A method for predicting deep-sea rare earth elements in three dimensions by a hybrid attention network fusing spatial and depth data

Abstract

This invention discloses a hybrid attention network method for 3D prediction of deep-sea rare earth elements, integrating spatial and depth data, and relating to the interdisciplinary fields of artificial intelligence and marine mineral resources. The method first acquires multi-source environmental and depth data, which are then integrated, cleaned, spatially aligned, interpolated, and standardized to construct a model input containing environmental feature image patches and standardized depth values. Spatial feature vectors are extracted using a convolutional neural network, and depth feature vectors are encoded using a multilayer perceptron. The two types of features are concatenated, and dynamic weights are generated and weighted using an attention network to obtain a weighted feature vector. This weighted feature vector is input into the prediction head module, and the model is trained using a loss function, optimizer, and early stopping mechanism. The trained model is applied to a 3D mesh of the target region, and after inference prediction and de-standardization, the predicted 3D rare earth element content is output. Simultaneously, the feature contribution is analyzed through an interpretable model.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and marine mineral resources, specifically involving the exploration and prediction technology of deep-sea mineral resources. Background Technology

[0002] With the development of marine exploration technology, the use of artificial intelligence to predict deep-sea mineral resources has become an emerging research field. The technical approach involves leveraging readily available, multi-source data to achieve intelligent prediction of resource distribution, thereby improving exploration efficiency. Deep-sea rare-earth-rich sediments possess unique advantages, including relatively high concentrations of heavy rare-earth elements and minimal environmental impact during development, making them an ideal source for future global supply of critical metals. However, their mineralization models, controlling factors, and occurrence environments differ significantly from those of terrestrial rare earth deposits, rendering prediction models and exploration theories applicable to terrestrial deposits unsuitable for direct application.

[0003] Current mainstream marine mineral exploration technologies, such as seabed optical imaging, hyperspectral imaging, and side-scan sonar, are limited by their physical detection principles. These technologies struggle to penetrate seabed sediments, only identifying surface minerals and failing to provide three-dimensional imaging or identification of the sediment interior. Consequently, they are ill-suited for effectively detecting deep-sea rare-earth element (REE)-rich sediments located in specific strata within sediment columns. Related geochemical studies have confirmed that the enrichment of REE-rich sediments in deep-sea environments is a three-dimensional process. The concentration is controlled not only by the macroscopic two-dimensional spatial environment but also by the vertical depth of the sample within the sediment core. Therefore, the prediction of such mineral deposits must rely on depth data.

[0004] Current applications of artificial intelligence in deep-sea mineral prediction are limited by a two-dimensional spatial paradigm, with the technological focus concentrated on quasi-two-dimensional deposits such as polymetallic nodules and cobalt-rich crusts. One type of existing technology utilizes convolutional neural networks to perform target detection and segmentation on two-dimensional seabed images, training models to identify the morphology, size, and coverage of minerals exposed on the seabed surface. Another type, represented by traditional machine learning algorithms, divides the study area into grid cells, extracts two-dimensional environmental variables to form a feature table, and trains the model by combining known mineral or non-mineral point labels to generate a two-dimensional resource potential distribution map. These existing technologies are essentially two-dimensional spatial models, simplifying the prediction problem to the analysis of seabed planar features. In terms of model structure, they completely ignore the depth variable that determines the enrichment of rare earth element-rich sediments in deep-sea environments, failing to effectively model the complex nonlinear relationship between their concentration and depth, thus losing important mineralization information. Meanwhile, existing technologies lack adaptive screening mechanisms, making it difficult to determine the relative importance of different spatial environmental features in specific geological contexts based on depth information. The simple splicing or linear combination of these heterogeneous features cannot reflect the dynamic changes in their contribution to mineralization under different geological contexts, and cannot achieve intelligent allocation of attention, resulting in biased prediction results and a lack of sufficient geological basis. Summary of the Invention

[0005] The purpose of this invention is to provide a hybrid attention network method for predicting three-dimensional rare earth elements in deep seas by integrating spatial and depth data. This method addresses the problems of existing technologies that use a two-dimensional paradigm to ignore depth variables, are unable to model the nonlinear relationship between the concentration of rare earth elements in deep seas and depth, and lack the ability to adaptively fuse heterogeneous features, resulting in inaccurate predictions.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A hybrid attention network method for 3D prediction of deep-sea rare earth elements, which integrates spatial and depth data, includes the following steps:

[0008] Acquire multi-source environmental data covering the target sea area and depth data from sediment cores, preprocess the multi-source environmental data and depth data to obtain standardized environmental feature image patches and standardized depth values;

[0009] Spatial feature vectors are extracted using the standardized environmental feature image patches, and depth feature vectors are obtained by encoding the standardized depth values.

