| Petroleu products ay leak into subterranean aquifers and conta inate groundwater during the anufacturing,storage,and transportation processes,har ing hu an health and environ ental quality.Because of the hidden nature of groundwater conta ination,the detection lag,and the long-ter nature of conta ination,it is difficult to quickly and accurately grasp the characteristics of groundwater conta ination sources,such as the nu ber of sources,spatial location,and release intensity,aking it difficult to assess the risk of groundwater conta ination,develop conta ination control technologies and progra s,and identify responsible parties.As a result,it is critical to do study on the features of groundwater petroleu conta ination sources.Previous groundwater conta ination source inversion studies ostly focus on the identification of a particular aspect of the source characteristics or aquifer para eters,and usually generalize the aquifer as ho ogeneous to si plify the real-world proble,which akes the inversion results have great uncertainty.In practical proble s,the transport rate,transport pathway and orphological distribution of conta inants are controlled by the spatial variability of aquifer para eters.Therefore,while inversion identification of conta ination source characteristics,it can finely portray the spatial variability of aquifer para eters,which is an i portant guidance for groundwater resource anage ent,groundwater conta ination risk control and conta ination re ediation.In this paper,a petroche ical conta inated site is used as an exa ple study area,and the characteristics of the conta ination source of Light Non-Aqueous Phase Liquids（LNAPLs）and the inho ogeneous per eability field of the aquifer are investigated by using athe atical physical equations forward and backward,nu erical si ulation of groundwater ultiphase flow,deep learning theory,data assi ilation algorith and stochastic statistical theory.The identification study was conducted and the ain findings are as follows.（1）In order to accurately characterize the conta ination source of groundwater LNAPLs and the actual aquifer inho ogeneity while reducing the co putational load of this high-di ensional nonlinear inversion identification proble,the KL（Karhunen-Lo?ve）expand ethod is used in this paper to stochastically characterize the high-di ensional inho ogeneous per eability field of the aquifer in the exa ple study area.This ethod resulted in a 94.89%reduction in the su of the variables to be inverted for the real-world inversion proble.The surrogate odel for the nu erical si ulation odel of groundwater ultiphase flow at the actual conta inated site of LNAPLs is established based on the Deep Convolutional Encoder-Decoder Neural Network structure,and the construction task of the surrogate odel is converted to the i age regression prediction task in the neural network,which can i prove the surrogate odeling capability of the surrogate odel for the nu erical si ulation odel of ultiphase flow with high-di ensional co plex nonlinear apping relationships.Co pared with the Deep Dense Convolutional Neural Network（DDCN）,the Deep Residual Dense Convolutional Neural Network（DRDCN）constructed in this paper has a higher approxi ation accuracy for the ultiphase flow nu erical si ulation odel,with 0.9877 and 3.9178 of R2 and RMSE on the test dataset and the sa ple size of the training dataset can effectively i prove the training effect of the two neural networks and i prove the approxi ation accuracy of the surrogate odel to the si ulation odel.（2）The Iterative Local Update Ense ble S oother（ILUES）algorith,a data assi ilation algorith that can solve a higher degree of nonlinearity and is ore applicable,can effectively identify jointly the unknown conta inant features and hydrogeological para eters such as high-di ensional inho ogeneous per eability fields of aquifers in real-world sites.The results show that all the posteriori inho ogeneous per eability fields identified by the algorith have si ilar spatial distribution characteristics,and the logarith ic per eability ean field can be used as the reco ended identification result of the aquifer inho ogeneous per eability field at this real-world site.The probability distributions of the identified conta inant source infor ation para eters and hydrogeological para eters other than per eability basically show single-peaked and double-peaked distributions,and the point esti ates and confidence interval esti ates of the identification results of each para eter can be given at the sa e ti e.A ong the,the edian identification result in the point esti ate can be taken as the reco ended identification result for the site conta inant source characteristics and hydrogeological para eters,i.e.,horizontal coordinatexL 6332.38,vertical coordinateyL 6 315.74,release intensity Q 6 5.37 3/d,release initial o ent T0 6 1559.52d,release final o ent T1 6 1822.18d,porosity of ediu n 6 0.351,The horizontal water phase dispersivity?T 68.53,and the longitudinal water phase dispersivity?L 640.33.Co pared with the ean and MAP values,there is a s aller fitting error between the si ulation results and the actual observation results,R2 and RMSE are 0.9808 and 3.3214,respectively.The results of probability distribution esti ation,interval esti ation and point esti ation of the above para eters can provide ore co prehensive reference for decision akers.（3）In the process of joint inversion of real-world site conta ination source characteristics and hydrogeological para eters using the ILUES inversion algorith,the co putational load of the inversion process can be significantly reduced by invoking the surrogate odel to predict the output response of the ultiphase flow nu erical si ulation odel.The results show that the total co putational load of the ILUES algorith inversion process based on the surrogate odel is only 1.92%of the ti e consu ed in the inversion process based on the si ulation odel（without the surrogate odel）,and the total co putational ti e of the inversion is reduced by98.08%. |
