Neural network based process coupling and parameter upscaling in reactive transport simulations. (15th December 2020)

• Record Type: Journal Article
• Authors: Prasianakis, Nikolaos I.; Haller, Robin; Mahrous, Mohamed; Poonoosamy, Jenna; Pfingsten, Wilfried; Churakov, Sergey V.
• Is Part Of: Geochimica et cosmochimica acta. Volume 291(2020)
• Page Start: 126
• Page End: 143
• Publication Date: 2020-12-15
• Subjects: Multiscale -- Neural networks -- Multiphysics -- Microfluidics -- Machine learning -- Reactive transport -- Dissolution -- Precipitation; Geochemistry -- Periodicals; Meteorites -- Periodicals; Géochimie -- Périodiques; Météorites -- Périodiques; Geochemie; Astrochemie; Electronic journals; 551.905
• DOI: 10.1016/j.gca.2020.07.019
• Languages: English
• ISSNs: 0016-7037

Abstract

Abstract: The multiscale modelling of geochemical processes requires efficient couplings between scales and physics. The use of machine learning techniques and neural networks has the potential to systematically improve the accuracy of models at acceptable computational costs. In this paper, we discuss an efficient framework to transfer information between multi-physics models across spatial scales. In the first example, we train a shallow neural network based on the results of microscopic geochemical reactive transport simulations, and integrate it in a Darcy-scale reactive transport code. In the second example, we train a neural network on geochemical speciation data produced from dedicated geochemical solvers, and adapted to the needs of a lab-on-a-chip microfluidic experiment, in order to accelerate the geochemical calculations. The reactive transport simulation benchmarks show that the neural network approach performs better than the full speciation reactive transport simulations or the look up table-based approaches, both in terms of computational efficiency and memory requirements. Based on these results we discuss the advantages and drawbacks of each simulation approach as well as the potential for further development of the modelling algorithms.