and can quantify each profile, providing a brand-new approach to quantifying the material contribution of the source region.

[0042] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description and the drawings. Attached Figure Description

[0043] Figure 1 This is a flowchart of a deep learning-based method for quantifying the provenance of detrital zircon according to the present invention.

[0044] Figure 2 This is a flowchart of the process for collecting detrital zircon U-Pb geochronological data according to the present invention;

[0045] Figure 3 This is a schematic diagram of the material-source clustering (DTCG-att) model structure of the deep temporal clustering of the present invention;

[0046] Figure 4 This is a schematic diagram of the Position-BiTCN model structure of the temporal convolutional network of the present invention;

[0047] Figure 5 This is a schematic diagram of the bidirectional causal dilated convolution structure of the present invention;

[0048] Figure 6 This is a schematic diagram of the basic residual block structure of the bidirectional temporal convolutional network of the present invention;

[0049] Figure 7 This is a schematic diagram of a deep learning-based detrital zircon provenance quantification system according to the present invention. Detailed Implementation

[0050] 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.

[0051] The design concept of this invention includes:

[0052] First, literature on the provenance of detrital zircon from typical regions was searched through manual collection methods. U-Pb geochronological data of detrital zircon were collected and organized, and a U-Pb geochronological database of detrital zircon from typical regions containing spatiotemporal dimensions was constructed.

[0053] Secondly, in the task of determining the source region of materials, based on the characteristics of detrital zircon U-Pb geochronology data, a source clustering model based on deep temporal clustering (DTCG-att) is proposed. This model is based on the DTC (Deep Temporal Clustering) model and introduces a global attention mechanism after the encoder to enhance the model's ability to extract absolute location information from time series data and further optimize the latent feature representation. At the same time, bidirectional GRU (Gated Recurrent Units) is used to replace LSTM (Long Short-Term Memory) network to accelerate the model training speed.

[0054] Finally, in the task of quantifying the proportion of material contribution sources, based on the characteristics of detrital zircon U-Pb geochronology data, a position-BiTCN model based on temporal convolutional networks is proposed. This model is based on the TCN (Temporal Convolutional Network) model, introduces a position embedding layer to embed absolute position information into the input data, and guides the model to extract absolute position information. At the same time, the feature extraction layer uses bidirectional TCN to improve the model's classification accuracy.

[0055] To address the shortcomings of existing technologies, this invention discloses a deep learning-based method for quantifying the provenance of detrital zircon. Figure 1 This is a flowchart of a deep learning-based method for quantifying the provenance of detrital zircon, as described in the present invention. Figure 1 As shown, the method includes the following steps:

[0056] Step S1: Construct a detrital zircon U-Pb geochronology database:

[0057] Relevant papers and corresponding zircon U-Pb geochronological data were systematically collected manually. The process of collecting detrital zircon U-Pb geochronological data is as follows: Figure 2As shown, it includes six steps: (1) Literature search mainly involves using the literature indexing tool to search for relevant papers using keywords, and collecting the title, abstract, keywords and DOI information of relevant papers for subsequent literature screening; (2) Literature screening mainly involves screening out the literature containing detrital zircon U-Pb geochronology data; (3) Data download mainly involves downloading the original text of the target literature and the corresponding data file; (4) Data collection mainly involves copying the data from the original text of the target literature and the data file into a transition file for subsequent data entry; Specifically, data collection includes: file conversion, table data collection, text data collection, image data collection and data screening; (5) Data organization mainly involves organizing the collected data, supplementing the field information in the database, and reviewing the data; (6) Data entry mainly involves entering the organized data into the database.

