LLM-based three-dimensional shear wave velocity prediction method

Abstract

The invention provides a method for predicting a three-dimensional shear wave speed based on LLM, namely, a prompt word template adaptive to a three-dimensional shear wave speed prediction task is designed, a prompt sequence data set is generated by utilizing known three-dimensional shear wave speed data, the LLM is finely adjusted in combination with a standard trainer, and the three-dimensional shear wave speed is predicted by means of the finely-adjusted LLM. The method has the three-dimensional shear wave velocity data complementing capability, can relieve the problems of space coverage blind areas and low resolution in the inversion of the three-dimensional shear wave velocity by a traditional surface wave imaging method, and provides a more complete data basis for subsequent geological structure analysis; according to the method, a part of known three-dimensional shear wave velocity is used for prediction instead of traditional full-region observation inversion of the three-dimensional shear wave velocity, so that the calculation cost of large-scale high-resolution prediction of the three-dimensional shear wave velocity is reduced. According to the invention, deep fusion of artificial intelligence and geological structure research is promoted, and development of the earth science field is better served.

Description

Technical Field

[0001] This invention relates to a three-dimensional shear wave velocity prediction method, specifically, a three-dimensional shear wave velocity prediction method based on LLM. This invention belongs to the field of artificial intelligence data processing technology. Background Technology

[0002] Shear wave velocity is one of the important physical parameters of rocks inside the Earth, reflecting properties such as rock density and elastic modulus. Traditional methods for extracting shear wave velocity include sonic logging and rock physics testing, but their drawback is that they can only extract one-dimensional (depth profile) and two-dimensional (along the survey line profile) shear wave velocities. To gain a deeper understanding of geodynamic processes and accurately characterize complex geological structures, it is necessary to master three-dimensional shear wave velocity.

[0003] Three-dimensional shear wave velocity is a spatially distributed parameter describing the elastic properties of subsurface media. Its core lies in locating the propagation velocity of shear waves using three-dimensional spatial coordinates (longitude, latitude, and depth). A three-dimensional shear wave velocity structure model can demonstrate how the shear wave velocity varies with latitude, longitude, and depth in specific regions within the Earth's interior. Three-dimensional shear wave velocity is of great significance for understanding the complex geological structures within the Earth's interior, assessing seismic hazards, and exploring oil, gas, and geothermal resources.

[0004] Most existing studies utilize surface wave imaging methods based on background noise to invert the three-dimensional shear wave velocity within the Earth. However, these methods still have some limitations and shortcomings in practical applications, mainly reflected in the following aspects: 1. Limited by the number of observation stations and the limited density of ray path coverage, the obtained three-dimensional shear wave velocity data is limited and cannot clearly depict the geological structure of the Earth's interior. 2. The resolution of the inversion results drops significantly in the peripheral regions, resulting in insufficient resolution. For example, in the study of the northeastern margin of the Tibetan Plateau, although the structure above 75 km in the central region was well restored, the resolution in the peripheral regions was reduced due to sparse data coverage. Similarly, in the inversion of the Hanzhong Basin, short-period surface wave data has weak constraints on the shallow layers, leading to unstable inversion results in areas with large lateral velocity variations in the shallow layers. 3. The high computational cost of surface wave imaging data processing limits its application in real-time monitoring or rapid response scenarios. Background noise data is usually recorded continuously over a long period, requiring a massive amount of data to be processed. Performing cross-correlation calculations, dispersion curve extraction, and three-dimensional tomographic inversion on this data requires enormous computational resources and time costs. This problem is particularly pronounced when performing large-scale, high-resolution imaging.

