A neural network-based approach to interpreting multi-parameter contributions in rock physics models

By training a fully connected network using a neural network, the contribution of each parameter in the rock physics model is explained, solving the problem of difficulty in separating reservoir physical parameters and improving the accuracy of reservoir prediction and the ranking of parameter importance.

CN116168783BActive Publication Date: 2024-11-05CENT SOUTH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310173344.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-02-28
Publication Date
2024-11-05
Estimated Expiration
2043-02-28

AI Technical Summary

Technical Problem

Existing rock physics models struggle to isolate reservoir physical parameters, leading to complex seismic wave propagation and low reservoir prediction accuracy.

Method used

A neural network-based approach is adopted. By training a fully connected network, using activation functions and batch normalization, a loss function and optimizer are defined to prevent overfitting, predict the P-wave velocity and the inverse quality factor of attenuation dissipation, and interpret the contribution of each parameter through neuron weight information.

Benefits of technology

This study enabled a qualitative interpretation of the effects of various parameters in the rock physics model on the prediction results, improved the accuracy of reservoir prediction, and determined the importance ranking of the sensitivity parameters.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116168783B_ABST
    Figure CN116168783B_ABST
Patent Text Reader

Abstract

The present invention provides a method for interpreting the contribution of multiple parameters in a rock physics model based on a neural network. The method comprises: first, using a neural network as a proxy for a rock physics model to predict the P-wave velocity and the attenuation dissipation inverse quality factor; then, using the network parameters saved by a trained neural network to extract the neuron weight information involved in each input parameter, and individually calculating the predicted value of each parameter; then, taking the difference between the predicted result obtained by adding each new parameter and the predicted result without adding the parameter as the contribution of the parameter, accumulating and summing up all the contributions involved in the parameter to obtain the contribution of the parameter to the final predicted result; finally, by summing up the contributions of all the parameters, the original neural network prediction result can be obtained; thus, the contribution of each parameter in the rock physics model represented by the neural network to the prediction result can be indirectly explained.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of rock physics and artificial intelligence, and in particular to a method for interpreting multi-parameter contributions in a rock physics model based on a neural network. Background Art

[0002] In oil and gas exploration, seismic exploration is usually used to obtain the location and structure information of underground reservoirs by using the difference in propagation speed and attenuation and dissipation laws of seismic waves in different underground media, and then further invert and interpret the reservoir physical parameters based on the seismic elastic parameters to finally achieve reservoir prediction. In fact, the velocity dispersion and attenuation dissipation of seismic waves are caused by the flow of fluids in the pores when seismic waves pass through porous media. In order to better establish the connection between seismic elastic parameters and reservoir physical parameters, the establishment of a rock physics model is crucial to studying the propagation laws of seismic waves.

[0003] However, the propagation of seismic waves is affected by many factors. The diversity of reservoir physical properties causes the complexity of seismic wave propagation, and the two are characterized by a strong nonlinear many-to-many mapping. Although specific rock physics models can be established for certain specific problems and assumptions to solve the wave propagation speed and attenuation dissipation, it is still impossible to separate a certain physical property parameter from the rock physics model to obtain its contribution to the calculation results. According to experience, the solid skeleton of reservoir lithology and its porosity are the key points involved in reservoir prediction. Explaining the importance of related physical property parameters is the key to improving the accuracy of reservoir prediction.

[0004] Neural networks provide new ideas and methods for rock physics analysis. They realize nonlinear mapping between input and output through linear superposition of multiple neurons, which is very suitable for solving complex strong nonlinear problems. Therefore, they can be used as proxy rock physics models to predict P-wave velocity and attenuation dissipation inverse quality factor. Moreover, compared with complex rock physics models, simple linear superposition is easier to separate a certain parameter, so that by finding a certain calculation method based on neural networks, the contribution of each input parameter to the predicted P-wave velocity and attenuation dissipation inverse quality factor can be obtained, and further, an indirect interpretation of the rock physics model can be achieved. Summary of the invention

