CN-122014214-A - Microseism reverse time imaging method, system, equipment and medium based on BF16 mixed precision

CN122014214ACN 122014214 ACN122014214 ACN 122014214ACN-122014214-A

Abstract

The invention belongs to the technical field of oil and gas geophysical exploration, and relates to a microseism reverse time imaging method, a system, equipment and a medium based on BF16 mixed precision. According to the invention, a BF16 and FP32 mixed storage format is used, a BF16 mixed precision-based microseism reverse-time imaging method is executed by adopting a heterogeneous parallel computing framework consisting of a CPU and a GPU, the computation speed of the seismic source reverse-time imaging can be improved by several times by introducing a BF16 mixed precision computing strategy and a CPU+GPU heterogeneous parallel, and the BF16 format is adopted and assisted with a high precision accumulator and a FP32 main copy storage strategy, so that the performance advantage brought by low precision computation is enjoyed, the numerical stability under long-term time iteration and the physical reliability of a final result are ensured, and the efficient computation and low storage use of the microseism reverse-time imaging are realized.

Inventors

ZHOU LIYE
HE ZHI
HU NAN
WANG YAOJUN
LIANG SHASHA
LI HAO
WANG LEI
LI XIAOYUE

Assignees

地球脉动(宁波)科技有限公司
宁波东方理工产业技术研究有限公司

Dates

Publication Date: 20260512
Application Date: 20260123

Claims (10)

