基于混合注意力机制的矿田尺度三维岩性建模:以焦家金矿田为例

李风 ,  王功文 ,  卢紫阳 ,  付超 ,  刘烊 ,  东玉龙 ,  龚天一 ,  张智强

地球科学 ›› 2026, Vol. 51 ›› Issue (03) : 955 -969.

PDF (3807KB)
地球科学 ›› 2026, Vol. 51 ›› Issue (03) : 955 -969. DOI: 10.3799/dqkx.2025.286

基于混合注意力机制的矿田尺度三维岩性建模:以焦家金矿田为例

作者信息 +

Three⁃Dimensional (3D) Lithological Modeling via Hybrid Attentional Mechanism Deep Learning Model: A Case Study of Jiaojia Gold Field

Author information +
文章历史 +
PDF (3897K)

摘要

矿田尺度深部地质结构“透明化”是矿产勘查与成矿预测的核心,三维岩性建模是实现这一目标的关键技术.然而,当前矿田尺度三维岩性建模主要依赖效率较低的显式建模方法,难以满足多阶段矿产勘查和矿山实时生产的需求,因此亟须研发高精度、高效率的三维岩性隐式建模方法.针对上述问题,本研究以三维卷积神经网络(3D Convolutional Neural Network, 3D CNN)为基础,融合卷积注意力机制(Convolutional Block Attention Module, CBAM)与自注意力机制(Self⁃Attention Module, SAM)构建混合注意力机制深度学习算法(Hybrid Attentional Mechanism deep learning model, HAM),并基于该算法挖掘多源地质‒地球物理数据中的深层次特征,确定建模所需地质体边界,实现既能捕捉局部上下文、又能表征全局上下文的三维岩性隐式建模方法.为验证HAM算法有效性,本研究选择胶东半岛焦家金矿田作为研究区,开展对比实验与消融实验.结果表明,相较于随机森林(Random Forest, RF)和3D CNN等基线算法,本次研究提出的HAM算法在三维岩性建模的准确率、宏平均精确率、召回率、宏平均F1分数和混淆矩阵上表现出显著优势,对推动深部找矿和矿山生产具有重要意义.

Abstract

Enhancing the transparency of deep geological structures at the ore-field scale is critical for subsurface mineral exploration and prospectivity modeling, and three-dimensional (3D) lithological modeling serves as a critical technology for this objective. However, existing ore-field-scale modeling workflows rely on explicit modeling approaches with relatively low efficiency, which can hardly meet the demands of multi-stage mineral exploration and real-time mining. Consequently, high-precision and high-efficiency implicit 3D lithological modeling methods are urgently needed. To address this issue, a Hybrid Attentional Mechanism deep learning model (HAM) is constructed on the basis of the 3D Convolutional Neural Network (3D CNN), integrating the Convolutional Block Attention Module (CBAM) and the Self-Attention Module (SAM). Based on this algorithm, deep representations within multi-source geological and geophysical data are mined to determine the boundaries of geological bodies required for modeling, thereby achieving a 3D lithological implicit modeling method capable of capturing both local details and long-range dependencies. To validate the effectiveness of the proposed hybrid attentional mechanism model, the Jiaojia gold field in the Jiaodong Peninsula was selected as the study area, and comparative and ablation experiments were conducted. Relative to baseline models‒Random Forest (RF) and a vanilla 3D-CNN, HAM markedly improves the macro-averaged accuracy, precision, recall, macro-averaged F1 score and confusion matrix of ore-field-scale implicit 3D lithological modeling, with direct implications for subsurface mineral exploration and mining operations.

Graphical abstract

关键词

三维岩性隐式建模 / 三维卷积神经网络 / 卷积注意力机制 / 自注意力机制 / 焦家金矿田 / 大数据 / 矿床地质.

Key words

3D lithological implicit modeling / 3D convolutional neural network / convolutional block attention module / self-attention module / Jiaojia gold field / big data / ore deposits

引用本文

引用格式 ▾
李风,王功文,卢紫阳,付超,刘烊,东玉龙,龚天一,张智强. 基于混合注意力机制的矿田尺度三维岩性建模:以焦家金矿田为例[J]. 地球科学, 2026, 51(03): 955-969 DOI:10.3799/dqkx.2025.286

