A method for optimizing an initial model in seismic data inversion

Abstract

The application belongs to the technical field of seismic exploration, and particularly relates to a seismic data inversion initial model optimization method, which comprises the following steps: randomly generating initial positions of all agents, simulating a cooperative mechanism of a seismic source and a geophone through a seismic exploration optimizer, randomly selecting part of agents in a population to be set as seismic source agents, and setting the rest of the agents as geophone agents, traversing each seismic source agent and non-seismic source agent, evaluating the fitness of each agent, and updating a global optimal agent; for each non-optimal agent, calculating a new moving direction, updating the position of the agent according to the new moving direction and a step factor, adopting an S wave propagation strategy, calculating a vibration factor and updating the position of the agent; continuing iteration, and when a maximum iteration number is reached or a convergence precision is met, terminating the cycle, and outputting the global optimal agent position and the fitness value thereof. The method can ensure that the agents are efficiently evaluated and optimized in the parameter space between the seismic source agents and the geophone agents.

Description

Technical Field

[0001] This application belongs to the field of seismic exploration technology, specifically a method for optimizing the initial model for seismic data inversion. Background Technology

[0002] Full waveform inversion (FWI) is a high-precision seismic data inversion method widely used for detailed exploration of subsurface media properties. Due to the nonlinear characteristics of seismic wave propagation, FWI is highly dependent on the initial model. Existing initial model-based inversion processes suffer from low convergence speed and low accuracy, making them prone to getting trapped in local minima, increasing computational costs, and resulting in distortion between the obtained subsurface structure image and the actual structure. Summary of the Invention

[0003] The technical problem to be solved by this application is to provide an initial model optimization method for seismic data inversion, which can solve the problem that the initial model is prone to getting trapped in local minima and that there is distortion between the underground structure image and the real structure.

[0004] An initial model optimization method for seismic data inversion according to an embodiment of this application includes:

[0005] Initialize the parameters of the seismic exploration optimizer, randomly generate the initial locations of all agents, evaluate the fitness of each agent, and mark the globally optimal agent;

[0006] The seismic exploration optimizer simulates the collaborative mechanism between the seismic source and the geophone. Some agents in the population are randomly selected and set as seismic source agents, while the rest are set as geophone agents. Each seismic source agent and non-seismic source agent is traversed, the fitness of each agent is evaluated, and the global optimal agent is updated.

[0007] For each non-optimal agent, calculate the direction vector pointing to the global optimal agent. Calculate a new movement direction using a random direction and the direction vector pointing to the global optimal agent. Update the agent's position based on the new movement direction and a step factor, where the step factor is calculated based on the direction vector pointing to the global optimal agent.

[0008] The S-wave propagation strategy is adopted to traverse each agent and determine whether the agent's position is the globally optimal agent. If not, the oscillation factor is calculated and the agent's position is updated. If it is, the original globally optimal agent position is overwritten.

[0009] Continue iterating until the maximum number of iterations is reached or the convergence accuracy is met. Then, terminate the loop and output the globally optimal agent position and its fitness value.

[0010] Furthermore, the parameters of the seismic exploration optimizer are initialized, including: setting the search space of the problem as... Dimension, define the number of agents as The position of each agent is represented as follows: ,in, Indicates the current iteration number Next The agent in the first The position of the dimension = The initial position is randomly generated within the given upper and lower bounds: in, For the first The lower bound of dimensionality For the first The upper bound of the dimension, Indicates the interval Uniform random numbers on the surface Indicates the current iteration number Next The agent in the first The initial position of the dimension.

[0011] Furthermore, the collaborative mechanism between the seismic source and the detector is simulated using a seismic exploration optimizer, including:

[0012] Based on the current iteration number and maximum number of iterations Determine the number of earthquake source proxies;

[0013] Each source agent interacts with all detector agents to construct new agents.

[0014] Furthermore, the calculation of the new movement direction using a random direction vector and a direction vector pointing to the globally optimal agent includes:

[0015] Randomness weights and deterministic weights are calculated to dynamically adjust the randomness and determinism in the search process;

[0016] Normalized random direction vector;

[0017] Calculate the new direction of movement: ,in, For random weights, For deterministic weights, This is a normalized random direction vector. This is the direction vector pointing to the globally optimal agent. It represents the Euclidean norm of the direction vector pointing to the globally optimal agent.

