【作者】 李阳；

【导师】 韩立国；

【作者基本信息】 吉林大学 ， 地球探测与信息技术， 2023， 博士

【摘要】 重力勘探作为一种被动源地球物理勘探的方法,相比其他地球物理方法具有更差的深度分辨率,在场源参数反演过程中存在反演精度低、深度分辨率低等问题。本文利用人工智能算法进行重力数据场源参数反演,通过建立基于深度学习的地质体水平边界位置、物性参数和界面起伏参数的系列反演算法,提高了重力数据反演结果的可靠性和精度。同时,以青海共和盆地作为方法应用示范区域,验证了深度学习技术在重力数据反演中的适用性;进而,利用重力数据对该区域的盖层、热储、热源和控热通道特征进行了定量分析,分析了共和盆地地区地热成因模式。这些结果为该地区的地热勘探和开发提供了理论支撑,有望对该地区的资源开发和利用产生积极影响。实现了一种基于深度学习的重力数据场源水平边界位置反演方法,以解决现有边界识别方法在深部弱异常识别能力弱,以及在正负叠加异常复杂条件下产生虚假边界的问题。在设计深度学习模型时,针对卷积神经网络结构特点,将重力异常转换量总水平导数(Thd)作为输入,提高深度学习网络对深部弱异常的识别能力。在数据预处理过程中,将重力异常转换量Thd转换为规则的矩阵数据输入到卷积神经网络中,然后通过构建好的卷积神经网络学习重力异常的特征值,完成训练后保存训练好的网络模型用于地质体边界位置的圈定。为了分析建立方法的有效性,将深度学习边界识别方法与传统的Thd、Theta图边界识别方法进行临近叠加模型、深浅叠加模型、含噪组合模型的方法应用测试,验证了方法的有效性。结果表明,与传统方法相比,基于深度学习的边界识别方法具有更高的准确性和稳定性,特别是在处理深部弱异常和正负叠加异常复杂条件下具有更好的表现。因此,该方法可以有效地提高重力数据场源水平边界位置反演的准确性和可靠性,为地质勘探和资源开发提供了新的思路和手段。本文引入表征重力异常拟合程度的约束项和深度加权函数到深度学习网络模型的建立中,构建了深度学习物性参数反演方法,解决了传统深度学习物性反演方法深度分辨率低、数据拟合差较大等问题,提高了反演的精度和深度分辨率。首先,通过对基于吉洪诺夫正则化的传统物性正反演理论算法进行分析,表明传统物性反演中的深度加权函数和数据拟合项的作用,为深度学习物性反演方法的建立和优缺点对比提供理论基础;其次,基于重力数据正演理论通过随机行走方法建立大规模数据集,为深度学习网络模型的训练提供更加完备的训练数据集;最后,在网络模型建立中,通过改进传统损失函数,将正演数据拟合差和深度加权函数约束到现有数据驱动的深度学习训练过程,提高了反演的精度和深度分辨率,并通过多种合成数据测试说明方法的应用效果。提出了一种基于深度学习U-net网络的重力数据密度界面反演方法。首先,对半椭球体界面模型进行随机抽取和组合进而形成地下起伏界面数据集,并基于Parker正演理论对界面数据集进行重力异常正演计算,为深度学习网络模型的训练提供特征完备的数据源;其次,设计了基于U-net网络模型的深度学习界面反演算法,在传统的损失函数基础上增加光滑损失项和过拟合抑制项,提高重力界面反演结果的光滑性和收敛效率;再次,通过测试样本集进行反演预测,验证建立深度学习网络模型的泛化性;最后,以Parker-Oldenburg界面正反演方法为对比方法,通过理论模型数据试验分析,验证了本文方法在密度界面反演中的有效性和实用性。青海省共和盆地的地热资源储量较为丰富,但我国目前对地热资源的开发和利用水平较低,对干热岩型地热资源的相关研究工作起步较晚且面临着诸多困难。目前,对于青海共和地区热储研究虽已有一定进展,但仍存在干热岩靶区或甜点位置不确定以及地热成因模式不清楚等问题。本文利用共和盆地的不同比例尺的重力数据,通过深度学习反演方法进行了地热资源的定量反演。具体来说,采用了深度学习重力数据参数反演方法,对青海共和盆地实测重力数据进行了断裂识别、三维密度空间结构分布反演、深部莫霍面起伏深度反演等分析,从而得到了与干热岩系统有关的盖层、储层、热源、控热通道等参数,并进一步建立了青海共和盆地地热成因模型。为该地区的干热岩资源的进一步勘探提供技术支撑。更多还原

