Physics-informed machine learning models for predicting the progress of reactive-mixing

This paper presents a physics-informed machine learning (ML) framework to construct reduced-order models (ROMs) for reactive-transport quantities of interest (QoIs) based on high-fidelity numerical simulations. QoIs include species decay, product yield, and degree mixing. The ROMs are applied quantify understand how the chemical evolve over time. First, high-resolution datasets constructing generated by solving anisotropic reaction–diffusion equations using non-negative finite element formulation different input parameters. reactive-mixing model parameters are: time-scale associated with flipping velocity, spatial-scale controlling small/large vortex structures perturbation parameter vortex-based dispersion contrast, molecular diffusion. Second, random forests, F-test, mutual information criterion used evaluate importance inputs/features respect QoIs. We observed that contrast is most important feature velocity least feature. Third, Support Vector Machines (SVM) Regression (SVR) inputs. constructed SVR-ROMs then predict scaling also present estimates inequalities QoIs, which inform mix, produce in an exponential fashion. These radial basis function suitable kernel SVM/SVR It R2-score unseen data greater than 0.9, implying able system state reasonably well. Finally, terms computational cost, proposed SVM-ROMs O(107) times faster running simulation evaluating makes ML-based attractive sensing real-time monitoring applications as they significantly yet accurate.