Groundwater contaminant source identification (GCSI) can provide support for the confirmation of responsibility and the remediation of pollution. This study has developed an innovative combination algorithm to recognize source properties and model parameters in groundwater contamination simultaneously. The research idea is based on Bayes theory and combines an ensemble smoother (ES) algorithm, differential evolutionary Markov chain (DEMC) algorithm, and adaptive kriging surrogate model (AKSM). Poor selection of initial estimates for unknown variables will slow down the convergence rate. Therefore, the initial points are not generated completely randomly, but are partially obtained by the ES algorithm. After the initial points have been determined, the DEMC algorithm is used to recognize source properties and model parameters in groundwater contamination. To improve the efficiency of the DEMC algorithm, the updating formula was adjusted by introducing information about the optimal chain into the iteration. However, the inversion process is time-consuming because both the ES and DEMC algorithms need to run the original simulation model frequently. To solve this problem, an AKSM was established for the original simulation model, which greatly accelerated the inversion process. Different hypothetical cases with different complexities were used to illustrate the validity of the combination algorithm. The identification results implied that the combination algorithm had not only faster convergence, but also higher accuracy. These improvements were more evident in the second case and third case. This implies that the proposed method will play a greater role with increasing problem complexity.

地下水污染物来源识别（GCSI）可以为责任的确认和污染的修复提供支持。本研究开发了一种创新的组合算法，可以同时识别地下水污染中的源属性和模型参数。研究思路基于贝叶斯理论，结合了集成平滑器（ES）算法、差分进化马尔可夫链（DEMC）算法和自适应克里金代理模型（AKSM）。对未知变量的初始估计选择不当会减慢收敛速度。因此，初始点不是完全随机生成的，而是部分通过ES算法获得的。初始点确定后，DEMC 算法用于识别地下水污染中的源属性和模型参数。为了提高DEMC算法的效率，通过在迭代中引入最优链信息来调整更新公式。然而，反演过程非常耗时，因为 ES 和 DEMC 算法都需要频繁运行原始模拟模型。为了解决这个问题，对原始模拟模型建立了AKSM，大大加快了反演过程。使用不同复杂度的不同假设案例来说明组合算法的有效性。识别结果表明组合算法不仅收敛速度更快，而且精度更高。这些改进在第二种情况和第三种情况下更为明显。