登录浏览全文

4963

注册一个新账户 忘记密码

0 引言

三维岩性建模旨在融合多源地学数据基础上,结合空间插值方法,实现深部地质体岩性类型与空间展布的定量识别与可视化.高质量的三维岩性模型对降低勘探风险、优化工程布设和实现深部找矿突破具有重要意义(邓浩等, 2020;王昊等, 2023;陈建平等, 2024;Zhang et al., 2024;成秋明, 2025;毛先成等, 2026;张明明等, 2026).目前矿田尺度三维岩性建模方法主要分为显式建模与隐式建模两类(郭甲腾等, 2019;陈麒玉等, 2025).其中显式建模通过手工圈连多源地学数据中的地质体边界实现三维建模(Min et al., 2019Guo et al., 2021).该方法支持在建模过程任意环节进行人工干预,因此适应性强,方便融入地质专家对研究区已有的地质认知,但建模效率低,更新成本高,难以满足多阶段矿产勘查和矿山实时生产的需求(Guo et al., 2016;陈进等, 2020;Zhang et al., 2024;陈麒玉等, 2025;叶舒婉等, 2025).隐式建模则通过构建多源地学数据中地质体边界之间的隐函数关系实现三维建模,该方法自动化程度高,支持模型的快速动态更新(娄渝明, 2025;王统荣等, 2025),但如果地质体边界数据稀疏或存在误差,则容易导致建模结果不准确(Jia et al., 2021;刘志斌等, 2024;陈麒玉等, 2025;肖凡等, 2025).在矿田尺度三维岩性建模中,作为关键输入的地质体边界,按其数据源与可靠性可分为两类.第一类边界来源于露头、实测剖面、探矿工程及采矿工程等直接观测手段.此类数据可靠度高,能够精确刻画已知地质体的岩石类型和空间形态,但这类数据通常集中于矿床内部及周边区域,覆盖范围和深度有限,难以满足矿田尺度三维岩性建模需求.第二类边界则通过多源地质‒地球物理解译获得.这类数据覆盖范围广、深度大,并能够填补第一类边界缺失的空白区,但其解译过程复杂,且往往带有较强的主观性和不确定性(Guo et al., 2022;袁峰等, 2024).因此如何智能化圈定多源地质‒地球物理数据中的地质体边界对提高矿田尺度三维岩性隐式建模效率和精度具有重要意义(陈建平等, 2024;叶舒婉等, 2025).针对以上问题,郭甲腾等(2019)、Jia et al.(2021)Zhang et al.(2023a)相继将支持向量机(Support Vector Machine, SVM)、随机森林(Random Forest, RF)、极限提升树(Extreme Gradient Boosting, XGBoost)、堆叠模型(Stacking model)和贝叶斯XGBoost等浅层机器学习算法应用于三维岩性隐式建模领域,并取得较好效果.然而,上述浅层机器学习算法在挖掘多源地质‒地球物理数据方面存在深层次特征提取不足和缺乏空间上下文信息的问题(邓浩等, 2021;左仁广等, 2021;师路易和左仁广, 2026).为此,Zhang et al.(2024)将3D CNN引入三维岩性隐式建模领域.值得指出的是,地质体除具有复杂三维几何形态特征外,还在一定尺度上呈现空间连续性,但这种连续性并非一成不变,例如早期花岗岩内部可能存在后期脉岩侵入,从而出现局部不连续现象(赵惊涛等, 2016;蔡志慧等, 2020).这种“整体连续‒局部突变”的复杂空间上下文关系与多源地球物理数据对不同岩性的敏感性差异,导致传统3D CNN应用于三维岩性隐式建模领域时面临如下挑战:(1)传统3D CNN难以挖掘地质体复杂空间上下文关系;(2)传统3D CNN在多源地球物理深层次特征融合中对不同特征通道与空间位置近似等权,难以突出对岩性分类更具判别指示意义的物性和准确刻画地质体复杂的空间变化.

针对上述问题,本次研究以3D CNN为基础,通过融合CBAM与SAM构建HAM算法.上述算法可以实现基于多源地质‒地球物理数据的深部地质体边界智能化圈定,并与现有主流三维地质建模商业软件(如Leapfrog、SKUA⁃GoCAD、Geomodeller、3D Mine等)相结合,实现高效精准的三维岩性建模.笔者选择胶东半岛焦家金矿田开展算法测试.结果表明,本次研究所提出的HAM算法可以有效提升三维岩性建模精度,对推动深部矿产资源勘查和服务矿山实时生产具有重要意义.

1 研究区地质背景与数据集

1.1 地质背景

焦家金矿田位于华北克拉通胶东半岛西北部(图1a和1b),焦家断裂带中段,拥有焦家、新城、寺庄超大型金矿床和上庄、河东、河西、马塘等大中小型金矿床,探明储量超1 500余吨,是我国最重要的金矿田之一(图1c)(Zhang et al., 2024).研究区主构造是北东走向、北西倾向的焦家断裂,同时存在望儿山、河西和侯家等焦家断裂派生的次级断裂(图1c).上述主次级断裂共同控制着成矿热液的运移与就位(朱日祥等, 2015;周明岭等, 2024).区内出露地层和岩石主要为第四系覆盖层、早白垩世郭家岭花岗闪长岩(锆石LA⁃ICP⁃MS U⁃Pb年龄为166~146 Ma)(姚晓峰等, 2020;李世勇等, 2022)、晚侏罗世玲珑花岗岩(锆石LA⁃ICP⁃MS U⁃Pb年龄为135~123 Ma)(Zhang et al., 2022;于晓卫等, 2023)和太古代胶东群(锆石LA⁃ICP⁃MS U⁃Pb年龄为2 860~2 820 Ma)(祝培刚等, 2022).其中太古代胶东群主要由角闪岩和片麻岩等组成,并在焦家断裂上盘的红布‒磁口‒姚家一带出露(图1c).研究区早白垩世郭家岭花岗闪长岩与金成矿(133~120 Ma)在时间上联系最为紧密,并且常以岩株方式侵入胶东群和玲珑花岗岩(Deng et al., 2020Li et al., 2023).空间上,金矿虽然与早白垩世郭家岭花岗闪长岩紧邻(杨进辉等, 2021).但其包含的金矿数量和资源量不及晚侏罗世玲珑花岗岩.此外,研究区第四系覆盖范围较大,因此亟须开展三维岩性建模工作,支撑深部找矿(图1c).

1.2 数据集

本次研究使用的多源地质地球物理数据主要包括:(1)地质数据:①矿田尺度1∶50 000的地质图,117个钻孔,其中约72%(84个)钻孔集中分布于焦家断裂带及其次级构造带2 km范围内,主要覆盖了焦家(38个)、新城(29个)、寺庄(17个)等超大型金矿床.其余钻孔分布于河东、马塘等中小型矿床,上述矿床外围区域钻孔相对稀疏(图2);(2)覆盖全区的地球物理数据:①矿田尺度1∶25 000重力数据(图3a);②矿田尺度 1∶50 000磁法数据(图3b);③矿田尺度1∶25 000大地电磁(Magnetotelluric Sounding,MT)数据(图3);④1 076块岩石物性测量数据(其中431块样本采集自41个钻孔)(图4).上述数据中的地质数据由山东黄金集团有限公司提供,地球物理数据由中国地质调查局天津地质调查中心提供,钻孔岩石物性测量在中国地质大学(北京)完成.