[0018] Furthermore, the agent's position is updated based on the new direction of movement and step factor, including:

[0019] Calculate the step size factor: ,in, Step size factor Update according to the following formula: , To adjust the parameters, used to control the changing trend of the step size factor, as the iteration progresses, A smooth transition from near 0 to near 1, Indicates the maximum number of iterations. Indicates the current iteration number;

[0020] Update the proxy location: , The location of the updated proxy. The location of the proxy before the update. For the new direction of movement.

[0021] Furthermore, the randomness weights are calculated as follows: The deterministic weights are calculated as follows: , Indicates the maximum number of iterations. This indicates the current iteration number.

[0022] Furthermore, the global optimal agent is decayed to obtain the current optimal agent.

[0023] Furthermore, for each suboptimal agent, the adjustment factor is calculated: ;

[0024] Calculate the agent's movement position using the adjustment factor: , The location of the agent after the move. As the globally optimal proxy, For the current agent location, when The fitness value is better than , As The position of the next iteration is updated.

[0025] Compared with the prior art, the advantages of this application are as follows:

[0026] This application's method optimizes the search process by introducing a cooperative mechanism and a moderate degree of randomness. First, it begins by randomly initializing a set of agents, combining random sampling and scattering movement strategies to introduce significant randomness, ensuring thorough exploration of the entire parameter space. This approach effectively avoids the trap of local optima, especially in later iterations, where the search process gradually shifts from global exploration to refined local development by gradually reducing the weight of randomness and increasing the weight of determinism. Furthermore, this method introduces periodic perturbations through an S-wave propagation strategy, effectively preventing agents from stagnating near a local optimum. Compared to traditional strategies that rely solely on the parent population, this method can more flexibly and adaptively transmit effective information, maintaining population diversity and thus improving overall performance, ensuring efficient evaluation and optimization of agents in the parameter space between source and detector agents. Attached Figure Description

[0027] Figure 1 A flowchart of the method proposed in the embodiments of this application;

[0028] Figure 2 Box plots of the method of this application and other comparative methods in the embodiments of this application on the 10-dimensional CEC2017 benchmark function;

[0029] Figure 3 The convergence curves of the method of this application and other comparative methods in the embodiments of this application are shown on the 10-dimensional CEC2017 benchmark function;

[0030] Figure 4 The convergence curves of the method of this application and other comparative methods in the embodiments of this application are shown on the CEC2019 benchmark function;

[0031] Figure 5 Box plots of the method of this application and other comparative methods on the CEC2019 benchmark function in the embodiments of this application;

[0032] Figure 6 The figures shown in this application are (a) velocity model diagram, (b) initial model diagram, and (c) inversion result diagram of the Marmousi model without an initial model in the embodiments of this application.

[0033] Figure 7 This is a schematic diagram of the network structure and the initial model in the interpolation mapping in the embodiments of this application, wherein (a) is the calculation mesh diagram of the interpolation mapping process, and (b) is the initial model diagram obtained by interpolation mapping;

[0034] Figure 8 The convergence curves of the method of this application and other methods for constructing the initial model in the embodiments of this application;

[0035] Figure 9The diagram shows (a) an initial model and (b) an inversion result based on the initial model, calculated by the method of this application in the embodiments of this application. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0037] See Figure 1 The method for optimizing an initial model for seismic data inversion, as shown, includes:

[0038] S1 initializes the parameters of the seismic exploration optimizer, randomly generates the initial positions of all agents, evaluates the fitness of each agent, and marks the globally optimal agent;

[0039] A seismic exploration optimizer is an algorithm used to execute initial model optimization methods for seismic data inversion, addressing global optimization problems in a continuous search space. It integrates physical mechanisms into the search process through mathematical modeling, thereby achieving a balance between exploration and development within the search space.

[0040] The parameters include population size, maximum number of iterations, problem dimension, lower bound vector, upper bound vector, and fitness function.

[0041] S2 simulates the collaborative mechanism between seismic sources and detectors through a seismic exploration optimizer. It randomly selects some agents in the population to be set as seismic source agents and sets the rest as detector agents. It iterates through each seismic source agent and non-seismic source agent, evaluates the fitness of each agent, and updates the global optimal agent.

[0042] During the iteration process, a portion of the randomly selected agents in the population are designated as "seismic sources," while the remaining agents are designated as "detectors." The relative numbers of these agents change with the number of iterations. This process involves a double loop: the outer loop processes each seismic source agent one by one, while the inner loop operates on all non-seismic source agents.

