混合超记忆梯度法多尺度全波形反演

Multi-scale full waveform inversion with hybrid super memory gradient method

超记忆梯度类优化算法具有全局收敛性和超线性收敛速度,计算内存需求小,适合求解大规模无约束优化问题。将超记忆梯度类优化算法应用到全波形反演中,结合超记忆梯度类方法优点,提出混合超记忆梯度法全波形反演策略,并给出详细的实施流程。数值试算结果表明,混合超记忆梯度法优于共轭梯度法。含不同强度噪声的地震数据及不同精度初始模型的反演结果表明,混合超记忆梯度法反演精度较高。反演效率分析结果表明,混合超记忆梯度法反演耗时较短,证明了该混合策略在全波形反演应用中有一定的优势。

The super memory gradient method is an optimization algorithm which has global convergence and super-linear convergence rate,which uses less memory and is suitable for solving large-scale unconstrained optimization problems.The super memory gradient optimization algorithm is applied to full waveform inversion.We propose a hybrid super memory gradient method inversion strategy by combining the advantages of super memory gradient method.The detailed implementation processes are given in the flow diagram.The tests with synthetic data show the hybrid super memory gradient optimization algorithm is superior to conjugate gradient method.For seismic data containing different intensity noises or different precision initial models,the inversion results denote that the hybrid super memory gradient method can get high resolution inversion results.Inversion efficiency analysis shows that the hybrid super memory gradient method costs less computational time,and it demonstrates that hybrid strategy has some advantages in full waveform inversion.