摘要
地球物理测井是探测地下油气储层流体类型和评价储层参数的重要手段,传统测井解释方法面临挑战。人工智能(AI)算法先进,精度高,在测井解释方面具有独特的优势,“测井+AI”是当前测井解释研究的新领域。然而,智能测井解释中,样本规模小、训练模型泛化能力弱导致单纯机器学习的测井解释方法难以推广应用。物理模型包含从测井数据到地质目标的内在机理,将数据驱动与机理驱动相结合是提高测井解释精度的有效途径。现有的数据—机理联合驱动缺乏范式遵循,为此,聚焦智能测井解释参数预测,提出了数据—机理联合驱动的概念、思路,并总结出两个范式:一是数据引导的物理建模,以物理建模为主导,其中关键步骤或参数采用数据驱动获得,数据驱动为辅;二是物理引导的机器学习,以机器学习为主导,知识模型或物理机理为辅助,对输入数据、损失函数、训练过程进行监督和约束。因此,提出了物理模型增广数据集、知识驱动样本加权和岩石物理知识迁移三种数模双驱模式。将上述数模双驱的范式或模式应用于致密砂岩、有机页岩储层参数预测和矿物含量预测。与单纯数据驱动的机器学习相比,数据—机理联合的双轮驱动范式能够显著提高测井解释模型对小样本、差样本的学习能力,模型的稳健性更好,泛化能力更强,解释精度更高。
Geophysical logging plays a crucial role in detecting fluid types in subsurface oil and gas reservoirs and evaluating reservoir parameters.Traditional log interpretation methods face significant challenges.Artificial intelligence(AI)algorithms offer advanced capabilities and high accuracy,which makes them highly advantageous for log interpretation.The integration of“logging+AI”has emerged as a new research direction.However,in intelligent log interpretation,the limited sample size and weak generalization ability of training models hinder the widespread application of purely machine learning‑based log interpretation methods.Physical models inherently capture the underlying mechanisms that connect logging data to geological targets.Combining data‑driven and mechanism‑driven approaches provides an effective way to enhance log interpretation accuracy.However,existing joint data‑mechanism driving lacks a well‑defined paradigm.In view of this,the study focuses on the prediction of intelligent log interpretation parameters,proposes the concept and methodology of joint data‑mechanism driving,and presents two key paradigms:data‑guided physical modeling,where physical modeling is the primary framework,with data‑driven methods assisting in obtaining key steps or parameters,and physics‑guided machine learning,where machine learning is the primary approach,while knowledge models or physical mechanisms provide supervision and constraints on input data,loss functions,and training processes.To implement these paradigms,three hybrid models are proposed:physics‑augmented datasets,knowledge‑driven sample weighting,and rock physics knowledge transfer.These approaches are applied to predict reservoir parameters and mineral composition in tight sandstone and organic shale reservoirs.Compared with purely data‑driven machine learning models,the proposed data‑mechanism jointly driving paradigms significantly im‑prove the ability of the log interpretation model to learn from small and low‑quality samples and make the model have enhanced robustness,generalization ability,and interpretation accuracy.
作者
谭茂金
白洋
张博栋
TAN Maojin;BAI Yang;ZHANG Bodong(Borehole Detection Technology Branch,National Engineering Research Center for Offshore Oil and Gas Exploration,China University of Geosciences,Beijing 100083,China;School of Geophysics and Information Technology,China University of Geosciences,Beijing 100083,China)
出处
《石油地球物理勘探》
北大核心
2025年第4期966-977,共12页
Oil Geophysical Prospecting
基金
国家自然科学基金项目“深层非均质页岩气藏岩石物理多尺度融合与储层品质井震智能评价方法”(42430810)
“井旁声波远探测三维成像与智能解释方法研究”(42174149)
海洋油气勘探国家工程研究中心开放基金联合资助。
关键词
测井解释
机理模型
机器学习
联合驱动
储层智能评价
log interpretation
mechanistic model
machine learning
joint driving
intelligent reservoir evaluation
