Data-driven Surrogate Model Approach for Improving the Performance of Reactive Transport Simulations

Jatnieks, Janis; Lucia, Marco De; Dransch, Doris; Sips, Mike

doi:10.1016/j.egypro.2016.10.047

Cited by 30 publications

(26 citation statements)

References 18 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The technical feasibility and tractability of such approach remain to be investigated though. Our RTM problems are also more complex than that considered by Jatnieks et al [18], with equilibrium speciation and kinetics, more than 10 aqueous components, 100 (problem 1) to 200 (problem 2) kinetic reactions and mixing of different boundary solutions [problem 2, see 12, for details]. The work by Sun et al [42] reports on the calibration of 4 first-order reaction rates using a rather simple RTM that does not account for thermodynamic equilibrium, inter-species interactions and sorption processes.…”

Section: Related Workmentioning

confidence: 98%

“…Although the resulting benefits are potentially huge in terms of computational savings, emulation of RTMs has been scarcely addressed so far. Jatnieks et al [18] present a short comparison of a range of nonlinear regression methods for the emulation of the geochemical component of a relatively simple RTM. Some important differences with our work are as follows.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Emulation of CPU-demanding reactive transport models: a comparison of Gaussian processes, polynomial chaos expansion, and deep neural networks

Laloy

Jacques

2019

Comput Geosci

View full text Add to dashboard Cite

This paper presents a detailed comparison between 3 methods for emulating CPU-intensive reactive transport models (RTMs): Gaussian processes (GPs), polynomial chaos expansion (PCE) and deep neural networks (DNNs). State-of-the-art open-source libraries are used for each emulation method while the CPU-time incurred by one forward run of the considered RTMs varies from 1h to between 1h30 and 5 days. Besides direct emulation of the simulated uranium concentration time series, replacing the original RTM by its emulator is also investigated for global sensitivity analysis (GSA), uncertainty propagation and probabilistic calibration using Markov chain Monte Carlo (MCMC) sampling. The selected DNN is found to be superior to both GPs and PCE in reproducing the input -output behavior of the considered 8-dimensional and 13-dimensional CPU-intensive RTMs. This even though the used training sets are small: from 75 to 500 samples. Furthermore, the two used PCE variants: standard PCE and sparse PCE (sPCE) appear to always provide the least accuracy while not differing much in performance. As a consequence of its better emulation capabilities, the DNN method outperforms the two other methods for uncertainty propagation. Regarding GSA, the DNN and GP methods offer equally good approximations to the true first-order and total-order Sobol' sensitivity indices while PCE does less well. Most surprisingly, despite its superior emulation skills the DNN approach leads to the worst solution of the considered synthetic inverse problem which involves 1224 measurement data with low noise. This apparently contradicting behavior of the used DNN is at least partially due to the small but complicated deterministic noise that affects the DNN-based predictions. Indeed, this complex error structure can drive the emulated solutions far away from the true posterior distribution in case of high quality measurement data. Among the considered 3 methods only the GP method allows for retrieving emulated posterior solutions that jointly (1) fit the high-quality measurement data to the appropriate noise level (log-likelihood value) and (2) most closely fit the true model parameter values. Overall, our findings indicate that when the available training set is relatively small (75 -to 500 input -output examples) and fixed beforehand, PCE is not the best choice for emulating CPU-intensive RTMs. Instead GPs or DNNs should be preferred. However, a DNN can deliver overly biased model calibration results. In contrast, the GP method performs fairly well across all considered tasks: direct emulation, global sensitivity analysis, uncertainty propagation, and calibration.

show abstract

Section: Related Workmentioning

confidence: 98%

Section: Related Workmentioning

confidence: 99%

Emulation of CPU-demanding reactive transport models: a comparison of Gaussian processes, polynomial chaos expansion, and deep neural networks

Laloy

Jacques

2019

Comput Geosci

View full text Add to dashboard Cite

show abstract

“…We found out that the Tweedie distribution is suited to reproduce many of the variables in the training dataset. The Tweedie distribution is a special case of exponential dispersion models introduced by Tweedie (1984) and toroughly described by Jørgensen (1987), which finds application in many actuarial and signal processing processes (Hassine et al, 2017). A Random Variable Y is a Tweedie distribution of parameter p if Y ≥ 0, E[Y ] = µ and V ar(Y ) = σ 2 µ p .…”

Section: Fully Data-driven Approachmentioning

confidence: 99%

Chemistry speedup in reactive transport simulations: purely data-driven and physics-based surrogates

Lucia¹,

Kühn²

2020

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…Regarding the use of a proxy model in reactive transport modeling, Guérillot [11,12] proposes a complete workflow to replace thermodynamic equilibriums by an artificial neural network designed and trained in advance to reproduce these flashes during compositional reservoir simulations. Jatnieks et al [19] propose to replace the geochemical model by a data-driven surrogate model. They compare 32 statistical and machine learning methods on a 1D case study corresponding to the injection of a reactive solution leading to the dissolution of calcite and the precipitation of dolomite.…”

Section: Previous Workmentioning

confidence: 99%

“…This approach leads to an impressive speed-up factor range of 60-125. This paper fits into the framework defined by [11,12] and [19] and focuses on the use of artificial neural network to replace the geochemical equilibrium package in a 3D fluid flow simulator.…”

Section: Previous Workmentioning

confidence: 99%

Geochemical equilibrium determination using an artificial neural network in compositional reservoir flow simulation

Guérillot

Bruyelle

2019

Comput Geosci

View full text Add to dashboard Cite

The fluid injection in sedimentary formations may generate geochemical interactions between the fluids and the rock minerals, e.g., CO 2 storage in a depleted reservoir or a saline aquifer. To simulate such reactive transfer processes, geochemical equations (equilibrium and kinetics equations) are coupled with compositional flows in porous media in order to represent, for example, precipitation/dissolution phenomena. The aim of the decoupled approach proposed consists in replacing the geochemical equilibrium solver with a substitute method to bypass the huge consuming time required to balance the geochemical system while keeping an accurate equilibrium calculation. This paper focuses on the use of artificial neural networks (ANN) to determine the geochemical equilibrium instead of solving geochemical equations system. To illustrate the proposed workflow, a 3D case study of CO 2 storage in geological formation is presented.

show abstract

Data-driven Surrogate Model Approach for Improving the Performance of Reactive Transport Simulations

Cited by 30 publications

References 18 publications

Emulation of CPU-demanding reactive transport models: a comparison of Gaussian processes, polynomial chaos expansion, and deep neural networks

Emulation of CPU-demanding reactive transport models: a comparison of Gaussian processes, polynomial chaos expansion, and deep neural networks

Chemistry speedup in reactive transport simulations: purely data-driven and physics-based surrogates

Geochemical equilibrium determination using an artificial neural network in compositional reservoir flow simulation

Contact Info

Product

Resources

About