Adaptive Kriging surrogate model for the optimization design of a dense non-aqueous phase liquid-contaminated groundwater remediation process

The optimization model needs to call the simulation model to calculate the response under different conditions for many times, and this is computationally expensive and time-consuming. To solve this problem, surrogate models can be used to yield insight into the functional relationship between the design variables and the responses, instead of simulation models in the optimization. In this paper, an integrated optimization method based on adaptive Kriging surrogate models was proposed and applied to the cost optimization of a surfactant enhanced aquifer remediation process for dense non-aqueous phase liquids (DNAPLs). First, the initial samples were created by Latin hypercube sampling, and then the responses corresponding to the initial samples were computed by a simulation model. The initial Kriging model was derived through these samples. Secondly, the adaptive Kriging surrogate model was proposed based on updating initial Kriging with new samples via infill sampling criteria. The results showed that it had improved the accuracy of the surrogate model, and the added samples had provided more information about the simulation model than the common samples. Even with the same number of samples, the adaptive Kriging surrogate model performed better than the common Kriging surrogate model, which was built only once. What's more, the integrated approach not only greatly reduced the computational burden, but also determined the actual optimal DNAPLs remediation strategy.