1. The microseism inverse time imaging method based on BF16 mixed precision is characterized by using a storage format of BF16 and FP32 mixed and adopting a heterogeneous parallel computing framework consisting of a CPU and a GPU to execute, and comprises the following steps of: Generating a three-dimensional arbitrary curve physical grid based on the surface elevation data, and storing the three-dimensional arbitrary curve physical grid in an FP32 format; reading microseism data recorded by a station in an FP32 format, and reversing the microseism data on a time axis to obtain a reverse time reconstruction wave field; Performing time iteration core calculation on all time steps by CUDA at a GPU end by adopting a curve grid finite difference method in BF16 format on a three-dimensional arbitrary curve physical grid by taking an inverse time reconstruction wave field as a boundary source item, and circularly solving a first-order speed-stress equation set to obtain a wave field main copy; Decoupling and separating longitudinal wave and transverse wave from the wave field main copy in BF16 format to obtain a longitudinal wave acceleration field and a transverse wave acceleration field; acquiring imaging values of BF16 format on grid points according to the longitudinal wave acceleration field and the transverse wave acceleration field and based on the imaging condition formula; Aggregating imaging values on the grid points in all time steps, and transmitting an aggregation result back to the CPU end in an FP32 format to construct a four-dimensional imaging data body; and globally searching the maximum value in the four-dimensional imaging data volume, wherein the maximum value is the position of the microseism event, and the time corresponding to the maximum value is the occurrence time of the microseism.
2. The method for inverse time imaging of microseism based on BF16 hybrid precision as in claim 1, wherein the inverse time reconstructed wave field is obtained as follows: Wherein, the Is from the station position To any point in space Is a green's function; Is the total duration of the data; Representing a convolution; reconstructing the wavefield for inverse time; Is a spatial position coordinate; Is a time variable; is the station location; Component data recorded for the station.
3. The BF16 hybrid precision-based microseism inverse time imaging method of claim 1, wherein the performing of the time iterative core computation for all time steps circularly solves a first-order velocity-stress equation set to obtain a wavefield master copy comprises: Establishing a mapping relation between a physical space coordinate system and a calculation space Cartesian coordinate system based on the three-dimensional arbitrary curve physical grid; converting the first-order speed-stress equation set into a curve grid coordinate system through the mapping relation between a physical space coordinate system and a calculation space Cartesian coordinate system to obtain a converted first-order speed-stress equation set; Based on the converted first-order speed-stress equation set, in each time step, the wave field value is read from the FP32 global video memory, the wave field value is converted into BF16 format in the FP32 local accumulator to execute multiplication and accumulation, an accumulator result is obtained, and the accumulator result is converted back into the FP32 format and written back into the global video memory to serve as a wave field main copy.
4. The BF16 hybrid precision-based microseismic inverse time imaging method of claim 1, wherein the first-order velocity-stress equation set is as follows: Wherein, the Is the density; Is of a speed A component; Is of a speed A component; Is of a speed A component; is a stress tensor component; And Is a pull Mei Canshu; Is that Coordinates of the direction; Is that Coordinates of the direction; Is that Coordinates of the direction.
5. The BF16 hybrid precision-based microseism inverse time imaging method of claim 3, further comprising the steps of: At the CPU end, a three-dimensional medium model is subjected to regional decomposition by using a message transfer interface, and calculation tasks are distributed to a plurality of GPU nodes, wherein model parameters of microseism data and three-dimensional arbitrary curve physical grids are stored and processed by using a standard 32-bit single-precision floating point format; At the GPU end, performing first-order speed-stress equation solving and imaging condition calculating by using CUDA; A local register of the FP32 type is declared in each thread as an accumulator, and when finite difference operation is carried out, a multiplication part is completed under BF16, and a multiplication accumulation process is carried out in the FP32 accumulator; After carrying out back-transfer iteration on each time step on a time axis, obtaining a reverse time reconstruction wave field at the next moment, and when writing the reverse time reconstruction wave field back to the GPU global video memory, converting the longitudinal wave acceleration field and the transverse wave acceleration field from the format of a BF16 or FP32 accumulator back to the FP32 format, and then storing.
6. The BF16 hybrid precision-based microseism inverse time imaging method of claim 1, wherein the acquiring the imaging values at the grid points from the longitudinal wave acceleration field and the transverse wave acceleration field and based on the imaging condition formula comprises: acquiring an energy propagation direction vector of the longitudinal wave acceleration field according to the longitudinal wave acceleration field; acquiring an energy propagation direction vector of the transverse wave acceleration field according to the transverse wave acceleration field; acquiring imaging values on grid points according to the energy propagation direction vector of the longitudinal wave acceleration field and the energy propagation direction vector of the transverse wave acceleration field and based on the imaging condition formula; The imaging condition formula is as follows: Wherein, the For grid points At the moment of time Is a function of the imaging value of (a); is the energy propagation direction vector of the longitudinal wave acceleration field; Is the energy propagation direction vector of the transverse wave acceleration field.
7. The BF16 hybrid precision-based microseism inverse time imaging method of claim 6, wherein the energy propagation direction vector of the longitudinal wave acceleration field is obtained by the following formula: the energy propagation direction vector of the transverse wave acceleration field is obtained by the following formula: Wherein, the Is a displacement tensor; is the energy propagation direction vector of the longitudinal wave acceleration field; Is the energy propagation direction vector of the transverse wave acceleration field; Is the divergence of the velocity field; Is a longitudinal wave acceleration wave field; And Is a displacement tensor; Is a transverse wave acceleration wave field; spatial derivative of the velocity field.
8. Microseism reverse time imaging system based on BF16 hybrid accuracy, characterized by comprising: The three-dimensional arbitrary curve physical grid acquisition module is used for generating a three-dimensional arbitrary curve physical grid based on the surface elevation data and storing the three-dimensional arbitrary curve physical grid in an FP32 format; the reverse time reconstruction wave field acquisition module is used for reading the microseism data recorded by the station in the FP32 format, and reversing the microseism data on a time axis to obtain a reverse time reconstruction wave field; the elastic wave propagation result acquisition module is used for performing time iteration core calculation on all time steps through CUDA at the GPU end by adopting a curve grid finite difference method in BF16 format on a three-dimensional arbitrary curve physical grid by taking a reverse time reconstructed wave field as a boundary source item, and circularly solving a first-order velocity-stress equation set to obtain a wave field main copy; the acceleration field decomposition module is used for decoupling and separating longitudinal waves and transverse waves of the wave field main copy in a BF16 format to obtain a longitudinal wave acceleration field and a transverse wave acceleration field; The imaging value acquisition module on the grid points is used for acquiring imaging values of BF16 format on the grid points according to the longitudinal wave acceleration field and the transverse wave acceleration field and based on the imaging condition formula; The four-dimensional imaging data body acquisition module is used for aggregating imaging values on the grid points in all time steps, and transmitting an aggregation result back to the CPU end in an FP32 format to construct a four-dimensional imaging data body; the microseism event confirming module is used for globally searching the maximum value in the four-dimensional imaging data volume, wherein the maximum value is the position of the microseism event, and the time corresponding to the maximum value is the occurrence time of the microseism.
9. An electronic device comprising a processor, a memory, the electronic device storing computer program instructions, characterized in that it implements the BF16 hybrid precision-based microseism inverse time imaging method of any of claims 1-7 when executing the computer program.
10. A storage medium storing computer program instructions which, when loaded and executed by a processor, perform the BF16 hybrid precision based microseismic reverse time imaging method of any of claims 1-7.