[0005] Currently, there are numerous studies on predicting shear wave velocities based on empirical formulas, machine learning methods, and deep learning methods. However, studies on further prediction using known 3D shear wave velocities obtained through inversion are relatively few. Therefore, the purpose of this invention is to propose a method based on LLM (Laser-Based Modeling) for further prediction of 3D shear wave velocities using inverted 3D shear wave velocities, particularly for predicting 3D shear wave velocities in areas not fully covered by observation stations. This addresses the spatial coverage blind spots and resolution limitations of traditional surface wave imaging methods, providing a more complete data foundation for subsequent geological structural analysis. Summary of the Invention

[0006] In view of the above reasons, the purpose of this invention is to provide a three-dimensional shear wave velocity prediction method based on LLM. This method designs cue word templates adapted to the three-dimensional shear wave velocity prediction task, utilizes a cue sequence dataset generated from known three-dimensional shear wave velocity data, combines a standard trainer to fine-tune a large language model (LLM), and then uses the fine-tuned LLM to predict the three-dimensional shear wave velocity. To achieve the above objectives, the present invention adopts the following technical solution: a three-dimensional shear wave velocity prediction method based on LLM, which includes the following steps: S1: Obtain known three-dimensional shear wave velocity structure data; S2: Construct a cue sequence dataset; S3: Fine-tune the LLM based on the cue sequence dataset and the standard trainer, and output the fine-tuned LLM; S3.1: Randomly divide the prompt sequence dataset generated in step S2 into training set, validation set and test set according to the proportion; The training set is used for learning model parameters, the validation set is used to monitor the training process and prevent overfitting, and the test set is used to finally evaluate the model's generalization ability. S3.2: Convert the training and validation set data into standard JSON format; Each JSON record contains "input hints" and "output answer" fields; S3.3: Load the pre-trained large language model weights; S3.4: Fine-tuning the LLM based on the standard trainer provided by HuggingFace; S3.5: Output the finely tuned LLM; After fine-tuning, save the final model weights, configuration file and word segmenter to form an independently deployable model for three-dimensional shear wave velocity prediction, namely LLM. S4: Construct the context and issues that guide the LLM's focus based on the prompt word template; During the prediction phase, for new and unseen 3D velocity field regions, input prompts are constructed using the same serpentine traversal strategy and prompt word template. These input prompts include the spatial coordinates, shear wave velocities, and parameters of known points in the region to be predicted, as well as the overall trend ΔS and average velocity change percentage ΔS% calculated based on the sequence of known points to describe the overall variation trend of the 3D shear wave velocity at the known points in the region to be predicted. S5: The fine-tuned LLM performs predictions based on the question input in step S4; The constructed input prompts are input into the fine-tuned LLM. The LLM uses the context and question constructed in step S4 as conditions to autoregressively generate text descriptions of the velocity values ​​for the subsequent Step_out points; the specific numerical prediction sequence is obtained through decoding. S6: Output the prediction results; The text output generated by LLM is parsed into a structured list of velocity values, and the one-dimensional prediction sequence is reassembled back into a three-dimensional spatial grid by inverse mapping of the original serpentine path. The final output is the predicted three-dimensional shear wave velocity structure data.

[0007] The method for constructing the cue sequence dataset in step S2 is as follows: S2.1: Traverse the three-dimensional shear wave velocity structure data obtained in step S1 according to the snake traversal strategy to form a numerical sequence dataset; S2.2: Construct a prompt word template that maps three-dimensional shear wave velocity structure data from the numerical domain to the text domain; The three-dimensional shear wave velocity prompt template includes an input prompt section and an output prompt section; the input prompt section further includes a context section and a question section, the context section provides the historical information required for prediction, the question section is a query for the future; the output prompt section is the answer to the question. S2.3: Convert the three-dimensional shear wave velocity numerical sequence dataset into an LLM-processable cue sequence dataset based on the cue word template; S2.3.1: Determine the input step size Step_in and the output step size Step_out; S2.3.2: Based on the prompt word template T, traverse the three-dimensional shear wave velocity numerical sequence dataset S and complete the context, question and answer for each data record; S2.3.3: Calculate the overall trend of change ΔS and the percentage change in average velocity ΔS% for the Step_in points. S2.3.4: Write the context and question of a single record into the input prompt section of the prompt word template, and write the answer into the output prompt section; S2.3.5: Determine whether the three-dimensional shear wave velocity numerical sequence dataset S has been traversed completely; If not completed, slide the window along the serpentine path, with Step_in as the window length and Step_out as the prediction step size, and generate training samples in sequence. Repeat this process until the entire numerical sequence is traversed to ensure that all possible continuous segments are sampled, so as to make full use of the data and increase sample diversity. Once the traversal is complete, the original LLM-processable cue sequence dataset is obtained.

