Skip to main content
Log in

Comparison of geostatistical and deep-learning inversion methods for DNAPL source zone architecture characterization

Comparaison des méthodes d’inversion géostatistiques et d’apprentissage profond pour la caractérisation de l’architecture de la zone source DNAPL

Comparación de métodos geoestadísticos y de inversión por aprendizaje profundo para la caracterización de la arquitectura de zonas fuente de DNAPL

基于地质统计学和深度学习方法刻画DNAPL源区结构: 方法对比

Comparação de métodos de inversão geoestatística e de aprendizagem profunda para caracterização da arquitetura de zona de origem de DNAPL

  • Paper
  • Published:
Hydrogeology Journal Aims and scope Submit manuscript

Abstract

High-resolution site characterization of hydraulic properties and source zone architecture (SZA) are essential for performing risk assessment and designing remediation strategies for dense nonaqueous phase liquid (DNAPL) contamination. DNAPL SZA characterization is challenging because of the highly correlated unknown states and parameters, namely the spatial distribution of DNAPL saturation (SN) and hydraulic conductivity (K). Two methods can be used for the inversion of highly correlated parameters, i.e., geostatistical inversion with and without parameter-state cross-correlation. In this study, numerical experiments considering weak and strong heterogeneity for SN and K are used to compare the performance of the two geostatistical inversion methods and the deep-learning-based correlation method. Results show that the inversion with combined parameter-state cross-correlation successfully estimated both the K and DNAPL SZA for a weakly heterogeneous K field, but fails to reproduce the overall morphology of the DNAPL SZA for a strongly heterogeneous K field. In comparison, the inversion without parameter-state cross-correlation can robustly capture but over-smooth the detailed features of the K and DNAPL SZA, due to the deviation of the highly correlated KSN fields from the simplified prior cross-covariance. The deep-learning-based method consistently outperformed the inversion strategies with and without the parameter-state cross-correlation, in terms of computational efficiency and estimation accuracy for both the simple and the complex DNAPL SZA, since it can implicitly capture the KSN interdependence and the physical patterns of the DNAPL SZA without explicitly coupling the multiphase model to account for the KSN correlation.

Résumé

La caractérisation à haute résolution des propriétés hydrauliques et de l’architecture de la zone source (SZA) d’un site est essentielle pour effectuer une évaluation des risques et concevoir des stratégies de remédiation pour la contamination par des liquides denses en phase non aqueuse (DNAPL). La caractérisation de la SZA des DNAPL est difficile en raison des états et paramètres inconnus fortement corrélés, à savoir la distribution spatiale de la saturation (SN) en DNAPL et de la conductivité hydraulique (K). Deux méthodes peuvent être utilisées pour l’inversion de paramètres fortement corrélés, c’est-à-dire l’inversion géostatistique avec et sans corrélation croisée paramètre-état. Dans cette étude, des expériences numériques considérant une hétérogénéité faible et forte pour SN et K sont utilisées pour comparer les performances des deux méthodes d’inversion géostatistique, et la méthode de corrélation basée sur l’apprentissage profond. Les résultats montrent que l’inversion avec la corrélation croisée paramètre-état combinée a permis d’estimer avec succès les K et la SZA du DNAPL pour un champ K faiblement hétérogène, mais ne parvient pas à reproduire la morphologie globale de la SZA du DNAPL pour un champ K fortement hétérogène. En comparaison, l’inversion sans corrélation croisée entre paramètre-état peut capturer de manière robuste mais trop lisse les caractéristiques détaillées de K et de la SZA du DNAPL, en raison de la déviation des champs KSN très corrélés par rapport à la covariance croisée simplifiée à priori. La méthode basée sur l’apprentissage profond a constamment surpassé les stratégies d’inversion avec et sans corrélation croisée entre les paramètres-états, en termes d’efficacité de calcul et de précision d’estimation pour la SZA du DNAPL simple et complexe, car elle peut capturer implicitement l’interdépendance KSN et les schémas physiques de la SZA du DNAPL sans coupler explicitement le modèle multi-phase pour tenir compte de la corrélation KSN.