[0058] Step S2: Use a source clustering model based on deep temporal clustering to delineate the direct source region:

[0059] Based on the database constructed in step S1, a source-direction clustering model based on deep temporal clustering is used to divide the direct source-direction regions. The source-direction clustering model based on deep temporal clustering (DTCG-att) consists of three parts: an encoder, a decoder, and a time-series clustering layer. The model structure is as follows: Figure 3 As shown, zircon age distribution data is processed by an encoder to extract latent feature vectors. These latent feature vectors are then reconstructed by a decoder to optimize the latent feature vectors. Simultaneously, the latent feature vectors are fed into a time-series clustering layer for cluster assignment. Euclidean distance and KL divergence are used to optimize the encoder parameters, resulting in latent feature vectors more suitable for clustering. Finally, the reconstruction loss and clustering loss of the model are jointly optimized to obtain the optimal clustering result. Table 1 below shows the performance of the DTCG-att model. As can be seen from Table 1, compared with other models, the DTCG-att model has the highest silhouette coefficient and the lowest Davidson-Burding index, resulting in better clustering performance. The DTCG-att model is highly effective in determining the source region of materials.

[0060] Table 1. DTCG-att Model Performance

[0061]

[0062] The encoder employs three network structures: convolutional network, bidirectional GRU network, and global attention. The convolutional network uses a 1D CNN to extract short-range dependency information from the time-series data. Simultaneously, pooling layers are used for downsampling to reduce feature dimensionality. Max pooling is used, and the calculation formula is shown in expression (1):

[0063]

[0064] In equation (1), Y(j) represents the j-th eigenvalue after pooling, and K j Let X(i) represent the j-th pooling region, and let X(i) represent the i-th feature input value of the input data in the j-th pooling region.

[0065] Because a single convolutional layer has a small receptive field, it can only extract local features from the time series data. To obtain long-range dependencies in the time series, an RNN (Recurrent Neural Network) is needed for subsequent feature extraction. To reduce model complexity, a bidirectional GRU is used instead of a bidirectional LSTM, reducing training time while maintaining the model's clustering performance. The bidirectional GRU uses two GRUs to capture the forward and backward features of the time series data respectively, and then concatenates them for output.

[0066] The main function of global attention is to focus the model on the feature information at a specific time step, thereby enhancing the model's ability to extract absolute positional information. Global attention calculates attention weights for the entire time series data over all time steps. A fully connected layer is used to train the global attention parameters, and the softmax function is used to obtain the weights at each time step. The calculation formulas are shown in expressions (2), (3), and (4):

[0067] u = tanh(W) a X G +b a (2)

[0068]

[0069] Z = β × XG; (4)

[0070] In equations (2) to (4), X G The input data represents global attention, u represents the output of the fully connected layer, and W represents the input data. a b represents the weights of the global attention parameters. a This represents the global attention parameter bias. α t u represents the weight at time step t. t denoted by t, β represents the output of the fully connected layer at time step t, β represents the weight vector after calculating the attention weights for the entire time series data, and Z represents the latent feature vector after global attention.

[0071] Normalize the output u of the fully connected layer using the softmax function to obtain the weight α at each time step. t Finally, global attention is input into the data X. GMultiplying the value by the weight vector α yields the latent feature vector Z. After global attention, the feature values ​​at certain time steps are amplified, and subsequent model training will focus on the feature information at the corresponding time steps, thereby enhancing the model's ability to extract absolute position information.

[0072] The main function of the decoder is to decode the latent feature vector output by the encoder into a reconstructed vector with the same dimension as the original data, calculate the difference between the reconstructed vector and the original data to evaluate the effectiveness of the latent feature vector, so that the model can learn truly effective latent feature vectors in an unsupervised manner. The decoder mainly consists of two layers: an upsampling layer and a deconvolution layer. The upsampling layer uses a double interpolation algorithm to copy and add to the latent feature vector in the sequence length and encoding dimension directions to restore it to data with the same dimension as before max pooling. The calculation formulas are shown in expressions (5), (6), and (7):

[0073]

[0074]

[0075]