[0008] The serpentine traversal strategy involves forming a spatially continuous zigzag path by alternately reversing the latitude direction within the same depth layer. This ensures that the end and start nodes of adjacent longitude columns are spatially continuous, and the layers are connected end-to-end by mirroring the longitude sequence. This flattens the three-dimensional grid-like structural data into a spatially closed, metrically continuous one-dimensional path.

[0009] The calculation methods for the overall change trend ΔS and the average velocity change percentage ΔS% are as follows: ; ; Where: V i V0 represents the velocity value of the i-th point in the sequence consisting of input step_in points, V0 represents the velocity value of the starting point in the sequence consisting of input step_in points, and T represents the length of the sequence consisting of input step_in points.

[0010] The method for predicting three-dimensional shear wave velocities based on LLM in this invention has the following advantages: 1. It has the ability to complete three-dimensional shear wave velocity data, which can alleviate the problems of spatial coverage blind spots and low resolution in the traditional surface wave imaging method for inverting three-dimensional shear wave velocity, and provide a more complete data foundation for subsequent geological structure analysis.

[0011] 2. Reducing the computational cost of large-scale, high-resolution prediction of 3D shear wave velocities. Traditional surface wave imaging methods require two core conditions to achieve 3D shear wave velocity inversion: a dense network of observation stations and a complex inversion calculation process, resulting in a significant trade-off between cost and efficiency. This invention utilizes a subset of known (already inverted) 3D shear wave velocity data combined with outward prediction capabilities to replace the traditional approach of inverting 3D shear wave velocities through full-area observation, thereby reducing the computational cost of large-scale, high-resolution prediction of 3D shear wave velocities.

[0012] 3. This invention can also inspire new research directions and application scenarios, such as conducting more refined dynamic assessments of strong earthquake risks, promoting the deep integration of artificial intelligence and geological structure research, and better serving the development of the field of earth sciences. Attached Figure Description

[0013] Figure 1 This is a flowchart of the method for predicting three-dimensional shear wave velocities based on LLM according to the present invention; Figure 2 This is a schematic diagram of the snake traversal strategy of the present invention; Figure 3 A flowchart illustrating the generation of the prompt sequence dataset for this invention; Figure 4 This is a flowchart of the LLM fine-tuning process based on the cue sequence dataset and standard trainer of the present invention. Detailed Implementation

[0014] The structure and features of the present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that various modifications can be made to the embodiments disclosed herein; therefore, the embodiments disclosed in this specification should not be considered as limitations on the present invention, but merely as examples to make the features of the present invention readily apparent.

[0015] Large Language Models (LLMs) are artificial intelligence models based on deep learning technology. They can learn and understand the complex structure and semantic meaning of human language, process and generate human language, and engage in effective dialogue and answer questions with humans. In recent years, LLMs have been widely used to generate high-quality text (such as news reports, blog posts, and creative writing), language translation (such as achieving accurate translation between multiple languages ​​and supporting cross-language communication), and dialogue generation (such as virtual voice assistants and chatbots).

[0016] The inventors of this invention have discovered that although three-dimensional shear wave velocity data are spatial geophysical parameters, they can be transformed into a spatial sequence data format that can be processed by LLM (Liquid Least Mesh) through temporal processing for spatial sequence modeling. The proposed method for predicting three-dimensional shear wave velocity based on LLM involves: designing cue word templates adapted to the three-dimensional shear wave velocity prediction task; generating a cue sequence dataset using known three-dimensional shear wave velocity data; fine-tuning the LLM using a standard trainer; and then using the fine-tuned LLM to predict the three-dimensional shear wave velocity.

[0017] like Figure 1 As shown, the three-dimensional shear wave velocity prediction method based on LLM of this invention is as follows: S1: Obtain known three-dimensional shear wave velocity structure data.