[0005] In order to solve the above technical problems or at least partially solve the above technical problems, the present invention provides a method for interpreting multi-parameter contributions in a rock physics model based on a neural network, the method comprising the following steps:

[0006] Step 1: Prepare data, randomly generate a batch of rock physics data, calculate the P-wave velocity and attenuation dissipation inverse quality factor according to the rock physics model, normalize each input and output parameter, and divide the whole into training set and validation set in proportion;

[0007] Step 2: Train the network, use a fully connected network, add activation functions and batch normalization to the hidden layer, define the loss function, select the optimizer, determine the initial learning rate and learning rate update strategy, add regularization and early stopping to prevent overfitting, and use the validation set to test the network after each training.

[0008] Step 3, predicting the results, normalizing the new rock physics data to be predicted and inputting them into the trained neural network, realizing the prediction of the P-wave velocity and the attenuation dissipation inverse quality factor by the neural network proxy rock physics model, and performing inverse normalization restoration on the prediction results;

[0009] Step 4: Extract weights. Take all input parameters as features, obtain the full set of features by combining them, extract the neuron weight information containing each feature subset based on the trained neural network, and calculate the predicted value corresponding to each feature subset separately.

[0010] Step 5: Calculate the contribution. The difference between the prediction result obtained when a new feature is added to the subset and the prediction result obtained without adding the feature is the local contribution of the feature to the prediction result. When all local contributions of a certain input parameter are summed up, it is the global contribution of the parameter to the prediction result.

[0011] Step 6, explanation and verification, summing up all global contributions, we can get the values ​​of P-wave velocity and attenuation dissipation inverse quality factor originally predicted by the neural network, which can indirectly explain the contribution of each parameter in the rock physics model represented by the neural network to the predicted P-wave velocity and attenuation dissipation inverse quality factor.

[0012] The present invention is further configured as follows: the rock physics model refers to the classical rock physics models of White, Johnson, Tuncay, CRM and BR.

[0013] The present invention is further configured as follows: the input parameters refer to the seismic wave frequency (f), rock porosity (φ), solid bulk modulus (K S ), solid shear modulus (μ S ), solid density (ρ), fluid saturation (S) and permeability (κ).

[0014] The present invention is further configured as follows: the normalization method refers to using min-max standardization to normalize each input and output parameter to the range of [0, 1] or [-1, 1].

[0015] The present invention is further configured as follows: the activation function refers to ReLU, Tanh, Sigmoid and Leaky ReLU.

[0016] The present invention is further configured as follows: the loss function refers to MAE loss, MSE loss and Huber loss.

[0017] The present invention is further configured as follows: the optimizer refers to SGD, RMSprop, Adam and AdamW.

[0018] The present invention is further configured as follows: the learning rate update strategy refers to StepLR, MultiStepLR, ExponentialLR and ReduceLROnPlateau combined with the early stopping method.

[0019] The present invention is further configured such that: the prediction result of the neural network is within the allowable error range, and the average error is usually less than 1%.

[0020] The present invention is further configured as follows: the full set of features also includes an empty set, and the prediction value corresponding to the empty set can be replaced by the mean value of the longitudinal wave velocity and the attenuation dissipation inverse quality factor in the obtained output result by randomly selecting part of the training data to input into the network.

[0021] The present invention is further configured such that the local contribution and global contribution of the parameters are both represented by partial values ​​of the predicted longitudinal wave velocity and the attenuation dissipation inverse quality factor, and therefore have the same units as the longitudinal wave velocity and the attenuation dissipation inverse quality factor.

[0022] The present invention is further configured as follows: all global contributions also include the contribution of the empty set. At this time, the contribution of the empty set serves as a reference value for predicting the longitudinal wave velocity and the attenuation dissipation inverse quality factor. The contribution of all parameters can be interpreted as a disturbance to the reference value, that is, on the basis of the reference value, the disturbances of all parameters are accumulated and summed to obtain the final predicted value.

