复杂地下介质下人工智能改进全波形反演：技术分类、性能对比、工程应用与发展趋势

Improvement of Full Waveform Inversion with Artificial Intelligence in Complex Subsurface Media: Technical Classification, Performance Comparison, Engineering Applications and Development Trends

摘要: 全波形反演(Full-Waveform Inversion, FWI)是重建地下高精度速度模型的关键技术，但传统模型驱动FWI在复杂介质中受限于周期跳变、计算成本高昂等固有缺陷。人工智能的引入为解决上述难题提供了全新求解框架。本文系统梳理了人工智能改进全波形反演的三大技术架构：纯数据驱动、两阶段模型–数据耦合及物理信息神经网络一体化方法。文章阐述了端到端直接反演、数据预处理辅助、预训练初值重构、生成对抗网络降阶正则及物理信息神经网络无网格反演等主流技术的原理与地质适用条件，并对比了其在复杂地层、盐体、断块等场景下的客观数值约束条件。综述覆盖了油气勘探、深部地质、地下碳封存监测及城市地震危险性定量评估四大应用领域，剖析了当前技术在泛化能力、CYCLE与工程化落地等方面的共性技术约束，并展望了无监督学习、多物理场耦合PINN及多模态联合反演等前沿发展方向。

Abstract: Full-waveform inversion (FWI) serves as a critical technique for acquiring high-precision subsurface velocity models. Nevertheless, conventional model-driven approaches suffer from inherent bottlenecks such as cycle jump and excessive computational cost when applied to complex media. The adoption of artificial intelligence offers an innovative paradigm to tackle the aforementioned challenges. This paper systematically sorts out three major technical frameworks for AI-enhanced full-waveform inversion, namely pure data-driven schemes, two-stage model-data integration strategies, and unified physics-informed neural network (PINN) approaches. It elaborates on the principles and applicable scenarios of mainstream technologies including end-to-end direct inversion, data preprocessing assistance, initial model reconstruction via pre-training, dimension reduction regularization using generative adversarial networks, and meshless inversion based on PINNs. Moreover, the performance and limitations of these technologies are compared across geological settings such as complex strata, salt bodies and fault blocks. This review encompasses four major application fields: oil and gas exploration, deep geology, underground carbon storage, and seismic hazard assessment. It analyzes core existing drawbacks concerning generalization capacity, depth of physical constraint integration, and 3D industrial deployment. Finally, cutting-edge research trends are prospected, including unsupervised learning, multi-physics coupled PINNs, and joint inversion with multi-modal data.