Groundwater contaminated source estimation based on adaptive correction iterative ensemble smoother with an auto lightgbm surrogate

Abstract

In this work, we have proposed an adaptive-correction iterative ensemble smoother (ACIES) to adaptively adjust the range of unknown variables, for improving the estimation accuracy. While solving the groundwater contaminated source estimation (GCSE), a prior range for the unknown variables must be provided according to the field investigation or manual auxiliary information. However, the prior range might not include the actual values, thus causing misleading estimation results. Moreover, when using ACIES to solve GCSE, it requires massive realizations of the simulation model, causing huge calculation cost. Here, a surrogate model has been utilized to alleviate the high calculation cost. Furthermore, the accurate estimation of GCSE relies on the high-fidelity surrogate models. The well-tuning of the hyper-parameters of the surrogate models can guarantee high fitting accuracy (fidelity) for the origin simulation model. Although the existing tuning methods such as the grid search method can satisfactorily tune the surrogate model, the long tuning time, which has been brought by the ergodic running with every potential combination of hyper-parameters, must not be neglected. Thus, an auto light gradient boosting machine (lightgbm), tuned with swarm evolutionary algorithm, has been utilized as a high-fidelity surrogate model with a swift tuning time. Both the coal gangue and high-dimensional estimation scenarios have been designed to evaluate the performance of the proposed method. The results indicated the following. (1) While encountered with the vague prior ranges, the adaptive ensemble smoother could adjust the potential search ranges to provide an accurate estimation. (2) When compared with the typical surrogate models, an auto-lightgbm surrogate could achieve the promising generalization accuracy with fast a tuning time. This guaranteed the high fidelity of the surrogate models, which further improved the accuracy of the estimated results.