[0018] Three-dimensional shear wave velocity structure data D are obtained through surface wave imaging, seismic tomography, full waveform inversion, or other geophysical inversion methods. Each three-dimensional shear wave velocity structure data includes longitude, latitude, and depth coordinates, as well as the corresponding shear wave velocity value. These three-dimensional shear wave velocity structure data are not only used to characterize the elastic properties of the subsurface medium, but also serve as the physical basis for subsequent serialization and prediction of other three-dimensional shear wave velocity structure data.

[0019] S2: Construct a cue sequence dataset.

[0020] S2.1: Perform spatial sequence modeling by traversing the three-dimensional shear wave velocity structure data D obtained in step S1 according to the serpentine traversal strategy to form a numerical sequence dataset S.

[0021] To effectively capture the potential spatiotemporal correlation features in the three-dimensional shear wave velocity field, this invention proposes and employs a serpentine traversal strategy to traverse the acquired three-dimensional shear wave velocity structure data D, transforming the three-dimensional shear wave velocity field into a spatiotemporal sequence, thereby forming a numerical sequence dataset S. Specifically, as... Figure 2 As shown, within the same depth layer, alternating latitudinal reversals create a spatially continuous zigzag path, ensuring spatial continuity between the beginning and end nodes of adjacent longitude columns. Interlayer connections are achieved by mirroring the longitude sequence, thus flattening the three-dimensional grid-like structural data into a spatially closed, metrically continuous one-dimensional path. Its advantages include: maximizing the preservation of spatial proximity relationships between adjacent points in the three-dimensional grid-like structural data, encoding three-dimensional spatial structural information into the order of a one-dimensional sequence, and laying the foundation for subsequent large-scale language models to understand spatial correlations.

[0022] S2.2: Construct a prompt word template T that maps three-dimensional shear wave velocity structure data from the numerical domain to the text domain.

[0023] Using template-based descriptions is an effective and efficient method for data-to-text conversion. This invention explicitly introduces templates for three-dimensional shear wave data, and Table 1 lists the template formats and corresponding examples. Through template-based conversion, efficient and consistent mapping of three-dimensional shear wave velocity structure data from the numerical domain to the text domain is achieved.

[0024] Table 1. Template for 3D Shear Wave Velocity Prompts As shown in Table 1, the three-dimensional shear wave velocity cue word template T includes an input cue part and an output cue part; the input cue part includes a context part and a question part. The context part provides the historical information required for prediction, and the question part is a query for the future; the output cue part is the answer to the question.

[0025] The principle for converting three-dimensional shear wave velocity numerical data into a three-dimensional shear wave velocity prompt template is: 1. Contextual Data Aggregation Spatial coordinate representation of three-dimensional shear wave velocity: The three-dimensional spatial coordinate range is described by placeholders {Start_Longitude}° to {End_Longitude}° to avoid listing each point.

[0026] Shear wave velocity and parameters: The sequence data is represented by {V1, V2, ..., Vstep_in / step_out}, and ΔS and ΔS% summarize the trend of data change.

[0027] 2. Predict coordinates Implicit spatial expansion: The range of predicted data in three-dimensional space is declared through parameters such as {Next_Start_Longitude} and {Next_End_Longitude}, ensuring that the model traverses the latitude and longitude grid in sequence.

[0028] Sliding window constraint: The entire dataset is divided into multiple sequences by a sliding window. The window size is equal to the observation length Step_in plus the prediction length Step_out. The sliding step size is set to 1 by default.

[0029] 3. Input / output alignment Structured mapping: The {Next_*} parameters in the output (Answer) are completely consistent with the 3D coordinate range in the input (Question), avoiding ambiguity.

[0030] Unit consistency: All predicted speed values ​​are uniformly converted to km / s, aligning with the units of the original data.

[0031] 4. Scalability The template supports dynamic replacement of placeholders, making it suitable for different actual prediction needs of three-dimensional shear wave velocities.

[0032] S2.3: Based on the cue word template T, transform the three-dimensional shear wave velocity numerical sequence dataset S into an LLM-processable cue sequence dataset P.

