講演情報
[SGC46-06]When Expectation–Maximization Meets Self-Distillation: A Unified Framework for Robust Geochemical Anomaly Recognition
*Shuyan Yu1,2、Hao Deng1 (1.School of Geosciences and Info-Physics, Central South University、2.Earthquake Research Institute, The University of Tokyo)
キーワード：
Geochemical anomaly、Expectation-Maximization、Self distillation、Transformer、Geochemical pattern
Unsupervised deep learning is increasingly used for geochemical anomaly recognition in data-scarce greenfield exploration, typically through reconstruction-based models that characterize background patterns and detect anomalies via reconstruction errors. However, geochemical datasets are often affected by strong heterogeneity, complex mineralization processes, and sparse sampling, which introduce disruptive variance and increase the risk of overfitting to local anomalies. To address these limitations and better separate geochemical trends from mineralization-related anomalies, we reformulate reconstruction-based deep learning as a Maximum Likelihood Estimation (MLE) problem in which the geochemical background is treated as a latent variable. An Expectation–Maximization (EM) algorithm is introduced to iteratively estimate and refine this latent background representation. This EM procedure can be interpreted as a self-distillation process, providing a mathematically interpretable training strategy that stabilizes model learning and suppresses anomaly-induced variance. A Transformer architecture is employed as the backbone network to capture long-range geochemical dependencies that may reflect large-scale tectono-magmatic controls on elemental distributions. Random masking is incorporated to further enhance robustness under limited data conditions. The proposed framework is applied to geochemical data from the northwestern Jiaodong gold province, eastern China. Results demonstrate improved robustness to noise and anomalous fluctuations, clearer delineation of background geochemical structures, and superior anomaly detection performance compared to existing methods. This study establishes a unified framework linking Expectation-Maximization and self-distillation, providing a theoretically interpretable approach for modeling latent geochemical backgrounds and enhancing anomaly recognition in complex geological settings.