[0010] The spatial feature vector and the depth feature vector are fused to obtain a weighted feature vector;

[0011] The weighted feature vector is input into the prediction head module, and the model is trained in combination with the training strategy to obtain the trained model.

[0012] The trained model is used to perform inference and prediction on the target area data to obtain standardized prediction values;

[0013] The standardized predicted values ​​are destandardized to obtain the final rare earth element content prediction results.

[0014] In one possible implementation, when preprocessing the multi-source environmental data and depth data, the multi-source environmental data is first spatially aligned and interpolated to obtain environmental data with a uniform resolution; the depth data is then logarithmically transformed to obtain transformed depth values; and the uniform resolution environmental data and transformed depth values ​​are then standardized to obtain standardized environmental feature image blocks and standardized depth values.

[0015] In one possible implementation, when extracting spatial feature vectors using the standardized environmental feature image blocks, the standardized environmental feature image blocks are subjected to convolution, batch normalization, and activation operations through a convolutional neural network, and the spatial feature vectors are obtained after pooling processing.

[0016] In one possible implementation, when obtaining the depth feature vector using the standardized depth value encoding, the standardized depth value is subjected to multiple linear transformations and activation processes through a multilayer perceptron to obtain the depth feature vector.

[0017] In one possible implementation, when fusing the spatial feature vector and the depth feature vector, the spatial feature vector and the depth feature vector are first concatenated along the feature dimension to obtain a combined feature vector; the combined feature vector is then linearly transformed and activated by an attention network to generate an attention weight vector; and the combined feature vector and the attention weight vector are multiplied element-wise to obtain the weighted feature vector.

[0018] In one possible implementation, when training the model using a training strategy, the preprocessed dataset is divided into a training set, a validation set, and a test set; the loss between the predicted value and the true value is calculated using a basic loss function or a weighted loss function; the model parameters are adjusted through backpropagation using an optimizer and a learning rate scheduling strategy, and the model training is completed using an early stopping mechanism, thus obtaining the trained model.

[0019] In one possible implementation, when using the trained model to perform inference prediction on target area data, a three-dimensional prediction grid of the target area is generated according to a preset geographical range depth range and resolution; the standardized environmental feature image patch and standardized depth value corresponding to each grid point in the three-dimensional prediction grid are extracted; the standardized environmental feature image patch and standardized depth value are input into the trained model to obtain the standardized prediction value of each grid point.

[0020] In one possible implementation, when using a weighted loss function, each sample is assigned a weight related to the true content of rare earth elements; the weights are multiplied by the base loss value of each sample to obtain a weighted loss value; and the model parameters are optimized based on the weighted loss value.

[0021] In one possible implementation, when de-standardizing the standardized predicted value, the mean and standard deviation of the rare earth element content in the training set are used to perform reverse calculation on the standardized predicted value to obtain the final rare earth element content prediction result with actual physical meaning.

[0022] Compared with the prior art, the beneficial effects of this invention are as follows: by integrating multi-source environmental data and depth data, and jointly encoding depth information input with spatial features, this invention can capture the nonlinear variation law of deep-sea rare earth concentration with depth, realize the quantitative prediction of rare earth content at any depth inside sediments, form a continuous and complete three-dimensional prediction result, and significantly improve the comprehensiveness and accuracy of the prediction.

[0023] Attention fusion mechanism can dynamically adjust the weights of spatial environment features based on depth information, achieve adaptive fusion of heterogeneous features, and improve the model's adaptability to different geological environments.

[0024] By using interpretability analysis techniques, we can quantify the contribution of each input feature to the prediction results, clearly identify key mineralization factors, and characterize the correlation between features. This not only improves the credibility of the prediction results but also provides a scientific basis for the study of mineralization mechanisms.

[0025] The entire technical solution realizes an end-to-end process from data preprocessing to model training and predictive analysis, which simplifies the operation of deep-sea rare earth resource exploration, reduces the blindness of offshore operations, helps to optimize the selection of exploration target areas, reduces exploration costs and environmental disturbance, and has outstanding engineering application value. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is an overall flowchart of the prediction method in this embodiment of the invention;

[0028] Figure 2 This is a schematic diagram of the hybrid neural network model structure in an embodiment of the present invention;

[0029] Figure 3 This is a scatter plot comparing the actual and predicted values ​​of the test set in an embodiment of the present invention.