[0033] like Figure 3 As shown, the method for converting the three-dimensional shear wave velocity numerical sequence dataset S into an LLM-processable cue sequence dataset P based on the cue word template T is as follows: S2.3.1: Determine the input step size Step_in and the output step size Step_out.

[0034] The input step size defines the length of the historical context that the LLM relies on for prediction, i.e., the number of known points. The output step size defines the number of future points that the LLM needs to continuously predict.

[0035] S2.3.2: Based on the prompt word template T, traverse the three-dimensional shear wave velocity numerical sequence dataset S and complete the context, question, and answer for each data record.

[0036] S2.3.3: Calculate the overall trend of change ΔS and the percentage change in average velocity ΔS for the Step_in points.

[0037] When constructing the context of each data record in the three-dimensional shear wave velocity numerical sequence dataset, in addition to providing the specific velocity value for each known point, two statistics are introduced as auxiliary cue information: the overall trend ΔS and the average velocity change percentage ΔS%. ​​These statistics are embedded in cue word templates in natural language form, providing the language model with a general description of the magnitude and direction of the velocity field change in the local region. This helps the model (i.e., LLM) to grasp the overall change pattern more quickly and improve the physical rationality of the prediction.

[0038] ; ; Where: V i V0 represents the velocity value of the i-th point in the sequence (a sequence consisting of input step length Step_in points), V0 represents the velocity value of the starting point in the sequence (a sequence consisting of input step length Step_in points), and T represents the length of the sequence (a sequence consisting of input step length Step_in points).

[0039] S2.3.4: Write the context and question of a single record into the input prompt section of the prompt word template, and write the answer into the output prompt section.

[0040] The context descriptions generated in steps S2.3.2 and S2.3.3 are combined with the prediction question to form the input prompt part of the complete prompt word template. The generated answer is written into the output prompt part. This "input-output" pairing constitutes a complete training sample. The question explicitly specifies predicting the transverse wave velocity of Step_out points backward, starting from the last known point, thus formalizing the three-dimensional velocity prediction task as an autoregressive sequence generation task. For example, Step_in=25, predicting Step_out points backward starting from the 26th point.

[0041] This invention organically connects numerical data with a language model through prompt word templates. The coordinate information incorporated into the templates, as well as derived features such as the calculated overall trend ΔS and the average velocity change percentage ΔS%, constitute the "context" that guides LLM to focus on physical laws. Using this structured prompt as meta-knowledge can help LLM quickly grasp the macroscopic change patterns of the local velocity field, improving the physical rationality and accuracy of predictions.

[0042] S2.3.5: Determine if the three-dimensional shear wave velocity numerical sequence dataset S has been traversed completely. If not, slide the window along the serpentine path, with Step_in as the window length and Step_out as the prediction step size, and generate training samples sequentially. Repeat this process until the entire numerical sequence has been traversed to ensure that all possible continuous segments are sampled, so as to make full use of the data and increase sample diversity.

[0043] If the traversal is complete, the original LLM-processable cue sequence dataset P is obtained.

[0044] S3: Fine-tune the LLM based on the cue sequence dataset and the standard trainer, and output the fine-tuned LLM.

[0045] like Figure 4 As shown, the specific method is as follows: S3.1: Randomly divide the prompt sequence dataset P generated in step S2 into training set, validation set and test set according to the proportion.

[0046] In a preferred embodiment of the present invention, 64% of the cue sequence dataset is allocated to the training set, 16% to the validation set, and 20% to the test set. This allocation follows standard machine learning practices to ensure the effectiveness of model training, the reliability of hyperparameter tuning, and the unbiasedness of the final performance evaluation. The training set is used for model parameter learning, the validation set is used to monitor the training process and prevent overfitting, and the test set is used to ultimately evaluate the model's generalization ability.

[0047] S3.2: Convert the training and validation set data into standard JSON format.

[0048] To adapt to the data loading requirements of mainstream large language model fine-tuning frameworks, the input-output hint pairs for the training and validation sets are converted into standard JSON format. Each JSON record typically contains "input hint" and "output response" fields, facilitating subsequent instruction fine-tuning.

[0049] S3.3: Load the weights of the pre-trained Large Language Model (LLM).