Description

Microseism reverse time imaging method, system, equipment and medium based on BF16 mixed precision Technical Field The invention belongs to the technical field of oil and gas geophysical exploration, and relates to a microseism reverse time imaging method, a system, equipment and a medium based on BF16 mixed precision. Background Hydraulic fracturing is a core technology for commercial exploitation of shale gas, which creates a fracture network by injecting high pressure liquids into the subsurface to increase reservoir permeability. During this process, the fracture of the rock formations induces a large number of weak seismic events, i.e. "microseisms". The time-space distribution of the microseism events is accurately monitored, and the method has important guiding significance for evaluating fracturing effects, describing fracture network morphology and predicting productivity. Source reverse Time imaging (Time-REVERSE IMAGING, TRI) is a leading edge technique in the current microseismic monitoring field. The technology is based on wave equation theory, and the complete microseism waveform data recorded by the ground station is reversely transmitted back to the underground medium as a wave source. By applying specific imaging conditions, the energy of different types of seismic waves (such as longitudinal waves and transverse waves) can be focused at the true source position and origin time, thereby completing the accurate positioning of microseismic events. Compared with the traditional positioning method relying on time-of-arrival pickup, the reverse-time imaging does not need manual intervention to pick up the earthquake phase, has stronger adaptability to low signal-to-noise ratio data, and can process any complex underground medium theoretically. However, existing source reverse time imaging techniques still face significant challenges in practical applications, mainly manifested by computational resource bottlenecks. The source reverse time imaging is a computationally and memory intensive process. The method involves fine mesh dissection of a three-dimensional space and iterative solution of wave equations over thousands to tens of thousands of time steps, requiring significant computational effort and memory (particularly GPU memory) overhead. The bottleneck not only makes the calculation cost high and the period long, and the calculation efficiency low, greatly limits the popularization and application of the method in production practice, but also makes the method difficult to be combined with artificial intelligent technologies such as deep learning and the like which need mass data for model training, and cannot meet the requirement of the artificial intelligent technologies such as deep learning and the like on the rapid processing of mass data. Disclosure of Invention The invention aims to provide a microseism reverse-time imaging method, a system, equipment and a medium based on BF16 mixed precision, which are used for solving the technical problems that the existing earthquake focus reverse-time imaging technology has low calculation efficiency in practical application and cannot meet the requirement of the artificial intelligence technology such as deep learning on fast processing of mass data. In order to achieve the above purpose, the invention is realized by adopting the following technical scheme: in a first aspect, the present invention provides a microseism inverse time imaging method based on BF16 hybrid precision, which uses a storage format of BF16 and FP32 hybrid, and is executed by adopting a heterogeneous parallel computing framework composed of a CPU and a GPU, and includes the following steps: Generating a three-dimensional arbitrary curve physical grid based on the surface elevation data, and storing the three-dimensional arbitrary curve physical grid in an FP32 format; reading microseism data recorded by a station in an FP32 format, and reversing the microseism data on a time axis to obtain a reverse time reconstruction wave field; Performing time iteration core calculation on all time steps by CUDA at a GPU end by adopting a curve grid finite difference method in BF16 format on a three-dimensional arbitrary curve physical grid by taking an inverse time reconstruction wave field as a boundary source item, and circularly solving a first-order speed-stress equation set to obtain a wave field main copy; Decoupling and separating longitudinal wave and transverse wave from the wave field main copy in BF16 format to obtain a longitudinal wave acceleration field and a transverse wave acceleration field; acquiring imaging values in BF16 format on grid points according to the longitudinal wave acceleration field and the transverse wave acceleration field and based on the imaging condition formula; Aggregating imaging values on the grid points in all time steps, and transmitting an aggregation result back to the CPU end in an FP32 format to construct a four-dimensional imaging dat