| Velocity plays an important role in geophysical exploration.Accurate velocity model is not only a prerequisite for migration imaging and other high-resolution exploration methods,but also an essential factor for time-depth inversion of seismic data.In the conventional method,the full-waveform inversion or other tomography techniques can be used to obtain velocity information,which are time consuming and computationally expensive,and they rely heavily on human interaction and quality control.To solve this problem,a velocity modeling method based on supervised machine learning is proposed to accurately transform seismic data from time domain to depth domain.The machine learning method is different from the traditional full-waveform inversion(FWI)method,which is based on a large number of training data to predict the results.In the training stage,the multi-shot data set obtained according to the velocity model is used as the input data.During the prediction stage,the trained network can be used to estimate the velocity models from the new input seismic data.A key feature of machine learning is that it can automatically extract multiple layers of useful features without human-curated activities and an initial velocity setup.The data-driven method usually requires more time during the training stage,and actual predictions take less time,with only seconds needed.The accuracy of the training model is the key to the estimation of the velocity model by using machine learning method.Therefore,it is necessary to obtain accurate input model,which requires accurate forward modeling to obtain multi-shot data sets.By testing the simulated velocity model,the computational time of geophysical inversions,including real-time inversions,can be dramatically reduced once a good generalized network is built,which has a good application prospect. |
