Method for predicting cross-band rock physical parameters under pressure condition

Abstract

The invention provides a cross-band rock physical parameter prediction method under a pressure condition. According to the method, unified and accurate prediction of key parameters such as longitudinal wave velocity, transverse wave velocity, Young modulus, Poisson's ratio and the like under a full-band and multi-pressure condition from a static state, namely a low frequency, to a dynamic state, namely a high frequency is realized by utilizing strong nonlinear mapping and feature correlation capturing capability of a deep neural network and a Transform architecture. The method effectively overcomes the limitation of traditional experiments and models, and provides complete and reliable rock physical parameter support for geophysical inversion and engineering design.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rock physics, geophysical exploration and bottom acoustic, in particular to a method for predicting rock physical parameters across frequency bands under pressure conditions, and more particularly to a method for predicting the physical parameters such as the longitudinal wave velocity, the transverse wave velocity, the Young's modulus and the Poisson's ratio of rocks across frequency bands under pressure conditions. BACKGROUND

[0002] Accurate acquisition of rock physical parameters is crucial for oil and gas reservoir evaluation and water body target detection. Rock physical parameters (such as wave velocity and elastic modulus) exhibit significant frequency dependence (dispersion) and pressure sensitivity. Traditional experimental measurement methods have limitations: the cost of measurement across a large frequency band (such as Hz to MHz) is high and the period is long; the data points under different pressures and different frequencies are sparse, making it difficult to form a continuous and complete parameter atlas. Existing empirical models or theories (such as the Biot theory) can describe some of the rules, but it is difficult to accurately depict the nonlinear response of multiple parameter coupling under complex geological conditions.

[0003] Machine learning methods provide a new way for rock physical parameter prediction, but traditional models often lack accuracy and generalization ability when dealing with complex nonlinear relationships across frequency bands, multiple pressures and multiple parameter couplings. In particular, how to build a unified model to achieve simultaneous high-precision prediction of multiple rock physical parameters under a wide frequency range of 6-7 orders of magnitude and different pressure conditions is still a technical challenge currently faced. SUMMARY

[0004] The present application proposes a method for predicting rock physical parameters across frequency bands under pressure conditions to address the technical problem that existing technologies cannot accurately and efficiently obtain rock physical parameters under wide frequency bands and different pressure conditions.

[0005] The present application is achieved by the following technical solutions. The present application proposes a method for predicting rock physical parameters across frequency bands under pressure conditions, which comprises: Step one, collect rock physical parameter experimental data under cross-frequency and multiple pressure conditions, and construct a standardized data set; Step two, clean, integrate and feature engineering process the data in the data set, specifically including logarithmic transformation of frequency, normalization processing of pressure and target parameters, and one-hot encoding of rock samples and saturation state; Step three, design and build a hybrid deep learning prediction model, which adopts differential architecture according to the characteristics of the prediction parameters; specifically: for the longitudinal wave velocity and the transverse wave velocity, a multi-input deep neural network model is constructed, which processes rock samples, pressure and frequency information through independent feature processing subnetworks, and performs deep nonlinear fitting after the fusion layer; for the Young's modulus and Poisson's ratio, a double-input model based on the Transformer mechanism is constructed, which deeply excavates the complex interaction and long-range dependence between pressure and frequency features using the self-attention mechanism; Step four, train and verify the hybrid deep learning prediction model using a loss function with physical constraints and optimization strategies to ensure that the model's prediction results conform to physical laws and have generalization ability; Step five, use the trained model to input the specified rock type, pressure value and frequency value, and output the predicted full-frequency rock physical parameter curve.

[0006] Further, the multi-input deep neural network model has a multi-input layer structure and can accept rock sample information processed by one-hot encoding, pressure values processed by normalization, and frequency values processed by logarithmic transformation and normalization.

[0007] Further, each input is first processed by a corresponding feature subnetwork; specifically: rock sample features are processed by two fully connected networks for dimension reduction; pressure features are processed by a fully connected network for transformation; frequency features are processed by two fully connected networks to capture nonlinear performance.

[0008] Further, the features processed by the subnetworks are connected in a fusion layer to form a feature representation containing all input features, which is then sent to a deep network containing three fully connected layers. The number of neurons in the fully connected layer network decreases gradually, and batch normalization and dropout layers are set in the network to prevent overfitting and improve network stability and model generalization ability. The single neuron value of the model output is used to represent the wave velocity value.

