A deep adaptive cycle generative adversarial neural network for inverse
estimation of groundwater contaminated source and model parameter
Abstract
In light of the challenges posed by groundwater contamination and the
urgent need for accurate and efficient groundwater contaminated source
estimation (GCSE), the present study proposes a novel approach for GCSE
using a deep adaptive cycle generative adversarial neural network
(DA-CGAN). Given the equifinality from different parameters (EFDP) often
associated with GCSE, we leveraged a bidirectional adversarial training
pattern involving a forward process and a recovery process to supervise
the inverse mapping relationship. Once trained, the forward process can
be utilized to provide estimation for GSCE. This bidirectional-training
strategy mitigates EFDP, thereby effectively enhancing the reliability
of GCSE. Moreover, the performance of DA-CGAN is closely related to the
quality of the training samples. To address this, we introduced a
significant enhancement through an adaptive sampling strategy. This
substantially improves the quality of training samples and consequently
increases the accuracy of the GCSE. Furthermore, the inherent
data-driven attribute of the deep cycle GAN considerably reduces
computational costs when conducting GCSE. The research unfolds in the
contexts of both hypothetical and real-world scenarios, with the goal of
providing an efficient, precise, and cost-effective solution for GCSE.
The results demonstrate that the DA-CGAN, an innovative model in the
hydrogeological domain, exhibits superior performance in both estimation
accuracy (Average Relative Error (ARE) of 4.91% and R of 0.998) and
computational efficiency (0.17 seconds per run). This is particularly
notable when compared with typical inverse methods such as the genetic
algorithm (GA) and the ensemble kalman filter (ENKF).