Study On Gravity Forward And Inversion Method Based On Deep Learning

The forward and inverse problem of gravity is one of the core contents of interpretation of gravity exploration data.Depth learning has developed rapidly in recent years,and has made great achievements in many disciplines.It has many applications in the field of geophysics,such as seismic exploration,interpretation and inversion.It is of pioneering significance to study the application of depth learning to gravity forward and inversion.As a machine learning method,deep learning needs a lot of data.It is not realistic to manually create tens of thousands of gravity models.It is observed that the similarity between the open data set MNIST data set and the gravity model.Therefore,in the deep learning forward calculation,MNIST data set is used as the gravity model,as the training data set of deep learning,the gravity and gravity gradient anomalies obtained from the gravity forward method based on the grid node density are used as labels,as output,and a full connection neural network is built on the Tensor Flow platform for training and learning.After training,the neural network model with adjusted parameters is obtained,and the test data is input into the trained neural network model.Compared with the gravity forward method based on the grid node density,the two results are almost the same.Similarly,the neural network model of gravity gradient anomaly can be trained by taking MNIST data set as input and gravity gradient anomaly as output.In addition,the neural network model of gravity and gravity gradient anomaly can also be obtained by taking MNIST data set as input and gravity anomaly and gravity gradient anomaly as output at the same time.Taking gravity anomaly as input and MNIST data set as output,building neural network and training,which is gravity inversion based on deep learning,the model obtained is not as accurate as forward modeling,but it is also relatively good.By taking gravity anomaly and gravity gradient anomaly as input at the same time,the effect of neural network constructed and trained is improved.By using MNIST data set as gravity model,the neural network model of gravity forward and reverse is trained.At the same time,the gravity gradient anomaly is added into the gravity forward and inverse,and the feasibility of gravity forward and inverse method based on depth learning can be found.