[0030] Figure 4 The curves showing the change of the loss function between the training set and the validation set in this embodiment of the invention;

[0031] Figure 5 This is a histogram of the prediction error distribution of the test set in an embodiment of the present invention;

[0032] Figure 6 This is a summary diagram of the SHAP feature importance analysis in the embodiments of the present invention;

[0033] Figure 7 This is a spatial distribution map showing the predicted rare earth element content at different depths in the Pacific Ocean in this embodiment of the invention. Detailed Implementation

[0034] 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 this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0035] Example:

[0036] It should be noted that the terms "comprising" and "having" and any variations thereof in the embodiments of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0037] See Figure 1 This figure illustrates the complete workflow of this invention from data to final prediction results. The workflow consists of three stages: A. Data preparation and preprocessing: This stage is responsible for integrating, cleaning, aligning, and standardizing the original multi-source heterogeneous data (REY samples and environmental data); B. Model building and training: Convolutional Neural Networks (CNN) and Multilayer Perceptrons (MLP) are used to extract spatial and depth features respectively, and the features are fused and weighted through an attention mechanism, then backpropagation is used to optimize the model parameters; C. Prediction, application, and analysis: The trained model is applied to the target area to generate a 3D prediction grid and output the prediction results.

[0038] This invention provides a method for 3D prediction of deep-sea rare earth elements using a hybrid attention network that integrates spatial and depth data, comprising the following steps:

[0039] Step 101: Acquire multi-source environmental data covering the target sea area and depth data from sediment cores. Preprocess the two types of data to obtain standardized environmental feature image blocks and standardized depth values.

[0040] Specifically, during the preprocessing of the multi-source environmental data and depth data, the multi-source environmental data is first spatially aligned and interpolated to obtain environmental data with a uniform resolution; the depth data is then logarithmically transformed to obtain transformed depth values; and the uniform resolution environmental data and transformed depth values ​​are then standardized to obtain standardized environmental feature image blocks and standardized depth values.

[0041] For example, multi-source environmental data is represented as sample point data. The input format of the sample point data is tabular data, specifically in XLSX or CSV format. This tabular data contains complete information for N sample points. Each sample point corresponds to an independent set of basic data, including the sample point's geographic coordinates, seabed depth, and the actual measured content of the target geochemical element. The geographic coordinates are used to locate the specific spatial position of the sample in the deep sea area, the seabed depth reflects the vertical distribution of the sample in the sediment core, and the actual measured content serves as the target label data for model training.

[0042] For example, the input format of multi-source environmental data can be M independent geospatial environmental data files, all in NetCDF format, specifically nc format. These geospatial environmental data files together constitute a comprehensive environmental feature set describing the spatial distribution of geochemical elements.

[0043] The detailed data list is shown in the table below:

[0044]

[0045] For example, when spatially aligning environmental data, a unified target spatial resolution is first determined, and then bilinear interpolation is used to resample each environmental variable, ensuring that data from different sources correspond on the same spatial grid and avoiding spatial offsets caused by resolution differences.

[0046] For example, logarithmic transformation and standardization are performed to eliminate the differences in units between different data and to adjust the data distribution to make it more suitable for model learning. For deep data, logarithmic transformation is required first, as shown in equation (1):

[0047]

[0048]

[0049] in: : The transformed depth value Original sediment depth value Standardized data : Raw data (environmental variables, transformed depth or rare earth element content). Original dataset The mean, Original dataset The standard deviation.

[0050] Logarithmic transformation avoids undefined depth values ​​(e.g., seabed surface samples) and compresses the numerical range of depth data, making the data distribution closer to the model's fit. After depth data transformation, all environmental variables, transformed depth values, and target element content must be standardized using Z-score, with the corresponding formula... .

[0051] For example, when constructing an environmental feature image patch, the geographic coordinates of sample point i can be used as the center to extract local regions of size h multiplied by w (such as 5×5 pixels) from M standardized environmental data grids, and the regions corresponding to each environmental variable can be stacked along the channel dimension into a tensor of M multiplied by h multiplied by w, which is used to characterize the local environmental features around sample point i.

