Neural Network-based Receiver Function Calculation

The receiver function method is an important seismic tool for studying the crustmantle structure beneath a seismic station,and is also a prerequisite for many related studies in the field of seismology.In recent years,deep learning methods have been widely applied in the field of seismology,mainly in the areas of arrival time picking,earthquake phase identification,magnitude determination,seismic data denoising,and multiple wave suppression.However,there have been relatively few studies on the use of neural networks to fit receiver function calculations.We can consider using neural networks to fit the task of calculating the receiver function from two perspectives.On the one hand,the calculation process of the receiver function deconvolution is similar to the convolution process of convolutional neural networks.On the other hand,considering seismic data itself as a time series,we can also consider recurrent neural networks.To solve this problem,we first created a custom dataset containing one million data points.Secondly,we tried various types of networks on this dataset and compared their effectiveness.Finally,we also explored the transfer and application of the task in the audio field.Through continuous experimentation during the training process,we found that adding random data was crucial for the network to converge.Finally,we proposed a receiver function calculation method based on a convolutional neural network,which we called RFNet(receiver function network).This method involves inputting the Z and R components into a trained network to directly obtain the separated P-wave receiver function.We compared the network output with the results obtained from traditional receiver function calculations at seismic stations such as IC＿BJT and TA＿SUSD,and found that the convolutional network achieved a 91% fit to the receiver function.The Moho depths obtained from the network output were consistent with those obtained from traditional methods.The network also demonstrated good generalization capabilities for new data.This further demonstrates the ability of neural networks to extract underground structure features and fit complex geophysical calculation processes.Unlike most other networks in the field of seismology that can be classified as image processing or classification tasks,RFNet is one of the few networks that processes one-dimensional time series data.In addition,using RFNet can improve the speed by more than 10 times compared to traditional time domain iteration methods,while maintaining similar accuracy.