为量化研究区广泛存在的蚀变作用对岩石物性造成的影响,本次分别对蚀变岩和原岩进行了系统采样和物性测量.研究区主要岩石的物性箱型图如图4所示,结合研究区以往区域金矿普查的岩石物性统计数据,本次研究对研究区主要岩石物性总结如下:

(1)新太古代胶东群主要由英云闪长岩、奥长花岗岩、花岗闪长岩(Trondhjemite⁃Tonalite⁃ Granodiorite,TTG)和斜长角闪岩构成,TTG岩石物性与斜长角闪岩相近(图4)(杨进辉等, 2021).因此,本研究将二者归为一个建模单元,即新太古代胶东群(图4a).该套地层密度(2.69~3.12 g/cm3)明显高于晚侏罗世玲珑花岗岩(2.42~2.71 g/cm3)和早白垩世郭家岭花岗闪长岩密度(2.42~2.72 g/cm3)(图4a).

(2)新太古代胶东群(0.000 2~0.008 0 SI)和早白垩世郭家岭花岗闪长岩(0.000 1~0.005 0 SI)的磁化率高于玲珑花岗岩(0.000 02~0.000 9 SI),前两者与玲珑花岗岩之间存在约0.000 1~

0.001 0 SI的磁化率差(图4b).新太古代胶东群的电阻率(800~3900 Ω·m)略低于晚侏罗世玲珑花岗岩(400~7 800 Ω·m)和早白垩世郭家岭花岗闪长岩(400~5 100 Ω·m)(图4c).

(3)物性测量结果表明,蚀变虽然降低了原岩磁化率(降低约0.001~0.010 SI)和电阻率(降低约600~2 000 Ω·m)(图4b和4c).但蚀变新太古代胶东群(0.000 1~0.008 0 SI)和早白垩世郭家岭花岗闪长岩的磁化率(0.000 1~0.006 0 SI)仍高于蚀变晚侏罗世玲珑花岗岩磁化率(0.000 01~0.000 90 SI)(邹艳红等, 2023;宋志勇等, 2024).

(4)蚀变岩与原岩之间以及不同岩性的蚀变岩之间的物性值存在较大范围的重叠(图4).鉴于当前比例尺重磁电数据难以在矿田尺度三维岩性建模中区分蚀变岩与原岩之间的密度、磁化率和电阻率差异,因此本次研究暂不将各类蚀变岩作为独立建模单元.综上所述,本次三维岩性建模对象为新太古代胶东群、晚侏罗世玲珑花岗岩和早白垩世郭家岭花岗闪长岩.

本研究采用非线性共轭梯度法(Nonlinear Conjugate Gradient, NLCG)对MT数据进行二维反演,并进一步利用离散光滑插值算法(Discrete Smooth Interpolation, DSI)将反演结果插值为三维电阻率模型.针对重磁数据,本次研究选择UBC⁃GIF软件实施三维反演.反演网格核心区设置为50 m×50 m×25 m,边界区域则采用逐步扩大的网格(最大外延至100 m×100 m×50 m),以有效抑制边界效应.由于对重磁数据误差信息掌握有限,因此本次重磁反演采用广义交叉验证(Generalized Cross⁃Validation, GCV)模式,并引入深度加权函数,以避免反演过程中密度与磁化率异常体过度集中于浅表(Zhu et al., 2025).最终获得的重磁三维反演结果如图5a和5b所示.三维反演获得的密度、磁化率和电阻率模型,将作为岩性识别的输入特征(图5),标签则通过对钻孔沿钻探深度延伸方向每隔25 m采样获得,上述采样间隔与地球物理反演网格尺寸相一致,上述地球物理特征和岩性标签将共同用于矿田尺度三维岩性隐式建模.

2 研究方法与原理

本研究构建了一种HAM算法,用于矿田尺度三维岩性隐式建模.以下从数据预处理、算法架构和损失函数及训练过程几个方面介绍研究方法与原理.

2.1 数据预处理

本次研究通过滑动距离为1的三维滑动窗口提取13×13×13的三维体元.每个体元包含密度、磁化率和电阻率三种物性特征.识别上述三维体元中心体素的岩性类别是本次研究的核心任务.训练集和测试集体元中心体素岩性类别根据钻孔和地质露头确定,训练集与测试集比例为8∶2.钻孔和地质露头未覆盖区域则作为本次研究的预测区.为防止深度学习算法过拟合,本次研究通过三维随机裁剪方法对训练数据进行数据增强(Zhong et al., 2020Mumuni and Mumuni, 2022).三维随机裁剪方法可以以一定概率随机将三维体元除中心体素外的部分区域的特征清除,模拟实际地球物理数据中可能存在的局部信息缺失,从而防止深度学习模型过度依赖局部区域或某类特征,增强深度学习模型的特征提取能力和鲁棒性(Mumuni and Mumuni, 2022).

2.2 HAM算法架构

3D CNN虽能有效提取局部深层次特征,但存在固有的局部感受野和深层次特征融合中对不同物性通道与空间位置近似等权的缺点,导致其难以满足复杂三维岩性建模需求(Yu and Ma, 2021Liu et al., 2024).因此本研究以3D CNN为基础,依次嵌入CBAM(Woo et al., 2018)和SAM(Wang et al., 2018),从而充分发挥多源地球物理数据在不同类型岩性三维建模中的互补性,并增强三维岩性建模过程中局部与全局空间上下文关系的挖掘能力.HAM算法架构如图6所示.

2.2.1 浅部特征提取网络

浅部特征提取部分由通道数逐步提升的五层三维卷积层和最大池化层组成(图6).每层三维卷积层后均引入批归一化与Leaky ReLU激活函数.其中批归一化通过规范化特征分布缓解训练过程中的内部协变量偏移,加快收敛速度(Ioffe and Szegedy, 2015).Leaky ReLU则通过允许小幅负值输入,避免神经元“死亡”,Leaky ReLU的关键超参数为σ,根据Huang et al.(2021)的研究,本次研究将其设置为0.01.尺寸2×2×2、步长为2的最大池化层位于第2、第4及第5三维卷积层之后,用于保留关键特征的同时控制计算复杂度(Zitouni et al., 2025).