[0023] The present invention is further configured as follows: since the benchmark value is the mean of a randomly selected portion of training data, the value will vary to a certain extent each time prediction is made, but it does not affect the qualitative interpretation of the contribution of each parameter, in other words, it does not affect the importance ranking of each parameter to the prediction result.

[0024] The present invention is further configured such that: the interpretation method is a qualitative interpretation method or a local quantitative interpretation method, and is not a global quantitative interpretation method.

[0025] The method of the present invention predicts the P-wave velocity and the attenuation dissipation inverse quality factor with the help of a neural network proxy rock physics model. At the same time, the contribution of each input parameter is calculated in the form of a newly added feature prediction difference based on the weight information in the trained neural network, thereby realizing the interpretation of the importance of each input parameter in the output result and solving the problem that the traditional rock physics model cannot solve. It indirectly explains the contribution of each parameter in the rock physics model to the predicted P-wave velocity and the attenuation dissipation inverse quality factor, and locks the sensitivity parameters to improve the reservoir prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0027] Figure 1 Schematic diagram of the method of the present invention.

[0028] Figure 2 Flow chart of the method of the present invention.

[0029] Figure 3 A schematic diagram of a fully connected neural network provided in an embodiment of the present invention.

[0030] Figure 4 A schematic diagram illustrating a process of extracting neural network weight calculation provided by an embodiment of the present invention.

[0031] Figure 5 A schematic diagram of the error between the prediction result and the true value provided by an embodiment of the present invention.

[0032] Figure 6 A schematic diagram of the error between the interpretation result provided in an embodiment of the present invention and the true value.

[0033] Figure 7 A schematic diagram of the error between the interpretation result and the prediction result provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0034] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] Figure 1Schematic diagram of a method for interpreting multi-parameter contributions in a rock physics model based on a neural network according to an embodiment of the present invention. Figure 2 As shown, the method specifically comprises the following steps:

[0036] Step 1: Prepare data, randomly generate a batch of rock physics data, calculate the P-wave velocity and attenuation dissipation inverse quality factor according to the rock physics model, normalize each input and output parameter, and divide the whole into training set and validation set in proportion;

[0037] In the embodiment of the present invention, the rock physics model described in step 1 is the White model, and the input parameters include seismic wave frequency (f), rock porosity (φ), solid bulk modulus (K S ), solid shear modulus (μ S ), solid density (ρ), fluid saturation (S w ) and permeability (κ), and their random ranges are defined as [10 2 ,10 8 ], [0.001, 0.4], [20, 60], [8, 50], [1500, 3000], [10 -5 ,1-10 -5 ],[10 -16 ,10 -9 ], where the frequency (f) and permeability (κ) are taken logarithmically, and the range becomes [2,8] and [-16,-9]; the output parameter is the longitudinal wave velocity (V P ) and the attenuation dissipation inverse quality factor (Q -1 )2 parameters, whose random ranges are defined as [0,9000], [10 -40 ,1], where the logarithm of the attenuation dissipation inverse quality factor is taken, and the range becomes [-40,0]. On this basis, each input parameter x and output parameter y is min-max standardized respectively, so that all parameters are normalized to [-1,1], that is:

[0038]

[0039]

[0040] Finally, 3.6 million sets of data were generated through the White model, and the data were divided into training set and validation set in a ratio of 9:1.

[0041] Step 2: Train the network, use a fully connected network, add activation functions and batch normalization to the hidden layer, define the loss function, select the optimizer, determine the initial learning rate and learning rate update strategy, add regularization and early stopping to prevent overfitting, and use the validation set to test the network after each training.

[0042] In the embodiment of the present invention, the fully connected neural network described in step 2 is as follows: Figure 3 As shown, the input layer size is 7, the output layer size is 2, the middle hidden layer contains 5 fully connected layers, each layer contains 128 neurons, and then passes through the ReLU activation function and batch normalization layer; the loss function adopts Huber loss, that is, smoothed mean absolute error:

[0043]

[0044] The optimizer adopts AdamW. Since the current training is for large amounts of data, the use of AdamW optimization is equivalent to adding L2 regularization during Adam optimization, which can effectively avoid overfitting during training; the initial learning rate adopts 0.001, and the learning rate update strategy adopts the adaptively adjusted ReduceLROnPlateau to better adjust the learning rate according to the actual decrease in training and validation losses, thereby improving training efficiency and avoiding training from falling into local extremes; the early stopping method is also to prevent the network from overfitting. The current network parameters will be saved when and only when the validation loss decreases, so as to ensure that the network saved when training stops can achieve the best test effect.

