Reservoir physical parameters serve as fundamental quantitative indices for characterizing the storage capacity and fluid percolation potential of subsurface reservoirs.Well logging interpretation,a critical methodology for accurately estimating these parameters,constitutes a sophisticated nonlinear regression challenge.To address the inherent limitations of existing petrophysical parameter interpretation techniques,particularly their inadequate generalization performance under few-shot learning conditions,this investigation systematically devises a dual-framework analytical approach.This study initially proposes a sample optimization methodology based on cluster analysis.The spatial configuration of samples is partitioned through the implementation of the K-means clustering algorithm,followed by selective sample curation according to spatial distribution char-acteristics to maximize learning sample diversity.Building upon this optimized sample architecture,the study further introduces a hierarchical residual neural network-based interpretation framework for petrophysical parameter estimation.The proposed methodology enhances conventional fully connected neural architecture through four innovative mechanisms:(1)Integration of cross-layer residual connections facilitates progressive refinement of residual mappings between multivariate logging inputs and target petrophysical outputs,thereby enabling hierarchical abstraction of complex petrophysical relationships from limited training instances.(2)The integration of ensemble learning paradigms amalgamates diverse machine learning methodologies,effectively mitigating overfitting risks through algorithmic diversity.(3)The implementation of a multi-task learning framework establishes intrinsic correlations between porosity and permeability interpretation tasks via shared latent representations,thereby enhancing individual task generalizability under data scarcity constraints.(4)The introduction of a quadratically weighted root mean square error loss function preferentially reduces interpretation errors in high-permeability reservoir intervals.Results from 90 rigorously designed comparative experimental configurations in the study area demonstrate that the cluster-based sample opti-mization methodology effectively enhances generalization performance across multiple machine learning models under few-shot learning constraints.Application of the proposed hierarchical residual neural network framework for well-logging interpretation of reservoir porosity and permeability within the investigated reservoir area achieves coefficients of determination of 88%and 94%,respectively,demonstrating statistically significant superiority over conventional methodologies in both petrophysical interpretation accuracy and generalization capability.Blind testing validation on cored wells reveals 12 and 20 percentage point improvements in predictive precision compared to other various existing methodologies,the proposed approach in this study demonstrates substantial advancements in addressing few-shot learning challenges through algorithm optimization strategies encompassing distribution-based sample selection and multi-task collaborative frameworks.This methodology significantly enhances feature representation fidelity in petrophysical datasets,exhibiting superior petrophysical interpretation accuracy and enhanced generalization capabilities.

Total organic carbon(TOC)content is a crucial geochemical parameter for assessing reservoir quality and hydro-carbon generation potential of source rocks.The accurate prediction of TOC content is important for optimizing the exploration and development processes of shale oil and gas.With the rapid development of artificial intelligence technologies,individual machine learning algorithms have been increasingly applied to evaluate TOC content in shale.Despite the promising results of the individual machine learning algorithms,they are often subject to several challenges including overfitting,underfitting,and getting trapped in local optima of objective function.To address these limitations,the ensemble learning models are developed.Ensemble learning models leverage the strengths of multiple individual intelligent algorithms to enhance prediction accuracy and stability.Among them,combination strategy is one of the key factors in optimizing the ensemble learning models.Arithmetic average method as the simplest combination strategy fails to fully use prediction performance of the best individual intelligent model,and it can be severely affected by the individual intelligent model with a large prediction error,which can interfere with prediction outcome of overall model.In comparison,weighted summation method as a common combination strategy assigns the weights to different individual intelligent models according to their performance on training data.This method will perform excellently on training set,but it tends to have a poor performance when applied to test set.This paper develops an ensemble model based on an intelligent matching technology(IMTEM).The proposed method utilizes a set of robust intelligent algorithms including extreme gradient boosting,random forest,support vector machine,and extreme learning machine as algorithm modules to initially process input data.Then,the processed feature information combined with original log responses is fed to feedforward neural network layer for nonlinear transformation and feature learning,thereby enabling accurate and continuous estimation of TOC content in shale.To validate effectiveness of the IMTEM,the proposed method is applied to the prediction of TOC content in the Longmaxi Formation shale in the Sichuan Basin.Test results indicate that,compared to two ensemble models,five baseline models,and the ΔlogR method,predictions of the IMTEM exhibit higher consistency with measured TOC content.This demonstrates that the IMTEM is more suitable for predicting TOC content in shale.