2.2.2 卷积注意力机制

在三维岩性建模中,不同物性特征和不同空间位置对不同类型岩石的识别贡献存在显著差异(徐璐平等, 2022).为防止深层次特征融合过程中对不同物性特征与空间位置近似等权(Hu et al., 2018),本研究在浅部特征提取网络之后嵌入CBAM(Woo et al., 2018).CBAM由通道注意力与空间注意力两个子模块顺序串联组成(Woo et al., 2018):

          F''=MsMcFFF',

其中,FRC×D×H×WC代表通道数,DHW代表空间维度)代表经过浅部特征提取网络输出特征,为逐元素乘法,Mc()为通道注意力机制,Ms()为空间注意力机制,F'=Mc(F)F为经过通道注意力机制提取的中间特征,F''表示经过CBAM提取的特征.通道注意力机制(McF)可以针对不同岩性分类任务,动态调整不同物性深层次特征的权重(公式(2)):

McF=σMLPAvgPoolF+                                    MLPMaxPoolF,

其中,σSigmoid函数,AvgPoolMaxPool分别表示全局平均池化操作和最大池化操作.空间注意力机制(MsF')则通过加强局部关键区域的权重分配,以增强模型对局部地质体空间上下文信息的识别:

MsF'=σconv3×3×3AvgPoolF'; MaxPool F',

其中,conv  3×3×3; 分别代表3×3×3三维卷积和通道拼接.

2.2.3 自注意力机制

3D CNN和CBAM受局部感受野的限制,难以有效刻画三维岩性建模过程中的全局空间上下文关系(Huang et al., 2023bZhao et al., 2024).针对这一问题,本研究在CBAM后嵌入SAM(图6). SAM通过计算13×13×13三维体元深层特征中任意两个体素ij之间的全局相关性Ai,  j,增强全局上下文建模能力(Woo et al., 2018et al., 2025).设F''RN×D×H×W为经过CBAM后提取的特征(N为通道数),SAM先通过三个1×1×1三维卷积convθ1×1×1convφ1×1×1convg1×1×1进行线性嵌入,得到13×13×13三维体元中每个体素的查询向量(Q)、键向量(K)与值向量(V):

          Q=convθ1×1×1F''K=convφ1×1×1F''V=convg1×1×1F'',

然后计算13×13×13三维体元中i体素和j体素的QK相似度Si,  j

          Si,  j=QiTKj,

并以此为基础,计算13×13×13三维体元中各个体素的注意力权重Ai,  j

         Ai,  j=exp Si,  jk=1Nexp Si,  j, i, i,j1,2,, 13×13×13,N=13×13×13,

得到Ai,  j后,SAM计算每个体素经过全局注意力权重加权后的特征Z公式(7)),为进一步保留原始特征信息,SAM还引入残差机制,因此最终经过SAM提取的特征F'''公式(7)所示:

          Z=j=1NAi,  jVjF'''=WzZ+F'',

其中,Wz是一个可学习的权重矩阵.

2.2.4 焦点损失函数

本次研究训练集中标签为胶东群的三维体元数据远少于标签为玲珑花岗岩的三维体元数据,其不平衡比约为1∶40,存在明显的类别不平衡问题(Buda et al., 2018).交叉熵损失会使模型偏向多数类(玲珑花岗岩),导致对少数类(胶东群)的分类性能下降.因此本次研究引入了焦点损失函数(Lin et al., 2017Zhang et al., 2023b):

          FL(pt)=-αt(1-pt)γlgpt,

其中,pt为模型对真实类别的预测概率,αt为平衡因子,每个类别αt值设定为其样本数量倒数的归一化值,确定为[17.53939 3, 0.429717.539 3, 0.429 7, 1.6244],分别对应胶东群、玲珑花岗岩、郭家岭花岗闪长岩,这一过程赋予样本量少的岩性类别更高的权重(Lin et al., 2017;Zhang et al., 2023b).聚焦参数γ用于调节困难样本的关注强度,是焦点损失函数的关键超参数,本次研究通过网格搜索确定焦点损失函数的最佳γ为2.0.

3 结果与讨论

3.1 模型训练

本次研究以三维岩性建模领域广泛应用的RF和3D CNN算法为基线算法.其中基线算法的3D CNN模型通过HAM模型移除CBAM和SAM模块获得.此外,为评价模型中关键组 件——CBAM与SAM的有效性,本研究开展包括移除SAM的3D CNN⁃CBAM、移除CBAM的3D CNN⁃SAM以及完整HAM在内的消融实验.

上述模型中,RF关键超参数通过网格搜索确定,其决策树数量、决策树最大深度、节点可分的最小样本数分别为150、15和12.在平衡模型效果和计算开销的前提下,经网格搜索反复测试,本次研究3D CNN⁃CBAM和HAM模型中CBAM模块的通道注意力机制降维比率为8,空间注意力机制卷积核尺寸为3×3×3(Ridnik et al., 2021).3D CNN⁃SAM和HAM模型中SAM模块的自注意力机制头数为8,每头维度为32,总隐藏维度256 (Woo et al., 2018Chen et al., 2022Zhang et al., 2025).本次深度学习模型训练使用的优化器为AdamW,初始学习率为1×10-4,权重衰减系数为1×10-5,训练批大小为256.在设定初始学习率基础上,本次研究进一步采用自适应调整策略对学习率进行动态调整,从而促进模型收敛至更优性能.此外,为有效防止过拟合并节省计算资源,本次深度学习模型训练过程引入了早停机制,其中早停机制的最大训练轮次为500,当验证集损失连续10个轮次下降幅度小于阈值1×10-3时,将提前终止模型训练.