[0052] Step 102: Extract spatial feature vectors using the standardized environmental feature image blocks, and obtain depth feature vectors using the standardized depth value encoding.

[0053] Specifically, when extracting spatial feature vectors using the standardized environmental feature image blocks, the standardized environmental feature image blocks are subjected to convolution, batch normalization, and activation operations through a convolutional neural network, and the spatial feature vectors are obtained after pooling processing.

[0054] For example, the function of a convolutional neural network is to extract high-level spatial abstract features from an environmental feature image patch of dimension M × h × w. In a specific embodiment, the convolutional neural network can be composed of several 3×3 convolutional blocks, each of which extracts local spatial features in the order of Conv2D → BatchNorm2D → ReLU; if necessary, MaxPool2D can be added after some convolutional blocks for downsampling. After all convolutional processing, a spatial feature vector of fixed dimensions (e.g., 256 dimensions) is obtained through adaptive average pooling. .

[0055] The depth feature vector is obtained by performing linear transformation and nonlinear activation on the standardized depth values ​​through a multilayer perceptron.

[0056] For example, a multilayer perceptron may include several fully connected layers and nonlinear activation functions to map one-dimensional normalized depth values ​​to high-dimensional vectors. For instance, Linear (1→32), Tanh, Linear (32→64), and Tanh may be used sequentially to obtain a 64-dimensional depth feature vector. .

[0057] Step 103: The spatial feature vector and the depth feature vector are fused to obtain a weighted feature vector.

[0058] When fusing the spatial feature vector and the depth feature vector, the spatial feature vector and the depth feature vector are first concatenated along the feature dimension to obtain a combined feature vector; the combined feature vector is then linearly transformed and activated by an attention network to generate an attention weight vector; and the combined feature vector and the attention weight vector are multiplied element-wise to obtain the weighted feature vector.

[0059] See Figure 2 The environmental feature image patch is processed by a convolutional neural network to extract spatial feature vectors, and the depth value is encoded into a depth feature vector by a multilayer perceptron. The two types of features are concatenated in the fusion module and weights are generated by an attention network to achieve weighted processing. The resulting weighted feature vector is input into the prediction head to output the element content prediction value.

[0060] For example, concatenation along the feature dimension specifically involves combining spatial feature vectors. With deep feature vectors By directly concatenating the features along the feature dimension, a combined feature vector containing all the original information is formed. The corresponding calculation relationship is:

[0061] ;

[0062] in: : The concatenated combined feature vector Spatial feature vectors extracted by the CNN module. : Depth feature vectors extracted by the depth coding MLP module : Indicates splicing along the feature dimension.

[0063] For example, generating the attention weight vector specifically involves concatenating the combined feature vectors. Input attention network. Its computation process involves multiple steps, firstly through... The fully connected layer performs a first linear transformation on the combined feature vector Vcombined, mapping the features to an appropriate dimension. Then, ReLU is used to modify the linear unit activation function to process the result of the linear transformation, increasing the non-linearity of the feature representation and allowing the network to capture more complex feature relationships. Afterwards... The fully connected layer undergoes a second linear transformation to further adjust the feature dimensions; finally, the output value of the second linear transformation is compressed to (0,1) using the Sigmoid activation function to obtain the attention weight vector. .

[0064]

[0065] in: The generated attention weight vector has elements with values ​​between 0 and 1. , : Represents two fully connected layers (linear transformation); Modify the activation function of the linear unit to add nonlinearity; The Sigmoid activation function compresses the output value to the (0,1) interval so that it can be used as a weight.

[0066] For example, the weighted feature vector is obtained by multiplying the combined feature vector element-wise with the attention weight vector. Specifically, this involves generating the attention weight vector. Combined feature vector with the original Element-wise multiplication, also known as the Hadamard product, is performed to obtain the attention-weighted feature vector Vattended. The corresponding calculation relationship is as follows:

[0067]

[0068] in: : Feature vectors after attention weighting; : indicates element-wise multiplication (Hadamard product).

[0069] Step 104: Input the weighted feature vector into the prediction head module, and train the model using the training strategy to obtain the trained model.

[0070] Specifically, when training the model using the training strategy, the preprocessed dataset is divided into a training set, a validation set, and a test set; the loss between the predicted value and the true value is calculated using a basic loss function or a weighted loss function; the model parameters are adjusted through backpropagation using an optimizer and a learning rate scheduling strategy, and the model training is completed using an early stopping mechanism, resulting in the trained model.