[0045] Step 3, predicting the results, normalizing the new rock physics data to be predicted and inputting them into the trained neural network, realizing the prediction of the P-wave velocity and the attenuation dissipation inverse quality factor by the neural network proxy rock physics model, and performing inverse normalization restoration on the prediction results;

[0046] In the embodiment of the present invention, the new rock physical data x to be predicted in step 3 will be normalized in the same way as in step 1. The prediction result y′ obtained after inputting into the network is the numerical result after normalization in step 1. It is also necessary to perform inverse normalization on y′ to obtain y, that is:

[0047]

[0048] Step 4: Extract weights. Take all input parameters as features, obtain the full set of features by combining them, extract the neuron weight information containing each feature subset based on the trained neural network, and calculate the predicted value corresponding to each feature subset separately.

[0049] In the embodiment of the present invention, since the calculation of the seven input parameters is relatively cumbersome and repetitive, it is inconvenient to present all of them in the specification. Here, the methods described in steps 4, 5, and 6 are introduced using three input parameters f, φ, and ρ as a simplified example. At this time, the full set of features described in step 4 is:

[0050]

[0051] The subset containing feature f is:

[0052] F={{f}, {f, φ}, {f, ρ}, {f, φ, ρ}}

[0053] The subset containing feature φ is:

[0054] Φ={{φ}, {f, φ}, {φ, ρ}, {f, φ, ρ}}

[0055] The subset containing feature ρ is:

[0056] P={{ρ}, {f, ρ}, {φ, ρ}, {f, φ, ρ}}

[0057] In order to illustrate the process of extracting neural network weight information, here we take Figure 4 The neural network shown is a simplified example. At this time, the neurons involved in the subset {f} and their weights are:

[0058] First layer:

[0059] y1=G1(W f1 f),

[0060] y2=G2(W f2 f),

[0061] y3=G3(W f3 f),

[0062] y4=G4(W f4 f),

[0063] Second layer:

[0064] y5=G5(W 15 y1+W 25 y2+W 35 y3+W 45 y4),

[0065] y6=G6(W 16 y1+W 26 y2+W 36 y3+W 46 y4),

[0066] y7=G7(W 17 y1+W 27 y2+W 37 y3+W 47 y4),

[0067] Third layer:

[0068] y8=G8(W 58 y5+W 68 y6+W78 y7),

[0069] y9=G9(W 59 y5+W 69 y6+W 79 y7),

[0070] Obviously, the prediction result corresponding to the subset {f} is:

[0071] V P =y8,

[0072] Q -1 =y9

[0073] Similarly, the neurons involved in the subset {f, φ} and their weights are:

[0074] First layer:

[0075] y1=G1(W f1 f+W φ1 φ),

[0076] y2=G2(W f2 f+W φ2 φ),

[0077] y3=G3(W f3 f+W φ3 φ),

[0078] y4=G4(W f4 f+W φ4 φ),

[0079] Second layer:

[0080] y5=G5(W 15 y1+W 25 y2+W 35 y3+W 45 y4),

[0081] y6=G6(W 16 y1+W 26 y2+W 36 y3+W 46 y4),

[0082] y7=G7(W 17 y1+W 27 y2+W 37 y3+W 47 y4),

[0083] Third layer:

[0084] y8=G(W 58 y5+W 68 y6+W78 y7),

[0085] y9=G9(W 59 y5+W 69 y6+W 79 y7),

[0086] Obviously, the prediction result corresponding to the subset {f, φ} is:

