As a fundamental approach in machine learning, neural networks have evolved through phases of emergence, relative dormancy, and subsequent flourishing since their inception. Over the past decade, the rapid advancement of computing power - particularly the accelerated computing capabilities of GPUs and the rapid development of cloud computing - has propelled deep learning technologies, with deep neural networks as their hallmark, to become one of the most effective solutions in both industrial and scientific computing domains. These technologies have outperformed traditional methods in numerous applications and now form a cornerstone of artificial intelligence. This study integrates deep learning with gravity anomaly detection, leveraging convolutional neural networks (CNNs) that have demonstrated exceptional performance in image recognition tasks. The approach treats gravity observation contour maps as 2D images for analysis, with spatial parameters of subsurface gravity anomalies serving as recognition outputs, thereby establishing a specialized CNN model for anomaly identification. During training, we randomly generate numerous 3D anomaly models with varying parameters, compute their 2D gravity observations through forward modeling, and train the CNN using both parameter labels and gravity data. Experimental validation on test cases demonstrates the model’s high accuracy. Notably, unlike traditional CNNs that only identify burial depths from 2D gravity lines, this convolutional network achieves comprehensive detection of both depth and size information for 3D anomalies. When applied to gravity observation data from Australia’s Kauring region, the model’s results align with previous research findings, confirming its generalization capability for real-world gravity anomaly identification with reliable outcomes. The proposed method offers several advantages over conventional techniques. Firstly, by utilizing CNNs, it automates the feature extraction process from gravity contour maps, eliminating the need for manual interpretation and reducing human subjectivity. Secondly, the end-to-end training framework enables simultaneous optimization of multiple anomaly parameters, enhancing detection accuracy and robustness. Thirdly, the model demonstrates strong adaptability to various geological settings, as evidenced by its successful application in the Kauring region with diverse subsurface structures. Furthermore, the computational efficiency of CNNs allows for rapid processing of large-scale gravity datasets, making it suitable for regional-scale anomaly mapping. These characteristics collectively establish the developed CNN-based approach as a promising tool for advancing gravity anomaly detection in both academic research and practical exploration applications.
Jiaxin Ji (Fri,) studied this question.