Xu, Ying, Zuo, Renguang, Chen, Zhiyi, Shi, Zixian, and Kreuzer, Oliver P. (2025) Recent advances and future research directions in deep learning as applied to geochemical mapping. Earth-Science Reviews, 270. 105209.

PDF (Published Version) - Published Version Restricted to Repository staff only

1

Abstract

Geochemical survey data are a key tool for identifying geochemical patterns and anomalies relevant to mineral exploration. In the past decade, artificial intelligence (AI) has been widely applied in geochemical data mining to compensate for the shortcomings of traditional methods. Here, we first reviewed the applications of five popular deep learning algorithms (DLAs) adopted in the past six years (i.e., from 2019 to 2025), namely deep belief network, recurrent neural network, convolutional neural network, autoencoder, and generative adversarial network. We then examined recent state-of-the-art applications of DLAs in geochemical spatial pattern recognition, which served to highlight their advantages over the five popular DLAs previously discussed. Subsequently, we flagged three critical challenges in DLA-based geochemical mapping: (i) inadequate representation of complex spatial heterogeneity patterns of geochemical survey data, (ii) development of innovative models to overcome the limitations imposed by insufficient training samples, and (iii) systematic integration of geological constraints to enhance model accuracy and interpretability. To address these limitations, we proposed two promising, novel architectures: (i) graph self-supervised learning and (ii) graph reinforcement learning (GRL). Graph self-supervised learning represents geochemical data as graph structures, using self-supervised techniques to address training data limitations. Furthermore, the model uses Transformer for modeling global spatial relationships and embeds knowledge nodes for ensuring geological consistency during model training. Like the above, GRL employs graph representations of geochemical data, also combining graph convolutional networks within a reinforcement learning system. The key advancement of GRL involves the creation of reward functions that incorporate geological rules, thereby linking expert knowledge and DLAs through dynamic environment feedback. A case study is presented to demonstrate the effectiveness of these approaches and highlights the potential for integrating advanced methodologies to enhance the accuracy and reliability of geochemical anomaly identification in complex geological settings.

Actions (Repository Staff Only)

Item Control Page