[0009] Further, the double-input model based on the Transformer mechanism, in which the double-input layer receives pressure values and frequency values, and then maps the original pressure and frequency to higher-quality representations in a 64-dimensional space through a feature embedding layer.

[0010] Further, in the process of encoding by the Transformer encoder, the embedded feature representation is converted into a sequence form and a self-attention mechanism is used; the encoder adopts a 4-way head attention mechanism, which can simultaneously place the stress and frequency features on different paths and exchange them, so that the correlation between the stress and frequency features can be viewed from different angles; the residual connection and layer normalization are used in the encoder to ensure that the model training process is smoother and more stable, and the output of the self-attention mechanism is further processed by a layer of feedforward network.

[0011] Further, the sequence feature processed by the Transformer encoder is converted into a fixed-length vector representation by global average pooling, and then input into an output network composed of three fully connected layers, and batch normalization and dropout regularization are added to the output network to output the predicted Young's modulus value or Poisson's ratio value.

[0012] Further, in the model training process, an early stopping strategy is used, and if there is no improvement for 50 consecutive cycles on the validation set, the training process is terminated in advance to avoid overfitting; a learning rate decay mechanism is used, and if the validation loss does not improve for 20 consecutive cycles, the current learning rate is halved to allow the model to make more fine adjustments in the later stage; and the weight with the best performance in each training cycle is saved during training, and it is used for final prediction and evaluation after training.

[0013] The application further provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method for predicting rock physical parameters across frequency bands under pressure conditions when executing the computer program.

[0014] The application further provides a computer readable storage medium for storing computer instructions, wherein the computer instructions implement the steps of the method for predicting rock physical parameters across frequency bands under pressure conditions when executed by a processor.

[0015] The application has the following beneficial effects: The present application aims to provide a cross-frequency rock physical parameter prediction method under pressure conditions. The present application proposes a differentiated modeling strategy for different rock physical parameter characteristics by constructing a special deep learning model that integrates rock properties, pressure and frequency information. This technology uses the powerful non-linear mapping and feature association capture ability of deep neural networks and Transformer architecture to achieve unified and accurate prediction of key parameters such as P-wave velocity, S-wave velocity, Young's modulus, and Poisson's ratio from static (low frequency) to dynamic (high frequency) under full frequency and multi-pressure conditions. This method effectively overcomes the limitations of traditional experiments and models, providing complete and reliable rock physical parameter support for geophysical inversion and engineering design. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0017] Figure 1 A flowchart of the rock physical parameter prediction method under pressure conditions described in the present application; Figure 2 A longitudinal wave Transformer model regression data statistical chart for four rock samples. The chart shows the performance of the longitudinal wave velocity prediction model on the test set. Through the scatter distribution and fitting of the predicted value and the measured value, the high precision of the model is shown.

[0018] Figure 3 A lateral wave Transformer model regression data statistical chart for four rock samples. The chart shows the performance of the lateral wave velocity prediction model on the test set. Similarly, through the scatter distribution of the predicted value and the measured value, the excellent prediction ability of the model for lateral wave velocity is verified.

[0019] Figure 4 A Transformer model regression data statistical chart for sample 1 Young's modulus. The chart shows the performance evaluation results of the Young's modulus prediction model for sample 1 on the test set.

[0020] Figure 5 A Transformer model regression data statistical chart for sample 2 Young's modulus. The chart shows the performance evaluation results of the Young's modulus prediction model for sample 2 on the test set. ).

[0021] Figure 6Dry sample 1 Poisson's ratio Transformer model regression data plot. This plot shows the performance evaluation results of the Poisson's ratio prediction model for sample 1 in dry state on the test set.

[0022] Figure 7 Oil-saturated sample 1 Poisson's ratio Transformer model regression data plot. This plot shows the performance evaluation results of the Poisson's ratio prediction model for sample 1 in oil-saturated state on the test set.

[0023] Figure 8 Dry sample 2 Poisson's ratio Transformer model regression data plot. This plot shows the performance evaluation results of the Poisson's ratio prediction model for sample 2 in dry state on the test set.

[0024] Figure 9 Oil-saturated sample 2 Poisson's ratio Transformer model regression data plot. This plot shows the performance evaluation results of the Poisson's ratio prediction model for sample 2 in oil-saturated state on the test set.