[0087] V P =y8,

[0088] Q -1 =y9

[0089] The neurons involved in the subset {f, ρ} and their weights are:

[0090] First layer:

[0091] y1=G1(W f1 f+W ρ1 ρ),

[0092] y2=G2(W f2 f+W ρ2 ρ),

[0093] y3=G3(W f3 f+W ρ3 ρ),

[0094] y4=G4(W f4 f+W ρ4 ρ),

[0095] Second layer:

[0096] y5=G5(W 15 y1+W 25 y2+W 35 y3+W 45 y4),

[0097] y6=G6(W 16 +y1+W 26 +y2+W 36 y3+W 46 y4),

[0098] y7=G7(W 17 y1+W 27 y2+W 37 y3+W 47 y4),

[0099] Third layer:

[0100] y8=G8(W 58 y5+W 68 y6+W78 y7),

[0101] y9=G9(W 59 y5+W 69 y6+W 79 y7),

[0102] Obviously, the prediction result corresponding to the subset {f, ρ} is:

[0103] V P =y8,

[0104] Q -1 =y9

[0105] The neurons involved and their weights of the subset {f, φ, ρ} are:

[0106] First layer:

[0107] y1=G1(W f1 f+W ρ1 ρ+W φ1 φ),

[0108] y2=G2(W f2 f+W ρ2 ρ+W φ2 φ),

[0109] y3=G3(W f3 f+W ρ3 ρ+W φ3 φ),

[0110] y4=G4(W f4 f+W ρ4 ρ+W φ4 φ),

[0111] Second layer:

[0112] y5=G5(W 15 y1+W 25 y2+W 35 y3+W 45 y4),

[0113] y6=G6(W 16 y1+W 26 y2+W 36 y3+W 46 y4),

[0114] y7=G7(W 17 y1+W 27 y2+W 37 y3+W 47 y4),

[0115] Third layer:

[0116] y8=G8(W 58 y5+W 68 y6+W 78 y7),

[0117] y9=G9(W 59 y5+W 69 y6+W 79 y7),

[0118] Obviously, the prediction result corresponding to the subset {f, φ, ρ} is:

[0119] V P =y8,

[0120] Q -1 =y9

[0121] It should be emphasized here that the prediction results corresponding to the output of the network weight information are all numerical results after normalization in step 1. Therefore, these numerical results need to be inversely normalized in step 3 before calculating the contribution.

[0122] Step 5: Calculate the contribution. The difference between the prediction result obtained when a new feature is added to the subset and the prediction result obtained without adding the feature is the local contribution of the feature to the prediction result. When all local contributions of a certain input parameter are summed up, it is the global contribution of the parameter to the prediction result.

[0123] In the embodiment of the present invention, taking feature f as an example, the contribution described in step 5 is mainly divided into subsets The feature f is added to the subset {f}, the feature f is added to the subset {φ} to become the subset {f, φ}, the feature f is added to the subset {ρ} to become the subset {f, ρ}, and the feature f is added to the subset {φ, ρ} to become the subset {f, φ, ρ}. In order to facilitate the explanation of the contribution calculation method, the following results are not normalized, so there is no need to perform inverse normalization. P For example, suppose that the empty set corresponds to V P The reference value is V P_base =3600m / s, and assume that the prediction results of each subset after extracting the weight information are as follows:

[0124] When only one feature is included:

[0125] y8(f)=600m / s

[0126] y8(φ)=2100m / s,

[0127] y8(ρ)=1200m / s

[0128] When two features are included:

[0129] y8(f,φ)=3300m / s,

[0130] y8(f,ρ)=1800m / s,

[0131] y8(φ,ρ)=2850m / s

[0132] When all three features are included:

[0133] y8(f,φ,ρ)=3750m / s

[0134] Obviously, the longitudinal wave velocity V of the neural network p The prediction result is V P_pred =3750m / s.

[0135] Assuming that the probability of adding each new feature is equal, then for the subset The probability of adding feature f to become a subset {f} is Then the local contribution of this part of feature f is:

[0136]

[0137] At this time, the negative sign represents that the disturbance of this part to the baseline value is a negative contribution.