Resumen

La caracterización de alta resolución de las propiedades hidráulicas del emplazamiento y de la arquitectura de una zona fuente (SZA) es esencial para llevar a cabo la evaluación de riesgos y diseñar estrategias de remediación de la contaminación por fase líquida densa no acuosa (DNAPL). La caracterización de la SZA para DNAPL es un desafío debido a los estados y parámetros desconocidos altamente correlacionados, concretamente la distribución espacial de la saturación de DNAPL (SN) y la conductividad hidráulica (K). Se pueden utilizar dos métodos para la inversión de parámetros altamente correlacionados, es decir, la inversión geoestadística con y sin correlación cruzada parámetro-estado. En este estudio, se utilizan experimentos numéricos considerando heterogeneidad débil y fuerte para SN y K para comparar el rendimiento de los dos métodos de inversión geoestadística, y el método de correlación basado en aprendizaje profundo. Los resultados muestran que la inversión con correlación cruzada parámetro-estado combinada estimó con éxito tanto la K como la SZA de la DNAPL para un campo K débilmente heterogéneo, pero no consigue reproducir la morfología global de la SZA de la DNAPL para un campo K fuertemente heterogéneo. En comparación, la inversión sin correlación cruzada parámetro-estado puede capturar de forma robusta, pero sobre-suavizar las características detalladas de la K y de la SZA de la DNAPL, debido a la desviación de los campos K-SN altamente correlacionados de la covarianza cruzada simplificada a priori. El método basado en aprendizaje profundo superó consistentemente a las estrategias de inversión con y sin correlación cruzada parámetro-estado, en términos de eficiencia computacional y precisión de estimación tanto para la SZA de DNAPL simple como para la compleja, ya que puede capturar implícitamente la interdependencia K-SN y los patrones físicos de la SZA de DNAPL sin acoplar explícitamente el modelo multifásico para dar cuenta de la correlación KSN.

摘要

含水层非均质性和污染源区结构(SZA)的高分辨率表征对于重质非水相液体(DNAPL)污染的修复方案设计及风险评估至关重要。由于高度相关的未知状态和参数,即DNAPL饱和度(SN)和导水率(K)的空间分布,因此DNAPL的SZA表征具有挑战性。有两种方法可用于反演高度相关的参数,即有和无参数状态交叉相关的地统计反演。在这项研究中,考虑到SNK的弱异质性和强异质性的数值实验被用来比较两种地质统计反演方法和基于深度学习的相关方法的性能。结果表明,对于弱异质性的K场,采用参数-状态交叉相关的反演方法成功地估计了K和DNAPL SZA,但对于强异质性的K场,未能再现DNAPL SZA的整体形态。相比之下,由于高度相关的KSN场与简化的先验交叉协方差的偏差,没有参数状态交叉协方差的反演可以稳健地捕获和过度平滑K和DNAPL SZA的细节特征。基于深度学习的方法在计算效率和对简单和复杂的DNAPL SZA的估计精度方面一直优于有和没有参数状态交叉相关的反演策略,因为它可以隐含地捕捉KSN的相互依赖性和DNAPL SZA的物理模式,而不需要明确地耦合多相模型来说明KSN的相关性。

Resumo

A caracterização das propriedades hidráulicas e da arquitetura de zona de origem (AZO) em alta resolução é essencial para realizar a avaliação de risco e projetar estratégias de remediação de contaminação por fase liquida densa não aquosa (DNAPL). A caracterização da AZO do DNAPL é um desafio devido aos estados e parâmetros desconhecidos altamente correlacionados, a saber, a distribuição espacial da saturação DNAPL (SN) e a condutividade hidráulica (K). Dois métodos podem ser usados para inversão de parâmetros altamente correlacionados, ou seja, a inversão geoestatística com e sem correlação cruzada de parâmetros. Neste estudo, experimentos numéricos considerando uma heterogeneidade fraca e forte para SN e K são usados para comparar o desempenho dos dois métodos de inversão geoestatística, e o método de correlação baseada em aprendizagem profunda. Os resultados mostram que a inversão com a correlação cruzada de parâmetros combinados estimou com sucesso tanto a K quanto a AZO do DNAPL para um campo de K pouco heterogêneo, mas não reproduziu a morfologia geral da AZO do DNAPL para um campo de K fortemente heterogêneo. Em comparação, a inversão sem correlação cruzada de parâmetros pode capturar com robustez, mas suaviza excessivamente as características detalhadas da K e da AZO do DNAPL, devido ao desvio dos campos KSN altamente correlacionados em relação à covariação cruzada anterior simplificada. O método baseado no aprendizado profundo superou consistentemente as estratégias de inversão com e sem a correlação entre os parâmetros de estado, em termos de eficiência computacional e precisão de estimativa tanto para a AZO do DNAPL simples como para o DNAPL complexo, uma vez que pode implicitamente capturar a interdependência KSN e os padrões físicos da AZO do DNAPL sem acoplar explicitamente o modelo multifásico para contabilizar a correlação KSN.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Fig. 1
Fig. 2
Fig. 3
Fig. 4
Fig. 5
Fig. 6
Fig. 7