[0076] In equations (5) to (7), f(R1) represents Q. 11 With Q 21 The result of a linear interpolation along the sequence length, f(R2) represents Q. 12 With Q 22 The result of a linear interpolation along the sequence length, f(P), represents the result of interpolation through Q. 11 Q 12 Q 21 Q 22 The interpolation point eigenvalues ​​obtained after performing the double interpolation algorithm, f(Q) 11 ) represents Q 11 eigenvalues, f(Q) 12 ) represents Q 12 eigenvalues, f(Q) 21 ) represents Q 21 eigenvalues, f(Q) 22 ) represents Q 22 eigenvalues, Q 11 Q 12 Q 21 Q 22 Let P represent the coordinates of the known data points, P represent the coordinates of the interpolation points, x and y represent the corresponding coordinates, and x1 represent Q. 11 Q 12 The coordinate value along the sequence length, x2 represents Q. 21 Q 22 The coordinate value along the sequence length, y1 represents Q.11 Q 21 The coordinate value in the encoding dimension direction, y2 represents Q. 12 Q 22 The coordinate value in the encoding dimension direction, R1 represents Q. 11 With Q 21 The interpolation point obtained by performing a linear interpolation along the sequence length direction, R2 represents Q. 12 With Q 22 The interpolation point is obtained by performing a linear interpolation along the length of the sequence.

[0077] A deconvolutional layer is a special type of convolutional layer that reconstructs a vector similar to the original data by transposing the convolution kernel during the convolution operation and then performing a convolution operation with the result of the forward convolution. The forward convolution can be expressed by matrix operations as formula (8):

[0078] Y = CX; (8)

[0079] In equation (8), X represents the data after deconvolution (i.e., the input data), Y represents the data after convolution, and C represents the weight matrix after the transformation of the convolution kernel in the convolution operation. The formula for calculating the deconvolution layer is shown in expression (9):

[0080] X = DY; (9)

[0081] In equation (9), Y represents the data after convolution, X represents the data after deconvolution, and D represents the weight matrix required for deconvolution.

[0082] The time series clustering layer mainly consists of three steps: first, initializing the cluster centers; second, calculating the distance between the input data and the cluster centers and converting it into a probability distribution; and finally, minimizing the KL divergence between the calculated probability distribution and the target distribution to update the cluster centers.

[0083] In the initialization of cluster centers, the K-means algorithm is used to cluster the latent feature vectors extracted by the deep autoencoder. Euclidean distance is used as the metric, allowing the model to focus more on absolute positional information. The Euclidean distance calculation formula is shown in expression (10):

[0084]

[0085] In equation (10), Z represents the latent feature vector after global attention, w represents the cluster center point, and z i w represents the latent feature vector extracted by the deep autoencoder (i.e., the eigenvalues ​​of the latent feature vector in dimension i). i This represents the feature value of the cluster center in dimension i, where i represents the dimension of the feature vector, i = 1, 2, 3, ..., n.

[0086] When the latent feature vector extracted by the deep autoencoder is fed into the time series clustering layer, the latent feature vector z is calculated using the Euclidean distance metric. i To the cluster center w j distance d ij Then the distance d ij Transform into a probability distribution to measure the eigenvector z i To the cluster center w j The similarity is determined using the Student's t-distribution function from the distance d. ij The probability distribution is obtained from the above, and the calculation formula is expression (11):

[0087]

[0088] In equation (11), z i w represents the latent feature vector extracted by the deep autoencoder. j q represents the cluster center. ij The latent feature vector z represents the input. i The probability of belonging to the j-th category, sinl() represents the distance between the latent feature vector and the cluster center, using Euclidean distance calculation, and α represents the degrees of freedom of the student t-distribution, usually taking a value of 1.

[0089] Step S3: Based on the partitioning results of the source clustering model based on deep temporal clustering, the source quantization model based on temporal convolutional network is used to learn the data features of the direct source region and perform source quantization on the sink region to obtain the quantization results:

[0090] Based on the partitioning result in step S2, a position-BiTCN-based source quantization model is used to learn the data features of the direct source region and perform source quantization on the sink region samples to obtain the quantization result of the direct source region to the sink region. The position-BiTCN-based source quantization model consists of three parts: a position embedding layer, a feature extraction layer, and an output layer. The model structure is as follows: Figure 4As shown, the zircon age distribution data is embedded into the input data through a position embedding layer to incorporate absolute positional information. Then, a bidirectional temporal convolutional network is used in the feature extraction layer to extract temporal features and absolute positional information from the input data from two directions. Finally, a softmax function is used in the output layer to output the probability distribution, yielding the classification result. Table 2 below shows the performance of the Position-BiTCN model. As can be seen from Table 2, the Position-BiTCN model has the best classification performance across the three datasets, improving the F1 score by 1.53%, 2.25%, and 2.45% compared to the TCN model, respectively. This demonstrates that the bidirectional structure can extract more accurate feature information from age distribution data, indicating that absolute positional information can guide the model to identify differences between age distribution data. The Position-BiTCN model is effective in the task of quantifying the proportion of material contribution sources.

[0091] Table 2. Performance of the Position-BiTCN Model

[0092]

[0093]

[0094] The position embedding layer uses trigonometric function-based absolute position encoding to embed absolute position information into the input data. The calculation formulas are shown in expressions (12) and (13):

[0095] p k,2i =sin(k / 10000) 2i / d (12)

[0096] p k,2i+1 =cos(k / 10000) 2i / d (13)

[0097] In equations (12) to (13), P k,2i P represents the even-numbered row encoding information at the k-th time step. k,2i+1 This represents the encoding information of the odd-numbered rows at the k-th time step, where d represents the dimension of the input data and k represents the time step corresponding to the input data.

[0098] Trigonometric function-based absolute position encoding is simple to calculate, exhibits explicit generation rules, and demonstrates a certain degree of extrapolation as the length of the input data sequence varies. The dimension of the trigonometric function-based absolute position vector is the same as the dimension of the input data; after calculation, it is directly added to the input data to complete the embedding of absolute position information.

[0099] The feature extraction layer employs a bidirectional temporal convolutional network to extract deep features and absolute positional information from the location embedding layer of the age distribution data. Two types of causal dilated convolutions are used in parallel to process the input data, with the structure as follows: Figure 5As shown, bidirectional causal dilated convolution includes two causal dilated convolution structures: a forward structure and a reverse structure. The reverse structure removes zero padding on the left, including present and future information while eliminating the influence of past information. This model uses four layers of bidirectional causal dilated convolution. During the convolution operation, the input data is spaced, and the dilation rate d increases exponentially by 2 with the number of layers.

[0100] The basic residual block structure of bidirectional temporal convolutional networks is as follows: Figure 6 As shown, this basic residual block uses two residual units to process data in parallel. After convolution, the residual units are normalized using WeightNorm, with ReLU as the activation function. Dropout is used to prevent overfitting. When the input and output dimensions are different, a 1×1 convolution is used to change the input data dimension. The two residual units have similar basic structures, mainly differing in their causal dilated convolutions. One uses a forward causal dilated convolution to extract input data features, while the other uses a backward causal dilated convolution to extract features. The input data for the residual units in the first layer of the basic residual block is the same. Afterward, the input data for each layer needs to be split into two parts according to the encoding dimension and fed into the corresponding residual unit. At output, the output data of the residual unit is concatenated according to the encoding dimension and then fed into the next layer.

[0101] The output layer employs a fully connected network and a softmax function. The fully connected network maps the extracted features to similarity scores between the input data and the features of the potential source region. During training, the softmax function assigns a specific category to the input data. During quantization, the softmax function normalizes the similarity scores to obtain the quantized contribution of the potential source region to the input sample. The calculation formula is shown in expression (14):

[0102] P = softmax(WH + b); (14)

[0103] In equation (14), P represents the quantization result obtained after normalization, W represents the weight of the fully connected network, b represents the bias of the fully connected network, and H represents the output result of the feature extraction layer.

[0104] This invention also discloses a deep learning-based system for quantifying the provenance of detrital zircon. Figure 7 This is a schematic diagram of a deep learning-based detrital zircon provenance quantification system according to the present invention, as shown below. Figure 7As shown, the system includes: a construction unit 11, a partitioning unit 12, and a quantization unit 13; wherein, the construction unit 11 is used to construct a detrital zircon U-Pb geochronology database; the partitioning unit 12 is used to partition the direct provenance region using a provenance clustering model based on deep temporal clustering under the database constructed by the construction unit 11; the quantization unit 13 is used to learn the data features of the direct provenance region and perform provenance quantization on the sink region to obtain the quantization result based on the partitioning result of the partitioning unit 12 using a provenance quantization model based on temporal convolutional network.

