人工智能驱动的陆架砂体地震沉积学表征——以珠江口盆地惠州凹陷新近系珠江组主力砂体为例

AI-driven seismic sedimentological characterization of shelf sand bodies： A case study on the Neogene Zhujiang Formation， Huizhou Sag， Pearl River Mouth Basin

葛家旺 陈聪 刘培 赵晓明 易震 甄艳 张安 唐小龙

Jiawang GE Cong CHEN Pei LIU Xiaoming ZHAO Zhen YI Yan ZHEN An ZHANG Xiaolong TANG

浅海陆架砂体具备优越的岩性圈闭条件，但因其储层厚度薄（小于1/8地震波长）、横向变化快且岩性复杂，导致砂体预测存在显著多解性。以珠江口盆地惠州凹陷新近系珠江组三段（SB—SB）主力产层ZJ3A和ZJ3B陆架砂体为研究对象，利用测井、岩心、薄片等资料，建立岩性定量解释标准，并优选出敏感地震属性。在此基础上，采用多个人工智能算法进行多属性拟合训练，建立人工智能算法驱动的地震沉积学表征方法体系。研究结果表明：① 随机森林算法地震多属性拟合的ZJ3A砂体预测效果最优，多发育南宽北窄的不对称状地貌单元，陆架砂体最长达17.15 km，平均面积为11.23 km；ZJ3B砂体采用增强回归树模型的预测效果最佳，多发育南北近似等宽的对称状地貌单元，且砂体规模小，平均面积6.21 km。② ZJ3A和ZJ3B两套砂体地震形貌分别指示了单向流动的沿岸流和双向流动的潮汐动力改造效应，综合反映了海平面升降期间沿岸流-潮汐水动力差异响应机制。③多种人工智能算法拟合及优选的地震沉积学研究提升了地震砂体预测准确度，定量化的地震地貌单元表征可揭示研究区沉积水动力学特征，可为后续岩性圈闭的目标优选提供良好的地质依据。

Neritic shelf sand bodies generally hold favorable geologic conditions for the formation of lithologic traps. However their small reservoir thickness （< 1/8 wavelength）， rapid lateral variation， and complex lithologies， lead to the sand body prediction results highly ambiguous. In this study， we investigate shelf sand bodies ZJ3A and ZJ3B， the major pay zones within the 3rd member of the Neogene Zhujiang Formation （also referred to as the Zhu 3 Member； confined by third-order sequence boundaries SB1 and SB2） in the Huizhou Sag， Pearl River Mouth Basin （PRMB）. Using well logs， as well as data from core and thin section observations， we establish quantitative lithological interpretation standards and select the optimal seismic attributes sensitive to sand body morphology. Accordingly， several artificial intelligence （AI） algorithms are employed for fitting and training based on multiple seismic attributes. As a result， an AI-driven seismic sedimentological characterization system is developed. The results indicate that the fitting of multiple seismic attributes using the random forest （RF） algorithm yielded the most accurate prediction for ZJ3A. The characterization results reveal that this sand body comprises predominantly asymmetric geomorphological units characterized by wide southern parts and narrow northern parts. These units exhibit a maximal length of up to 17.15 km and an average area of 11.23 km. In contrast， the boosted regression tree （BRT） model yielded the optimal prediction for ZJ3B. The characterization results show that this sand body consists primarily of symmetric geomorphological units characterized by roughly the same widths in the northern and southern parts. These units are small in scale， with an average area of 6.21 km. The seismic geomorphologies of ZJ3A and ZJ3B indicate the dynamic modification effects of unidirectional coastal currents and bidirectional tides， respectively. Therefore， their seismic geomorphologies jointly reflect the differential hydrodynamic response mechanisms of coastal currents and tidal dynamics during sea-level fluctuations. Seismic sedimentological research combined with fitting and selecting the optimal seismic attributes using multiple AI algorithms can enhance the seismic prediction accuracy of sand bodies. Furthermore， the quantitative characterization of seismic geomorphological units holds implications for the depositional hydrodynamic research of the study area. The results of this study will provide a robust geological basis for selecting the optimal lithologic trap targets subsequently.

10.11743/ogg20250311