【Abstract】 Gravity exploration is a passive source geophysical method with poorer depth resolution compared to other geophysical methods.In addition,source field parameter inversion often encounters problems such as low inversion accuracy and poor depth resolution.To address these issues,this study utilized artificial intelligence algorithms for gravity data source parameter inversion.By establishing a series of inversion algorithms based on deep learning of the horizontal boundary position,physical property parameters,and interface undulation parameters of geological bodies,the reliability and accuracy of gravity data inversion results were improved.Furthermore,the applicability of deep learning technology in gravity data inversion was verified by using Gonghe Basin in Qinghai province as a demonstration area for method application.Through quantitative analysis of the characteristics of the cover,thermal reservoir,heat source,and heat control channel in this area using gravity data,the geothermal genesis mode in the Republic Basin was also analyzed.These results provide theoretical support for geothermal exploration and development in this region,with the potential to have a positive impact on resource development and utilization in the area.A deep learning-based method was developed for the inversion of the horizontal boundary location of gravity data sources to address the weak recognition ability of existing boundary identification methods for deep weak anomalies and the problem of false boundaries under complex positive-negative superimposed anomalies.In designing the deep learning model,the total horizontal derivative of the gravity anomaly(Thd)was used as input to improve the deep learning network’s ability to recognize deep weak anomalies,considering the characteristics of the convolutional neural network structure.During the data preprocessing process,the Thd conversion of the gravity anomaly was transformed into regular matrix data and input into the convolutional neural network.The feature values of the gravity anomaly were learned through the constructed convolutional neural network,and the trained network model was saved for geological body boundary location delineation.To analyze the method’s effectiveness,the deep learning boundary recognition method was tested using adjacent stacking models,deep-shallow stacking models,and noisy combination models compared to traditional Thd and Theta graph boundary recognition methods,demonstrating the method’s effectiveness.The results show that the deep learning-based boundary recognition method has higher accuracy and stability than traditional methods,especially in dealing with deep weak anomalies and complex positive-negative superimposed anomalies.Therefore,this method can effectively improve the accuracy and reliability of gravity data source horizontal boundary location inversion,providing new ideas and means for geological exploration and resource development.A deep learning-based physical property inversion method was developed by introducing a constraint term that characterizes the fitting degree of gravity anomalies and a depth weighting function into the construction of the deep learning network model.This method solves the problems of low depth resolution and poor data fitting in traditional deep learning physical property inversion methods and improves the accuracy and depth resolution of inversion.Firstly,by analyzing the traditional physical property forward and inverse theory algorithms based on the Tikhonov regularization,the role of the depth weighting function and data fitting term in traditional physical property inversion is demonstrated,providing a theoretical basis for the establishment and comparative advantages and disadvantages of deep learning physical property inversion methods.Secondly,based on the gravity data forward theory and using the random walk method,a large-scale data set is established to provide more complete training data for the deep learning network model.Finally,in the construction of the network model,by improving the traditional loss function and constraining the fitting error of the forward data and the depth weighting function in the existing data-driven deep learning training process,the accuracy and depth resolution of the inversion are improved.The application effect of the method is demonstrated through various synthetic data tests.A gravity data density interface inversion method based on the deep learning U-net network is proposed.Firstly,a dataset of underground undulating interfaces is formed by randomly extracting and combining the ellipsoidal interface model,and the gravity anomaly forward calculation is performed on the interface dataset based on the Parker forward theory,providing a feature-complete data source for the deep learning network model training.Secondly,a deep learning interface inversion algorithm based on the Unet network model is designed,adding smoothness loss and overfitting suppression terms to the traditional loss function to improve the smoothness and convergence efficiency of the gravity interface inversion results.Thirdly,the generalization of the deep learning network model is validated by inverting prediction on the test sample set.Finally,by comparing with the Parker-Oldenburg interface forward and inversion method and analyzing the theoretical model data experimentally,the effectiveness and practicality of the proposed method in density interface inversion are verified.The Gonghe Basin in Qinghai Province has a relatively abundant reserve of geothermal resources.However,China’s development and utilization of these resources remain at a low level,with research on dry hot rock geothermal resources starting relatively late and facing many difficulties.Although some progress has been made in studying thermal storage in the Gonghe area,there are still issues with uncertainty in the location of dry hot rock target areas or sweet spots and unclear geothermal genesis models.This study used different scales of gravity data from the Gonghe Basin to quantitatively invert geothermal resources through a deep learning inversion method.Specifically,a deep learning gravity data parameter inversion method was used to analyze measured gravity data from the Gonghe Basin,including fault identification,three-dimensional density spatial distribution inversion,and deep Moho depth inversion,to obtain parameters related to the dry hot rock system,such as cap rock,reservoir,heat source,and heat control channel parameters.Additionally,a geothermal genesis model was established for the Gonghe Basin,providing technical support for further exploration of dry hot rock resources in the region.更多还原