Similar content being viewed by others

References

  • Aghasi A, Mendoza-Sanchez I, Miller EL, Ramsburg CA, Abriola LM (2013) A geometric approach to joint inversion with applications to contaminant source zone characterization. Inverse Probl 29(11):115014

    Article  Google Scholar 

  • Archie GE (1942) The electrical resistivity log as an aid in determining some reservoir characteristics. Trans AIME 146(01):54–62

    Article  Google Scholar 

  • Arshadi M, De Paolis Kaluza MC, Miller EL, Abriola LM (2020) Subsurface source zone characterization and uncertainty quantification using discriminative random fields. Water Resour Res 56(3):e2019WR026481

    Article  Google Scholar 

  • Bear J (1972) Dynamics of fluids in porous media. Elsevier, New York, 764 pp

  • Benoit N, Marcotte D, Molson J (2021) Stochastic correlated hydraulic conductivity tensor calibration using gradual deformation. J Hydrol 594:125880

    Article  Google Scholar 

  • Brewster ML, Annan AP, Greenhouse JP, Kueper BH, Olhoeft GR, Redman JD, Sander KA (1995) Observed migration of a controlled DNAPL release by geophysical methods. Groundwater 33(6):977–987

    Article  Google Scholar 

  • Calvello M, Finno RJ (2004) Selecting parameters to optimize in model calibration by inverse analysis. Comput Geotech 31(5):410–424

    Article  Google Scholar 

  • Camporese M, Cassiani G, Deiana R, Salandin P, Binley A (2015) Coupled and uncoupled hydrogeophysical inversions using ensemble Kalman filter assimilation of ERT-monitored tracer test data. Water Resour Res 51(5):3277–3291

    Article  Google Scholar 

  • Christ JA, Ramsburg CA, Pennell KD, Abriola LM (2006) Estimating mass discharge from dense nonaqueous phase liquid source zones using upscaled mass transfer coefficients: an evaluation using multiphase numerical simulations. Water Resour Res 42(W1142011). https://doi.org/10.1029/2006WR004886

  • Christensen S, Doherty J (2008) Predictive error dependencies when using pilot points and singular value decomposition in groundwater model calibration. Adv Water Resour 31(4):674–700

    Article  Google Scholar 

  • Cohen RM, Mercer JW (1993) DNAPL site evaluation: project summary. EPA/600/R-93/022, EPA, Washington, DC

  • Dietrich CR, Newsam GN (1993) A fast and exact method for multidimensional Gaussian stochastic simulations. Water Resour Res 29(8):2861–2869. https://doi.org/10.1029/93WR01070

    Article  Google Scholar 

  • Doherty JE, Hunt RJ (2010) Approaches to highly parameterized inversion: a guide to using PEST for groundwater-model calibration, vol 2010. US Geological Survey, Reston, VA

  • Domenico PA, Mifflin MD (1965) Water from low-permeability sediments and land subsidence. Water Resour Res 1(4):563–576

    Article  Google Scholar 

  • Emerick AA, Reynolds AC (2013) Ensemble smoother with multiple data assimilation. Comput Geosci 55:3–15

    Article  Google Scholar 

  • Ewing RP, Berkowitz B (1998) A generalized growth model for simulating initial migration of dense non-aqueous phase liquids. Water Resour Res 34(4):611–622

    Article  Google Scholar 

  • Ferrari A, Jimenez Martinez J, Borgne TL, Méheust Y, Lunati I (2015) Challenges in modeling unstable two-phase flow experiments in porous micromodels. Water Resour Res 51(3):1381–1400

    Article  Google Scholar 

  • Ghorbanidehno H, Kokkinaki A, Lee J, Darve E (2020) Recent developments in fast and scalable inverse modeling and data assimilation methods in hydrology. J Hydrol 591:125266. https://doi.org/10.1016/j.jhydrol.2020.125266

    Article  Google Scholar 

  • Hill MC, Tiedeman CR (2006) Effective groundwater model calibration: with analysis of data, sensitivities, predictions, and uncertainty. Wiley, Chichester, UK

