Frequency Domain Multi-scale Full Waveform Inversion Based On Multi-GPU Computing

The exploration of conventional oil and gas reservoirs has been intensive.The exploration targets turn to complex oil and gas reservoirs in recent years.However,the seismic wave fields in the complex reservoir are very commingled with various interference.Therefore,the seismic records can not be governed by a simple law.This makes the inversion and imaging of complex media difficult with conventional inversion methods like stacking velocity analysis,migration velocity and tomography velocity analysis.FWI(Full Waveform Inversion)is a modeling technique with high precision.With the development of high performance computing,FWI becomes more and more popular in Geophysics in recent years.Although FWI has a complete mathematics of physics and modeling capability,it still faces many challenges in actual applications.In this paper,we discussed the multi-scale inversion strategy and high performance computing for FWI in detail.A high precision forward modling is a necessary condition for the success of waveform inversion.In this paper,we first discussed the finite difference method of acoustic wave equation in first order stress-velocity form to achieve a high-precision forward modling.Adjoint state method is an effective method for calculating the gradient of FWI with an adjoint operator for the wave equation.We transform the operator of one order stress-velocity equation to pseudo-conservative form.This makes the state equation and adjoint equation be in same form.Multi-scale inversion is a powerful strategy to weaken the nonlinearity of waveform inversion.Based on previous studies,this paper discussed the frequency selection strategy of multi-scale inversion based on the observation system.It is found the wavenumber illumination is inadequate at both ends of model so that the inversion results are not reliable.Lack of computing power is an important factor restricting the development of waveform inversion in the last century.Because of independent parallelization of multi-shot computation in waveform inversion,we designed a parallelization strategy for FWI based on a computing platform with multi-node and multi-GPU to improve the computational efficiency of inversion.