3.2 模型对比

本次研究的评价指标包括准确率、宏平均精确率、召回率、宏平均F1分数和混淆矩阵.对比实验与消融实验结果如表1所示.深度学习模型在测试集上的宏平均准确率、精确率、召回率和宏平均F1分数均优于RF模型(表1).相较于3D CNN,3D CNN⁃CBAM在测试集上的准确率、宏平均精确率、召回率和宏平均F1分数分别提升0.76%、2.72%、2.81%和2.77%,3D CNN⁃SAM在测试集上的准确率、宏平均精确率、召回率和宏平均F1分数分别提升0.84%、2.21%、2.54%和2.31%(表1).添加CBAM和SAM的HAM模型在测试集上的准确率、宏平均精确率、召回率和宏平均F1分数分别提升1.25%、5.44%、2.97%和4.40%(表1).训练过程(图7)表明,HAM模型收敛速度最快,且收敛后的训练与验证损失最为稳定,反映出高效的学习效率与稳健的泛化能力.混淆矩阵(图8)进一步印证,HAM在岩性类别判别中显著优于RF与基础3D CNN模型.

3.3 讨论

在上述岩性分类基础上,本次研究通过DSI算法完成最终三维岩性隐式建模(图9).相较于RF算法,基于深度学习算法识别的各类岩体边界不仅更为精确,而且更为连续(表1图9),体现了深度学习在深层次提取与融合复杂空间上下文特征方面的优势.相较于3D CNN,添加CBAM或SAM模块可以提升基础3D CNN的性能,但仅添加CBAM或SAM模块模型的宏平均F1分数分别比包含CBAM和SAM的HAM低1.63%和2.09%(表1).通过融合CBAM和SAM,HAM在焦家金矿田的岩性识别中,既能聚焦于关键判别部位,又能构建全局上下文关系,从而提升地质体三维空间形态和分布的合理性,例如在识别研究区内出露的关键地质体郭家岭花岗闪长岩在焦家断裂中段的出露时,HAM模型的预测边界与地质图露头边界达到了高度吻合(图1c和图9e),实现了对该岩体空间形态的高精度三维刻画.作为对比,其他模型的预测边界在该区域则与实地露头存在明显偏差(图1c和图9),这进一步印证了HAM模型在三维岩性隐式建模过程中的优越性.当然,本次研究的HAM算法除与SKUA⁃GoCAD软件的DSI算法结合外,还可以与其他各类主流三维地质建模商业软件(如Leapfrog、Geomodeller、3D Mine等)的三维插值模块相结合,以实现高效精准的三维岩性建模.值得指出的是,HAM模型在处理空间分布连续、规模较大的地质体时具有显著优势,但在处理如断裂带等薄层地质体时则面临固有挑战,其原因在于:(1)薄层地质体在三维空间中厚度与横向延伸尺度相差悬殊,标准卷积核受限于规则网格上的局部感受野,难以有效适配此类非规则、各向异性显著的薄层结构(黄珏璇, 2022);(2)薄层地质体的地球物理响应信号通常较弱,且易与围岩信号混杂,使得深度学习模型从多源地质‒地球物理数据中提取的深层次特征更倾向于大规模地质体,对这类薄层地质体弱异常敏感性不足(Zhang et al., 2024).针对上述问题,可以通过多尺度特征融合方法挖掘从局部薄层地质体到全局大规模地质体的多尺度特征,或者结合图结构方法对节点和边的三维拓扑关系进行建模,从而不受固定规则卷积核形状限制,自然适配任意形态的薄层地质体(Huang et al., 2023a).

4 结论与展望

本研究报道了一种矿田尺度三维岩性隐式建模深度学习方法——HAM.在焦家金矿田开展的三维岩性建模实验表明,相较于传统RF算法和3D CNN算法,HAM算法不仅可以挖掘建模所需的深层次地球物理特征,还可以更为有效地构建由“整体连续‒局部突变”引起的复杂局部‒全局空间上下文关系,从而明显提升三维岩性建模的精度和效率.然而应当注意的是,本次研究所采用的深度学习算法仍主要依赖于数据驱动,尚未充分融合地层产状、构造格架等先验地质知识.因此,亟须研发融合地质知识与观测数据的“知识‒数据”双驱动深度学习算法.此外,HAM模型在刻画断裂等薄层地质体时存在局限,因此如何将多尺度特征融合或将图结构建模等方法引入三维岩性建模中,以突破薄层地质体的建模瓶颈,实现全要素高精度三维隐式建模,亦是后续需要重点探索的方向.另一方面,本研究尚未与图神经网络(Graph Neural Network, GNN)等先进方法展开系统对比,这也是未来研究中需要加强的部分.

参考文献

[1]

Buda, M., Maki, A., Mazurowski, M. A., 2018. A Systematic Study of the Class Imbalance Problem in Convolutional Neural Networks. Neural Networks, 106: 249-259. https://doi.org/10.1016/j.neunet.2018.07.011

[2]

Cai, Z. H., He, B. Z., Liu, R. H., 2020. Emplacement of Granitic Pluton and Cenozoic Deformation in the Wenquan Region, Tashkorgan, Xinjiang: The Implications for the Miocene Tectonic Evolution of the Northeast Pamir. Acta Petrologica Sinica, 36(10): 3137-3151 (in Chinese with English abstract).

[3]

Chen, J. N., Sun, S., He, J., et al., 2022. Transmix: Attend to Mix for Vision Transformers. The IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans. https://doi.org/10.48550/arXiv.2111.09833

[4]

Chen, J. P., Zhou, G. Y., Chu, Z. Y., et al., 2024. Geological Big Data Three⁃Dimensional Modelling and Mineralization Prediction of Diamond Deposit in Mengyin, Shandong, China. Mineral Deposits, 43(4): 802-820 (in Chinese with English abstract).

