An unsupervised GNN-transformer model for gold geochemical anomaly detection

Effectively identifying mineral anomalies from geochemical data is one of the key problems in mineral exploration. Some existing models remain limited in their ability to capture the mutual relationship between neighborhood data points and broader spatial context. Recent deep learning frameworks, including Graph Convolutional Networks (GCNs), Transformers, and hybrid variants, have shown notable potential for geochemical anomaly detection. In this study, we propose a hybrid model, GeoTransGNN, to detect gold anomalies from spatial information and Au concentration. Our model integrates the strengths of a graph neural network (GNN) and a Transformer. This hybrid design combines GNN-based local topology learning with Transformer-based broader contextual modeling, improving anomaly detection capability and interpretability. We evaluate GeoTransGNN on a sample-level geochemical dataset from the Hatu gold field in Xinjiang, China, using spatial coordinates and Au concentration as inputs. Because the available input metadata do not reliably specify the original sampling medium, the dataset is conservatively described as a sample-level geochemical dataset. Interpolated 2D maps are used only for visualization and map-level evaluation. The proposed hybrid model is trained in an unsupervised reconstruction setting with separate training, validation, and testing subsets and is compared with GNN and Transformer baselines under the same protocol. The revised experiments show that GeoTransGNN provides the strongest overall performance among the compared deep-learning models and preserves anomaly-map structure more effectively at the map level. These results indicate that GeoTransGNN is a competitive unsupervised framework for Au anomaly detection in this case study and has potential practical value for mineral exploration.