[0071] When using a weighted loss function, each sample is assigned a weight related to the true content of rare earth elements; the weight is multiplied by the base loss value of each sample to obtain the weighted loss value; and the model parameters are optimized based on the weighted loss value.

[0072] For example, the function of the prediction head module is to transform the attention-weighted feature vector into the final prediction result. The prediction head module receives the weighted feature vector. The final prediction is output through a standard multilayer perceptron. In a typical embodiment, its structure can be Linear(d attended->128)->ReLU->Dropout(p=0.5)->Linear(128->64)->ReLU->Linear(64->1), finally outputting a single scalar value, namely the standardized target element content prediction value.

[0073] For example, before model training begins, all sample data points need to be randomly divided into a training set (approximately 75%), a validation set (approximately 10%), and a test set (approximately 15%).

[0074] For example, the base loss function can be the mean squared error (MSE) or the mean squared logarithmic error (MSLE). In this example, MSLE is used (Formula (6)).

[0075]

[0076] in: Mean squared logarithmic error (MSLE) loss value; Total number of samples; : The true target value (REY content) of the i-th sample; : The model's target prediction value for the i-th sample.

[0077] For example, to make the model focus more on samples with high content, the weighted loss function can assign a value to each sample i that corresponds to its true content. Related weights (Formula (7)). The final weighted loss is shown in Formula (8).

[0078]

[0079]

[0080] in: The weight of the i-th sample. : The preset maximum weight hyperparameter is used to control the weighting degree of samples with high content. : The true REY content of the i-th sample The maximum REY content across all samples in the training set. The final weighted loss value. The weight of the i-th sample (calculated by formula (7)). : The base loss value of the i-th sample (e.g., the MSLE calculated by formula (6)).

[0081] Step 105: Use the trained model to perform inference and prediction on the target area data to obtain standardized prediction values.

[0082] Specifically, when using the trained model to perform inference and prediction on target area data, a three-dimensional prediction grid of the target area is generated according to a preset geographical range, depth range, and resolution; the standardized environmental feature image patch and standardized depth value corresponding to each grid point in the three-dimensional prediction grid are extracted; the standardized environmental feature image patch and standardized depth value are input into the trained model to obtain the standardized prediction value of each grid point.

[0083] For example, when generating the prediction grid, the preset geographical range, depth range, and resolution are first determined based on the actual needs of the target exploration area. The geographical range needs to cover the deep-sea area to be explored, the depth range needs to include strata where rare earth elements may be enriched in sediment cores, and the resolution needs to balance prediction accuracy and computational efficiency. Based on the determined parameters, a regular three-dimensional discrete grid is generated, with each grid point corresponding to a unique set of geographical coordinates (lon). j ,lat j depth k These grid points together constitute the complete prediction space covering the target area.

[0084] For each grid point in the 3D prediction grid, the corresponding environmental feature image patch and depth value are first extracted. The method for extracting the environmental feature image patch is consistent with the training phase, i.e., a fixed-size image patch is extracted from the standardized environmental data grid centered on the geographic coordinates of the grid point. The extracted depth value is the actual vertical depth of the grid point in the sediment core. Subsequently, the extracted environmental feature image patch and depth value are standardized using the exact same statistical parameters as in the training phase, ensuring that the processed input data format completely matches the input format used during model training. The standardized environmental feature image patch and depth value are then input into the trained hybrid neural network model. The model outputs the standardized prediction value corresponding to the grid point through internal calculations. This value is the model's preliminary prediction result of the rare earth element content at that grid point.

[0085] Step 106: Perform destandardization on the standardized predicted values ​​to obtain the final rare earth element content prediction results.

[0086] Specifically, when performing destandardization on the standardized predicted values, the mean and standard deviation of the rare earth element content in the training set are used to perform reverse calculation on the standardized predicted values ​​to obtain the final rare earth element content prediction result with actual physical meaning.

[0087] For example, the denormalization operation requires calling the rare earth element content statistical parameters retained during the training phase, reversing the normalized predicted values ​​output by the model to eliminate the influence of the normalization process on the data, and obtaining predicted content values ​​with actual physical meaning. This value can directly reflect the true content level of rare earth elements at the grid points. The predicted content values ​​of all grid points are combined according to their corresponding three-dimensional coordinate positions to form a complete three-dimensional predicted data volume. This data volume can be stored as a NetCDF format file for easy subsequent visualization analysis and data retrieval.