[5]

Chen, J., Mao, X. C., Deng, H., 2020. 3D Quantitative Mineral Prediction in the Depth of the Dayingezhuang Gold Deposit, Shandong Province. Acta Geoscientica Sinica, 41(2): 179-191 (in Chinese with English abstract).

[6]

Chen, Q. Y., Xun, L., Cui, Z. S., et al., 2025. Recent Progress and Development Trends of Three⁃ Dimensional Geological Modeling. Bulletin of Geological Science and Technology, 44(3): 373-387 (in Chinese with English abstract).

[7]

Cheng, Q. M., 2025. A New Paradigm for Mineral Resource Prediction Based on Human Intelligence⁃Artificial Intelligence Integration. Earth Science Frontiers, 32(4): 1-19 (in Chinese with English abstract).

[8]

Deng, H., Wei, Y. F., Chen, J., et al., 2021. Three⁃ Dimensional Prospectivity Mapping and Quantitative Analysis of Structural Ore⁃Controlling Factors in Jiaojia Au Ore⁃Belt with Attention Convolutional Neural Networks. Journal of Central South University (Science and Technology), 52(9): 3003-3014 (in Chinese with English abstract).

[9]

Deng, H., Zheng, Y., Chen, J., et al., 2020. Deep Learning⁃Based 3D Prediction Model for the Dayingezhuang Gold Deposit, Shandong Province. Acta Geoscientica Sinica, 41(2): 157-165 (in Chinese with English abstract).

[10]

Deng, J., Yang, L. Q., Groves, D. I., et al., 2020. An Integrated Mineral System Model for the Gold Deposits of the Giant Jiaodong Province, Eastern China. EarthScience Reviews, 208: 103274. https://doi.org/10.1016/j.earscirev.2020.103274

[11]

Guo, J. T., Liu, Y. H., Han, Y. F., et al., 2019. Implicit 3D Geological Modeling Method for Borehole Data Based on Machine Learning. Journal of Northeastern University (Natural Science), 40(9): 1337-1342 (in Chinese with English abstract).

[12]

Guo, J. T., Wang, X. L., Wang, J. M., et al., 2021. Three⁃Dimensional Geological Modeling and Spatial Analysis from Geotechnical Borehole Data Using an Implicit Surface and Marching Tetrahedra Algorithm. Engineering Geology, 284: 106047. https://doi.org/10.1016/j.enggeo.2021.106047

[13]

Guo, J. T., Wang, Z. X., Li, C. L., et al., 2022. Multiple⁃Point Geostatistics⁃Based Three⁃Dimensional Automatic Geological Modeling and Uncertainty Analysis for Borehole Data. Natural Resources Research, 31(5): 2347-2367. https://doi.org/10.1007/s11053⁃022⁃10071⁃6

[14]

Guo, J. T., Wu, L. X., Zhou, W. H., et al., 2016. Towards Automatic and Topologically Consistent 3D Regional Geological Modeling from Boundaries and Attitudes. ISPRS International Journal of GeoInformation, 5(2): 17. https://doi.org/10.3390/ijgi5020017

[15]

Hu, J., Shen, L., Sun, G., 2018. Squeeze⁃ and ⁃Excitation Networks. The IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City. https://doi.org/10.48550/arXiv.1709.01507

[16]

Huang, J. X., Deng, H., Chen, J., et al., 2023a. Assessing Geometrical Uncertainties in Geological Interface Models Using Markov Chain Monte Carlo Sampling via Abstract Graph. Tectonophysics, 864: 230032. https://doi.org/10.1016/j.tecto.2023.230032

[17]

Huang, J. X., 2022. MCMC Simulation and Variational Reconstruction for 3D Geological Interface of Metallogenic Structure (Dissertation). Central South University, Changsha (in Chinese with English abstract).

[18]

Huang, S. Y., Wang, T. Y., Xiong, H. Y., et al., 2021. Semi⁃Supervised Active Learning with Temporal Output Discrepancy. The IEEE/CVF International Conference on Computer Vision, Montreal. https://doi.org/10.48550/arXiv.2107.14153

[19]

Huang, Z. L., Wang, X. G., Wei, Y. C., et al., 2023b. CCNet: Criss⁃Cross Attention for Semantic Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6): 6896-6908. https://doi.org/10.1109/tpami.2020.3007032

[20]

Ioffe, S., Szegedy, C., 2015. Batch Normalization: Accelerating Deep Network Training by reducing Internal Covariate Shift. International Conference on Machine Learning. Lille. https://doi.org/10.48550/arXiv.1807.06521

[21]

Jia, R., Lü, Y. K., Wang, G. W., et al., 2021. A Stacking Methodology of Machine Learning for 3D Geological Modeling with Geological⁃Geophysical Datasets, Laochang Sn Camp, Gejiu (China). Computers & Geosciences, 151: 104754. https://doi.org/10.1016/j.cageo.2021.104754

[22]

Li, J. J., Dang, Z. C., Fu, C., et al., 2023. Genesis of the Yangjiakuang Gold Deposit, Jiaodong Peninsula, China: Constraints from S⁃He⁃Ar⁃Pb Isotopes, and Sm⁃Nd and U⁃Pb Geochronology. Frontiers in Earth Science, 11: 1048509. https://doi.org/10.3389/feart.2023.1048509

[23]

Li, S. Y., Li, J., Song, M. C., et al., 2022. Metallogenic Characteristics and Mineralization of the Linglong Gold Field, Jiaodong Peninsula. Acta Geologica Sinica, 96(9): 3234-3260 (in Chinese with English abstract).

