In the contaminant remediation of groundwater, the release history of contaminant sources and hydraulic conductivity field are two key parameters that need to know, but their actual values are difficult to obtain and can only be inversely identified by limited measured data. However, the process of solving the inverse problem needs to repeatedly call the forward model of contaminant transport, which is very time-consuming, especially for the high-dimensional inverse problems. In this study, based on the training data generated from a prior range of parameters (the release strength of contaminant sources and hydraulic conductivity at pilot points), the self-organizing maps (SOM) algorithm was employed to construct the surrogate model for the numerical model of contaminant transport in a simplified hypothetical aquifer, then the surrogate model was used to retrieve jointly the contaminant strength of sources and the hydraulic conductivity at pilot points, and the Kriging method of geostatistics was further used to process the estimated K-values at pilot points to obtain the hydraulic conductivity field. Also, to investigate the ability of the SOM-based surrogate model for retrieving both contaminant source strengths and hydraulic conductivity, we gradually expanded the prior range and increased the number of inversion terms in each prior range. Moreover, the robustness of the SOM-based surrogate model for inversion was illustrated by proposing the scarcity of data and different degrees of measurement error in the limited actual observation data. When the actual observation data is reduced by 2/3, the Root Mean Square Error (RMSE) of retrieving source strengths and hydraulic conductivity at pilot points are 1.07 and 0.09, respectively. The results indicated the SOM-based surrogate model shows remarkable inversion precision and robustness, and an accurate estimation of the actual hydraulic conductivity field could be obtained by the Kriging method based on that.

在地下水污染物修复中，污染物源的释放历史和导水率场是两个需要知道的关键参数，但它们的实际值很难获得，只能通过有限的实测数据进行反识别。但是，求解逆问题的过程需要反复调用污染物运移正演模型，非常耗时，尤其是对于高维逆问题。在本研究中，基于从先验参数范围（污染物源的释放强度和先导点的水力传导率）生成的训练数据，采用自组织图 (SOM) 算法构建数值模拟模型。简化假设中的污染物迁移模型含水层，然后使用代理模型联合反演源的污染物强度和先导点的水力传导率，以及地质统计学的克里金方法进一步用于处理先导点的估计 K 值以获得水力传导率场。此外，为了研究基于 SOM 的替代模型检索污染物源强度和水力传导率的能力，我们逐渐扩大了先验范围并增加了每个先验范围中的反演项数。此外，通过在有限的实际观测数据中提出数据的稀缺性和不同程度的测量误差，说明了基于SOM的反演替代模型的稳健性。当实际观测数据减少2/3时，在先导点反演源强度和导水率的均方根误差（RMSE）分别为1.07和0.09。结果表明，基于 SOM 的代理模型具有显着的反演精度和鲁棒性，