[0043] S3 For each non-optimal agent, calculate the direction vector pointing to the global optimal agent. Calculate a new movement direction using the random direction and the direction vector pointing to the global optimal agent. Update the agent's position based on the new movement direction and a step factor, where the step factor is calculated based on the direction vector pointing to the global optimal agent.

[0044] Here, non-optimal agents refer to agents other than the globally optimal agent. For each non-optimal agent, a direction vector pointing towards the globally optimal agent is calculated. This vector, combined with a random direction and the direction vector pointing towards the globally optimal agent, is then used to calculate a new movement direction for the non-optimal agent. This process belongs to a scattering movement strategy used to simulate the scattering characteristics of seismic waves.

[0045] S4 uses an S-wave propagation strategy to traverse each agent and determine whether the agent's position is the globally optimal agent. If not, it calculates the oscillation factor and updates the agent's position. If it is, it overwrites the original globally optimal agent position.

[0046] S5 continues to iterate. When the maximum number of iterations is reached or the convergence accuracy is met, the loop terminates and the globally optimal agent position and its fitness value are output.

[0047] In one embodiment, initializing the parameters of the seismic exploration optimizer includes: setting the search space of the problem as... Dimension, define the number of agents as The position of each agent is represented as follows: ,in, Indicates the current iteration number Next The agent in the first The position of the dimension = The initial position is randomly generated within the given upper and lower bounds: in, For the first The lower bound of dimensionality For the first The upper bound of the dimension, Indicates the interval Uniform random numbers on the surface Indicates the current iteration number Next The agent in the first The initial position of the dimension.

[0048] In one embodiment, the coordination mechanism between the seismic source and the detector is simulated using a seismic exploration optimizer, including:

[0049] Based on the current iteration number and maximum number of iterations Determine the number of earthquake source proxies;

[0050] Each source agent interacts with all detector agents to construct new agents.

[0051] The collaborative mechanism is represented as: ,in It is based on the current iteration number and maximum number of iterations Determine the number of earthquake source proxies; This indicates the number of agents. As iterations progress, the number of source agents gradually decreases. For each source agent, it interacts with all detector agents: ,in, This indicates the location of the new agent. Let represent a D-dimensional vector, where each element is a uniformly random number in the interval [0,1]. Indicates the epicenter agent, This indicates a detector agent.

[0052] In one embodiment, the attenuation process of seismic waves is simulated by reducing the current value of the globally optimal agent, thereby optimizing the globally optimal agent and performing local fine-tuning. The update formula is as follows: ,in, The attenuation coefficient is... It is the globally optimal proxy. This is the globally optimal proxy after attenuation optimization.

[0053] In one embodiment, a new movement direction is calculated using a random direction vector and a direction vector pointing to the globally optimal agent, including:

[0054] Randomness weights and deterministic weights are calculated to dynamically adjust the randomness and determinism in the search process;

[0055] Normalized random direction vector;

[0056] Calculate the new direction of movement: ,in, For random weights, For deterministic weights, This is a normalized random direction vector. This is the direction vector pointing to the globally optimal agent. It represents the Euclidean norm of the direction vector pointing to the globally optimal agent.

[0057] Update the agent's position based on the new direction of movement and step factor, including:

[0058] Calculate the step size factor: ,in, Step size factor Update according to the following formula: , To adjust the parameters, used to control the changing trend of the step size factor, as the iteration progresses, A smooth transition from near 0 to near 1, Indicates the maximum number of iterations. Indicates the current iteration number;

[0059] Update the proxy location: , The location of the updated proxy. The location of the proxy before the update. For the new direction of movement.

[0060] Randomness weights and deterministic weights are calculated to dynamically adjust the randomness and determinism in the search process: , As the iteration proceeds, the randomness weights Gradually decrease, deterministic weight Gradually increase. Normalized random direction vectors are used to increase the diversity of the search. , Represents a normalized random direction vector.

[0061] In the early stages of iteration, randomness weights dominate, and the agent's movement direction is mainly influenced by random factors. As the iteration progresses, the scattering intensity gradually weakens. In the later stages of iteration, deterministic weights dominate, and the agent's movement direction moves towards the optimal agent.

[0062] In one embodiment, for each suboptimal agent, an adjustment factor is calculated: ;

[0063] Calculate the agent's movement position using the adjustment factor: , The location of the agent after the move. As the globally optimal proxy, The current agent position is set as [position name]. If the fitness value of [position name] is better than [position name], the position of [position name] in the next iteration is updated. This application employs an S-wave propagation strategy, traversing each agent, determining whether the agent's position is the optimal solution, and if not, calculating the oscillation factor and updating the agent position, evaluating the fitness of each agent, and marking the globally optimal agent. If the new solution is better, the original globally optimal agent position is overwritten. By employing the S-wave propagation strategy, periodic fluctuations are induced by a sine function, guiding agents out of local optima and thus developing more promising positions.

