基于深度学习的地震岩相反演方法

复杂岩性预测是地震储层预测的难题,基于机器学习的非线性反演是识别岩性的有效手段。常规方法多以测井特征曲线(伽马曲线等)为学习目标,利用BP神经网络建立非线性映射预测岩性体,但这种方法存在两个问题,一是井震分辨率不匹配,二是BP神经网络在反演过程中存在局部收敛、效果不稳定以及非线性表征能力弱的问题。为解决这些问题,一是通过引入地震岩相概念解决井震分辨率不匹配问题,二是将深度学习引入到地震岩相反演中,经过优化样本采样、抽取相控伪井解决大样本集的构建问题,采用增量学习的策略进一步提高预测模型的精度和稳定性。以分频地震数据作为预测模型的输入,井岩相曲线为反演目标,实现了基于深度学习的地震岩相反演,有效解决了复杂岩性预测的难题。将该方法应用于海上某深水陆坡水道沉积研究区(该区发育灰岩、钙质砂岩、砂岩和泥岩4种岩相,岩石物理规律复杂,区分困难)岩性预测,结果表明,基于深度学习的地震岩相反演结果与井资料吻合,与地质认识相符。与叠前反演方法和BP神经网络学习岩相反演方法相比,基于深度学习的地震岩相反演方法准确度和分辨率更高,证明该方法是复杂岩性预测的有效手段。

 Complex lithology prediction is a difficult problem in seismic reservoir prediction.Nonlinear inversion based on machine learning is an effective method for identifying lithology.Conventional method is to establish nonlinear mapping to predict lithologic bodies with logging characteristic curves (such as gamma) as the learning objectives.But there are two limitations,one is the mismatch of well seismic resolution,the other is that the BP neural network has local convergence,unstable effect,and weak nonlinear characterization in the inversion process.To overcome these limitations,the concept of seismic lithofacies is introduced to solve the mismatch of well seismic resolution,deep learning is introduced into seismic lithofacies inversion to construct large sample sets by optimizing sampling and extracting facies-controlled pseudo wells,and also an incremental learning strategy is adopted to further improve the accuracy and stability of prediction model.Using frequency division seismic data as the input to the prediction model and the well lithofacies curve as the inversion target,seismic lithofacies inversion based on deep learning was realized,and found to be effectively solving the problem with complex lithology prediction.The study area is a deep-water continental slope waterway in the sea,where four lithofacies develop,including limestone,calcareous sandstone,sandstone,and mudstone;the lithofacies are complicated and difficult to distinguish.Yet,the seismic lithofacies inversion results based on deep learning were consistent with both the well data and the geological knowledge.The proposed method is superior to prestack inversion and BP neural network learning method for its higher accuracy and resolution,proving its effectiveness for complex lithology prediction.