基于卷积神经网络的重力异常反演

重力异常反演是地球勘探中常用的方法，它是通过地表观测重力异常推断地下介质的密度分布。针对传统反演方法存在的多解性、初始模型依赖和计算时间较长等问题，本文提出一种基于卷积神经网络(CNN)的重力异常反演方法，该方法首先通过大量正演计算获得训练数据集；然后采用该数据集训练CNN网络，使其建立从地表观测重力异常到地下密度模型之间的映射关系；最后将重力异常数据输入到训练好的卷积神经网络，得到对应的地下密度模型。实验结果表明，该方法能快速、准确的反演出地下重力异常体的密度、位置和形状，具有较强的泛化能力，能有效解决重力异常反演问题。

Inversion of Gravity Anomaly Based on Convolutional Neural Network

Gravity anomaly inversion is commonly used as a way to explore the earth. The density of underground materials can be inferred by observing gravity anomalies on the surface. The following problems are commonly addressed in the traditional gravity anomaly inversion methods: diversification, initial model dependence, and long calculation time. A gravity anomaly inversion method is proposed from this paper. This method is based on a convolutional neural network (CNN): The dataset was obtained through numerous forward calculations. Then the data set was applied to train the CNN network. Thereby, the mapping relationship would be established between the gravity anomaly observed on the surface and the density distribution of the underground medium. Finally, the gravity anomaly data was put into the trained convolutional neural network. Through this, the corresponding underground Medium density distribution was obtained. The following results can be observed in this experiment: the density, location, and boundary of underground gravity anomalies are quickly and accurately inversed by this method. A strong generalization ability and effective problem solving for gravity anomaly inversion are also indicated by this experiment.