RF-PINNs: Reactive flow physics-informed neural networks for field reconstruction of laminar and turbulent flames using sparse data

Abstract

Physics-Informed Neural Networks (PINNs) have emerged as a promising tool to model flow fields by embedding physical laws into neural networks and thereby reducing the dependency on data. While PINNs have shown substantial success in modeling non-reactive flows, their application for chemically reacting flows is less well understood. The overarching objective of this work is to propose reactive flow physics-informed neural networks (RF-PINNs) that reconstruct all flow quantities of interest solely based on sparse velocity profiles. We present RF-PINNs to reconstruct steady-state mean fields of premixed laminar and turbulent flames with sparse data. First, we extensively study a prototype flame to determine the best combination of governing equations to find a balance between prediction accuracy and optimization speed. We then demonstrate the RF-PINN prediction capabilities by utilizing measured velocity profiles to reconstruct the entire velocity, temperature, and density mean fields of laminar and turbulent flames. To highlight the bidirectional reconstruction capability, we utilize in the second step measured temperature profiles for both flames and, again, successfully reconstruct the flow fields of interest. For both scenarios, the subsequent comparison of predicted field quantities based on only 0.2% of the entire data set is sufficient for accurate field reconstruction. This study underscores the potential of RF-PINNs to complement comprehensive field data of laminar and turbulent reacting flows using both sparse and noisy experimental data.