Search

Reactive transport numerical modelling is used to explore chemical interactions between fluids and solids associated with a broad range of subsurface processes. It is a powerful tool extensively used in geoscientific applications. However, simulations often require large computational times. This is due to the fact that they typically involve slow and iterative calculations of geochemical reactions, even when these are based on a very similar set of input values. In this work, we present a generic pipeline to implement, train and validate Machine Learning algorithms to be used as surrogate models of the geochemical reactions. The resulting surrogate models are subsequently used in a novel reactive transport modelling framework. As a proof of concept, a reactive transport model with Machine Learning-based geochemistry is used to simulate the three-dimensional hydrothermal dolomitization of a fractured carbonate reservoir. The model encompasses fluid flow, heat transfer, solute transport and chemical reactions. The accuracy and efficiency of the proposed approach are evaluated by comparison with a conventional reactive transport modelling tool. It is shown that the reactive transport model based on artificial neural networks provides a substantial reduction of the computational burden, with a speedup of one order of magnitude, while providing considerable accuracy. It is concluded that the proposed framework is particularly appealing for large-scale simulations of thermo-hydro-geochemical systems. The limitations of the proposed methodology are discussed, and potential mitigation strategies are proposed for future work. Article highlights: Artificial neural networks are trained to reproduce chemical equilibrium reactions in a carbonate reservoir under non-isothermal conditions Trained neural networks are used as surrogate models in a reactive transport modelling framework Using surrogate models in a reactive transport setup resulted in a speedup of one order of magnitude [ABSTRACT FROM AUTHOR]