[0064] To validate the method proposed in this application, 12 benchmark functions (F1, F3, F4, F5, F6, F7, F10, F17, F19, F21, F24, and F29) were selected from the CEC2017 benchmark set (the benchmark function set released at the 2017 IEEE Congress on Evolutionary Computation, CEC). Furthermore, the method was further evaluated on the benchmark functions from the CEC2019 benchmark set (the benchmark function set released at the 2019 IEEE Congress on Evolutionary Computation, CEC). The CEC2019 benchmark functions focus on multimodal global optimization problems and include 10 highly complex multimodal test functions, designed to evaluate the global search capability, convergence performance, and robustness of the method proposed in this application when dealing with a large number of local extrema and high dimensionality.

[0065] Ten optimization methods were selected for comparison, including ATOA (arithmetic-trigonometric optimization algorithm), CDO (Chernobyl disaster optimizer), SABO (Subtraction-Average-Based Optimizer), LCA (Liver Cancer Algorithm), SCSO (Sand Cat swarm optimization), AO (Artemisinin Optimization), GOOSE, NRBO (Newton-Raphson-Based Optimizer), CPO (Crested Porcupine Optimizer), and FATA (Fata Morgana algorithm). To ensure fairness and consistency in the comparison, all methods used randomized population initialization with a population size of 50 and a maximum number of iterations of 500. Each method was run independently for 30 times, and its average performance was recorded to reduce the impact of randomness on the results.

[0066] This application uses the mean and standard deviation as performance evaluation metrics. The mean measures the global convergence of the method on the test function, i.e., the average level of the convergent function values ​​obtained in 30 independent runs. The standard deviation reflects the stability of the results; a lower standard deviation indicates that the method performed more stably. Statistical analysis was performed using the Wilcoxon signed-rank test and the Friedman test. The Wilcoxon signed-rank test is a non-parametric statistical method suitable for comparing differences between two paired samples, used to test the significance of the results between this application's method and each comparative method. The Friedman test is used for multi-method comparisons, evaluating the overall performance differences of each method on multiple benchmark functions. It can effectively analyze the significance level between multiple sets of samples and comprehensively evaluate the global performance of this application's method compared to other methods on multiple benchmark functions.

[0067] The performance of our method on the 10-dimensional CEC2017 benchmark function is compared with other methods, and the results are shown in Table 1. Experimental results show that our method achieves the lowest average value among the 12 benchmark functions. Furthermore, our method has the lowest standard deviation in the 19th benchmark function. Specifically, in the optimization process of unimodal functions (F1, F3), our method employs a decay mechanism to introduce perturbations near the current optimal solution to improve search accuracy. For multimodal functions (F4, F5, F6, F7, F10), our method uses a cooperative mechanism to generate surrogate positions and combines a scattering shift strategy to introduce search sample diversity, preventing the method from prematurely falling into local optima. In the optimization of mixed functions (F17, F19), our method utilizes cooperative and decay mechanisms to conduct multi-scale searches, determine the global optimum region within the search space, and gradually approach the global optimum. For combined functions (F21, F24, F29), our method employs an S-wave propagation strategy to adjust the surrogate positions to adapt to the high-dimensional complex search space. Figure 2 Box plots of all methods on the 10-dimensional CEC2017 benchmark functions F1, F3, F4, F5, F6, and F7 are presented. The results show that the proposed method has lower median and interquartile range in the corresponding box plots for each benchmark function compared to other methods.

[0068] Table 1. Optimization results of the method in this application and other comparative methods on the 10-dimensional CEC2017 benchmark function.

[0069]

[0070] The convergence of the method in this application is analyzed. Figure 3Convergence curves of all methods on the 10-dimensional CEC2017 benchmark functions F1, F10, F17, F21, F24, and F29 are presented. The results show that the average fitness value of the proposed method is significantly lower than that of other methods in the early stages of iteration, indicating that its exploration strategy can efficiently identify potential optimal solution regions. Specifically, on the benchmark functions F1, F10, F17, F21, F24, and F29, the convergence speed of the proposed method is significantly faster than the comparative methods, enabling it to quickly enter the development stage with fewer iterations. Furthermore, on the benchmark functions F10, F21, F24, and F29, the proposed method continuously improves the accuracy of the solution. More importantly, the proposed method achieves fitness values ​​in the early iterations that other methods can only achieve with the maximum number of iterations. In summary, the analysis of the convergence curves of the proposed method demonstrates its superior convergence performance.

