<p>Data-driven deep learning methodologies have been extensively employed in the domain of geochemical anomaly identification. These approaches utilize either supervised or unsupervised learning techniques to extract geochemical anomalies. Nonetheless, several challenges substantially impede the effective application of deep learning methods, notably: (1) the variability in geochemical distribution patterns across different study areas often results in suboptimal performance when applying a trained model; (2) the variability in the number of input geochemical elements across study areas can compromise a model’s applicability; and (3) the scarcity of mineral deposit samples limits the efficacy of supervised learning frameworks. To overcome these limitations, this study adopted a hybrid methodology that integrates unsupervised pre-training with supervised fine-tuning to develop a foundation model for geochemical anomaly identification. In the unsupervised pre-training stage, a masked autoencoder architecture based on the vision transformer was utilized to train an encoder capable of robustly capturing geochemical spatial patterns, thereby establishing a foundation model. Subsequently, a classifier was appended to this foundation model for supervised fine-tuning, facilitating the transformation of high-dimensional features to identify geochemical anomaly. To assess the model’s effectiveness, two fine-tuning strategies were implemented: (1) transfer learning, wherein two study areas with analogous geochemical spatial patterns served as source and target domains for training and prediction, respectively, thereby assessing the model’s ability to generalize learned geochemical patterns across domains; and (2) few-shot learning, which involved training the model on an extremely limited subset of mineral deposits within a given study area and subsequently predicting the remaining mineral deposits, thereby testing the model’s capacity to infer geochemical patterns from sparse data and identify potential mineral deposits. This study employed all available geochemical survey data from Hubei Province, China, for unsupervised pre-training to develop the foundation model. Transfer learning experiments were conducted using the western and southeastern regions of Hubei Province as the source and target domains, respectively. For evaluating cross-regional applicability, few-shot learning experiments utilized two known mineral deposits in western Henan Province of China, with the remaining known deposits serving as evaluation data. Comparative analyses with traditional deep learning algorithms demonstrated that the proposed foundation model effectively captures spatial distribution patterns in geochemical survey data.</p>

A Foundation Model for Geochemical Anomaly Identification

期刊论文
发表日期：

2026年3月16日

查看全文

Luyi Shi,
Renguang Zuo

摘要

Data-driven deep learning methodologies have been extensively employed in the domain of geochemical anomaly identification. These approaches utilize either supervised or unsupervised learning techniques to extract geochemical anomalies. Nonetheless, several challenges substantially impede the effective application of deep learning methods, notably: (1) the variability in geochemical distribution patterns across different study areas often results in suboptimal performance when applying a trained model; (2) the variability in the number of input geochemical elements across study areas can compromise a model’s applicability; and (3) the scarcity of mineral deposit samples limits the efficacy of supervised learning frameworks. To overcome these limitations, this study adopted a hybrid methodology that integrates unsupervised pre-training with supervised fine-tuning to develop a foundation model for geochemical anomaly identification. In the unsupervised pre-training stage, a masked autoencoder architecture based on the vision transformer was utilized to train an encoder capable of robustly capturing geochemical spatial patterns, thereby establishing a foundation model. Subsequently, a classifier was appended to this foundation model for supervised fine-tuning, facilitating the transformation of high-dimensional features to identify geochemical anomaly. To assess the model’s effectiveness, two fine-tuning strategies were implemented: (1) transfer learning, wherein two study areas with analogous geochemical spatial patterns served as source and target domains for training and prediction, respectively, thereby assessing the model’s ability to generalize learned geochemical patterns across domains; and (2) few-shot learning, which involved training the model on an extremely limited subset of mineral deposits within a given study area and subsequently predicting the remaining mineral deposits, thereby testing the model’s capacity to infer geochemical patterns from sparse data and identify potential mineral deposits. This study employed all available geochemical survey data from Hubei Province, China, for unsupervised pre-training to develop the foundation model. Transfer learning experiments were conducted using the western and southeastern regions of Hubei Province as the source and target domains, respectively. For evaluating cross-regional applicability, few-shot learning experiments utilized two known mineral deposits in western Henan Province of China, with the remaining known deposits serving as evaluation data. Comparative analyses with traditional deep learning algorithms demonstrated that the proposed foundation model effectively captures spatial distribution patterns in geochemical survey data.