[0088] Optionally, to facilitate researchers and exploration personnel's intuitive understanding of the information in the 3D prediction data volume, the results of the 3D prediction data volume need to be visualized. The most common visualization method is to slice along the depth axis, select the depth layer of interest according to the analysis requirements, and for each selected depth layer, map the predicted content values ​​of all grid points in that depth layer onto a 2D plane to generate a 2D planar map. In the planar map, different colors or contour lines are used to represent the content distribution of the target element at that depth. The shade of the color or the density of the contour lines can intuitively reflect the content value. The higher the content value, the darker the color or the denser the contour lines, allowing users to quickly identify the rare earth element enrichment areas at that depth layer.

[0089] Optionally, to enhance the transparency and credibility of model prediction results and avoid model predictions becoming an unexplainable black box operation, this invention further integrates SHAP (SHapley Additive exPlanations) technology, which can be used to quantitatively interpret model prediction results.

[0090] Its general additivity interpretation model can be expressed as:

[0091]

[0092] in: : An interpretive model used to approximate the output of the original complex model; : A simplified binary input vector, where each element represents the presence or absence of a feature; The total number of input features; The baseline output of the model, i.e. the predicted value when there are no features input (usually the mean of the predicted values ​​of all training samples). : The contribution of the i-th feature to the final prediction result, i.e., the SHAP value.

[0093] In the REY content prediction task of this invention, for any prediction point j, its final predicted value is REY. pred,j SHAP can be decomposed into the sum of the baseline value and the SHAP values ​​of all input features (depth and M environment variables):

[0094]

[0095] in: : The model's final REY content prediction for the j-th prediction point; The average of the predicted values ​​of all training samples, used as the baseline; : The contribution of deep features to the predicted value of the j-th point; The contribution of the i-th environmental variable feature to the predicted value of the j-th point; The total number of environment variables.

[0096] By calculating the SHAP value of each feature, we can clearly identify the globally important features that have the greatest impact on model predictions. For example, we can clarify which depth features and environmental variables are key factors affecting rare earth element content. We can also understand the positive or negative impact of feature values ​​on prediction results, such as determining whether an increase in a certain environmental variable will lead to an increase or decrease in the predicted content value. Using these analytical results, we can verify whether the model's predictions conform to known geological laws. If the analysis results are consistent with geological theories, it indicates that the model's predictions are scientific and reliable. If we discover feature influence relationships that differ from existing geological theories, it can provide new scientific insights into the study of deep-sea rare earth mineralization mechanisms, promoting the improvement and development of theories in related fields.

[0097] To verify the effectiveness of the method of the present invention, a series of experiments were conducted to test the application effect of the model in real-world scenarios. The following will show the results of a specific experiment in detail. The test objective of this experiment was to predict the rare earth element content in the Pacific Ocean. This experiment can intuitively demonstrate the predictive performance of the method of the present invention.

[0098] The input features are the 15 environmental variables defined above; the model structure adopts the aforementioned hybrid neural network, and the key hyperparameters are set as follows: CNN filter is [64, 256, 256], the output dimension of the deep MLP is 64, and the attention bottleneck factor is 2. The MSLE loss function is used during training, along with a linear weighting mechanism (maximum weight 5.0). The dataset is split into 75% for training, 10% for validation, and 15% for testing.

[0099] On a completely independent test set, the model's coefficient of determination (R²Score) in this embodiment is approximately 0.82, indicating that the model can explain most of the variations in REY content in the test set and has good predictive performance.

[0100] Figure 3 The model visually presents the comparison between predicted and actual values ​​on the test set, with the scatter points closely distributed around the diagonal of y=x, further confirming the model's high accuracy.

[0101] See Figure 3 , Figure 3 This is a scatter plot of the actual and predicted values ​​on the test set. The scatter plot presents the model's predictive performance on completely independent test sets. Each point in the plot represents a test sample, with the x-axis representing the actual measured REY content and the y-axis representing the model's predicted REY content. Ideally, all points should fall on the diagonal of y=x (represented by the dashed line in the plot). As can be seen from the plot, the scatter points are tightly clustered around the diagonal, indicating a high degree of consistency and correlation between the model's predictions and the actual values. The test set coefficient of determination (R²Score) is 0.8235, meaning the model can explain approximately 82% of the variance in the test data, further confirming the model's high predictive accuracy.