[0105] Furthermore, the construction unit 11 is specifically used to search for literature on the provenance of detrital zircon in typical areas through manual collection, collect and organize detrital zircon U-Pb geochronological data, and construct a typical area detrital zircon U-Pb geochronological database containing spatiotemporal dimensions.

[0106] Furthermore, the partitioning unit 12 is specifically used to extract potential feature vectors from the zircon age distribution data obtained from the database constructed by the construction unit 11 through an encoder; the potential feature vectors are reconstructed into reconstruction vectors by the decoder for optimizing the potential feature vectors, and at the same time, the potential feature vectors are sent to the time series clustering layer for clustering assignment, and the encoder parameters are optimized using Euclidean distance and KL divergence to obtain potential feature vectors that are more suitable for clustering; finally, the reconstruction loss and clustering loss of the model are optimized simultaneously by joint training to obtain the optimal clustering result.

[0107] Furthermore, the quantization unit 13 is specifically used to embed the absolute position information into the input data through the position embedding layer using the zircon age distribution data obtained from the partitioning result of the source clustering model based on deep temporal clustering by the partitioning unit 12; then the feature extraction layer uses a bidirectional temporal convolutional network to extract the temporal features and absolute position information of the input data from two directions; then the softmax function is used in the output layer to output the probability distribution and obtain the classification result.

[0108] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for quantifying the provenance of detrital zircon based on deep learning, characterized in that, The method includes the following steps: Step S1: Construct a detrital zircon U-Pb geochronology database; Step S2: Under the database constructed in step S1, the direct material source region is divided using a material source clustering model based on deep temporal clustering. The material source clustering model based on deep temporal clustering consists of three parts: an encoder, a decoder, and a time series clustering layer. Step S3: Based on the partitioning results of step S2, a source quantization model based on a temporal convolutional network is used to learn the data features of the direct source region, and source quantization is performed on the sink region to obtain the quantization result. Specifically, this includes: The zircon age distribution data obtained from the partitioning results of the provenance clustering model based on deep temporal clustering in step S2 is used to embed absolute position information into the input data through a position embedding layer; then, the feature extraction layer uses a bidirectional temporal convolutional network to extract temporal features and absolute position information from the input data from two directions; finally, the softmax function is used in the output layer to output the probability distribution and obtain the classification result. The location embedding layer uses trigonometric function-based absolute location encoding to embed absolute location information into the input data. The calculation formula is as follows: ； ； In the formula, P k,2i Indicates the first k Even-numbered rows of encoded information at each time step. P k,2i+1 Indicates the first k The odd-numbered rows of encoded information are located at each time step. Indicates the dimension of the input data. This indicates the time step corresponding to the input data; The feature extraction layer uses a bidirectional temporal convolutional network to extract deep features and absolute position information of age distribution data from the position embedding layer, and uses bidirectional causal dilated convolution to process the input data in parallel. The output layer employs a fully connected network and a softmax function; wherein, the fully connected network maps the extracted features to similarity scores between the input data and the features of the potential source region; the softmax function normalizes the similarity scores to obtain a quantified result of the proportion of material contribution of the potential source region to the input sample.

2. The method for quantifying the provenance of detrital zircon based on deep learning according to claim 1, characterized in that, Step S1, constructing a detrital zircon U-Pb geochronology database, specifically includes: By manually collecting literature on the provenance of detrital zircon from typical regions, and by collecting and organizing detrital zircon U-Pb geochronological data, a typical regional detrital zircon U-Pb geochronological database containing spatiotemporal dimensions was constructed.