[24]

Lin, T. Y., Goyal, P., Girshick, R., et al., 2017. Focal Loss for Dense Object Detection. The IEEE International Conference on Computer Vision. Venice. https://doi.org/10.48550/arXiv.1708.02002

[25]

Liu, Y., Zhang, Y., Wang, Y. X., et al., 2024. A Survey of Visual Transformers. IEEE Transactions on Neural Networks and Learning Systems, 35(6): 7478-7498. https://doi.org/10.1109/TNNLS.2022.3227717

[26]

Liu, Z. B., Zhang, J. Q., Du, X. F., et al., 2024. Implicit 3D Integrated Modeling of Complex Geological Structures in Mining Areas. Journal of Northeastern University (Natural Science), 45(9): 1317-1325 (in Chinese with English abstract).

[27]

Lou, Y. M., Kang, X., Lai, Y. P., et al., 2025. Application of Implicit Modeling and Machine Learning Algorithm to 3D Metallogenic Prediction of the Julong Porphyry Copper⁃Molybdenum Deposit, Xizang. Earth Science Frontiers, 32(5): 440-455 (in Chinese with English abstract).

[28]

Lü, P. F., Chen, W. Y., Zou, X. Y., 2025. Precision Recognition of Rock Thin Section Images with Multi⁃Head Self⁃Attention Convolutional Neural Networks. Journal of Geophysical Research: Machine Learning and Computation, 2(2): e2025JH000617. https://doi.org/10.1029/2025JH000617

[29]

Mao, X. C., Duan, M., Deng, H., et al., 2026. Intelligent 3D Prediction of Deep Mineral Resources: Theory, Methods, and Challenges. Earth Science, 51(3): 793-815 (in Chinese with English abstract).

[30]

Min, Q. F., Lu, Y. G., Liu, Z. Y., et al., 2019. Machine Learning Based Digital Twin Framework for Production Optimization in Petrochemical Industry. International Journal of Information Management, 49: 502-519. https://doi.org/10.1016/j.ijinfomgt.2019.05.020

[31]

Mumuni, A., Mumuni, F., 2022. Data Augmentation: A Comprehensive Survey of Modern Approaches. Array, 16: 100258. https://doi.org/10.1016/j.array.2022.100258

[32]

Ridnik, T., Lawen, H., Noy., A., et al., 2021. Tresnet: High Performance GPU⁃Dedicated Architecture. The IEEE/CVF Winter Conference on Applications of Computer Vision. Hawaii. https://doi.org/10.48550/arXiv.2003.13630

[33]

Shi, L. Y., Zuo, R. G., 2026. Foundation Model for Mineral Prospectivity Mapping. Earth Science, 51(3): 832-848 (in Chinese with English abstract).

[34]

Song, Z. Y., Xiang, Y. H., Liu, Z. K., et al., 2024. Lithogeochemistry of Altered Rocks and Mineralization in Xindongzhuang Gold Deposit, Northwest Jiaodong Peninsula. Gold, 45(8): 89-93, 98 (in Chinese with English abstract).

[35]

Wang, H., Yan, J. Y., Qi, G., et al., 2023. Metallogenic Prediction Method Based on Gravity and Magnetic Three⁃ Dimensional Modeling and Machine Learning: A Case Study of Zhuxi. Progress in Geophysics, 38(2): 734-747 (in Chinese with English abstract).

[36]

Wang, T. R., Ji, X. B., Wang, J. B., et al., 2025. Implicit 3D Geological Modeling Based on Machine Learning: A Case Study of Lazigou Gold Deposit in Muping⁃Rushan Metallogenic Belt. Earth Science, 50(8): 3167-3181 (in Chinese with English abstract).

[37]

Wang, X., Girshick, R., Gupta, A., et al., 2018. Non⁃local Neural Networks. The IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City. https://doi.org/10.48550/arXiv.1711.07971

[38]

Woo, S., Park, J., Lee, J. Y., et al., 2018. CBAM: Convolutional Block Attention Module. arXiv: 1807.06521. https://arxiv.org/abs/1807.06521

[39]

Xiao, F., Chen, X. Y., 2025. Numerical Modeling and Exploration Data Coupled⁃Driven Mineral Prospectivity Mapping: A Case Study of Fankou Pb⁃Zn Deposit. Geotectonica et Metallogenia, 49(2): 298-316 (in Chinese with English abstract).

[40]

Xu, L. P., Zhu, W. P., Zhu, H. W., et al., 2022. Physical Property Characteristics of Rocks in Hanzhong and Ankang Areas at the Southern Foot of Qinling Mountains and Their Application. Geophysical and Geochemical Exploration, 46(5): 1167-1179 (in Chinese with English abstract).

[41]

Yang, J. H., Xu, L., Sun, J. F., et al., 2021. Geodynamics of Decratonization and Related Magmatism and Mineralization in the North China Craton. Scientia Sinica (Terrae), 51(9): 1401-1419 (in Chinese with English abstract).

[42]

Yao, X. F., Cheng, Z. Z., Du, Z. Z., et al., 2020. U⁃Pb Age of Post⁃Ore Dykes in the Xiejiagou Gold Deposit and Its Constraints on Ore⁃Forming Age, Northwest Jiaodong, China. Geological Bulletin of China, 39(8): 1153-1162 (in Chinese with English abstract).

[43]

Ye, S. W., Hou, W. S., Yang, J., et al., 2025. Advance of 3D Smart Geological Modeling. Earth Science Frontiers, 32(4): 182-198 (in Chinese with English abstract).

[44]

Yu, S. W., Ma, J. W., 2021. Deep Learning for Geophysics: Current and Future Trends. Reviews of Geophysics, 59(3): e2021RG000742. https://doi.org/10.1029/2021rg000742

[45]

Yu, X. W., Wang, L. M., Ren, T. L., et al., 2023. Geochemistry, Zircon U⁃Pb Age and Lu⁃Hf Isotope of the Concealed Guojialing Granite Revealed by Boreholes in the Northwestern Jiaodong Region. Acta Geologica Sinica, 97(2): 417-432 (in Chinese with English abstract).