[0102] To further diagnose the model, its training process and error distribution were analyzed. Figure 4 The graph presents the curves showing the change of the loss function on the training and validation sets over the training period. The graph clearly shows that both curves decrease smoothly and eventually converge to a low level, with no significant separation or rebound between the validation loss and the training loss. This fully demonstrates that the training strategies employed in this invention (such as early stopping, learning rate scheduling, and regularization) are effective, the model training process is stable, serious overfitting is successfully avoided, and the model's generalization ability is ensured.

[0103] Figure 4 The figure shows the loss function curves during the training process. It illustrates how the loss function value changes with each training epoch. The blue curve represents the training loss, and the orange curve represents the validation loss. Both curves show a steady decline and eventual convergence, indicating that the model effectively learns the patterns in the data during training. More importantly, there is no significant separation or rebound between the validation loss and the training loss. This confirms that the regularization, learning rate scheduling, and early stopping strategies employed in this invention successfully prevent overfitting and ensure the model possesses good generalization ability.

[0104] Figure 5The graph illustrates the distribution of the residuals between predicted and actual values ​​on the test set. It clearly shows that the distribution of the residuals (i.e., prediction errors) approximates a normal distribution with a mean of zero, with the vast majority of errors concentrated around zero, showing no significant bias or skewness. This strongly suggests that the model's predictions are unbiased, and the errors are randomly generated, not stemming from systematic flaws. An unbiased and random error distribution is an important indicator of the performance of a regression model.

[0105] Figure 5 This plot shows the prediction residuals for the test set. It illustrates the distribution of the model's prediction error (i.e., the residuals, which are the actual values ​​minus the predicted values) on the test set. The horizontal axis represents the residual values, and the vertical axis represents the sample frequency or density corresponding to those residual values. As can be seen from the plot, the residual distribution approximates a normal distribution centered at 0 (a bell-shaped curve), with the prediction errors for the vast majority of samples concentrated around 0. This distribution characteristic indicates that the model's predictions are unbiased, meaning there is no systematic overestimation or underestimation trend, and the error exhibits randomness. This is an important indicator of the performance of a regression model.

[0106] The SHAP technique was used to conduct an in-depth interpretability analysis of the model. Figure 6 The global impact of each input feature on the model's predictions is presented. The analysis shows that depth is the most important single feature affecting predictions, with shallower depths contributing positively to higher REY content. Regarding environmental features, vertical gravity gradient, total organic carbon content, and bottom water flow velocity difference are the most influential factors. The SHAP analysis provides direct evidence for the effectiveness of the proposed adaptive screening mechanism. Figure 6 As shown, the model, after training, has learned to consistently and primarily focus its attention on the "depth" feature itself, which plays a decisive role in mineralization, as well as key environmental variables closely related to it, such as "water depth" and "sediment thickness." This indicates that the invented attention module is not an unexplainable "black box," but has successfully learned an attention allocation pattern that conforms to the laws of Earth science and has practical significance, thus verifying the scientific nature and reliability of its decision-making process.

[0107] Figure 6This is a summary plot of SHAP feature importance. (a) Deep features, (b) Contextual features. This plot uses the SHAP (SHapley Additive exPlanations) method to present the global impact of each input feature on the model's prediction results. Each row in the plot represents a feature, arranged from top to bottom according to its importance (average absolute SHAP value). Each point represents a sample, and the color of the point indicates the original value of that feature in that sample (red represents high value, blue represents low value). The position of the point on the horizontal axis indicates the contribution of that feature value to the current prediction (positive value indicates a higher prediction, negative value indicates a lower prediction).

[0108] To further demonstrate the application effect of the method of the present invention, Figure 7 This paper presents spatial distribution maps of REY content predicted by the model at three different depths (0m, 3m, and 6m) below the seabed within the global Pacific Ocean. Figure 7 It can be clearly seen that the enriched regions of REY have significant spatial clustering and depth variability.

[0109] At a depth of 0 meters, high-concentration areas (bright yellow areas in the figure) are mainly concentrated in the southeastern Pacific basin. At a depth of 3 meters, the distribution pattern of REY changes somewhat. New high-concentration hotspots appear near the Mariana Trench in the western Pacific and in the Clarion-Clipperton Fault Zone (CCZ) region in the northeastern Pacific, indicating vertical migration and re-enrichment of the element. At a depth of 6 meters, the enrichment trend in the southern Pacific basin is more significant and concentrated, forming large-scale contiguous high-value areas, while the high-value areas in the central and western Pacific are relatively weakened or shifted.