  • Hu LY (2000) Gradual deformation and iterative calibration of Gaussian-related stochastic models. Math Geol 32(1):87–108

    Article  Google Scholar 

  • Hunt RJ, Doherty J, Tonkin MJ (2007) Are models too simple? Arguments for increased parameterization. Groundwater 45(3):254–262

    Article  Google Scholar 

  • Illman WA, Berg SJ, Liu X, Massi A (2010) Hydraulic/partitioning tracer tomography for DNAPL source zone characterization: small-scale sandbox experiments. Environ Sci Technol 44(22):8609–8614

    Article  Google Scholar 

  • James AI, Graham WD, Hatfield K, Rao PSC, Annable MD (2000) Estimation of spatially variable residual nonaqueous phase liquid saturations in nonuniform flow fields using partitioning tracer data. Water Resour Res 36(4):999–1012 h10.1029/2000WR900004

    Article  Google Scholar 

  • Jardani A, Revil A, Dupont JP (2013) Stochastic joint inversion of hydrogeophysical data for salt tracer test monitoring and hydraulic conductivity imaging. Adv Water Resour 52(52):62–77. https://doi.org/10.1016/j.advwatres.2012.08.005

    Article  Google Scholar 

  • Kang X, Shi X, Deng Y, Revil A, Xu H, Wu J (2018) Coupled hydrogeophysical inversion of DNAPL source zone architecture and permeability field in a 3D heterogeneous sandbox by assimilation time-lapse cross-borehole electrical resistivity data via ensemble Kalman filtering. J Hydrol 567:149–164

    Article  Google Scholar 

  • Kang X, Kokkinaki A, Kitanidis PK, Shi X, Revil A, Lee J, Soueid Ahmed A, Wu J (2020) Improved characterization of DNAPL source zones via sequential hydrogeophysical inversion of hydraulic-head, self-potential and partitioning tracer data. Water Resour Res 56(8):e2020WR027627

    Article  Google Scholar 

  • Kang X, Kokkinaki A, Kitanidis PK, Shi X, Lee J, Mo S, Wu J (2021a) Hydrogeophysical characterization of nonstationary DNAPL source zones by integrating a convolutional variational autoencoder and ensemble smoother. Water Resour Res 57(2):e2020WR028538

    Article  Google Scholar 

  • Kang X, Kokkinaki A, Power C, Kitanidis PK, Shi X, Duan L, Liu T, Wu J (2021b) Integrating deep learning-based data assimilation and hydrogeophysical data for improved monitoring of DNAPL source zones during remediation. J Hydrol 601:126655

    Article  Google Scholar 

  • Kingma DP, Welling M (2014) Auto-encoding variational Bayes. Paper presented at the International Conference on Learning Representations (ICLR), Banff, BC, April 2014

  • Kitanidis PK, Lee J (2014) Principal Component Geostatistical Approach for large-dimensional inverse problems. Water Resour Res 50(7):5428–5443

    Article  Google Scholar 

  • Koch J, Nowak W (2015) Predicting DNAPL mass discharge and contaminated site longevity probabilities: conceptual model and high-resolution stochastic simulation. Water Resour Res 51(2):806–831

    Article  Google Scholar 

  • Koch J, Nowak W (2016) Identification of contaminant source architectures: a statistical inversion that emulates multiphase physics in a computationally practicable manner. Water Resour Res 52(2):1009–1025. https://doi.org/10.1002/2015WR017894

    Article  Google Scholar 

  • Kueper BH, Stroo HF, Vogel CM, Ward CH (eds) (2014) Chlorinated solvent source zone remediation. Springer, New York

    Google Scholar 

  • Hörning S, Bárdossy A (2018) Phase annealing for the conditional simulation of spatial random fields. Comput Geosci 112:101–111

    Article  Google Scholar 

  • Lauzon D, Marcotte D (2020) Calibration of random fields by a sequential spectral turning bands method. Comput Geosci 135:104390

    Article  Google Scholar 

  • Lerner DN, Kueper BH, Wealthall GP, Smith JWN, Leharne SA (2003) An illustrated handbook of DNAPL transport and fate in the subsurface. Environment Agency, Bristol, UK

  • National Research Council (2013) Alternatives for managing the nation’s complex contaminated groundwater sites. National Academies Press, Washington, DC

  • Nowak W (2009) Best unbiased ensemble linearization and the quasi-linear Kalman ensemble generator. Water Resour Res 45(4). https://doi.org/10.1029/2008WR007328

