(287c) Machine Learning to Speed up Dynamic Flux Balance Analysis (FBA) and FBA-Based Reactive Transport Simulations

Advancement in experimental and instrumental technologies has generated an increasingly large body of omics data for environmental systems. Genome-scale metabolic networks are considered an ideal tool to integrate these molecular data for predictive biogeochemical modeling. We recently developed a workflow using the DOE’s KBase (www.kbase.us) modeling pipeline for biogeochemical and reactive transport modeling based on genome-scale metabolic networks built from metagenomes and other omics data. Flux balance analysis of genome-scale metabolic networks is an established method for predicting microbial flux distributions and growth, but its coupling with reactive transport models is challenging due to the significant computational burden. To overcome this barrier, here we present neural network-based reduced-order modeling as a new component of our genome-scale network-based biogeochemical and reactive transport modeling pipeline. We demonstrated the effectiveness of this new pipeline in 0-dimensional batch/continuous reactor and 1-dimensional column configurations. In case studies using metagenomes and high-resolution metabolomic data of river corridor samples collected by the PNNL’s River Corridor Scientific Focus Area (SFA) team, the neural network-based reduced-order models achieved significant time reduction, e.g., from hours to less than a second in the column simulation. This development therefore enables an unprecedented level of detail in representing biogeochemistry in reactive transport models while reducing the computational time to the point that the incorporation of molecular data and insights into multi-scale biogeochemical and reactive transport modeling is possible in practical settings. Ultimately this development will lead to improved understanding and enhanced predictions at larger scales.