[0138] Similarly, to add feature f to the subset {φ} to become the subset {f, φ}, we first need to Add feature φ to become the subset {φ}, the probability of this process is still Next, the probability of adding feature f and feature ρ to the subset {φ} is also equal. The probability of adding feature f to the subset {φ} to become the subset {f, φ} is Then the local contribution of this part of feature f is:

[0139]

[0140] At this time, the positive sign represents that the disturbance of this part to the baseline value is a positive contribution.

[0141] Similarly, when feature f is added to subset {ρ} to become subset {f, ρ}, the local contribution of this part of feature f is:

[0142]

[0143] The probability that the subset {φ, ρ} becomes the subset {f, φ, ρ} by adding feature f is 1, and the subset {φ, ρ} can be transformed into the subset {f, φ, ρ} by adding feature f. First add feature φ to become subset {φ} and then add feature ρ and subset First add feature ρ to become the subset {ρ} and then add feature φ. Then the local contribution of this part of feature f is:

[0144]

[0145] Finally, the feature f in the new data predicts the longitudinal wave velocity V P The global contribution of is:

[0146] C f =C f1 +C f2 +C f3 +C f4 =-400m / s

[0147] Using the above method to calculate the characteristic contribution, we can obtain the characteristic φ in the new data to predict the longitudinal wave velocity V P The global contribution of is:

[0148] C φ =C φ1 +C φ2 +C φ3 +C φ4 =+875m / s,

[0149] And the characteristic ρ in the new data predicts the longitudinal wave velocity V P The global contribution of is:

[0150] C ρ =C ρ1 +C ρ2 +C ρ3 +C ρ4 =-325m / s

[0151] Step 6, explanation and verification, summing up all global contributions, we can get the values ​​of P-wave velocity and attenuation dissipation inverse quality factor originally predicted by the neural network, which can indirectly explain the contribution of each parameter in the rock physics model represented by the neural network to the predicted P-wave velocity and attenuation dissipation inverse quality factor.

[0152] In the embodiment of the present invention, all global contributions described in step 6 include not only the global contributions of each feature calculated in step 5, but also the contribution V of the empty set. P_base ,Right now:

[0153] C=V P_base +C f +C φ +C ρ =3750m / s

[0154] At this time, C=V P_pred, which verifies the correctness of the interpretation results. At the same time, according to the obtained characteristic contribution values, it can be seen that the characteristics f and ρ in the new data are negative contributions, and the characteristic φ is a positive contribution. According to the absolute value of the contribution value, it can be seen that the three parameters in the new data have a significant impact on the prediction of the longitudinal wave velocity V P The order of importance is: φ>f>ρ.

[0155] Similar calculation processes will not be described in detail in this embodiment. It should be noted again that all numerical values ​​in the above embodiments are assumed values, and the obtained prediction results and interpretation results are only used for method description.

[0156] The following is a set of actual results of the complete implementation of the method of the present invention. The results are generated using 36,000 sets of data newly generated by the White model. Figure 2 The neural network gets V P and Q -1 The prediction result Y P ', after the reverse normalization reduction in the above embodiment, V P and Q -1 The prediction result Y P The error with the true value Y is Figure 5 As shown in , the average errors are 0.27% and 0.41%, both less than 1%, meeting the accuracy requirement. P and Q -1 Explanation of the results Y C ′, at this time due to

[0157]

[0158] The inverse normalization process needs to be decomposed as follows:

[0159]

