Deep learning algorithms (DLAs) have achieved better results than traditional methods in the field of multivariate geochemical anomaly recognition because of their strong ability to extract feature from nonlinear data. However, most of DLAs are black-box approaches because of the high nonlinearity characteristics of the hidden layer. In addition, the integration of domain knowledge into the DLAs to ensure physical consistency is a challenge for DLAs in geoscience. In this study, we adopted the adversarial autoencoder (AAE) algorithm for geochemical anomaly detection. The interpretability of the model is improved by visualizing features and integrating geological domain knowledge into the loss function of the AAE. The feature visualization method was used to display the changes of information in the model calculation process to further understand the inherent operation law and principle of the neural network. The penalty term was added to the optimized loss function, and the spatiotemporal and genetic relationships between felsic intrusions and mineralization were integrated into the AAE with the aim of improving the geological interpretability of the network. The added penalty item can guide the changes in the stage of data reconstruction and improve the understandability of the results of geologically constrained AAE. In addition, the effectiveness of injecting the concept of physical constraints into the AAE can be verified via feature visualization. A case study in the southern Jiangxi Province and its surrounding areas was performed to identify multivariate geochemical anomalies. The results obtained by the geologically constrained AAE demonstrated a strong spatial correlation with the outcrop of intrusions in the study area, and most of the known mineral deposits are located in or near the highly anomalous areas.

深度学习算法（DLA）由于其强大的非线性数据特征提取能力，在多元地球化学异常识别领域取得了优于传统方法的效果。然而，由于隐藏层的高非线性特性，大多数 DLA 都是黑盒方法。此外，将领域知识整合到 DLA 中以确保物理一致性是地球科学 DLA 面临的挑战。在本研究中，我们采用对抗性自动编码器 (AAE) 算法进行地球化学异常检测。通过将特征可视化并将地质领域知识集成到 AAE 的损失函数中来提高模型的可解释性。采用特征可视化的方法展示模型计算过程中信息的变化，进一步了解神经网络的内在运行规律和原理。在优化损失函数中增加了惩罚项，并将长英质侵入体与矿化之间的时空和成因关系整合到AAE中，以提高网络的地质可解释性。增加的惩罚项可以指导数据重建阶段的变化，提高地质约束AAE结果的可理解性。此外，可以通过特征可视化验证将物理约束概念注入 AAE 的有效性。在江西省南部及其周边地区进行了案例研究，以识别多元地球化学异常。受地质约束的 AAE 获得的结果表明，与研究区侵入体露头有很强的空间相关性，大多数已知矿床位于高度异常区域内或附近。