An artificial-neural-network-based surrogate modeling workflow for reactive transport modeling

Abstract

Process-based reactive transport modeling (RTM) integrates thermodynamic and kinetically controlled fluid-rock interactions with fluid flow through porous media in the subsurface and surface environment. RTM is usually conducted through numerical programs based on the first principle of physical processes. However, the calculation for complex chemical reactions in most available programs is an iterative process, where each iteration is in general computationally intensive. A workflow of neural network-based surrogate model as a proxy for process-based reactive transport simulation is established in this study. The workflow includes (1) base case RTM design, (2) development of training experiments, (3) surrogate model construction based on machine learning, (4) surrogate model validation, and (5) prediction with the calibrated model. The training experiments for surrogate modeling are generated and run prior to the predictions using RTM. The results show that the predictions from the surrogate model agree well with those from processes-based RTM but with a significantly reduced computational time. The well-trained surrogate model is especially useful when a large number of realizations are required, such as the sensitivity analysis or model calibration, which can significantly reduce the computational time compared to that required by RTM. The benefits are (1) it automatizes the experimental design during the sensitivity analysis to get sufficient numbers and coverage of the training cases; (2) it parallelizes the calculations of RTM training cases during the sensitivity analysis to reduce the simulation time; (3) it uses the neural network algorithm to rank the sensitivity of the parameters and to search the optimal solution for model calibration.

Keywords

• Reactive transport modeling
• Surrogate model
• Machine learning
• Dolomitization
• Carbonate reservoir