[0025] Figure 10 Transformer model regression statistical data plot for Poisson's ratio of each sample. This plot comprehensively compares the performance indicators of the Poisson's ratio prediction model under different samples and saturation states ).

[0026] Figure 11 Longitudinal wave velocity prediction plot for rock sample S1 under different pressures. This plot shows a family of curves of the model-predicted longitudinal wave velocity of sample S1 as a function of frequency, with each curve corresponding to a specific pressure condition, intuitively presenting the prediction ability across frequency bands and multiple pressures.

[0027] Figure 12 Transverse wave velocity prediction plot for rock sample S1 under different pressures. This plot shows a family of curves of the model-predicted transverse wave velocity of sample S1 as a function of frequency, also showing the dispersion characteristics under different pressures.

[0028] Figure 13 Cross-frequency prediction curve for the Young's modulus of sample 1. This plot shows the curve of the model-predicted Young's modulus of sample 1 as a function of frequency, embodying the continuous change of parameters within the full frequency band.

[0029] Figure 14 Cross-frequency prediction curve for the Young's modulus of sample 2. This plot shows the curve of the model-predicted Young's modulus of sample 2 as a function of frequency.

[0030] Figure 15 Cross-frequency prediction curve for the Poisson's ratio of dry sample 1. This plot shows the curve of the model-predicted Poisson's ratio of sample 1 in dry state as a function of frequency.

[0031] Figure 16 Cross-band prediction curve for Poisson's ratio of oil-saturated sample 1. This graph shows the model-predicted curve of Poisson's ratio of sample 1 as a function of frequency in the oil-saturated state.

[0032] Figure 17 Cross-band prediction curve for Poisson's ratio of dry sample 2. This graph shows the model-predicted curve of Poisson's ratio of sample 2 as a function of frequency in the dry state.

[0033] Figure 18 Cross-band prediction curve for Poisson's ratio of oil-saturated sample 2. This graph shows the model-predicted curve of Poisson's ratio of sample 2 as a function of frequency in the oil-saturated state. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0035] In combination Figures 1-18 , the present application proposes a cross-band rock physical parameter prediction method under pressure conditions, which comprises: Step one, collecting rock physical parameter experimental data under cross-band, multi-pressure conditions, and constructing a standardized data set; Step two, cleaning, integrating and feature engineering processing the data in the data set, specifically including logarithmic transformation of frequency, normalization processing of pressure and target parameters, and one-hot encoding of rock samples and saturation state; Step three, designing and constructing a hybrid deep learning prediction model, which adopts a differentiated architecture according to the characteristics of the prediction parameters; specifically, for P-wave velocity and S-wave velocity, a multi-input deep neural network model is constructed, which processes rock sample, pressure and frequency information through independent feature processing subnetworks, and performs deep nonlinear fitting after the fusion layer; for Young's modulus and Poisson's ratio, a double-input model based on the Transformer mechanism is constructed, which deeply mines the complex interaction and long-range dependence between pressure and frequency features using the self-attention mechanism; Step four, training and verifying the hybrid deep learning prediction model using a loss function with physical constraints and an optimization strategy to ensure that the prediction results of the model conform to the physical laws and have good generalization ability; Step five, using the trained model, input the specified rock type, pressure value and frequency value, output the predicted full-band rock physical parameter curve.

[0036] The main principle followed in the design of the model is: although each rock mechanics parameter (wave velocity, modulus, Poisson's ratio) has commonality but also differences, and each given parameter is an independent module, so under the condition of using a unified model architecture framework, some adjustments can be made locally for a certain parameter; in the process of predicting the longitudinal wave velocity and the transverse wave velocity, the influence of the combined action of rock samples, pressure and frequency needs to be considered, while in the prediction of Young's modulus and Poisson's ratio, more emphasis is placed on the influence of pressure and frequency, so the model structure that pays more attention to the frequency-parameter relationship modeling is selected. In addition, the size of the experimental data set is not very large (only about 480 points), so the designed model cannot be too complex to avoid overfitting problems, and it is also hoped that the model has strong generalization ability.

[0037] In order to complete the prediction task of longitudinal wave velocity and transverse wave velocity, a deep neural network model with multiple inputs is constructed. The multiple input deep neural network model has a multi-input layer structure, which can accept rock sample information input after one-hot encoding processing, pressure value input after normalization processing and frequency value input after logarithmic transformation and normalization processing. Each input is first processed by the corresponding feature subnetwork; specifically: rock sample features are processed by two fully connected networks for dimension reduction; pressure features are processed by a fully connected network for transformation; frequency features are processed by two fully connected networks to capture non-linear performance.

