Ultra-fast geochemical calculations in reactive transport modeling with on-demand learning algorithms

Abstract

Reactive transport simulations are in general time-consuming due to costly geochemical equilibrium and/or kinetics calculations. These may account for over 99% of all computing costs when complex chemical systems are considered, because those computations are needed one or more times in each cell of high-resolution meshes, at every time step of the simulation. To reduce their computing cost by orders of magnitude, we present an on-demand learning strategy that enables geochemical calculations to be rapidly and accurately predicted using previously learned geochemical states. We use sensitivity derivatives combined with first-order Taylor estimations to achieve these fast computations. These derivatives enable a complete bypass of expensive operations such as evaluation of thermodynamic properties (e.g., species activities, fugacities, equations of state), solution of matrix equations in each Newton iteration, time integration of ordinary differential equations, and more. We present reactive transport simulations, considering realistic chemical systems and strong non-ideal thermodynamic behavior, in which geochemical calculations were speed up by a factor of 100 to 200 using this on-demand learning algorithm.