首页|期刊导航|Green and Smart Mining Engineering|A weighted ML-Kriging 3D geological model integrated with deep iterative optimization

A weighted ML-Kriging 3D geological model integrated with deep iterative optimizationOA

中文摘要

“Transparent geology”underpins intelligent mine blasting by enabling the precise classification of rock mass blastability and optimal matching of explosive energy.To advance transparency of geological models,we proposed a six-stage workflow comprising data synthesis,classifier-based labeling,stable point selection,machine learning‑based probability field prediction,conditional Kriging interpolation,and 3D geological reconstruction.First,we generated multiple synthetic borehole datasets with invariant geological features by applying random skeleton reconstruction fused with generative adversarial networks(GANs)under selected prior constraints.We then used an optimal classifier to predict the lithologies of each synthetic batch and tally the prediction frequencies,retaining only points that recurred across runs as“stable high‑precision points.”Next,these stable points were merged with the original borehole logs to form a unified set of“high‑confidence points.”For each query location,we computed two neighborhood descriptors:(1)the number of high-confidence points within a predefined radius N_(S)and(2)the mean Euclidean distance to those points D_(A).Following this,these neighborhood descriptions were concatenated with the spatial coordinates to construct a five-dimensional feature vector(x,y,z,N_(S),D_(A)).This vector was inputted into a deep neural network that predicted the lithological class probabilities at each spatial query point.Simultaneously,conditional Kriging was performed on the original logs to obtain a smoothly varying geostatistical probability field.Finally,we fused the machine learning-derived and Kriging‑derived probability fields via a weighted scheme and used the fused field to build a 3D lithological model.Cross-sectional slices at specified locations were then extracted to illustrate both thin interbeds and large-scale geological structures.Comparative experiments showed that embedding stable anchor points and spatial neighborhood features enabled the model to capture local heterogeneity more effectively and replace stand-alone Kriging.Even though the addition of spatial features in the machine learning model led to a drop in the training borehole accuracy compared with the pure machine learning model,it substantially improved generalization to unseen boreholes and enhanced thin layer detection.

Huaqiang Liu;Zhongwen Yue;Wei Liu;Qingyu Jin;Kejun Xue;Jiayao Chen

School of Mechanics and Civil Engineering,China University of Mining and Technology(Beijing),Beijing 100083,ChinaSchool of Mechanics and Civil Engineering,China University of Mining and Technology(Beijing),Beijing 100083,ChinaSchool of Mechanics and Civil Engineering,China University of Mining and Technology(Beijing),Beijing 100083,ChinaSchool of Mechanics and Civil Engineering,China University of Mining and Technology(Beijing),Beijing 100083,ChinaSchool of Mechanics and Civil Engineering,China University of Mining and Technology(Beijing),Beijing 100083,ChinaSchool of Mechanics and Civil Engineering,China University of Mining and Technology(Beijing),Beijing 100083,China

矿业与冶金

Mine blastingGeological modelingSpatially constrained anchor pointsSpatial feature engineeringMachine learning

《Green and Smart Mining Engineering》 2026 (2)

P.166-179,14

supported by the National Natural Science Foundation of China(Nos.52174094,52474117,and 12002366).

10.1016/j.gsme.2026.01.004

评论