Well logging curves are essential for recording the physical parameters of formations during drilling,providing vital information for analyzing rock properties,evaluating hydrocarbon reservoirs,and understanding reservoir distribution.As oil and gas exploration continues to progress,the complexity of subtle and hidden reservoirs has increased,posing challenges for traditional exploration techniques.Despite their importance,conventional well logging data suffer from low resolution,which significantly limits their ability to address the requirements of detailed reservoir characterization.In particular,the inability to precisely identify modification points in thin interbedded reservoirs remains a critical bottleneck in reservoir analysis.To overcome these limitations,developing high-resolution interpretation methods for well logging data has become an urgent priority in the field of reservoir analysis and geological exploration.This study proposes a novel reservoir prediction model based on the ResNet50 regression algorithm.By integrating vertically continuous optical thin-section data,which can capture fine-scale and complex vertical geological features,with five conventional well logging parameters,the proposed model aims to improve the resolution and accuracy of reservoir analysis.This combination leverages the strengths of image-based geological analysis and traditional well logging to deliver a more precise interpretation of subsurface formations.The model was validated using data collected from five intervals of the Permian formation in a specific well area.A total of 570 continuous geological image samples,combined with their corresponding well logging data,were utilized for model training and prediction.The results demonstrate that the model effectively enhances the resolution of well logging data,improving it from the traditional 12.5 cm to 6.25 cm.This significant improvement not only increases the precision of well logging interpretation but also provides a more detailed under-standing of reservoir characteristics.The model's performance was rigorously evaluated using three widely recognized metrics:the coefficient of determination(R2),root mean square error(RMSE),and mean absolute error(MAE).The results revealed that the model excels in predicting parameters such as acoustic time(AC),compensated neutron(CNL),resistivity(RT),and gamma ray(GR),achieving an average prediction error below 0.094.This highlights the model's reliability and superior performance in reservoir prediction tasks.However,challenges remain in predicting density(DEN),where the model's accuracy is impacted in intervals with significant lithological heterogeneity or complex geological conditions.

The progressive application of artificial intelligence technology within oil and gas exploration has resulted in an inevitable shift towards the transformation of geomechanical parameter prediction from a traditional to an intelligent approach.This paper presents a comprehensive review and critical analysis of machine learning algorithms in the direct and indirect pre-diction of rock mechanics parameters,pre-drilling prediction,monitoring while drilling and post-drilling evaluation of formation pore pressure,1D in-situ stresses and 3D in-situ stresses field prediction.Furthermore,the paper compared machine learning models,input parameters,sample data volume,output parameters,and model prediction performance under different tasks.It has been demonstrated that machine learning algorithms exhibit superior performance in terms of accuracy,timeliness,and applicability in geomechanical parameter prediction compared to laboratory tests,field tests,and empirical model calculations.The current research emphasis is on hybrid models,deep learning models,and physical-constrained neural network models,which have been validated as highly accurate,robust,capable of generalization,and easily interpretable.However,the existing research primarily concerns the prediction of 1D geomechanical parameters post-drilling.Consequently,it is not possible to effectively predict 3D geomechanical parameters prior to drilling or during the drilling process.In order to facilitate the digital and intelligent transformation of geomechanical parameters,an intelligent prediction framework for geomechanical parameters is proposed in this paper.This framework considers the influence of multi-source data,including seismic,logging,and mud log data on the prediction of geomechanical parameters.The machine learning model,which is driven by data and physics,enables the prediction of 3D geomechanical parameters.This model is updated in real-time through the most recent drilling data,thus allowing for the pre-drilling prediction,monitoring while drilling and post-drilling evaluation of regional 3D geomechanical parameters.In addition,the key technical problems facing the intelligent prediction of geomechanical parameters are identified:(1)The transformation of unstructured data types should be minimized,the complexity of the data set should be reduced,and the consistency and comparability of the data should be ensured.(2)Multi-source data fusion should be conducted,and multi-source data sets,including seismic,logging,mud log,laboratory tests,and field test data,should be constructed.Subsequently,data processing and feature selection should be performed.(3)Machine learning models should be enhanced to improve performance,integrated models should be adopted to improve prediction accuracy,and mechanism models and domain knowledge should be integrated to enhance model robustness and explainability.