[0050] We select an autoregressive language model pre-trained on a large-scale text corpus as the base model, load its pre-trained weights, and leverage the powerful capabilities the model has learned in general language understanding and sequence generation tasks to provide a high starting point for the downstream transverse wave velocity prediction task.

[0051] In a preferred embodiment of the present invention, the pre-trained large language model selected by the present invention is Text-to-TextTransfer Transformer, abbreviated as T5.

[0052] S3.4: Fine-tuning the LLM based on the standard trainer provided by HuggingFace.

[0053] Using the standard trainer provided by the Hugging Face Transformers library, the LLM is fine-tuned in a supervised manner on the prepared training set. The training objective is to minimize the loss function (such as cross-entropy loss) between the predicted sequences generated by the LLM and the true velocity value sequences. By setting appropriate hyperparameters such as learning rate, batch size, and number of training epochs, the model is adapted to the mapping pattern from geological description text to velocity value sequences.

[0054] S3.5: Output the finely tuned LLM.

[0055] After fine-tuning, the final model weights, configuration file, and word segmenter are saved, forming an independently deployable model for 3D shear wave velocity prediction, namely LLM. This model has the ability to generate subsequent shear wave velocity prediction sequences based on a given 3D coordinate context.

[0056] S4: Construct the context and issues that guide the LLM's focus based on the prompt word template.

[0057] During the prediction phase, for new and unseen 3D velocity field regions, input prompts are constructed using the same serpentine traversal strategy and prompt word template. This input prompt includes a sequence of known points in the region to be predicted, along with their statistical auxiliary information: the overall trend ΔS and the average velocity change percentage ΔS%. ​​Its format is consistent with the prompt word template used in the training phase.

[0058] S5: LLM performs predictions based on the questions input in step S4.

[0059] The constructed input prompts are input into the fine-tuned LLM. The LLM uses the context and question constructed in step S4 as conditions to autoregressively generate text descriptions of the velocity values ​​for the subsequent Step_out points; the specific numerical prediction sequence is obtained through decoding.

[0060] S6: Output the prediction results.

[0061] The text output generated by LLM is parsed into a structured list of velocity values, and the one-dimensional prediction sequence is reassembled back into a three-dimensional spatial grid by inverse mapping of the original serpentine path. The final output is the predicted three-dimensional shear wave velocity structure data.

[0062] This invention defines the three-dimensional shear wave velocity prediction task as a sequence-to-sequence (Seq2Seq) problem of "given a text describing spatial location and velocity values, generating subsequent location and velocity value texts via autoregression". It makes full use of the pre-training capabilities of LLM in long context modeling and conditional text generation, and fine-tunes it with instructions to adapt it to specific geophysical laws, thereby completing the prediction of three-dimensional shear wave velocity.

[0063] The following specific examples further illustrate the method for predicting three-dimensional shear wave velocities based on LLM according to the present invention.

[0064] S1: Obtain known three-dimensional shear wave velocity structure data.

[0065] This embodiment is based on 13 months (August 2017 - August 2018) of continuous background noise waveform data recorded by 105 broadband seismic stations in the Sichuan, Chongqing, Gansu, Qinghai, and Shaanxi seismic networks (data service platform of the International Data Center for Earthquake Science). The fundamental Rayleigh surface wave dispersion curves within 5-40 seconds were extracted, and a three-dimensional shear wave velocity structure data with a depth up to 120 km, a longitude range of 98°~110°, a latitude range of 39°~27°, and a spatial resolution of 0.5°×0.5° was obtained using the direct inversion method of surface wave dispersion data, totaling 8125 sets. The screenshot below shows the three-dimensional shear wave velocity structure data D at a depth of 0 km, a longitude range of 101°~101.5°, and a latitude range of 36.5°~37.5°.

[0066] S2: Construct a cue sequence dataset.