[0038] After subnetwork processing, the features are connected in a fusion layer to form a feature expression form containing all input features, which will then be sent to a deep network containing three fully connected layers. The number of neurons in the fully connected layer network is gradually decreasing (64-32-16), and batch normalization and dropout layers are set in the network to prevent overfitting problems and improve the stability of the network and the generalization ability of the model. The single neuron value of the model output is used to represent the wave velocity value of the target. This structure helps to realize the input of various features and their interaction, and to a certain extent, it can also achieve the purpose of wave velocity prediction. This structure of multiple inputs is conducive to the processing of information in each dimension and its interaction, and is very suitable for wave velocity prediction tasks that need to fuse multiple features and integrate various information.

[0039] For the prediction of Young's modulus and Poisson's ratio, a dual-input model structure based on the Transformer mechanism is proposed due to the cross-action of frequency and pressure. The dual-input layer receives the pressure value (normalized) and the frequency value (log-transformed and normalized), respectively, and then maps the original pressure and frequency to a higher-quality representation in a 64-dimensional space through a feature embedding layer (full connection layer).

[0040] During the encoding process of the Transformer encoder, the embedded feature representation is converted into a sequence form and the self-attention mechanism is used. The encoder uses a 4-way head attention mechanism that can place the pressure and frequency features on different paths and exchange them, allowing for a different perspective on the relationship between pressure and frequency. The encoder uses residual connections and layer normalization to ensure a smoother and more stable training process. The output of the self-attention mechanism is then processed through a feed-forward network, which often involves dimension expansion and compression operations to change the dimensions of the vectors.

[0041] The sequence features processed by the Transformer encoder are converted into a fixed-length vector representation using global average pooling, and then input into an output network consisting of three fully connected layers. Batch normalization and dropout regularization are added to the output network to output the predicted Young's modulus or Poisson's ratio. The network structure using the Transformer self-attention mechanism can capture long-range complex relationships between input features, making it suitable for predicting parameters that are sensitive to both frequency and pressure.

[0042] Developing a reasonable training method is an important prerequisite for the model to learn effective things from data and obtain stable results. The full data set is randomly divided into training and test sets at a ratio of 8:2, which allows the use of samples not present in the training set for testing and evaluation. The loss functions for the three models are differentiated. For the wave velocity prediction model, the mean squared error is chosen as the loss function. For the Young's modulus prediction model and the Poisson's ratio prediction model, the Huber loss is used instead of the mean squared error loss function. The reason for this is that the Huber loss is less sensitive to outliers than the MSE. When there are many abnormal points in the sample, using the Huber loss can make the model more robust.

[0043] The AdamW optimizer was selected, which features some minor improvements over the original Adam optimizer and often achieves better generalization results in practical tests. The initial learning rate was 0.001, ensuring training speed and stable performance. To control training and prevent problems, several training control measures were implemented: an early stopping strategy was used, terminating training early if there was no improvement on the validation set for 50 consecutive epochs to avoid overfitting; a learning rate decay mechanism was used, halving the learning rate if the validation loss did not improve for 20 consecutive epochs, allowing for more refined adjustments later; and the best-performing weights (model checkpoints) from each training epoch were saved and used for final prediction and evaluation after training. Because various mechanical parameters and different rock samples may vary, models may not achieve good generalization. Therefore, this invention employs a classification training method: training the model separately for each mechanical parameter and different rock samples (Young's modulus and Poisson's ratio prediction).

[0044] Multiple evaluation metrics are used to comprehensively and objectively measure the performance of the trained model. MSE represents the root mean squared error, which measures the average of the squared errors between the actual and predicted values. MAE refers to the mean absolute error, which is the mean of the absolute values ​​of the differences between the actual and predicted values. This refers to the variance explained by the model for the target variable, and the closer it is to 1, the better. We then analyze whether the distribution of the prediction error (i.e., the residual) has a strong regularity to determine whether there are problems with the model. By combining the results of different evaluation indicators, we evaluate the prediction effect of the model from different perspectives. Finally, we select the optimization method according to the actual requirements.