  • Phelan TJ, Lemke LD, Bradford SA, O’Carroll DM, Abriola LM (2004) Influence of textural and wettability variations on predictions of DNAPL persistence and plume development in saturated porous media. Adv Water Resour 27(4):411–427

    Article  Google Scholar 

  • Poeter EP, Hill MC (1997) Inverse models: a necessary next step in ground-water modeling. Groundwater 35(2):250–260

    Article  Google Scholar 

  • Pollock D, Cirpka OA (2012) Fully coupled hydrogeophysical inversion of a laboratory salt tracer experiment monitored by electrical resistivity tomography. Water Resour Res 48(1). https://doi.org/10.1029/2011WR010779

  • Poulsen MM, Kueper BH (1992) A field experiment to study the behavior of tetrachloroethylene in unsaturated porous media. Environ Sci Technol 26(5):889–895

    Article  Google Scholar 

  • Power C, Gerhard JI, Tsourlos P, Giannopoulos A (2013) A new coupled model for simulating the mapping of dense nonaqueous phase liquids using electrical resistivity tomography. Geophysics 78(4):N1–N15. https://doi.org/10.1190/geo2012-0395.1

    Article  Google Scholar 

  • Revil A, Qi Y, Ghorbani A, Coperey A, Soueid Ahmed A, Finizola A, Ricci T (2019) Induced polarization of volcanic rocks: 3. imaging clay cap properties in geothermal fields. Geophys J Int 218(2):1398–1427. https://doi.org/10.1093/gji/ggz207

    Article  Google Scholar 

  • Soueid Ahmed A, Jardani A, Revil A, Dupont JP (2013) SP2DINV: a 2D forward and inverse code for streaming potential problems. Comput Geosci 59:9–16. https://doi.org/10.1016/j.cageo.2013.05.008

    Article  Google Scholar 

  • Soueid Ahmed A, Revil A, Abdulsamad F, Steck B, Vergniault C, Guihard V (2020) Induced polarization as a tool to non-intrusively characterize embankment hydraulic properties. Eng Geol 271:105604. https://doi.org/10.1016/j.enggeo.2020.105604

    Article  Google Scholar 

  • Tikhonov AN, Arsenin VI (1977) Solutions of ill-posed problems. Trans. (from Russian) Fritz John, Wiley, Chichester, UK

  • Yeh TJ, Zhu J (2007) Hydraulic/partitioning tracer tomography for characterization of dense nonaqueous phase liquid source zones. Water Resour Res 43(6). https://doi.org/10.1029/2006WR004877

  • Zheng F, Gao Y, Sun Y, Shi X, Xu H, Wu J (2015) Influence of flow velocity and spatial heterogeneity on DNAPL migration in porous media: insights from laboratory experiments and numerical modelling. Hydrogeol J 23(8):1703–1718

    Article  Google Scholar 

  • Zhu J, Yeh TCJ (2005) Characterization of aquifer heterogeneity using transient hydraulic tomography. Water Resour Res 41(7):W07028

    Article  Google Scholar 

Download references

Acknowledgements

We are grateful to the High-Performance Computing Center (HPCC) of Nanjing University for doing the numerical calculations in this paper on its blade cluster system.

Funding

This work was supported by the National Natural Science Foundation of China (41730856, 41977157 and 42202267) and the Key Laboratory of Earth Fissures Geological Disaster, Ministry of Natural Resources (Geological Survey of Jiangsu Province). Xiaoqing Shi was partly supported by the Fundamental Research Funds for the Central Universities (020614380159). Xueyuan Kang was supported by China Postdoctoral Science Foundation (2022M711565) and Jiangsu Funding Program for Excellent Postdoctoral Talent (20220ZB15).

Author information

Authors and Affiliations

Authors

Corresponding author

Correspondence to Xueyuan Kang.

Ethics declarations

Conflicts of interest

On behalf of all authors, the corresponding author states that there is no conflict of interest.

Additional information

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Published in the special issue “Geostatistics and hydrogeology”

Supplementary information

ESM 1

(PDF 476 kb)

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Shi, X., Kokkinaki, A., Kang, X. et al. Comparison of geostatistical and deep-learning inversion methods for DNAPL source zone architecture characterization. Hydrogeol J 31, 1679–1693 (2023). https://doi.org/10.1007/s10040-023-02606-9

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1007/s10040-023-02606-9

Keywords

Navigation