[0067] First, spatial sequence modeling of the three-dimensional shear wave velocity structure data D is performed according to a serpentine traversal strategy, forming a numerical sequence dataset S. An example of the formed numerical sequence is shown in the screenshot below. Specifically, within the same depth layer, a spatially continuous zigzag path is formed by alternately reversing the latitudinal direction, ensuring spatial continuity between the end and start nodes of adjacent longitude columns. Between layers, the beginning and end are connected by mirroring the longitude sequence, thus flattening the three-dimensional grid-like structure data into a spatially closed, metrically continuous one-dimensional path.

[0068] Spatial sequence modeling of D is performed using a serpentine traversal strategy to form a numerical sequence dataset S. Secondly, a prompt template T is constructed to map the three-dimensional shear wave velocity structure data from the numerical domain to the text domain. The prompt template is shown in the screenshot below. The input prompt includes context and the question, and the output prompt includes the answer.

[0069] Mapping three-dimensional shear wave velocity structure data from the numerical domain to the text domain prompt word template T Finally, the numerical sequence dataset S is transformed into an LLM-processable prompt sequence dataset P based on the prompt word template T. An example of the transformed prompt sequence data is shown in the screenshot below. The first row of the figure shows the input prompt as a three-dimensional shear wave velocity context sequence with a depth of 0.00 km, a longitude range of 101.00°~101.50°, and a latitude range of 36.00°~36.50°. The second row shows the input prompt as a three-dimensional shear wave velocity question sequence with a depth of 0.00 km, a longitude range of 101.50°~102.00°, and a latitude range of 36.00°~36.50°. The second row also shows the output prompt as a three-dimensional shear wave velocity response sequence with a depth of 0.00 km, a longitude range of 101.50°~101.20°, and a latitude range of 36.00°~36.50°.

[0070] Example of a cue sequence dataset P that LLM can process. S3: Fine-tune the LLM based on the cue sequence dataset and the standard trainer, and output the fine-tuned LLM.

[0071] In this embodiment, the pre-trained large language model selected is the Text-to-Text Transfer Transformer, abbreviated as T5. First, T5 is fine-tuned based on the training set of the cue sequence dataset P and the standard trainer. After fine-tuning, the output of the fine-tuned model M and the performance of model M on the validation set are shown in the following screenshots. Then, model M is loaded and its performance on the test set is tested, as shown in the following screenshots.

[0072] Fine-tuning LLM based on cue sequence dataset P and standard trainer Performance of model M on the test set S4: Construct the context and issues that guide the LLM's focus based on the prompt word template.

[0073] This embodiment constructs the following context and issues to guide the LLM's attention based on the prompt word template: Context: From longitude 102.50° to 102.50°, latitude 31.00° to 29.00°, depth 130.00 km to 130.00 km, the S-wave velocity are 4.5336, 4.5370, 4.5415,4.5387, 4.5379 km / s. The general trend value of S-wave velocity change is -0.5193 km / s, and the average velocity change rate is -2.2372%. Question: Based on the known coordinates, S-wave velocity, and auxiliaryinformation for the 5 points, what is the S-wave velocity for the next 5points (longitude 102.50° to 103.00°, latitude 28.50° to 39.00°, depth 130.00km to 130.00 km)? S5: LLM performs predictions based on the questions input in step S4.

[0074] S6: Output the prediction results.