[0071] The performance of the proposed method was further evaluated through experiments on the CEC2019 benchmark function test, and the results are shown in Tables 2 and 3. In Table 2, the proposed method achieved the lowest mean fitness on all benchmark functions. Specifically, on the F1 function, the ATOA method, CDO method, NRBO method, FATA method, and the proposed method all successfully converged to the global optimum. Regarding stability, the proposed method achieved the lowest standard deviation on five benchmark functions, but on the F2 benchmark function, the CDO method showed slightly better stability. Furthermore, the CDO method and FATA method achieved the lowest standard deviation on three functions, respectively. The FT test results show that the proposed method ranked first, the AO method and NRBO method performed similarly, ranking second and third respectively, while the CPO method and LCA method performed poorly, ranking tenth and eleventh respectively. The Rank in Table 2 indicates the ranking. Table 3 shows that the significant difference rate between the method of this application and other methods is 92%, and the significant advantage ratio is 23 / 25. This indicates that the method of this application is significantly better than other methods on most benchmark functions and has a statistical advantage.

[0072] Table 2 shows the optimization results of the method in this application and other comparative methods on the CEC2019 benchmark function.

[0073]

[0074] Table 3. WSR test results of the method in this application and other comparative methods on the CEC2019 benchmark function.

[0075]

[0076] In Table 3, NaN indicates that the data cannot be calculated, + indicates that it is better than the benchmark method, - indicates that it is worse than the benchmark method, ~ indicates that it is close to the benchmark method, and 9 / 0 / 1 means that in 10 experiments, the current method is better than the benchmark method 9 times, worse than the benchmark method 0 times, and close to the benchmark method 1 time.

[0077] Figure 4 The convergence curves of each method on the CEC2019 benchmark function are shown. From Figure 4 As can be seen, the proposed method exhibits high efficiency in the early stages of optimization for most benchmark functions, gradually progressing into the development phase. Specifically, in the early iterations of the F1 and F2 benchmark functions, the proposed method has already located the global optimum region. On the F3, F4, F6, F7, and F8 benchmark functions, the proposed method also demonstrates a relatively fast convergence speed. Although the AO method performs similarly to the proposed method on these benchmark functions, its convergence accuracy is slightly inferior. Notably, in the F10 benchmark function, the proposed method achieved the convergence level of other methods at the maximum number of iterations in the early stages of iteration. Figure 5 Box plots of the proposed method versus other methods on the CEC2019 benchmark functions F1, F2, F3, F4, F5, and F6 are presented. It can be seen that the proposed method has a narrower box plot and is positioned lower in F1, F2, F4, F5, and F6, indicating good stability and high convergence accuracy. Overall, the proposed method demonstrates superior performance compared to other methods on the CEC2019 benchmark functions.