[0045] The trained model was applied to a reserved test set, and the main validation results show that the model achieved satisfactory performance in transverse and longitudinal wave prediction. All models... The values ​​are all above 0.98, indicating that the model fits the data very well and has strong predictive ability. Specifically, the prediction models for P-wave velocity and S-wave velocity... The values ​​reached 0.9899 and 0.9961 respectively. The prediction model also performed well on Young's modulus; the model prediction for sample 1 was [value missing]. =0.8346, while the model prediction for sample 2 is... =0.9752, which may be attributed to the Transformer model's advantage in capturing the complex interaction between stress and frequency. However, the model performs poorly in predicting Poisson's ratio: the prediction for four sets of Poisson's ratio data... The difference is significant, ranging from 0.8718 for the best and 0.5964 for the worst.

[0046] Embodiments The present application will be described in detail below with reference to the accompanying drawings and embodiments.

[0047] As Figure 1 shown, the present application is a cross-frequency rock physical parameter prediction method under pressure conditions, comprising the following steps: (1) Data preparation: Collect experimental data from 4 kinds of rock samples, measure their P-wave velocity, S-wave velocity, Young's modulus and Poisson's ratio at multiple pressure points from 5 MPa to 30 MPa, and frequency range from 2 Hz to 1 MHz, and record the saturation state (dry / water saturation / oil saturation) at the same time. Clean the data and remove abnormal points.

[0048] (2) Feature engineering: Logarithmic transformation is performed on the frequency. Z-score normalization is performed on the pressure value, transformed frequency value and each target parameter (wave velocity, modulus). The rock sample number and saturation state are one-hot encoded.

[0049] (3) Model construction and training: For P-wave velocity and S-wave velocity prediction, the multi-input DNN model is constructed. The processed rock sample code, pressure value and frequency value are respectively input into the corresponding subnetwork, and the predicted wave velocity is output after fusion through the deep network. Train using mean square error loss and AdamW optimizer. For Young's modulus and Poisson's ratio prediction, construct a Transformer model. The processed pressure value and frequency value are respectively embedded and input into the Transformer encoder, and the self-attention mechanism is used to capture feature interaction, and finally the predicted modulus or Poisson's ratio is output. Train using Huber loss and AdamW optimizer to enhance robustness. Early stopping strategy and learning rate decay are used in training.

[0050] (4) Model verification: Use the reserved test set data to verify the model. As Figures 2 to 10 shown, the P-wave and S-wave velocity prediction are both higher than 0.98, the Young's modulus prediction is as high as 0.9752, and the Poisson's ratio prediction performance varies with samples and states.

[0051] (5) Parameter prediction: input the sample type, target pressure value P and target frequency range (logarithmic transformation and normalization) of the rock to be measured into the trained corresponding model, and the model can output the continuous rock physical parameter prediction curve across the entire frequency range under the pressure P, as Figures 11 to 18 shown.

[0052] The application further provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method for predicting rock physical parameters across frequency bands under pressure conditions.

[0053] The application further provides a computer readable storage medium for storing computer instructions, wherein the computer instructions are executed by a processor to implement the steps of the method for predicting rock physical parameters across frequency bands under pressure conditions.

[0054] The memory in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. The non-volatile memory can be a read only memory (ROM), a programmable read only memory (PROM), an erasable programmable read only memory (EPROM), an electrically EPROM (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as a static random access memory (SRAM), a dynamic random access memory (DRAM), a synchronous dynamic random access memory (SDRAM), a double data rate SDRAM (DDR SDRAM), an enhanced SDRAM (ESDRAM), a synchlink DRAM (SLDRAM), and a direct rambus RAM (DR RAM). It should be noted that the memory of the method described in the present application is intended to include, but not limited to, these and any other suitable types of memory.

[0055] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (such as floppy disk, hard disk, magnetic tape), optical media (such as high-density digital video disc (digital video disc, DVD)), or semiconductor media (such as solid state disc (solid state disc, SSD)) and the like.

[0056] In the implementation process, each step of the above method can be completed by integrated logic circuit of hardware in the processor or instruction in the form of software. The steps of the method disclosed in the embodiments of the present application can be directly embodied as hardware processor execution or executed by combination of hardware and software modules in the processor. The software module can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, register or other mature storage medium in the art. The storage medium is located in the memory, and the processor reads the information in the memory and combines the hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0057] It should be noted that the processor in the embodiments of the present application can be an integrated circuit chip with a signal processing capability. In the implementation process, each step of the method embodiments can be completed by the integrated logic circuit of hardware in the processor or the instructions in the form of software. The processor mentioned above can be a general processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. Each method, step and logic block diagram disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register or other mature storage medium in the art. The storage medium is located in the memory, and the processor reads the information in the memory and combines the hardware to complete the steps of the above method.