Random noise and coherent noise interfere with seismic data collected in the field,resulting in the reduction of the signal-to-noise ratio,which affects the subsequent processing of seismic data,such as seismic migration and imaging.There-fore,it is necessary to develop an efficient and adaptive method to attenuate random and coherent noise in real seismic data.Conventional supervised learning algorithms need to manually generate a large number of labels to train the network,which is very difficult in the field of seismic exploration where the data volume is small.In addition,supervised learning-based noise attenuation methods are expensive in terms of computation and labor costs.To solve this problem,this paper constructs an adaptive deep learning framework based on unsupervised learning strategies to attenuate random and irregular(erratic)noise in multi-dimensional seismic data.This method uses the corresponding structure of encoding and decoding to compress and reconstruct data features.In order to improve the network's attention to important waveform features,this paper uses a soft attention mechanism to assign more weight to important waveform features in a weighted way.In this paper,the multi-di-mensional noisy data is segmented into a large number of one-dimensional noisy signals,which are input into the network for iteration,so as to adaptively attenuate random and erratic noise in seismic data.This small-scale signal denoising method can effectively improve the noise attenuation performance of the network and help to avoid artifacts.In this paper,we use a more robust Huber loss function to attenuate random and erratic noise,which combines the root-mean-square error with l2 norm and the average absolute error loss with l1 norm.In addition,a Total Variation(TV)regularization term is added to the constructed network to capture the local smooth structure of the seismic data.By adjusting the weight of Huber loss function and TV regularization term,the network can obtain the best denoising performance.The method constructed in this paper can be directly used for attenuation of random and erratic noise of multi-dimensional seismic data,and ensure transverse continuity of seismic signals after reconstruction.We compare the proposed framework with classical seismic data denoising methods and noise attenuation methods based on unsupervised learning to analyze the advantages and disadvantages of each method.The test results of 2D and 3D synthetic data and actual seismic data show that the proposed method has better noise attenuation and useful signal protection capabilities.

高分辨率地震数据在地震数据处理中扮演着关键角色,特别是当地震勘探目标变得越来越复杂时,它可以提供更准确的储层识别和描绘.近年来,随着深度学习技术的快速发展,它越来越多地被引入到高分辨率地震数据处理中.基于大量标记数据,建立了低分辨率地震数据和高分辨率地震数据之间的复杂非线性关系.然而,深度学习在高分辨率数据处理中的精度与稳定性高度依赖于训练集的准确性与多样性.深度学习技术在生产中实际应用的主要挑战之一是稀疏的井数据,这经常导致训练集受限.为了解决这个问题,本文提出了一种基于深度学习的高分辨率处理方法,通过使用大量逼真的训练集,将井数据所表示的分层结构与地震数据所表示的空间地质结构相结合.建立训练集包括三个步骤:(1)使用井数据计算波阻抗序列,并利用高斯匹配函数拟合波阻抗高频部分的振幅分布,得到一个概率密度函数,最后生成一系列符合井数据统计分布的波阻抗序列.(2)在波阻抗序列的基础上,建立二维水平分布的波阻抗模型,并逐步添加折叠变形、倾角变形和断层变形,生成包含各种地质模式的二维阻抗模型.(3)使用阻抗模型计算反射系数,然后用反射系数模型分别卷积低频和高频子波,得到训练集.通过自动生成具有地下地质信息的大量训练集,训练的网络可以估计稳定而准确的高分辨率结果.深度学习的框架由两个部分组成:提取输入数据特征的编码部分和通过提取的特征重建输出的解码部分.此外,残差模块被整合到框架中,使网络更有效地从训练集中提取特征进而提高网络性能,从而实现计算精度和效率之间更好的平衡.通过模型数据和实际数据的测试,本文提出的方法相比于传统深度学习方法对噪声具有更好的鲁棒性,可以产生更精确且横向连续性更好的高分辨率结果.