[0160] After this inverse normalization reduction, V P and Q -1 Explanation of the results Y C The error with the true value Y is Figure 6 As shown, the average errors are 0.27% and 0.41%, both less than 1%, which also meets the accuracy requirements. P and Q -1 Explanation of the results Y C And the predicted result Y P The error is as Figure 7 As shown, the error between the two is basically 0, that is, the interpretation result can be verified. Finally, the method provided by the present invention achieves the indirect interpretation of the contribution of each parameter in the rock physics model represented by the neural network to the prediction of the P-wave velocity and the attenuation dissipation inverse quality factor.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for interpreting multi-parameter contributions in a rock physics model based on a neural network, characterized in that: The following steps are involved: Step 1: Prepare data, randomly generate a batch of rock physics data, calculate the P-wave velocity and attenuation dissipation inverse quality factor according to the rock physics model, normalize each input and output parameter, and divide the whole into training set and validation set in proportion; Step 1: The rock physics model is White, Johnson, Tuncay, CRM or BR; the input parameters include seismic wave frequency , rock porosity , solid bulk modulus K S , solid shear modulus μ S , solid density ρ, fluid saturation S w and permeability κ; the normalization method is min-max normalized linear function normalization; Step 2: Train the network, use a fully connected network, add activation functions and batch normalization to the hidden layer, define the loss function, select the optimizer, determine the initial learning rate and learning rate update strategy, add regularization and early stopping to prevent overfitting, and use the validation set to test the network after each training. Step 3, predicting the results, normalizing the new rock physics data to be predicted and inputting them into the trained neural network, realizing the prediction of the P-wave velocity and the attenuation dissipation inverse quality factor by the neural network proxy rock physics model, and performing inverse normalization restoration on the prediction results; Step 4: Extract weights. Take all input parameters as features, obtain the full set of features by combining them, extract the neuron weight information containing each feature subset based on the trained neural network, and calculate the predicted value corresponding to each feature subset separately. Step 5: Calculate the contribution. The difference between the prediction result obtained when a new feature is added to the subset and the prediction result obtained without adding the feature is the local contribution of the feature to the prediction result. When all local contributions of a certain input parameter are summed up, it is the global contribution of the parameter to the prediction result. Step 6, explanation and verification, sum up all global contributions to obtain the values ​​of P-wave velocity and attenuation dissipation inverse quality factor originally predicted by the neural network, which can indirectly explain the contribution of each parameter in the rock physics model represented by the neural network to the predicted P-wave velocity and attenuation dissipation inverse quality factor.

2. A method for interpreting multi-parameter contributions in a rock physics model based on a neural network according to claim 1, characterized in that: In step 2, the activation function is ReLU, Tanh, Sigmoid or Leaky ReLU; the loss function is MAE loss, MSE loss or Huber loss; the optimizer is SGD, RMSprop, Adam or AdamW; the learning rate update strategy is StepLR, MultiStepLR, ExponentialLR or ReduceLROnPlateau.

3. A method for interpreting multi-parameter contributions in a rock physics model based on a neural network according to claim 1, characterized in that: The allowed range of the prediction result error in step 3 is an average error of less than 1%.

4. A method for interpreting multi-parameter contributions in a rock physics model based on a neural network according to claim 1, characterized in that: The full set of features described in step 4 also includes an empty set. At this time, the predicted value corresponding to the empty set is replaced by the mean value of the longitudinal wave velocity and the attenuation dissipation inverse quality factor in the obtained output result by randomly selecting part of the training data to input into the network.

5. The method for interpreting multi-parameter contributions in a rock physics model based on a neural network according to claim 1, characterized in that: The local contribution and global contribution of the parameters described in step 5 are both represented by partial values ​​of the predicted P-wave velocity and the attenuation dissipation inverse quality factor, and have the same units as the P-wave velocity and the attenuation dissipation inverse quality factor.

6. A method for interpreting multi-parameter contributions in a rock physics model based on a neural network according to claim 1 or 5, characterized in that: All global contributions also include the contribution of the empty set. At this time, the contribution of the empty set is used as the benchmark value for predicting the longitudinal wave velocity and the attenuation dissipation inverse quality factor. The contribution of all parameters is the disturbance to the benchmark value, that is, on the basis of the benchmark value, the disturbances of all parameters are accumulated and summed to obtain the final predicted value.

7. A method for interpreting multi-parameter contributions in a rock physics model based on a neural network according to claim 6, characterized in that: The reference value is the mean of a randomly selected portion of training data.