面向复杂地下介质：人工智能缓解全波形反演难题综述

A Review on Alleviating Full-Waveform Inversion Problems Using Artificial Intelligence for Complex Subsurface Media

摘要: 全波形反演(Full-Waveform Inversion, FWI)能够利用地震记录中的全波形信息实现高精度成像，是当今复杂地下介质速度建模与储层刻画中最受关注的地球物理技术之一，已广泛应用于油气勘探、深部地质探测、地下储碳及地震灾害评估等领域。然而FWI本质上是一个强非线性、非凸且病态的逆问题。当面对复杂地层、强非均质介质、速度剧烈变化以及观测数据稀疏含噪等情况时，周期跳跃、局部最优、收敛缓慢、成像畸变等问题往往十分突出，给实际工程应用带来了很大困难。近年来，人工智能技术凭借其在非线性拟合、特征提取、自适应优化及数据降噪等方面的能力，为改善复杂介质FWI的效果开辟了新途径。本文围绕复杂地下介质条件下传统FWI面临的主要困难，对现有的人工智能改良方法——包括物理驱动、数据驱动与混合驱动三大类——的技术原理、适用场景和各自优劣进行了分类梳理，同时归纳了当前研究的不足之处，并对未来发展方向做了展望，以期为复杂地质条件下波形反演成像的进一步发展提供参考。

Abstract: Full-Waveform Inversion (FWI) achieves high-precision subsurface imaging by utilizing full waveform information contained in seismic records. As one of the most popular geophysical technologies for velocity modeling and reservoir characterization of complex subsurface media, FWI has been widely applied in hydrocarbon exploration, deep geological survey, underground carbon storage, seismic hazard assessment and other fields. Nevertheless, FWI is inherently a strongly nonlinear, non-convex and ill-posed inverse problem. When dealing with complex strata, strongly heterogeneous media, drastic velocity variations, as well as sparse and noisy observed data, severe issues such as cycle skipping, local optima, slow convergence and imaging distortion frequently emerge, posing substantial obstacles to practical engineering applications. In recent years, artificial intelligence (AI) technologies, with outstanding capabilities in nonlinear fitting, feature extraction, adaptive optimization and data denoising, have opened up new avenues for improving the performance of FWI for complex media. Focusing on the core challenges encountered by conventional FWI under complex subsurface geological conditions, this paper systematically categorizes and reviews existing AI-enhanced FWI methods into three major branches: physics-driven, data-driven and hybrid-driven frameworks. It elaborates on their technical mechanisms, applicable scenarios, respective advantages and limitations, summarizes the deficiencies of current research, and prospects future development trends, aiming to provide references for the further advancement of waveform inversion imaging under complicated geological conditions.