Developing a Machine Learning Framework to Reduce the Computational Burden of Reactive-Transport Simulations

Reactive transport modeling is required to make informed decisions about sequestration and disposal in the subsurface. Geochemical reactions in multi-component reactive-transport modeling significantly increase execution times, but computational burdens can be reduced by including only the geologically dominant chemical reactions and ensuring that the primary characteristics governing mixing are included. This can be done by applying machine learning (ML) to the inputs and outputs of reactive-transport models. But reactive-transport simulation requires multiple inputs and yields multiple outputs, which result in a large, multidimensional signature database. Matrix factorization can discover the hidden features (e.g., dominant geochemical reactions) in a database, but tensor factorization is superior for multidimensional databases. Non-negative tensor factorization combined with a customized k-means clustering method, called NTFk, were applied to these reactive-transport simulations to rank dominant geochemical reactions, understand the underlying mixing processes of each reaction, and quantify uncertainties in reactive transport simulations. NTFk significantly pared down the number of chemical reactions to reduce the computational burden of the reactive-transport simulations. This method will benefit practical applications related to CO₂ sequestration as well as nuclear, municipal, and industrial waste storage, and deep-borehole brine injection by reducing computational burdens.