湖相重力流是目前沉积学研究的热点与难点,也是致密油、页岩油富集的有利场所,鄂尔多斯盆地庆城油田三叠系延长组作为致密油、页岩油的典型代表已显示出巨大的勘探开发前景.然而,由于湖相重力流砂体分布认识不清,导致油田并未达到预期的开发效率.本文采用支持向量机(SVR)的机器学习方法,先优选频段再优选属性,建立分频属性与测井解释砂体厚度的非线性映射关系,实现了致密砂岩的定量预测.研究结果表明,低频地震属性适合预测厚层砂体,高频地震属性适合预测薄层砂体;采用机器学习的方法,将不同频率的地震属性智能融合,能够兼顾预测不同厚度砂体,既提高了地震属性的解释精度,又降低了地震解释的多解性,实现了砂体厚度的定量预测.检验结果显示,智能融合属性与砂体厚度的分布趋势与值域区间基本一致,智能融合属性预测砂体分布的可靠性明显提高,与测井解释砂体厚度的相关性由0.60提高至0.79,大多数井点处预测的砂体厚度误差小于5 m.继而,根据融合属性与测井解释,刻画了研究区的沉积微相展布特征:研究区目的层发育湖底扇沉积,细分为分支水道、朵叶主体、朵叶侧缘、滑塌体与朵叶间/水道间5种沉积微相;砂体主体呈扇形连片式沉积,厚度顺物源方向逐渐减薄;分支水道呈窄条带状树形分叉,下切发育于朵叶体之上;朵叶体沉积为研究区的沉积主体;滑塌体为湖底扇前端失稳滑塌形成的小规模孤立砂体,长轴方向多平行于湖底扇前端.研究成果对油田下一步高效开发具有重要意义.

水平地应力是井壁稳定分析和水力压裂改造的关键基础参数,但深部地层地质环境复杂且隐蔽,使得水平地应力的准确快速预测难度较大.考虑到传统测井解释和神经网络模型难以描述测井数据与地应力之间的空间相关性,提出采用一种基于双向长短期记忆神经网络(BiLSTM)的水平地应力预测方法;以四川盆地CL气田两口直井为例,将两口直井分别作为训练井和测试井,通过训练井建立测井参数与地应力之间的非线性映射关系,实现对测试井水平地应力的预测;结合测井参数相关性和实际地质含义,分析了不同测井参数组合模式下水平地应力的预测效果.研究结果表明:①对比测井解释和岩心差应变测试结果发现,垂向地应力测井解释误差为0.39％,最大水平地应力测井解释误差为0.18％～0.64％,最小水平地应力测井解释误差为0.29％,说明测井解释与实际地应力吻合较好;②工区地应力大小排序为垂向地应力>最大水平地应力>最小水平地应力,属于潜在正断层应力状态;③水平地应力与垂深、密度和自然伽马呈较强的正相关关系,与纵波时差、井径、补偿中子和横波时差呈负相关关系;④不同的测井参数组合对水平地应力的预测效果不同,其中最优的测井参数组合为垂深、井径、密度、补偿中子、自然伽马、纵波时差;⑤通过正交设计实验,确定了最优超参数取值方案,其预测得到的最大和最小水平地应力平均绝对百分比误差分别为0.48‰和0.50‰.结论认为,BiLSTM模型能够有效捕捉测井参数随深度的变化趋势和测井参数的前后关联信息,可以实现水平地应力的精准预测.

地层孔隙压力是油气井从设计到完钻过程中重要的基础数据,准确计算地层孔隙压力是保障钻井安全、提高钻井效率的重要前提.为了克服传统地层孔隙压力计算方法精度不足、计算效率不高的问题,本文考虑到钻井与地层沉积均为序列性和非线性过程,提出了将长短期记忆神经网络(LSTM)和误差反向传播神经网络(BP)相结合计算地层孔隙压力的方法,利用LSTM层提取钻-测-录多源数据中的序列性特征信息,经过BP层构建特征信息与地层孔隙压力之间的非线性映射关系.通过对油田现场钻测录数据进行清洗并综合相关性分析和钻井经验知识优选了18种输入参数,对LSTM-BP地层孔隙压力计算模型进行训练和测试,并采用网格搜索法对LSTM-BP神经网络模型的5种模型超参数进行了优选,效果最优的单井计算模型和邻井计算模型的平均绝对误差分别为4.92 MPa和2.34 Mpa,均方根误差分别为6.65 Mpa和3.03 Mpa,平均相对误差分别为4.36％和8.31％.最后与传统BP模型、LSTM模型和支持向量机(SVM)模型的最优结果进行对比,结果显示,本文所建立的LSTM-BP神经网络模型精度均高于BP模型、LSTM模型和SVM模型,表明所提出的LSTM-BP地层孔隙压力计算模型具有较高的计算精度.