[0078] Seismic exploration technology is widely used in mineral exploration, engineering construction, and environmental protection. Full waveform inversion (FWI) technology can be used to invert subsurface media and obtain the wave velocity matrix, which reflects the propagation characteristics of seismic waves in the subsurface medium. Analysis of the wave velocity matrix helps infer the physical properties of the subsurface medium. Full waveform inversion technology is based on gradient optimization methods, and the accuracy of its results is affected by the initial model. For example... Figure 6 The diagram shown is a velocity model diagram, an initial model diagram, and an inversion result diagram of the Marmousi model without an initial model in this embodiment of the application. Figure 6 (a) in the diagram is the velocity model diagram. Figure 6 (b) in the diagram represents the initial model diagram and Figure 6 Figure (c) shows the inversion results. It can be seen that without an initial model, the inversion results fail to accurately reflect the wave velocity distribution in the actual model. In engineering practice, a global optimization method is typically used to calculate the initial model, which is then used as the basis for subsequent inversions. Since the initial model calculation involves solving a large number of parameters, it is prone to the curse of dimensionality problem. Therefore, a computational grid is constructed, and interpolation mapping is used to correlate the initial model. Figure 7 The interpolation mapping process is shown, where Figure 7 (a) in the diagram is a 13×12 computational grid with a total of 156 solution parameters; Figure 7 In Figure (b), the initial model diagram obtained through interpolation mapping has a size of 61×220. The objective function is defined as the L2 norm between the seismic wave observation records in the real model and the simulated records in the initial model. Figure 8 The convergence curves of the method in this application and other methods are shown. Figure 8 As can be seen, the convergence curve of the method in this application continues to decrease during the iteration. In contrast, other methods get stuck in a local optimum at the 30th generation. Figure 9 (a) in the figure is the initial model diagram calculated by the method of this application. Figure 9 Figure (b) shows the inversion results based on this initial model. The inversion results reflect the wave velocity distribution of the true Marmousi model.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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 optimizing an initial model for seismic data inversion, characterized in that, include: Initialize the parameters of the seismic exploration optimizer, randomly generate the initial locations of all agents, evaluate the fitness of each agent, and mark the globally optimal agent; The seismic exploration optimizer simulates the collaborative mechanism between the seismic source and the geophone. Some agents in the population are randomly selected and set as seismic source agents, while the rest are set as geophone agents. Each seismic source agent and non-seismic source agent is traversed, the fitness of each agent is evaluated, and the global optimal agent is updated. For each non-optimal agent, calculate the direction vector pointing to the global optimal agent. Calculate a new movement direction using a random direction and the direction vector pointing to the global optimal agent. Update the agent's position based on the new movement direction and a step factor, where the step factor is calculated based on the direction vector pointing to the global optimal agent. The S-wave propagation strategy is adopted to traverse each agent and determine whether the agent's position is the globally optimal agent. If not, the oscillation factor is calculated and the agent's position is updated. If it is, the original globally optimal agent position is overwritten. Continue iterating until the maximum number of iterations is reached or the convergence accuracy is met. Then, terminate the loop and output the globally optimal agent position and its fitness value.

2. The method for optimizing an initial model for seismic data inversion according to claim 1, characterized in that, Initialize the parameters of the seismic exploration optimizer, including: setting the search space of the problem as follows: Dimension, define the number of agents as The position of each agent is represented as follows: ,in, Indicates the current iteration number Next The agent in the first The position of the dimension = The initial position is randomly generated within the given upper and lower bounds: in, For the first The lower bound of dimensionality For the first The upper bound of the dimension, Indicates the interval Uniform random numbers on the surface Indicates the current iteration number Next The agent in the first The initial position of the dimension.

3. The method for optimizing an initial model for seismic data inversion according to claim 1, characterized in that, The seismic exploration optimizer simulates the collaborative mechanism between the seismic source and the detector, including: Based on the current iteration number and maximum number of iterations Determine the number of earthquake source proxies; Each source agent interacts with all detector agents to construct new agents.

4. The method for optimizing an initial model for seismic data inversion according to claim 1, characterized in that, The calculation of the new movement direction using a random direction vector and a direction vector pointing to the globally optimal agent includes: Randomness weights and deterministic weights are calculated to dynamically adjust the randomness and determinism in the search process; Normalized random direction vector; Calculate the new direction of movement: ,in, For random weights, For deterministic weights, This is a normalized random direction vector. This is the direction vector pointing to the globally optimal agent. It represents the Euclidean norm of the direction vector pointing to the globally optimal agent.

5. The method for optimizing an initial model for seismic data inversion according to claim 4, characterized in that, Update the agent's position based on the new direction of movement and step factor, including: Calculate the step size factor: ,in, Step size factor Update according to the following formula: , To adjust the parameters, used to control the changing trend of the step size factor, as the iteration progresses, A smooth transition from near 0 to near 1, Indicates the maximum number of iterations. Indicates the current iteration number; Update the proxy location: , The location of the updated proxy. The location of the proxy before the update. For the new direction of movement.

6. The method for optimizing an initial model for seismic data inversion according to claim 4, characterized in that, The randomness weights are calculated as follows: The deterministic weights are calculated as follows: , Indicates the maximum number of iterations. This indicates the current iteration number.

7. The method for optimizing an initial model for seismic data inversion according to claim 1, characterized in that, The globally optimal proxy is attenuated to obtain an attenuated globally optimal proxy.

8. The method for optimizing an initial model for seismic data inversion according to claim 1, characterized in that, Calculate the oscillation factor and update the agent position, including: Calculate the adjustment factor: ; Calculate the agent's movement position using the adjustment factor: , The location of the agent after the move. As the globally optimal proxy, This is the current agent location; If the fitness value of the moved agent's position is better than the fitness value of the current agent's position, then the moved agent's position is used as the position for the next iteration of the current agent's position.

Images