CN-122017987-A - LLM-based three-dimensional transverse wave speed prediction method

CN122017987ACN 122017987 ACN122017987 ACN 122017987ACN-122017987-A

Abstract

The invention provides a method for predicting three-dimensional transverse wave speed based on LLM, namely, a prompting word template adapting to a three-dimensional transverse wave speed prediction task is designed, a prompting sequence data set generated by using known three-dimensional transverse wave speed data is utilized, LLM is finely adjusted by combining a standard trainer, and three-dimensional transverse wave speed is predicted by means of the finely adjusted LLM. The method has the capacity of supplementing three-dimensional transverse wave speed data, can alleviate the problems of space coverage blind areas and low resolution in inversion of the three-dimensional transverse wave speed by the traditional surface wave imaging method, provides a more complete data basis for subsequent geological structure analysis, utilizes a part of known three-dimensional transverse wave speeds to predict, replaces the traditional full-area observation inversion of the three-dimensional transverse wave speed, and reduces the calculation cost of predicting the three-dimensional transverse wave speed in a large scale and high resolution. The invention promotes the deep fusion of artificial intelligence and geological structure research, and better serves for the development of the field of earth science.

Inventors

ZHAO GUOAN
LAN XIAOJUAN
WANG SHAOBO
GAO ZEHUA
LAN CHUWEN
YUAN YI
GAO YUANLI

Assignees

北京邮电大学

Dates

Publication Date: 20260512
Application Date: 20260121

Claims (4)

1. A three-dimensional transverse wave speed prediction method based on LLM is characterized by comprising the following steps: s1, acquiring known three-dimensional transverse wave speed structure data; s2, constructing a prompt sequence data set; s3, fine-tuning the LLM based on the prompt sequence data set and the standard trainer, and outputting the fine-tuned LLM; s3.1, randomly dividing the prompting sequence data set generated in the step S2 into a training set, a verification set and a test set according to a proportion; The training set is used for model parameter learning, the verification set is used for monitoring the training process and preventing overfitting, and the test set is used for finally evaluating the generalization capability of the model; s3.2, converting the training set and the verification set data into a standard JSON format; each JSON record contains an "input hint" and an "output answer" field; S3.3, loading the weight of the pre-trained large language model; S3.4, fine-tuning LLM based on the standard trainer provided by HuggingFace; s3.5, outputting the fine-tuned LLM; after the fine tuning is finished, a final model weight, a configuration file and a word segmentation device are saved, and a model which can be independently deployed and is used for three-dimensional transverse wave speed prediction, namely LLM, is formed; S4, constructing a context and a problem for guiding the attention of the LLM based on the prompt word template; in the prediction stage, aiming at a new and unseen three-dimensional speed field area, constructing an input prompt according to the same serpentine traversal strategy and prompt word template, wherein the input prompt part comprises the space coordinates of known points of the area to be predicted, the transverse wave speed and parameters, and the input prompt part further comprises an overall change trend delta S and an average speed change percentage delta S, wherein the overall change trend delta S is calculated based on the known point sequence and is used for describing the overall change trend of the three-dimensional transverse wave speed of the known points of the area to be predicted; S5, performing prediction by the fine-tuned LLM according to the problem input in the step S4; Inputting the constructed input prompt into the fine-tuned LLM, and taking the constructed context and the problems in the Step S4 as conditions, automatically regressively generating a speed value text description of subsequent step_out points by the LLM; S6, outputting a prediction result; and analyzing the text output generated by LLM into a structured speed value list, recombining the one-dimensional prediction sequence back to the three-dimensional space grid according to the inverse mapping of the original serpentine path, and finally outputting the three-dimensional prediction sequence as predicted three-dimensional transverse wave speed structure data.
2. The method for predicting three-dimensional shear wave velocity based on LLM according to claim 1, wherein the method for constructing the cue sequence data set in step S2 comprises the following steps: S2.1, traversing the three-dimensional transverse wave speed structure data obtained in the step S1 according to a snake-shaped traversing strategy to form a numerical sequence data set; s2.2, constructing a prompt word template for mapping the three-dimensional transverse wave speed structure data from a numerical domain to a text domain; The three-dimensional transverse wave speed prompt word template comprises an input prompt part and an output prompt part, wherein the input prompt part also comprises a context part and a question part, the context part provides history information required by prediction, and the question part is a future query; S2.3, converting the three-dimensional transverse wave speed numerical value sequence data set into a prompt sequence data set which can be processed by LLM according to the prompt word template; S2.3.1 determining an input Step step_in and an output Step step_out; S2.3.2 traversing the three-dimensional transverse wave speed numerical sequence data set S based on the prompt word template T to complement the context, the questions and the answers of each data record; s2.3.3 calculating the overall change trend delta S and the average speed change percentage delta S of the step_in points; S2.3.4 writing single record context and questions into an input prompting part of a prompting word template, and writing answers into an output prompting part; s2.3.5, judging whether the three-dimensional transverse wave velocity numerical sequence data set S is traversed; If not, sliding a window along the serpentine path, taking step_in as a window length, and step_out as a prediction Step length, sequentially generating training samples, repeating the process until the whole numerical sequence is traversed, and ensuring that all possible continuous fragments are sampled so as to fully utilize data and increase sample diversity; and if the traversal is completed, obtaining the prompting sequence data set which can be processed by the original LLM.
3. The method of predicting three-dimensional shear wave velocity based on LLM according to claim 2, wherein the serpentine traversal strategy is to form spatially continuous zigzag paths in the same depth layer by alternately reversing latitude directions, so that the tail ends of adjacent longitude columns are spatially continuous with the starting nodes, and the end-to-end connection is realized by mirroring longitude sequences between layers, so that three-dimensional latticed structure data are flattened into structure data of a spatially closed one-dimensional path with continuous measurement.
4. The LLM-based three-dimensional shear wave velocity prediction method according to claim 1, wherein the overall trend of variation ΔS of the overall trend of variation and the calculation method of the average percent change ΔS% of velocity; ; ; Wherein V i represents the speed value of the ith point in the sequence of input Step step_in points, V 0 represents the speed value of the starting point in the sequence of input Step step_in points, and T represents the length of the sequence of input Step step_in points.

