Groundwater contaminant source identification is normally a prerequisite for contaminant remediation. This study proposes a new two-stage surrogate-assisted Markov chain Monte Carlo (MCMC)-based Bayesian framework to identify contaminant source parameters for groundwater polluted by dense nonaqueous phase liquid. In the framework, an adaptive update feedback process is proposed to construct a locally accurate surrogate model over posterior distributions to replace the time-consuming multiphase flow model. To increase the efficiency of the MCMC simulation, a multiobjective feasibility-enhanced particle swarm optimization algorithm (MOFEPSO) is adopted to generate the initial guess of the contamination source parameters. The accuracy and efficiency of the proposed framework are confirmed via a synthetic study. The contaminant source parameters generated by the proposed approach are compared with those computed by the one-stage surrogate-assisted MCMC-based Bayesian approach. The results demonstrate that the root mean squared error (RMSE) between true value of parameters and maximum a-posteriori density values (MAP) obtained by the proposed method decreased by 71.3% compared with those obtained by one-stage surrogate-based framework. To further assess the efficiency of MOFEPSO, the same inversion problem is solved with random values as the initial guesses of the unknown parameters during MCMC simulation; the other conditions are the same as the proposed framework. The results indicate that adopting MOFEPSO improves the efficiency of MCMC simulation. Therefore, the proposed approach can accurately and effectively identify the contaminant source parameters with achieving about 148 times of speed-up compared to the simulation-based MCMC simulation.

地下水污染物源识别通常是污染物修复的先决条件。这项研究提出了一种新的两阶段替代辅助马尔可夫链蒙特卡罗（MCMC）的贝叶斯框架，以识别被稠密非水相液体污染的地下水的污染物源参数。在该框架中，提出了一种自适应更新反馈过程，以构造后验分布上的局部精确替代模型，以取代耗时的多相流模型。为了提高MCMC仿真的效率，采用了多目标可行性增强的粒子群优化算法（MOFEPSO）来生成污染源参数的初始估计值。通过综合研究证实了所提出框架的准确性和效率。将所提出的方法所产生的污染物源参数与基于MCMC的一阶段代理辅助贝叶斯方法所计算的那些相比较。结果表明，与一阶段基于代理的框架相比，所提方法获得的参数真实值与最大后验密度值（MAP）之间的均方根误差（RMSE）降低了71.3％。为了进一步评估MOFEPSO的效率，使用MCMC模拟过程中未知参数的初始猜测来解决相同的反问题。其他条件与建议的框架相同。结果表明，采用MOFEPSO可以提高MCMC仿真的效率。因此，