[0058] The above describes in detail the method for predicting cross-band rock physical parameters under pressure conditions according to the present application. The principles and implementation manners of the present application are described by using specific examples. The above embodiment is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for predicting cross-frequency rock physical parameters under pressure conditions, characterized in that, The method includes: Step 1: Collect experimental data on rock physical parameters under multiple frequency bands and pressure conditions, and construct a standardized dataset; Step 2: Clean, integrate, and perform feature engineering on the data in the dataset. Specifically, this includes logarithmic transformation of the frequency, normalization of the pressure and target parameters, and unique thermal encoding of the rock samples and saturation state. Step 3: Design and construct a hybrid deep learning prediction model. This model adopts a differentiated architecture based on the characteristics of the prediction parameters. Specifically, for P-wave velocity and S-wave velocity, a multi-input deep neural network model is constructed. This model processes rock samples, pressure, and frequency information separately through independent feature processing subnetworks, and performs deep nonlinear fitting after the fusion layer. For Young's modulus and Poisson's ratio, a dual-input model based on the Transformer mechanism is constructed. This model utilizes a self-attention mechanism to deeply explore the complex interaction relationship and long-range dependence between pressure and frequency features. Step 4: Train and validate the hybrid deep learning prediction model using a loss function and optimization strategy with physical constraints to ensure that the model's prediction results conform to physical laws and have generalization ability. Step 5: Using the trained model, input the specified rock type, pressure value, and frequency value, and output the predicted full-band rock physical parameter curve.

2. The method according to claim 1, characterized in that, The multi-input deep neural network model has a multi-input layer structure and can accept rock sample information that has undergone unique thermal encoding, pressure value input that has undergone normalization, and frequency value input that has undergone logarithmic transformation and normalization.

3. The method according to claim 2, characterized in that, Each input is first processed by a corresponding feature sub-network; specifically: rock sample features are dimensionality reduced through a two-layer fully connected network; pressure features are transformed through a one-layer fully connected network; and frequency features are captured by a two-layer fully connected network to capture nonlinear characteristics.

4. The method according to claim 3, characterized in that, The features processed by the sub-network are connected in a fusion layer to form a feature representation that includes all input features. After that, it will be fed into a deep network containing three fully connected layers. The number of neurons in the fully connected layer network gradually decreases. Batch normalization and dropout layers are set in the network to prevent overfitting and improve the stability of the network and the generalization ability of the model. The single neuron value of the model output is used to characterize the magnitude of the target wave velocity value.

5. The method according to claim 1, characterized in that, The dual-input model based on the Transformer mechanism has two input layers that receive pressure and frequency values ​​respectively. Then, the original pressure and frequency are mapped to a higher quality representation in 64-dimensional space through feature embedding layers.

6. The method according to claim 5, characterized in that, During the Transformer encoder encoding process, the embedded feature representation is converted into a sequence form and a self-attention mechanism is used. The encoder adopts a 4-way head attention mechanism, which can simultaneously place pressure and frequency features on different paths and cross-interchange them, so that the relationship between pressure and frequency features can be viewed from different perspectives. The encoder uses residual connections and layer normalization to ensure that the model training process is smoother and more stable, and the output of the self-attention mechanism will continue to be processed by a feedforward network layer.

7. The method according to claim 6, characterized in that, The sequence features processed by the Transformer encoder are transformed into fixed-length vector representations through global average pooling, and then input into an output network consisting of three fully connected layers. Batch normalization and dropout regularization are added to the output network to output the predicted Young's modulus or Poisson's ratio.

8. The method according to claim 1, characterized in that, During model training, an early stopping strategy is used: if there is no improvement on the validation set for 50 consecutive epochs, the training process is terminated early to avoid overfitting. A learning rate decay mechanism is used: if the validation loss does not improve for 20 consecutive epochs, the current learning rate is halved, allowing the model to be finely adjusted later. In addition, the weights with the best model performance in each training epoch are saved during training and used for final prediction and evaluation after training.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-8.

Images