3. The method for quantifying the provenance of detrital zircon based on deep learning according to claim 2, characterized in that, The collection and organization of the detrital zircon U-Pb geochronological data includes: literature search, literature screening, data download, data collection, data organization, and data entry; among which, The literature search involves using a literature indexing tool to find relevant papers using keywords, and collecting the titles, abstracts, keywords, and DOI information of the relevant papers for subsequent literature screening. The literature screening refers to selecting literature that contains detrital zircon U-Pb geochronological data. The data download refers to downloading the original text of the target document and the corresponding data file; The data collection process involves copying data from the original target document and data files into a transition file. The data processing involves organizing the collected data, supplementing the field information in the database, and reviewing the data. If the review is successful, the next step of data entry is performed; if the review fails, the process returns to the data collection step to organize and modify the parts that failed. The data entry refers to the process of inputting the organized data into the database.

4. The method for quantifying the provenance of detrital zircon based on deep learning according to claim 1, characterized in that, Step S2, which involves using a deep temporal clustering-based source clustering model to delineate direct source regions within the database constructed in step S1, specifically includes: The zircon age distribution data obtained from the database constructed in step S1 is used to extract latent feature vectors through an encoder. The latent feature vectors are then reconstructed into reconstructed vectors by a decoder to optimize the latent feature vectors. At the same time, the latent feature vectors are also fed into a time series clustering layer for clustering. Euclidean distance and KL divergence are used to optimize the encoder parameters to obtain latent feature vectors that are more suitable for clustering. Finally, the reconstruction loss and clustering loss of the model are optimized simultaneously through joint training to obtain the optimal clustering results.

5. The method for quantifying the provenance of detrital zircon based on deep learning according to claim 4, characterized in that, The encoder employs three network structures: convolutional network, bidirectional GRU network, and global attention. The decoder adopts a two-layer structure, namely an upsampling layer and a deconvolution layer; The time series clustering layer consists of three steps: first, initializing the cluster centers; second, calculating the distance between the input data and the cluster centers and converting it into a probability distribution; and finally, minimizing the KL divergence between the calculated probability distribution and the target distribution, and updating the cluster centers.

6. The method for quantifying the provenance of detrital zircon based on deep learning according to claim 1, characterized in that, The source quantization model based on temporal convolutional networks consists of three parts: a location embedding layer, a feature extraction layer, and an output layer.

7. A deep learning-based system for quantifying the provenance of detrital zircon, characterized in that, The system includes: a construction unit, a partitioning unit, and a quantization unit; wherein... Building blocks are used to construct a detrital zircon U-Pb geochronology database; The partitioning unit is used to partition the direct material source region under the database constructed by the construction unit using a material source clustering model based on deep temporal clustering. The material source clustering model based on deep temporal clustering consists of three parts: an encoder, a decoder, and a time series clustering layer. The quantization unit, based on the partitioning results, utilizes a temporal convolutional network-based source quantization model to learn the data features of the direct source region and perform source quantization on the sink region to obtain the quantization result. Specifically, it includes: Zircon age distribution data obtained from the partitioning results of the provenance clustering model based on deep temporal clustering is embedded into the input data through a location embedding layer to embed absolute position information; then, a bidirectional temporal convolutional network is used in the feature extraction layer to extract temporal features and absolute position information of the input data from two directions; finally, a softmax function is used in the output layer to output the probability distribution and obtain the classification result. The location embedding layer uses trigonometric function-based absolute location encoding to embed absolute location information into the input data. The calculation formula is as follows: ； ； In the formula, P k,2i Indicates the first k Even-numbered rows of encoded information at each time step. P k,2i+1 Indicates the first k The odd-numbered rows of encoded information are located at each time step. Indicates the dimension of the input data. This indicates the time step corresponding to the input data; The feature extraction layer uses a bidirectional temporal convolutional network to extract deep features and absolute position information of age distribution data from the position embedding layer, and uses bidirectional causal dilated convolution to process the input data in parallel. The output layer employs a fully connected network and a softmax function; wherein, the fully connected network maps the extracted features to similarity scores between the input data and the features of the potential source region; the softmax function normalizes the similarity scores to obtain a quantified result of the proportion of material contribution of the potential source region to the input sample.

Images