Description

LLM-based three-dimensional transverse wave speed prediction method Technical Field The invention relates to a three-dimensional transverse wave speed prediction method, in particular to a three-dimensional transverse wave speed prediction method based on LLM. The invention belongs to the technical field of artificial intelligence data processing. Background The shear wave velocity is one of the important physical parameters of the rock in the earth, and can reflect the characteristics of rock density, elastic modulus and the like. The traditional transverse wave speed extraction method has the defects of acoustic logging, petrophysical testing and the like, and only one-dimensional (depth profile) transverse wave speed and two-dimensional (along-line profile) transverse wave speed can be extracted. When the earth dynamics process is deeply understood and the complex geological structure is accurately described, the three-dimensional transverse wave speed needs to be mastered. The three-dimensional transverse wave velocity is a spatially distributed parameter describing the elastic properties of an underground medium, and is centered on locating the propagation velocity value of the transverse wave by three-dimensional spatial coordinates (longitude, latitude, depth). The three-dimensional shear wave velocity structural model can be displayed in a specific area inside the earth, and how the shear wave velocity changes with the changes of longitude, latitude and depth. The three-dimensional transverse wave velocity has important significance for understanding complex geological structures in the earth, evaluating earthquake disasters, exploring oil gas, geothermal resources and the like. The existing research mostly utilizes a surface wave imaging method based on background noise to invert the three-dimensional transverse wave velocity in the earth, but the method still has some limitations and defects in practical application, and is mainly characterized in that 1, the method is limited by the distribution number of observation stations, the ray route coverage density is limited, the three-dimensional transverse wave velocity data obtained by inversion is limited, and the geological structure in the earth cannot be clearly described. 2. The resolution of the inversion results is significantly reduced in the edge region, and the resolution is insufficient. For example, in the north of Qinghai-Tibet Gao Yuandong, although the structure recovery is better over 75km in the center region, the resolution is reduced in the edge region due to the sparse data coverage. In another example, in inversion of a basin in the Han, short-period surface wave data has weak constraint on shallow layers, resulting in unstable inversion results in areas with large lateral changes in shallow velocities. 3. The calculation cost of the surface wave imaging data processing is high, and the application of the surface wave imaging data processing in a real-time monitoring or quick response scene is limited. Background noise data is usually recorded continuously for a long time, the amount of data to be processed is huge, and huge calculation resources and time cost are required for performing cross-correlation calculation, dispersion curve extraction and three-dimensional tomography inversion on the data. This problem is particularly pronounced when large-scale, high-resolution imaging is performed. Currently, there are a great deal of research on the prediction of transverse wave velocity based on "empirical formula method", "machine learning method" and "deep learning method", however, there are relatively few research on further prediction of known three-dimensional transverse wave velocity by inversion. In view of the above, the invention aims to provide a method for further predicting three-dimensional transverse wave speed by utilizing three-dimensional transverse wave speed obtained by inversion based on LLM, in particular to predict the three-dimensional transverse wave speed of an incomplete coverage area of an observation station, solve the problems of space coverage blind area and limited resolution of the traditional surface wave imaging method, and provide a more complete data base for subsequent geological structure analysis. Disclosure of Invention In view of the above, the present invention aims to provide a three-dimensional shear wave velocity prediction method based on LLM. The method comprises the steps of designing a prompt word template adapting to a three-dimensional transverse wave speed prediction task, utilizing a prompt sequence data set generated by known three-dimensional transverse wave speed data, combining a standard trainer to finely tune a large language model (LLM for short), and predicting the three-dimensional transverse wave speed by means of the finely tuned LLM In order to achieve the purpose, the invention adopts the following technical scheme that the three-dimensional transverse wave speed pr