[0110] This series of prediction maps not only visually demonstrates the complex distribution patterns of target elements in three-dimensional space, but also verifies the powerful ability of the model in this invention to capture such depth-dependent spatial heterogeneity. The model successfully learned and expressed the nonlinear trend of REY content with depth under different geological units and environmental conditions. Its prediction results are highly consistent with known geochemical exploration patterns and geological background, providing crucial, high-resolution three-dimensional data support for delineating mineral exploration target areas and conducting resource assessment.

[0111] Figure 7 This is a spatial distribution map of predicted REY content at three different depths (0 meters (a), 3 meters (b), and 6 meters (c)) below the seabed in a portion of the Pacific Ocean, as an embodiment of the invention.

[0112] The above method steps can be implemented by computer program instructions. These instructions can be provided to the processor of a general-purpose computer, special-purpose computer, or other programmable data processing device to generate a program, such that the instructions, executed by the processor, create means for implementing the functions / steps specified in the flowchart.

[0113] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0114] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A hybrid attention network method for 3D prediction of deep-sea rare earth elements, integrating spatial and depth data, characterized in that... Includes the following steps: Acquire multi-source environmental data covering the target sea area and vertical depth data from sediment cores. Preprocess the multi-source environmental data and depth data to obtain standardized environmental feature image patches and standardized depth values. A convolutional neural network is used to extract spatial feature vectors from the standardized environmental feature image patch, and a multilayer perceptron is used to encode the standardized depth values ​​to obtain depth feature vectors. The spatial feature vector and the depth feature vector are fused to obtain a weighted feature vector. Specifically, the spatial and depth feature vectors are first concatenated along their feature dimensions to obtain a combined feature vector. Then, an attention network is used to perform a linear transformation and activation process on the combined feature vector to generate an attention weight vector. Finally, the combined feature vector and the attention weight vector are multiplied element-wise to obtain the weighted feature vector. The weighted feature vector is input into the prediction head module, and the model is trained in combination with the training strategy to obtain the trained model. The trained model is used to infer and predict data in the target area to obtain standardized prediction values. Specifically, when using the trained model to infer and predict data in the target area, a three-dimensional prediction grid for the target area is generated based on a preset geographical range, depth range, and resolution. Standardized environmental feature image patches and standardized depth values ​​corresponding to each grid point in the three-dimensional prediction grid are extracted. The standardized environmental feature image patches and standardized depth values ​​are then input into the trained model to obtain the standardized prediction value for each grid point. The standardized predicted values ​​are destandardized to obtain the final rare earth element content prediction results.

2. The method according to claim 1, characterized in that, When preprocessing the multi-source environmental data and depth data, the multi-source environmental data is first spatially aligned and interpolated to obtain environmental data with uniform resolution; the depth data is then logarithmically transformed to obtain transformed depth values; and the uniform resolution environmental data and transformed depth values ​​are then standardized to obtain standardized environmental feature image blocks and standardized depth values.

3. The method according to claim 1, characterized in that, When extracting spatial feature vectors using the standardized environmental feature image blocks, the standardized environmental feature image blocks are subjected to convolution, batch normalization, and activation operations through a convolutional neural network, and the spatial feature vectors are obtained after pooling processing.

4. The method according to claim 1, characterized in that, When obtaining the depth feature vector using the standardized depth value encoding, the standardized depth value is subjected to multiple linear transformations and activation processes through a multilayer perceptron to obtain the depth feature vector.

5. The method according to claim 1, characterized in that, When training the model using the training strategy, the preprocessed dataset is divided into a training set, a validation set, and a test set; the loss between the predicted value and the true value is calculated using a basic loss function or a weighted loss function; the model parameters are adjusted through backpropagation using an optimizer and a learning rate scheduling strategy, and the model training is completed using an early stopping mechanism, resulting in the trained model.

6. The method according to claim 5, characterized in that, When using a weighted loss function, each sample is assigned a weight related to the true content of rare earth elements; the weight is multiplied by the base loss value of each sample to obtain the weighted loss value; and the model parameters are optimized based on the weighted loss value.

7. The method according to claim 1, characterized in that, When de-standardizing the standardized predicted values, the mean and standard deviation of the rare earth element content in the training set are used to perform reverse calculation on the standardized predicted values ​​to obtain the final rare earth element content prediction result with actual physical meaning.

Images