[0075] Model M performs predictions based on the context and question input in step S4, and the output results are as follows: Answer: In the range of longitude 102.50° to 103.00°, latitude 28.50° to39.00°, depth 130.00 km to 130.00 km, the predicted S-wave velocity are: 4.5716, 4.5716, 4.5716, 4.5717, 4.5717 km / s. Finally, it should be noted that the above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended 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 therein. Such 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 three-dimensional shear wave velocity prediction method based on LLM, characterized in that: It includes the following steps: S1: Obtain known three-dimensional shear wave velocity structure data; S2: Construct a cue sequence dataset; S3: Fine-tune the LLM based on the cue sequence dataset and the standard trainer, and output the fine-tuned LLM; S3.1: Randomly divide the prompt sequence dataset generated in step S2 into training set, validation set and test set according to the proportion; The training set is used for learning model parameters, the validation set is used to monitor the training process and prevent overfitting, and the test set is used to finally evaluate the model's generalization ability. S3.2: Convert the training and validation set data into standard JSON format; Each JSON record contains "input hints" and "output answer" fields; S3.3: Load the pre-trained large language model weights; S3.4: Fine-tuning the LLM based on the standard trainer provided by HuggingFace; S3.5: Output the finely tuned LLM; After fine-tuning, save the final model weights, configuration file and word segmenter to form an independently deployable model for three-dimensional shear wave velocity prediction, namely LLM. S4: Construct the context and issues that guide the LLM's focus based on the prompt word template; During the prediction phase, for new and unseen 3D velocity field regions, input prompts are constructed using the same serpentine traversal strategy and prompt word template. These input prompts include the spatial coordinates, shear wave velocities, and parameters of known points in the region to be predicted, as well as the overall trend ΔS and average velocity change percentage ΔS% calculated based on the sequence of known points to describe the overall variation trend of the 3D shear wave velocity at the known points in the region to be predicted. S5: The fine-tuned LLM performs predictions based on the question input in step S4; The constructed input prompts are input into the fine-tuned LLM. The LLM uses the context and question constructed in step S4 as conditions to autoregressively generate text descriptions of the velocity values ​​for the subsequent Step_out points; the specific numerical prediction sequence is obtained through decoding. S6: Output the prediction results; The text output generated by LLM is parsed into a structured list of velocity values, and the one-dimensional prediction sequence is reassembled back into a three-dimensional spatial grid by inverse mapping of the original serpentine path. The final output is the predicted three-dimensional shear wave velocity structure data.

2. The three-dimensional shear wave velocity prediction method based on LLM according to claim 1, characterized in that: The method for constructing the cue sequence dataset in step S2 is as follows: S2.1: Traverse the three-dimensional shear wave velocity structure data obtained in step S1 according to the snake traversal strategy to form a numerical sequence dataset; S2.2: Construct a prompt word template that maps three-dimensional shear wave velocity structure data from the numerical domain to the text domain; The three-dimensional shear wave velocity prompt template includes an input prompt section and an output prompt section; The input prompt section includes a context section and a question section. The context section provides historical information needed for prediction, while the question section is a query for the future. The output prompt section provides the answer to the question; S2.3: Convert the three-dimensional shear wave velocity numerical sequence dataset into an LLM-processable cue sequence dataset based on the cue word template; S2.3.1: Determine the input step size Step_in and the output step size Step_out; S2.3.2: Based on the prompt word template T, traverse the three-dimensional shear wave velocity numerical sequence dataset S and complete the context, question and answer for each data record; S2.3.3: Calculate the overall trend of change ΔS and the percentage change in average velocity ΔS% for the Step_in points. S2.3.4: Write the context and question of a single record into the input prompt section of the prompt word template, and write the answer into the output prompt section; S2.3.5: Determine whether the three-dimensional shear wave velocity numerical sequence dataset S has been traversed completely; If not completed, slide the window along the serpentine path, with Step_in as the window length and Step_out as the prediction step size, and generate training samples in sequence. Repeat this process until the entire numerical sequence is traversed to ensure that all possible continuous segments are sampled, so as to make full use of the data and increase sample diversity. Once the traversal is complete, the original LLM-processable cue sequence dataset is obtained.

3. The three-dimensional shear wave velocity prediction method based on LLM according to claim 2, characterized in that: The serpentine traversal strategy involves forming a spatially continuous zigzag path within the same depth layer by alternately reversing the latitude direction. This ensures that the end and start nodes of adjacent longitude columns are spatially continuous, while the layers are connected end-to-end by mirroring the longitude sequence. This flattens the three-dimensional grid-like structural data into a spatially closed, metrically continuous one-dimensional path.

4. The three-dimensional shear wave velocity prediction method based on LLM according to claim 1, characterized in that: The calculation methods for the overall change trend ΔS and the average velocity change percentage ΔS% of the overall change trend; ； ； Where: V i V0 represents the velocity value of the i-th point in the sequence consisting of input step_in points, V0 represents the velocity value of the starting point in the sequence consisting of input step_in points, and T represents the length of the sequence consisting of input step_in points.

Images