[46]

Yuan, F., Li, X. H., Tian, W. D., et al., 2024. Key Issues in Three⁃Dimensional Predictive Modeling of Mineral Prospectivity. Earth Science Frontiers, 31(4): 119-128 (in Chinese with English abstract).

[47]

Zhang, B. Y., Xu, Z. H., Wei, X. Z., et al., 2024. Deep Subsurface Pseudo⁃Lithostratigraphic Modeling Based on Three⁃Dimensional Convolutional Neural Network (3D CNN) Using Inversed Geophysical Properties and Shallow Subsurface Geological Model. Lithosphere, 2024(1): lithosphere_2023_273. https://doi.org/10.2113/2024/lithosphere_2023_273

[48]

Zhang, M. M., Chen, C., Huang, Y. Q., et al., 2026. Three⁃Dimensional Mineral Prospectivity Modeling of Skarn⁃Type Copper Deposits in the Anqing Area Based on Causal Inference and Graph Attention Networks. Earth Science, 51(3): 909-920 (in Chinese with English abstract).

[49]

Zhang, Q. B., Song, M. C., Ding, Z. J., et al., 2022. Exhumation History and Preservation of the Jiaojia Giant Gold Deposit, Jiaodong Peninsula. Science China Earth Sciences, 65(6): 1161-1177. https://doi.org/10.1007/s11430⁃021⁃9887⁃1

[50]

Zhang, Z. Q., 2022. Research of Machine Learning for District⁃Scale Three⁃Dimensional Implicit Geological Modeling and Mineral Potential Mapping (Dissertation). China University of Geosciences, Beijing (in Chinese with English abstract).

[51]

Zhang, Z. Q., Li, Y. J., Wang, G. W., et al., 2023a. Supervised Mineral Prospectivity Mapping via Class⁃ Balanced Focal Loss Function on Imbalanced Geoscience Datasets. Mathematical Geosciences, 55(7): 989-1010. https://doi.org/10.1007/s11004⁃023⁃10065⁃x

[52]

Zhang, Z. Q., Wang, G. W., Carranza, E. J. M., et al., 2023b. An Integrated Machine Learning Framework with Uncertainty Quantification for Three⁃Dimensional Lithological Modeling from Multi⁃Source Geophysical Data and Drilling Data. Engineering Geology, 324: 107255. https://doi.org/10.1016/j.enggeo.2023.107255

[53]

Zhang, Z. Q., Wang, G. W., Carranza, E. J. M., et al., 2025. Three⁃Dimensional Mineral Prospectivity Mapping Using a Residual Convolutional Neural Network with Lightweight Attention Mechanisms. Ore Geology Reviews, 185: 106797. https://doi.org/10.1016/j.oregeorev.2025.106797

[54]

Zhao, J. T., Yu, C. X., Peng, S. P., et al., 2016. Seismic Sparse Inversion Method Implemented on Image Data for Detecting Discontinuous and Inhomogeneous Geological Features. Chinese Journal of Geophysics, 59(9): 3408-3416 (in Chinese with English abstract).

[55]

Zhao, L., Zhang, H. L., Sun, X. D., et al., 2024. Application of ResUNet⁃CBAM in Thin⁃Section Image Segmentation of Rocks. Information, 15(12): 788. https://doi.org/10.3390/info15120788

[56]

Zhong, Z., Zheng, L., Kang, G. L., et al., 2020. Random Erasing Data Augmentation. Proceedings of the AAAI Conference on Artificial Intelligence, 34(7): 13001-13008. https://doi.org/10.1609/aaai.v34i07.7000

[57]

Zhou, M. L., Sun, L. L., Lyu, J. Y., et al., 2024. Exploration and Scientific Research of the Jiaojia⁃Type Gold Deposit. Journal of Geomechanics, 30(5): 747-767 (in Chinese with English abstract).

[58]

Zhu, P. G., Zhang, W. J., Chi, N. J., et al., 2022. Geochemical Characteristics and Zircon U⁃Pb Age of Concealed Granitoids in the Footwall of Jiaojia Fault Zone in Jincheng Area, Shandong Province. Science Technology and Engineering, 22(15): 5976-5987 (in Chinese with English abstract).

[59]

Zhu, R. X., Fan, H. R., Li, J. W., et al., 2015. Decratonic Gold Deposits. Scientia Sinica (Terrae), 45(8): 1153-1168 (in Chinese with English abstract).

[60]

Zhu, Y. J., Wang, D. D., Liu, J., et al., 2025. 3⁃D Gravity and Magnetic Inversion with a Modified Generalized Depth Weighting. IEEE Transactions on Geoscience and Remote Sensing, 63: 5916313. https://doi.org/10.1109/TGRS.2025.3586257

[61]

Zitouni, M. S., Alkhatib, M. Q., Aburaed, N., et al., 2025. A Comparative Analysis of Attention Mechanisms in 3D CNN⁃Based Hyperspectral Image Super⁃Resolution. 2024 14th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS). Helsinki. https://doi.org/10.1109/WHISPERS65427.2024.10876507

[62]

Zou, Y. H., Zhang, W. Q., Mao, X. C., et al., 2023. Numerical Simulation of Hydrothermal Alteration Chemical Reactions during Ore⁃Forming Process of the Jiaojia Gold Deposit, Jiaodong Peninsula, China. Geotectonica et Metallogenia, 47(5): 1158-1172 (in Chinese with English abstract).

[63]

Zuo, R. G., 2021. Data Science⁃Based Theory and Method of Quantitative Prediction of Mineral Resources. Earth Science Frontiers, 28(3): 49-55 (in Chinese with English abstract).

基金资助

国家科技重大专项(2024ZD1001900)

国家自然科学基金项目(42402301)

AI Summary AI Mindmap
PDF (3807KB)

0

访问

0

被引

详细

导航
相关文章

AI思维导图

/