机器学习算法已经成为工程领域建模的有力工具,这些方法通过使用复杂结构或多重非线性变换从更高的维度拟合多个变量之间的非线性关系,适用于解决工程中由于变量关系复杂,传统物理模型或经验模型无法有效解决的问题.由于传统的测井资料解释方法是基于岩石物理机理和模型为基础,需要许多假设条件,在实际应用中存在一定偏差.因此,利用机器学习算法对测井资料进行处理和解释,尤其是对储层流体进行识别具有重要意义.现有的储层流体识别方法没有充分挖掘测井数据多维空间的关联性.此外,还存在储层类别分布严重不均衡、相似物性储层的识别容易混淆的问题.本文针对储集层流体识别任务提出了基于机器学习算法的高效解决方案.采用长短期记忆网络(LSTM)和卷积神经网络(CNN)分别表征测井曲线时序特征以及多条测井曲线之间的相互关联关系;考虑到油气储集层识别任务的类别分布不均衡性问题以及不同储层的价值排序有所差异,本文采用加权交叉熵损失函数,在模型训练中更注重学习少样本类别的特征,进一步提升含油储层的识别准确度.依据储层物性差异和相似度,设计了多层级储层流体识别方法,将LSTM和CNN的模型结构应用于层级II(含油储层、含水储层和干层)和层级III(油层、油水同层、差油层和水层、含油水层)的识别.本方法在真实油田测井数据上加以验证,其中数据分布极不均衡,高价值的含油类储层占比9％,符合真实工业场景.通过大量对比实验证明,LSTM和CNN的并行网络结构能够充分挖掘测井数据多维空间的相关特征;加权交叉熵损失函数显著提高了高价值含油储层的识别准确度,多层级储层流体识别方法对于物性相近的易混淆储层,如油水同层和含油水层的区分更加准确.实验结果表明,该方法能有效克服储层流体识别任务中的各种问题,对于帮助地质专家和工程师寻找地下储层,完成储层评价具有一定的实用价值.

地下储层地质建模对油气和水资源的开发以及CO2地质封存(CCS)具有重要意义.传统基于地质统计学的建模方法(如基于变差函数或多点统计学的方法)产生的储层地质模型可在一定程度上与地质模式保持一致,但当模式特征变得复杂时,则具有明显的缺陷.深度学习中的生成对抗网络(GANs)能够抽象和再现复杂的空间模式特征,而在许多领域得到了成功的应用.近年来,学者将生成对抗网络与地质建模相结合,利用由卷积神经网络(CNN)构成的生成器(模拟器)去学习复杂的地质模式特征,进而产生非常逼真的地质模型.基于生成对抗网络的地质建模方法在许多方面得到了深入研究,该方法甚至已被应用于实际油田复杂储层的三维地质建模,取得了很好的效果.本文综述了基于生成对抗网络地质建模方法的研究进展,主要包括无条件约束和有条件约束两大类.无条件约束即只需产生吻合期望模式特征的地质模型而无须与条件数据一致.根据生成对抗网络训练方式的不同,其可分为传统生成对抗网络地质建模方法和渐进增长的生成对抗网络地质建模方法.前者对模拟器和判别器中的所有卷积层同时进行训练,而后者则是从浅到深逐层进行训练.渐进增长的方法允许模拟器从大到小逐尺度学习地质模式特征,因而在地质模型效果和训练时间方面优于传统方法.为了构建同时吻合模式特征和给定条件数据的地质模型,有条件约束的生成对抗网络地质建模方法被提出并得到广泛研究.其中,基于模拟器输入向量搜寻的条件化方法主要应用梯度下降法或马尔可夫链蒙特卡罗法(MCMC)去搜寻适当的潜在向量,使其通过预训练模拟器生成与给定条件数据一致的地质模型.但是,当条件数据发生变化时,则需要重新搜寻另一组合适的潜在向量,需要大量的时间和计算资源.为此,模拟器直接条件化的地质建模方法(GANSim)被提出.该方法中,模拟器被训练以同时学习模式特征和各类条件数据与地质模型之间关系的知识,进而利用这两类知识将给定的任意条件数据直接转化为与模式特征和条件数据都吻合的地质模型.GANSim方法已被拓展到三维,形成GANSim-3D框架,并成功应用于塔河油田溶洞储层的三维地质建模.最后,笔者就GANSim框架、生成对抗网络地质建模方法所耗资源、该方法的工业落地和地质知识数字化研究提出了几点展望.